├── calc_f_stat.py ├── normalize_0_mean_1_std.py ├── brier_loss.py ├── recovery_performance_mod.py ├── multivariate_split.py ├── run_all ├── cao_min_embedding_dimension.py ├── utils_mod.py ├── causal_ccm.py ├── README.md ├── make_all_multi_plots.py ├── LICENSE └── run_all_models.py /calc_f_stat.py: -------------------------------------------------------------------------------- 1 | #!/usr/bin/env python3 2 | # -*- coding: utf-8 -*- 3 | """ 4 | Input: RSS_R,RSS_U,n,pu,pr 5 | Output: f_GC as values of f-statistic 6 | 7 | - "RSS_R" and unrestricted "RSS_U" models. 8 | - "pu" and "pr" are the number of parameters to be estimated for the unrestricted and restricted model, respectively. 9 | - "n" is the number of time-delayed vectors. 10 | 11 | @Reference: 12 | Wismüller, A., Dsouza, A.M., Vosoughi, M.A., and Abidin, Anas. 13 | Large-scale nonlinear Granger causality for inferring directed dependence from short multivariate time-series data. Sci Rep 11, 7817 (2021). 14 | 15 | """ 16 | 17 | 18 | 19 | def calc_f_stat(RSS_R,RSS_U,n,pu,pr): 20 | f_GC = ((RSS_R-RSS_U)/(RSS_U))*((n-pu-1)/(pu-pr)); 21 | return f_GC 22 | -------------------------------------------------------------------------------- /normalize_0_mean_1_std.py: -------------------------------------------------------------------------------- 1 | #!/usr/bin/env python3 2 | # -*- coding: utf-8 -*- 3 | """ 4 | 5 | 6 | @Reference: 7 | Wismüller, A., Dsouza, A.M., Vosoughi, M.A. et al. 8 | Large-scale nonlinear Granger causality for inferring directed dependence from short multivariate time-series data. Sci Rep 11, 7817 (2021). 9 | """ 10 | import numpy as np 11 | 12 | def normalize_0_mean_1_std(inp_series): 13 | inp_series=inp_series.copy() 14 | mean_ts=np.array([inp_series.mean(axis=1)]).transpose() 15 | mean_ts_mtrx = mean_ts*np.ones((1,inp_series.shape[1])); 16 | unb_data_mtrx = inp_series - mean_ts_mtrx 17 | p = np.power(unb_data_mtrx,2) 18 | s=np.array([p.sum(axis=1)]).transpose() 19 | sc=np.sqrt(s/p.shape[1]) 20 | sc2=sc*(np.ones((1,p.shape[1]))) 21 | nrm= np.divide(unb_data_mtrx,sc2) 22 | return nrm 23 | -------------------------------------------------------------------------------- /brier_loss.py: -------------------------------------------------------------------------------- 1 | import numpy as np 2 | from sklearn import metrics 3 | from copy import deepcopy 4 | 5 | 6 | def brier_loss_func(proba, label): 7 | proba = proba.copy() 8 | label = label.copy() 9 | N = len(proba) 10 | brier_all = np.zeros((N), dtype=np.float32) 11 | for i in range(N): 12 | label_matrix = deepcopy(label[i]).astype(np.float32) 13 | np.fill_diagonal(label_matrix, np.nan) 14 | label_matrix = label_matrix.flatten() 15 | label_matrix = label_matrix[~np.isnan(label_matrix)] 16 | label_matrix = label_matrix.astype(int) 17 | proba_matrix = deepcopy(proba[i]) 18 | proba_matrix = np.round(proba_matrix, 16) 19 | proba_matrix = proba_matrix.astype(np.float32) 20 | np.fill_diagonal(proba_matrix, np.nan) 21 | proba_matrix = proba_matrix.flatten() 22 | proba_matrix = proba_matrix[~np.isnan(proba_matrix)] 23 | proba_matrix = np.clip(proba_matrix, a_min=0, a_max=1) 24 | brier_all[i] = metrics.brier_score_loss(label_matrix, proba_matrix, pos_label=0) 25 | return brier_all 26 | -------------------------------------------------------------------------------- /recovery_performance_mod.py: -------------------------------------------------------------------------------- 1 | #!/usr/bin/env python3 2 | # -*- coding: utf-8 -*- 3 | """ 4 | The code returns performance measures (in terms of AUROC) given a list of 5 | "Adjs": data matrices and 6 | "label": label matrices. 7 | 8 | @Reference: 9 | 10 | Originally created by: 11 | Wismüller, A., Dsouza, A.M., Vosoughi, M.A. et al. 12 | Large-scale nonlinear Granger causality for inferring directed dependence from short multivariate time-series data. Sci Rep 11, 7817 (2021). 13 | 14 | Modified by Wojciech (Victor) Fulmyk to account for the diagonal which is always labelled correctly. 15 | """ 16 | import numpy as np 17 | from sklearn import metrics 18 | from copy import deepcopy 19 | 20 | 21 | def recovery_performance(Adjs,label): 22 | Adjs=Adjs.copy() 23 | label=label.copy() 24 | N=len(Adjs) 25 | auc_all = np.zeros((N), dtype=np.float32) 26 | for i in range(N): 27 | label_matrix = deepcopy(label[i]).astype(float) 28 | np.fill_diagonal(label_matrix, np.nan) 29 | label_matrix = label_matrix.flatten() 30 | label_matrix = label_matrix[~np.isnan(label_matrix)] 31 | Adj_matrix = deepcopy(Adjs[i]).astype(float) 32 | np.fill_diagonal(Adj_matrix, np.nan) 33 | Adj_matrix = Adj_matrix.flatten() 34 | Adj_matrix = Adj_matrix[~np.isnan(Adj_matrix)] 35 | auc_all[i] = metrics.roc_auc_score(label_matrix, Adj_matrix) 36 | return auc_all 37 | 38 | -------------------------------------------------------------------------------- /multivariate_split.py: -------------------------------------------------------------------------------- 1 | #!/usr/bin/env python3 2 | # -*- coding: utf-8 -*- 3 | """ 4 | The function separates input time-series data into pieces of data, based on Taken's theorem to represent system dynamics within the state-space. 5 | 6 | @Reference: 7 | Wismüller, A., Dsouza, A.M., Vosoughi, M.A. et al. 8 | Large-scale nonlinear Granger causality for inferring directed dependence from short multivariate time-series data. Sci Rep 11, 7817 (2021). 9 | """ 10 | import torch 11 | import numpy as np 12 | 13 | 14 | def multivariate_split(X,ar_order, valid_percent=0): 15 | X=X.copy() 16 | TS=np.shape(X)[1] 17 | n_vars=np.shape(X)[0] 18 | val_num=int(valid_percent*TS) 19 | my_data_train=torch.zeros((TS-ar_order-val_num,ar_order,n_vars)) 20 | my_data_y_train=torch.zeros((TS-ar_order-val_num,1,n_vars)) 21 | my_data_val=torch.zeros((val_num,ar_order,n_vars)) 22 | my_data_y_val=torch.zeros((val_num,1,n_vars)) 23 | for i in range(TS-ar_order-val_num): 24 | my_data_train[i]=torch.from_numpy(X.transpose()[i:i+ar_order,:]) 25 | my_data_y_train[i]=torch.from_numpy(X.transpose()[i+ar_order,:]) 26 | 27 | for i in range(TS-ar_order-val_num, TS-ar_order,1): 28 | my_data_val[i-(TS-ar_order-val_num)]=torch.from_numpy(X.transpose()[i:i+ar_order,:]) 29 | my_data_y_val[i-(TS-ar_order-val_num)]=torch.from_numpy(X.transpose()[i+ar_order,:]) 30 | return my_data_train, my_data_y_train, my_data_val, my_data_y_val 31 | -------------------------------------------------------------------------------- /run_all: -------------------------------------------------------------------------------- 1 | python run_all_models.py '5_linear' '500' '100' 2 | python run_all_models.py '5_linear' '1000' '200' 3 | python run_all_models.py '5_linear' '1500' '300' 4 | python run_all_models.py '5_linear' '2000' '400' 5 | 6 | python run_all_models.py '5_nonlinear' '500' '500' 7 | python run_all_models.py '5_nonlinear' '1000' '600' 8 | python run_all_models.py '5_nonlinear' '1500' '700' 9 | python run_all_models.py '5_nonlinear' '2000' '800' 10 | 11 | python run_all_models.py '7_nonlinear' '500' '900' 12 | python run_all_models.py '7_nonlinear' '1000' '1000' 13 | python run_all_models.py '7_nonlinear' '1500' '1100' 14 | python run_all_models.py '7_nonlinear' '2000' '1200' 15 | 16 | python run_all_models.py '9_nonlinear' '500' '1300' 17 | python run_all_models.py '9_nonlinear' '1000' '1400' 18 | python run_all_models.py '9_nonlinear' '1500' '1500' 19 | python run_all_models.py '9_nonlinear' '2000' '1600' 20 | 21 | python run_all_models.py '11_nonlinear' '500' '1700' 22 | python run_all_models.py '11_nonlinear' '1000' '1800' 23 | python run_all_models.py '11_nonlinear' '1500' '1900' 24 | python run_all_models.py '11_nonlinear' '2000' '2000' 25 | 26 | python run_all_models.py '34_zachary1' '500' '2100' 27 | python run_all_models.py '34_zachary1' '1000' '2200' 28 | python run_all_models.py '34_zachary1' '1500' '2300' 29 | python run_all_models.py '34_zachary1' '2000' '2400' 30 | 31 | python run_all_models.py '34_zachary2' '500' '2500' 32 | python run_all_models.py '34_zachary2' '1000' '2600' 33 | python run_all_models.py '34_zachary2' '1500' '2700' 34 | python run_all_models.py '34_zachary2' '2000' '2800' 35 | 36 | 37 | python make_all_multi_plots.py 'auc' 38 | python make_all_multi_plots.py 'brier' 39 | python make_all_multi_plots.py 'accuracy' 40 | python make_all_multi_plots.py 'balanced_accuracy' 41 | python make_all_multi_plots.py 'f1wgt' 42 | python make_all_multi_plots.py 'sensitivity' 43 | python make_all_multi_plots.py 'specificity' 44 | python make_all_multi_plots.py 'gmean' 45 | python make_all_multi_plots.py 'accuracy2' 46 | python make_all_multi_plots.py 'balanced_accuracy2' 47 | python make_all_multi_plots.py 'f1wgt2' 48 | python make_all_multi_plots.py 'sensitivity2' 49 | python make_all_multi_plots.py 'specificity2' 50 | python make_all_multi_plots.py 'gmean2' 51 | -------------------------------------------------------------------------------- /cao_min_embedding_dimension.py: -------------------------------------------------------------------------------- 1 | import numpy as np 2 | import numba 3 | 4 | 5 | # Liangyue Cao's method for finding the minimum embedding dimension 6 | # https://doi.org/10.1016/S0167-2789(97)00118-8 7 | # Python implementation by Wojciech Fulmyk, 2023 8 | # Returns d (a list of embedding dimensions), E1, and E2 9 | @numba.jit(nopython=True) 10 | def cao_min_embedding_dimension( 11 | x, maxd=10, tau=1, threshold=0.95, max_relative_change=0.1 12 | ): 13 | if maxd < 3: 14 | raise ValueError("maxd must be greater than or equal to 3") 15 | v = np.zeros(maxd) 16 | a = np.zeros(maxd) 17 | ya = np.zeros(maxd) 18 | E1 = np.zeros(maxd - 1) 19 | E2 = np.zeros(maxd - 1) 20 | ndp = x.flatten().shape[0] 21 | for d in range(1, maxd + 1): 22 | a[d - 1] = 0.0 23 | ya[d - 1] = 0.0 24 | ng = ndp - d * tau 25 | for i in range(ng): 26 | v0 = 1.0e30 27 | for j in range(ng): 28 | if j == i: 29 | continue 30 | for k in range(d): 31 | v[k] = abs(x[i + k * tau] - x[j + k * tau]) 32 | if d != 1: 33 | for k in range(d - 1): 34 | v[k + 1] = max(v[k], v[k + 1]) 35 | vij = v[d - 1] 36 | if vij < v0 and vij != 0.0: 37 | n0 = j 38 | v0 = vij 39 | a[d - 1] += max(v0, abs(x[i + d * tau] - x[n0 + d * tau])) / v0 40 | ya[d - 1] += abs(x[i + d * tau] - x[n0 + d * tau]) 41 | a[d - 1] /= float(ng) 42 | ya[d - 1] /= float(ng) 43 | if d >= 2: 44 | E1[d - 2] = a[d - 1] / a[d - 2] 45 | E2[d - 2] = ya[d - 1] / ya[d - 2] 46 | d = np.array(list(range(1, maxd))) 47 | embedding_dim = np.nan 48 | for dimension in range(3, maxd + 1): 49 | relative_error = abs(E1[dimension - 2] - E1[dimension - 3]) / E1[dimension - 3] 50 | if ( 51 | np.isnan(embedding_dim) 52 | and not np.isnan(E1[dimension - 2]) 53 | and E1[dimension - 2] >= threshold 54 | and relative_error < max_relative_change 55 | ): 56 | embedding_dim = dimension - 2 57 | break 58 | return d, E1, E2, embedding_dim 59 | 60 | 61 | # Confirm implementation works by reconstructing Fig 1 62 | # from Cao's paper, which plots E1 and E2 for the 63 | # Hénon attractor 64 | # Note that some small discrepancies in Fig 1 of 65 | # Cao's paper and the implementation here are normal 66 | # because the points used here may not be the 67 | # exact same ones used by Cao. Nonetheless, 68 | # the conclusion is the same: the minimum 69 | # embedding dimension is 2 and the figures 70 | # are very similar, which confirms that the 71 | # python implementation is valid. 72 | def cao_plot_henon(): 73 | import matplotlib.pyplot as plt 74 | 75 | def henon_map(xn, yn, a, b): 76 | return yn + 1 - (a * xn**2), b * xn 77 | 78 | xt = [] 79 | # Initial x and y for henon_map 80 | x = 0 81 | y = 0 82 | for i in range(10000): 83 | xn, yn = henon_map(x, y, 1.4, 0.3) 84 | xt.append(x) 85 | x, y = xn, yn 86 | xt_10t = np.array(xt) 87 | xt_1t = xt_10t[-1000:] 88 | result_1t = cao_min_embedding_dimension(xt_1t, maxd=10, tau=1) 89 | result_10t = cao_min_embedding_dimension(xt_10t, maxd=10, tau=1) 90 | plt.plot(result_1t[0], result_1t[1], "D-", color="green", label="E1-1T") 91 | plt.plot(result_10t[0], result_10t[1], "+:", color="red", label="E1-10T") 92 | plt.plot(result_1t[0], result_1t[2], "s--", color="blue", label="E2-1T") 93 | plt.plot(result_10t[0], result_10t[2], "x-.", color="black", label="E2-10T") 94 | plt.legend(loc="lower left") 95 | print("result_1t embedding dimension: " + str(result_1t[3])) 96 | print("result_10t embedding dimension: " + str(result_10t[3])) 97 | plt.show(block=False) 98 | plt.pause(0.001) 99 | -------------------------------------------------------------------------------- /utils_mod.py: -------------------------------------------------------------------------------- 1 | import numpy as np 2 | from scipy import stats, io, signal, linalg 3 | import matplotlib.pyplot as plt 4 | from scipy.spatial.distance import squareform, pdist 5 | from sklearn.model_selection import StratifiedShuffleSplit 6 | from sklearn.svm import SVC 7 | from sklearn import metrics 8 | from scipy.stats import f as scipyf 9 | import warnings 10 | from os.path import join 11 | import os 12 | import math 13 | from sklearn.cluster import KMeans 14 | import torch 15 | from normalize_0_mean_1_std import normalize_0_mean_1_std 16 | from calc_f_stat import calc_f_stat 17 | from multivariate_split import multivariate_split 18 | 19 | """ 20 | This code includes the main functionality of the proposed method. 21 | """ 22 | 23 | """ 24 | The file includes the primary function which calculates Granger causality using lsNGC. 25 | """ 26 | 27 | def lsNGC(inp_series, ar_order=1, k_f=3, k_g=2, normalize=1): 28 | if normalize: 29 | X_normalized=normalize_0_mean_1_std(inp_series) 30 | else: 31 | X_normalized=inp_series.copy() 32 | 33 | X_train, Y_train , X_test, Y_test=multivariate_split(X=X_normalized,ar_order=ar_order) 34 | 35 | X_train=torch.flatten(X_train, start_dim=1) 36 | 37 | km= KMeans(n_clusters= k_f, max_iter= 100, random_state=123) 38 | km.fit(X_train) 39 | cent= km.cluster_centers_ 40 | 41 | 42 | max=0 43 | 44 | for i in range(k_f): 45 | for j in range(k_f): 46 | d= np.linalg.norm(cent[i]-cent[j]) 47 | if(d> max): 48 | max= d 49 | d= max 50 | 51 | sigma= d/math.sqrt(2*k_f) 52 | 53 | sig_d=np.zeros((np.shape(X_normalized)[0],np.shape(X_normalized)[0])); 54 | sig=np.zeros((np.shape(X_normalized)[0],np.shape(X_normalized)[0])); 55 | 56 | # Z_train_label=Y_train 57 | for i in range(X_normalized.shape[0]): 58 | Z_temp=X_normalized.copy() 59 | Z_train, Z_train_label , _ , _=multivariate_split(X=Z_temp,ar_order=ar_order) 60 | Z_train=torch.flatten(Z_train, start_dim=1) 61 | Z_train_label=torch.flatten(Z_train_label, start_dim=1) 62 | 63 | # Obtain phase space Z_s by exclusing time series of of x_s 64 | Z_s_train, Z_s_train_label , _ , _=multivariate_split(X=np.delete(Z_temp,[i],axis=0),ar_order=ar_order) 65 | # Obtain phase space reconstruction of x_s 66 | W_s_train, W_s_train_label , _ , _=multivariate_split(X=np.array([Z_temp[i]]),ar_order=ar_order) 67 | 68 | # Flatten data 69 | Z_s_train=torch.flatten(Z_s_train, start_dim=1) 70 | Z_s_train_label=torch.flatten(Z_s_train_label, start_dim=1) 71 | 72 | W_s_train=torch.flatten(W_s_train, start_dim=1) 73 | W_s_train_label=torch.flatten(W_s_train_label, start_dim=1) 74 | # Obtain k_g number of cluster centers in the phase space W_s with k-means clustering, will have dim=(k_g * d) 75 | kmg= KMeans(n_clusters= k_g, max_iter= 100, random_state=123) 76 | kmg.fit(W_s_train) 77 | cent_W_s= kmg.cluster_centers_ 78 | # Calculate activations for each of the k_g neurons 79 | shape= W_s_train.shape 80 | row= shape[0] 81 | column= k_g 82 | G= np.empty((row,column), dtype= float) 83 | maxg=0 84 | 85 | for ii in range(k_g): 86 | for jj in range(k_g): 87 | dg= np.linalg.norm(cent_W_s[ii]-cent_W_s[jj]) 88 | if(dg> maxg): 89 | maxg= dg 90 | dg= maxg 91 | 92 | sigmag= dg/math.sqrt(2*k_g) 93 | if sigmag==0: 94 | sigmag=1 95 | for ii in range(row): 96 | for jj in range(column): 97 | dist= np.linalg.norm(W_s_train[ii]-cent_W_s[jj]) 98 | G[ii][jj]= math.exp(-math.pow(dist,2)/math.pow(2*sigmag,2)) 99 | # Generalized radial basis function 100 | g_ws=np.array([G[ii]/sum(G[ii]) for ii in range(len(G))]) 101 | # Calculate activations for each of the k_f neurons 102 | shape= Z_s_train.shape 103 | row= shape[0] 104 | column= k_f 105 | F= np.empty((row,column), dtype= float) 106 | for ii in range(row): 107 | for jj in range(column): 108 | cent_temp=cent.copy() 109 | cent_temp=np.delete(cent_temp,np.arange(jj,jj+ar_order),axis=1) 110 | dist= np.linalg.norm(Z_s_train[ii]-cent_temp) 111 | F[ii][jj]= math.exp(-math.pow(dist,2)/math.pow(2*sigma,2)) 112 | # Generalized radial basis function 113 | f_zs=np.array([F[ii]/sum(F[ii]) for ii in range(len(F))]) 114 | 115 | # Prediction in the presence of x_s 116 | num_samples=f_zs.shape[0] 117 | 118 | f_new=np.concatenate((0.5*f_zs,0.5*g_ws),axis=1) 119 | GTG= np.dot(f_new.T,f_new) 120 | GTG_inv= np.linalg.pinv(GTG) 121 | fac= np.dot(GTG_inv,f_new.T) 122 | W_presence= np.dot(fac,Z_train_label) 123 | 124 | prediction_presence= np.dot(f_new,W_presence) 125 | error_presence=prediction_presence-np.array(Z_train_label) 126 | sig[i,:]=np.diag(np.cov(error_presence.T)) 127 | 128 | # Prediction without x_s 129 | GTG= np.dot(f_zs.T,f_zs) 130 | GTG_inv= np.linalg.pinv(GTG) 131 | fac= np.dot(GTG_inv,f_zs.T) 132 | W_absence= np.dot(fac,Z_train_label) 133 | 134 | prediction_absence= np.dot(f_zs,W_absence) 135 | error_absence=prediction_absence-np.array(Z_train_label) 136 | sig_d[i,:]=np.diag(np.cov(error_absence.T)) 137 | # Comupte the Granger causality index 138 | 139 | Aff=np.log(np.divide(sig_d,sig)) 140 | Aff=(Aff>0)*Aff 141 | np.fill_diagonal(Aff,0) 142 | f_stat=calc_f_stat(sig_d, sig, n=num_samples+1, pu=k_f+k_g, pr=k_f) 143 | np.fill_diagonal(f_stat,0) 144 | n=num_samples+1 145 | pu=k_f+k_g 146 | pr=k_f 147 | f_dfn = pu-pr 148 | f_dfd = n-pu-1 149 | pvalue = np.empty_like(f_stat) 150 | for i in range(f_stat.shape[0]): 151 | for j in range(f_stat.shape[1]): 152 | pvalue[i,j] = scipyf.sf(f_stat[i,j], f_dfn, f_dfd) 153 | 154 | return Aff, f_stat, pvalue 155 | -------------------------------------------------------------------------------- /causal_ccm.py: -------------------------------------------------------------------------------- 1 | import numpy as np 2 | import pandas as pd 3 | import matplotlib.pyplot as plt 4 | import seaborn as sns 5 | from scipy.spatial import distance 6 | from scipy.stats import pearsonr 7 | 8 | import itertools 9 | 10 | class ccm: 11 | """ 12 | We're checking causality X -> Y 13 | Args 14 | X: timeseries for variable X that could cause Y 15 | Y: timeseries for variable Y that could be caused by X 16 | tau: time lag. default = 1 17 | E: shadow manifold embedding dimension. default = 2 18 | L: time period/duration to consider (longer = more data). default = length of X 19 | """ 20 | def __init__(self, X, Y, tau=1, E=2, L=None): 21 | ''' 22 | X: timeseries for variable X that could cause Y 23 | Y: timeseries for variable Y that could be caused by X 24 | tau: time lag 25 | E: shadow manifold embedding dimension 26 | L: time period/duration to consider (longer = more data) 27 | We're checking for X -> Y 28 | ''' 29 | self.X = X 30 | self.Y = Y 31 | self.tau = tau 32 | self.E = E 33 | if L == None: 34 | self.L = len(X) 35 | else: 36 | self.L = L 37 | self.My = self.shadow_manifold(Y) # shadow manifold for Y (we want to know if info from X is in Y) 38 | self.t_steps, self.dists = self.get_distances(self.My) # for distances between points in manifold 39 | 40 | def shadow_manifold(self, V): 41 | """ 42 | Given 43 | V: some time series vector 44 | tau: lag step 45 | E: shadow manifold embedding dimension 46 | L: max time step to consider - 1 (starts from 0) 47 | Returns 48 | {t:[t, t-tau, t-2*tau ... t-(E-1)*tau]} = Shadow attractor manifold, dictionary of vectors 49 | """ 50 | V = V[:self.L] # make sure we cut at L 51 | M = {t:[] for t in range((self.E-1) * self.tau, self.L)} # shadow manifold 52 | for t in range((self.E-1) * self.tau, self.L): 53 | v_lag = [] # lagged values 54 | for t2 in range(0, self.E-1 + 1): # get lags, we add 1 to E-1 because we want to include E 55 | v_lag.append(V[t-t2*self.tau]) 56 | M[t] = v_lag 57 | return M 58 | 59 | # get pairwise distances between vectors in the time series 60 | def get_distances(self, M): 61 | """ 62 | Args 63 | M: The shadow manifold from the time series 64 | Returns 65 | t_steps: timesteps 66 | dists: n x n matrix showing distances of each vector at t_step (rows) from other vectors (columns) 67 | """ 68 | 69 | # we extract the time indices and vectors from the manifold M 70 | # we just want to be safe and convert the dictionary to a tuple (time, vector) 71 | # to preserve the time inds when we separate them 72 | t_vec = [(k, v) for k,v in M.items()] 73 | t_steps = np.array([i[0] for i in t_vec]) 74 | vecs = np.array([i[1] for i in t_vec]) 75 | dists = distance.cdist(vecs, vecs) 76 | return t_steps, dists 77 | 78 | def get_nearest_distances(self, t, t_steps, dists): 79 | """ 80 | Args: 81 | t: timestep of vector whose nearest neighbors we want to compute 82 | t_teps: time steps of all vectors in the manifold M, output of get_distances() 83 | dists: distance matrix showing distance of each vector (row) from other vectors (columns). output of get_distances() 84 | E: embedding dimension of shadow manifold M 85 | Returns: 86 | nearest_timesteps: array of timesteps of E+1 vectors that are nearest to vector at time t 87 | nearest_distances: array of distances corresponding to vectors closest to vector at time t 88 | """ 89 | t_ind = np.where(t_steps == t) # get the index of time t 90 | dist_t = dists[t_ind].squeeze() # distances from vector at time t (this is one row) 91 | 92 | # get top closest vectors 93 | nearest_inds = np.argsort(dist_t)[1:self.E+1 + 1] # get indices sorted, we exclude 0 which is distance from itself 94 | nearest_timesteps = t_steps[nearest_inds] # index column-wise, t_steps are same column and row-wise 95 | nearest_distances = dist_t[nearest_inds] 96 | 97 | return nearest_timesteps, nearest_distances 98 | 99 | def predict(self, t): 100 | """ 101 | Args 102 | t: timestep at manifold of y, My, to predict X at same time step 103 | Returns 104 | X_true: the true value of X at time t 105 | X_hat: the predicted value of X at time t using the manifold My 106 | """ 107 | eps = 0.000001 # epsilon minimum distance possible 108 | t_ind = np.where(self.t_steps == t) # get the index of time t 109 | dist_t = self.dists[t_ind].squeeze() # distances from vector at time t (this is one row) 110 | nearest_timesteps, nearest_distances = self.get_nearest_distances(t, self.t_steps, self.dists) 111 | 112 | # get weights 113 | #u = np.exp(-nearest_distances/np.max([eps, nearest_distances[0]])) # we divide by the closest distance to scale 114 | # Fix FloatingPointError: underflow encountered in exp 115 | u = np.exp(np.clip( -nearest_distances/np.max([eps, nearest_distances[0]]), -500, 500 )) # we divide by the closest distance to scale 116 | w = u / np.sum(u) 117 | 118 | # get prediction of X 119 | X_true = self.X[t] # get corresponding true X 120 | X_cor = np.array(self.X)[nearest_timesteps] # get corresponding Y to cluster in Mx 121 | X_hat = (w * X_cor).sum() # get X_hat 122 | 123 | # DEBUGGING 124 | # will need to check why nearest_distances become nan 125 | # if np.isnan(X_hat): 126 | # print(nearest_timesteps) 127 | # print(nearest_distances) 128 | 129 | return X_true, X_hat 130 | 131 | 132 | def causality(self): 133 | ''' 134 | Args: 135 | None 136 | Returns: 137 | (r, p): how much X causes Y. as a correlation between predicted X and true X and the p-value (significance) 138 | ''' 139 | 140 | # run over all timesteps in M 141 | # X causes Y, we can predict X using My 142 | # X puts some info into Y that we can use to reverse engineer X from Y via My 143 | X_true_list = [] 144 | X_hat_list = [] 145 | 146 | for t in list(self.My.keys()): # for each time step in My 147 | X_true, X_hat = self.predict(t) # predict X from My 148 | X_true_list.append(X_true) 149 | X_hat_list.append(X_hat) 150 | 151 | x, y = X_true_list, X_hat_list 152 | r, p = pearsonr(x, y) 153 | 154 | return r, p 155 | 156 | def visualize_cross_mapping(self): 157 | """ 158 | Visualize the shadow manifolds and some cross mappings 159 | """ 160 | # we want to check cross mapping from Mx to My and My to Mx 161 | 162 | f, axs = plt.subplots(1, 2, figsize=(12, 6)) 163 | 164 | for i, ax in zip((0, 1), axs): # i will be used in switching Mx and My in Cross Mapping visualization 165 | #=============================================== 166 | # Shadow Manifolds Visualization 167 | 168 | X_lag, Y_lag = [], [] 169 | for t in range(1, len(self.X)): 170 | X_lag.append(self.X[t-self.tau]) 171 | Y_lag.append(self.Y[t-self.tau]) 172 | X_t, Y_t = self.X[1:], self.Y[1:] # remove first value 173 | 174 | ax.scatter(X_t, X_lag, s=5, label='$M_x$') 175 | ax.scatter(Y_t, Y_lag, s=5, label='$M_y$', c='y') 176 | 177 | #=============================================== 178 | # Cross Mapping Visualization 179 | 180 | A, B = [(self.Y, self.X), (self.X, self.Y)][i] 181 | cm_direction = ['Mx to My', 'My to Mx'][i] 182 | 183 | Ma = self.shadow_manifold(A) 184 | Mb = self.shadow_manifold(B) 185 | 186 | t_steps_A, dists_A = self.get_distances(Ma) # for distances between points in manifold 187 | t_steps_B, dists_B = self.get_distances(Mb) # for distances between points in manifold 188 | 189 | # Plot cross mapping for different time steps 190 | timesteps = list(Ma.keys()) 191 | for t in np.random.choice(timesteps, size=3, replace=False): 192 | Ma_t = Ma[t] 193 | near_t_A, near_d_A = self.get_nearest_distances(t, t_steps_A, dists_A) 194 | 195 | for i in range(self.E+1): 196 | # points on Ma 197 | A_t = Ma[near_t_A[i]][0] 198 | A_lag = Ma[near_t_A[i]][1] 199 | ax.scatter(A_t, A_lag, c='b', marker='s') 200 | 201 | # corresponding points on Mb 202 | B_t = Mb[near_t_A[i]][0] 203 | B_lag = Mb[near_t_A[i]][1] 204 | ax.scatter(B_t, B_lag, c='r', marker='*', s=50) 205 | 206 | # connections 207 | ax.plot([A_t, B_t], [A_lag, B_lag], c='r', linestyle=':') 208 | 209 | ax.set_title(f'{cm_direction} cross mapping. time lag, tau = {self.tau}, E = 2') 210 | ax.legend(prop={'size': 14}) 211 | 212 | ax.set_xlabel('$X_t$, $Y_t$', size=15) 213 | ax.set_ylabel('$X_{t-1}$, $Y_{t-1}$', size=15) 214 | plt.show() 215 | 216 | def plot_ccm_correls(self): 217 | """ 218 | Args 219 | X: X time series 220 | Y: Y time series 221 | tau: time lag 222 | E: shadow manifold embedding dimension 223 | L: time duration 224 | Returns 225 | None. Just correlation plots between predicted X|M_y and true X 226 | """ 227 | X_My_true, X_My_pred = [], [] 228 | for t in range((self.E-1) * self.tau, self.L): 229 | true, pred = self.predict(t) 230 | X_My_true.append(true) 231 | X_My_pred.append(pred) 232 | 233 | # predicting X from My 234 | r, p = np.round(pearsonr(X_My_true, X_My_pred), 4) 235 | 236 | plt.scatter(X_My_true, X_My_pred, s=10) 237 | plt.xlabel('$X(t)$ (observed)', size=15) 238 | plt.ylabel('$\hat{X}(t)|M_y$ (estimated)', size=15) 239 | plt.title(f'tau={self.tau}, E={self.E}, L={self.L}, Correlation coeff = {r}') 240 | 241 | plt.show() 242 | 243 | 244 | def loop_causal_cmm(data, tau=1, E=2, L=None): 245 | results_list = [] 246 | permute_list = list(itertools.permutations(range(data.shape[1]),2)) 247 | for X_idx, y_idx in permute_list: 248 | ccm1 = ccm(X=data[:,X_idx], Y=data[:,y_idx], tau=tau, E=E, L=L) 249 | score, pval = ccm1.causality() 250 | results_list.append([X_idx,y_idx,score,pval]) 251 | out_df = pd.DataFrame(results_list, columns=['X','y','score','pvalue']) 252 | return out_df 253 | -------------------------------------------------------------------------------- /README.md: -------------------------------------------------------------------------------- 1 | 2 | 3 | 4 | 5 | 12 | 18 | 19 | 20 | 21 | 22 | 29 | 30 |

mlcausality-krr-paper-replication

31 | 32 |

33 | Replication code for the paper "Nonlinear Granger Causality using Kernel Ridge Regression" 34 |
35 | Report Bug 36 |

37 | 38 | 39 | 40 | 41 | 42 |
43 | Table of Contents 44 |
    45 |
  1. 46 | About The Project 47 |
    2. 48 |
  2. 49 | Getting Started 50 | 55 |
    3. 56 |
  3. File Descriptions
    4. 57 |
63 |
64 | 65 | 72 | 73 | 74 | ## About The Project 75 | 76 | 77 | 78 | This repository stores replication code for the paper "Nonlinear Granger Causality using Kernel Ridge Regression" that introduces and highlights the use of the mlcausality package. 79 | 80 | The replication files contained herein use functions from the following projects: 81 | 96 | 97 | 98 |

(back to top)

99 | 100 | 101 | ## Getting Started 102 | ### Prerequisites 103 | 104 | Replication requires the installation of the following Python libraries: 105 | * [NumPy](https://numpy.org) 106 | * [SciPy](https://scipy.org) 107 | * [pandas](https://pandas.pydata.org) 108 | * [statsmodels](https://www.statsmodels.org) 109 | * [scikit-learn](https://scikit-learn.org) 110 | * [tqdm](https://github.com/tqdm/tqdm) 111 | * [matplotlib](https://matplotlib.org/) 112 | * [networkx](https://networkx.org/) 113 | * [psutil](https://github.com/giampaolo/psutil) 114 | * [tigramite](https://github.com/jakobrunge/tigramite) 115 | * [mlcausality](https://github.com/WojtekFulmyk/mlcausality) 116 | 117 | Please install these libraries into your Python system before proceeding. 118 | 119 | ### Downloading the Replication Files 120 | 121 | In order to replicate the results of the paper, you must download all of the files in this repository onto your local computer. You can do so, for instance, using the following command: 122 | 123 | git clone --depth 1 https://github.com/WojtekFulmyk/mlcausality-krr-paper-replication.git 124 | 125 | ### Replicate results 126 | 127 | You can replicate results by running all the codes in the `run_all` file on a terminal on your local machine. 128 | 129 |

(back to top)

130 | 131 | 132 | 133 | 134 | ## File Descriptions 135 | 136 | ### Files from the lsNGC project 137 | The following repository files are taken from the
  • lsNGC
    • project. Note that these files were forked from the originals in order to perform some modifications. In particular, the `utils_mod.py` was modified to calculate p-values, and the `recovery_performance_mod.py` file was modified to ignore the diagonal of the adjacency matrix. 138 | 139 | * `utils_mod.py` 140 | * `recovery_performance_mod.py` 141 | * `normalize_0_mean_1_std.py` 142 | * `multivariate_split.py` 143 | * `calc_f_stat.py` 144 | 145 | ### Files from the causal_ccm project 146 | The following file was taken and modified from the causal_ccm project. In particular, an error involving underflow in exp was fixed. 147 | 148 | * `causal_cmm.py` 149 | 150 | ### Files written by myself with the help of some code from the spectral-connectivity library 151 | The following files were largely written by myself (Wojciech "Victor" Fulmyk) with the help of some code from the spectral-connectivity project. In particular, spectral-connectivity was used to generate the 5_linear network; see the `run_all_models.py` file for details. The relevant functions from the spectral-connectivity package were forked herein as opposed to using the original package in order to ensure that the seed was set correctly for replication. 152 | 153 | * `run_all_models.py` 154 | * `make_all_multi_plots.py` 155 | * `cao_min_embedding_dimension.py` 156 | * `brier_loss.py` 157 | * `run_all` 158 | * `Readme.md` (this file) 159 | 160 | 161 | ## License 162 | 163 | Distributed under the GPLv3 License because the spectral-connectivity library, whose functions are used in this repository, are also released under GPLv3. See [LICENSE](https://github.com/WojtekFulmyk/mlcausality-krr-paper-replication/blob/master/LICENSE) for more information. 164 | 165 |

    (back to top)

    166 | 167 | 168 | 169 | 170 | 179 | 180 | 181 | 182 | 191 | 192 | 193 | 194 | 195 | [contributors-shield]: https://img.shields.io/github/contributors/WojtekFulmyk/mlcausality.svg?style=for-the-badge 196 | [contributors-url]: https://github.com/WojtekFulmyk/mlcausality/graphs/contributors 197 | [forks-shield]: https://img.shields.io/github/forks/WojtekFulmyk/mlcausality.svg?style=for-the-badge 198 | [forks-url]: https://github.com/WojtekFulmyk/mlcausality/network/members 199 | [stars-shield]: https://img.shields.io/github/stars/WojtekFulmyk/mlcausality.svg?style=for-the-badge 200 | [stars-url]: https://github.com/WojtekFulmyk/mlcausality/stargazers 201 | [issues-shield]: https://img.shields.io/github/issues/WojtekFulmyk/mlcausality.svg?style=for-the-badge 202 | [issues-url]: https://github.com/WojtekFulmyk/mlcausality/issues 203 | [license-shield]: https://img.shields.io/github/license/WojtekFulmyk/mlcausality.svg?style=for-the-badge 204 | [license-url]: https://github.com/WojtekFulmyk/mlcausality/blob/master/LICENSE 205 | [linkedin-shield]: https://img.shields.io/badge/-LinkedIn-black.svg?style=for-the-badge&logo=linkedin&colorB=555 206 | [linkedin-url]: https://linkedin.com/in/linkedin_username 207 | [product-screenshot]: images/screenshot.png 208 | [Next.js]: https://img.shields.io/badge/next.js-000000?style=for-the-badge&logo=nextdotjs&logoColor=white 209 | [Next-url]: https://nextjs.org/ 210 | [React.js]: https://img.shields.io/badge/React-20232A?style=for-the-badge&logo=react&logoColor=61DAFB 211 | [React-url]: https://reactjs.org/ 212 | [Vue.js]: https://img.shields.io/badge/Vue.js-35495E?style=for-the-badge&logo=vuedotjs&logoColor=4FC08D 213 | [Vue-url]: https://vuejs.org/ 214 | [Angular.io]: https://img.shields.io/badge/Angular-DD0031?style=for-the-badge&logo=angular&logoColor=white 215 | [Angular-url]: https://angular.io/ 216 | [Svelte.dev]: https://img.shields.io/badge/Svelte-4A4A55?style=for-the-badge&logo=svelte&logoColor=FF3E00 217 | [Svelte-url]: https://svelte.dev/ 218 | [Laravel.com]: https://img.shields.io/badge/Laravel-FF2D20?style=for-the-badge&logo=laravel&logoColor=white 219 | [Laravel-url]: https://laravel.com 220 | [Bootstrap.com]: https://img.shields.io/badge/Bootstrap-563D7C?style=for-the-badge&logo=bootstrap&logoColor=white 221 | [Bootstrap-url]: https://getbootstrap.com 222 | [JQuery.com]: https://img.shields.io/badge/jQuery-0769AD?style=for-the-badge&logo=jquery&logoColor=white 223 | [JQuery-url]: https://jquery.com 224 | 225 | [NumPy-url]: https://numpy.org 226 | [SciPy-url]: https://scipy.org 227 | [pandas-url]: https://pandas.pydata.org 228 | [statsmodels-url]: https://www.statsmodels.org 229 | [scikit-learn-url]: https://scikit-learn.org 230 | [XGBoost-url]: https://xgboost.readthedocs.io 231 | [LightGBM-url]: https://lightgbm.readthedocs.io 232 | [CatBoost-url]: https://catboost.ai 233 | [cuML-url]: https://github.com/rapidsai/cuml 234 | 235 | 236 | [NumPy-sheild]: https://img.shields.io/badge/numpy-%23013243.svg?style=for-the-badge&logo=numpy&logoColor=white 237 | [SciPy-sheild]: https://img.shields.io/badge/SciPy-%230C55A5.svg?style=for-the-badge&logo=scipy&logoColor=%white 238 | [Pandas-sheild]: https://img.shields.io/badge/pandas-%23150458.svg?style=for-the-badge&logo=pandas&logoColor=white 239 | [scikit-learn-shield]: https://img.shields.io/badge/scikit--learn-%23F7931E.svg?style=for-the-badge&logo=scikit-learn&logoColor=white 240 | 241 | -------------------------------------------------------------------------------- /make_all_multi_plots.py: -------------------------------------------------------------------------------- 1 | # Get plot type to make 2 | import sys 3 | 4 | plot_type = sys.argv[1] 5 | if plot_type.lower() == "brier": 6 | plot_type_cap = "Brier Score" 7 | elif plot_type.lower() == "auc": 8 | plot_type_cap = "AUC" 9 | elif plot_type.lower() == "accuracy": 10 | plot_type_cap = "Accuracy (thresh = 0.05)" 11 | elif plot_type.lower() == "balanced_accuracy": 12 | plot_type_cap = "Bal.Accuracy (thresh = 0.05)" 13 | elif plot_type.lower() == "f1wgt": 14 | plot_type_cap = "Weighted F1 (thresh = 0.05)" 15 | elif plot_type.lower() == "sensitivity": 16 | plot_type_cap = "Sensitivity (thresh = 0.05)" 17 | elif plot_type.lower() == "specificity": 18 | plot_type_cap = "Specificity (thresh = 0.05)" 19 | elif plot_type.lower() == "gmean": 20 | plot_type_cap = "G-mean (thresh = 0.05)" 21 | elif plot_type.lower() == "accuracy2": 22 | plot_type_cap = "Accuracy (max G-mean thresh)" 23 | elif plot_type.lower() == "balanced_accuracy2": 24 | plot_type_cap = "Bal. Accuracy (max G-mean thresh)" 25 | elif plot_type.lower() == "f1wgt2": 26 | plot_type_cap = "Weighted F1 (max G-mean thresh)" 27 | elif plot_type.lower() == "sensitivity2": 28 | plot_type_cap = "Sensitivity (max G-mean thresh)" 29 | elif plot_type.lower() == "specificity2": 30 | plot_type_cap = "Specificity (max G-mean thresh)" 31 | elif plot_type.lower() == "gmean2": 32 | plot_type_cap = "G-mean (max G-mean thresh)" 33 | 34 | import numpy as np 35 | import pandas as pd 36 | import matplotlib.pyplot as plt 37 | import matplotlib.ticker as ticker 38 | import seaborn as sns 39 | import pickle 40 | 41 | from copy import deepcopy 42 | from decimal import Decimal, localcontext, ROUND_DOWN 43 | 44 | 45 | # This truncate function is based on: 46 | # https://stackoverflow.com/a/28323804 47 | def truncate(number, places): 48 | if not isinstance(places, int): 49 | raise ValueError("Decimal places must be an integer.") 50 | if places < 1: 51 | raise ValueError("Decimal places must be at least 1.") 52 | # If you want to truncate to 0 decimal places, just do int(number). 53 | with localcontext() as context: 54 | context.rounding = ROUND_DOWN 55 | exponent = Decimal(str(10**-places)) 56 | return Decimal(str(number)).quantize(exponent) 57 | 58 | 59 | with open(r"pickled_data/5_linear_" + plot_type + "_500.pickle", "rb") as input_file: 60 | df_init1 = pickle.load(input_file) 61 | 62 | df_init1["Length of time-series"] = 500 63 | 64 | 65 | with open(r"pickled_data/5_linear_" + plot_type + "_1000.pickle", "rb") as input_file: 66 | df_init2 = pickle.load(input_file) 67 | 68 | df_init2["Length of time-series"] = 1000 69 | 70 | with open(r"pickled_data/5_linear_" + plot_type + "_1500.pickle", "rb") as input_file: 71 | df_init3 = pickle.load(input_file) 72 | 73 | df_init3["Length of time-series"] = 1500 74 | 75 | with open(r"pickled_data/5_linear_" + plot_type + "_2000.pickle", "rb") as input_file: 76 | df_init4 = pickle.load(input_file) 77 | 78 | df_init4["Length of time-series"] = 2000 79 | df1 = pd.concat([df_init1, df_init2, df_init3, df_init4]).reset_index(drop=True) 80 | df1 = df1.melt("Length of time-series", var_name="Model", value_name="Val") 81 | 82 | with open(r"pickled_data/5_nonlinear_" + plot_type + "_500.pickle", "rb") as input_file: 83 | df_init1 = pickle.load(input_file) 84 | 85 | df_init1["Length of time-series"] = 500 86 | 87 | 88 | with open( 89 | r"pickled_data/5_nonlinear_" + plot_type + "_1000.pickle", "rb" 90 | ) as input_file: 91 | df_init2 = pickle.load(input_file) 92 | 93 | df_init2["Length of time-series"] = 1000 94 | 95 | with open( 96 | r"pickled_data/5_nonlinear_" + plot_type + "_1500.pickle", "rb" 97 | ) as input_file: 98 | df_init3 = pickle.load(input_file) 99 | 100 | df_init3["Length of time-series"] = 1500 101 | 102 | with open( 103 | r"pickled_data/5_nonlinear_" + plot_type + "_2000.pickle", "rb" 104 | ) as input_file: 105 | df_init4 = pickle.load(input_file) 106 | 107 | df_init4["Length of time-series"] = 2000 108 | df2 = pd.concat([df_init1, df_init2, df_init3, df_init4]).reset_index(drop=True) 109 | df2 = df2.melt("Length of time-series", var_name="Model", value_name="Val") 110 | 111 | with open(r"pickled_data/7_nonlinear_" + plot_type + "_500.pickle", "rb") as input_file: 112 | df_init1 = pickle.load(input_file) 113 | 114 | df_init1["Length of time-series"] = 500 115 | 116 | 117 | with open( 118 | r"pickled_data/7_nonlinear_" + plot_type + "_1000.pickle", "rb" 119 | ) as input_file: 120 | df_init2 = pickle.load(input_file) 121 | 122 | df_init2["Length of time-series"] = 1000 123 | 124 | with open( 125 | r"pickled_data/7_nonlinear_" + plot_type + "_1500.pickle", "rb" 126 | ) as input_file: 127 | df_init3 = pickle.load(input_file) 128 | 129 | df_init3["Length of time-series"] = 1500 130 | 131 | with open( 132 | r"pickled_data/7_nonlinear_" + plot_type + "_2000.pickle", "rb" 133 | ) as input_file: 134 | df_init4 = pickle.load(input_file) 135 | 136 | df_init4["Length of time-series"] = 2000 137 | df3 = pd.concat([df_init1, df_init2, df_init3, df_init4]).reset_index(drop=True) 138 | df3 = df3.melt("Length of time-series", var_name="Model", value_name="Val") 139 | 140 | with open(r"pickled_data/9_nonlinear_" + plot_type + "_500.pickle", "rb") as input_file: 141 | df_init1 = pickle.load(input_file) 142 | 143 | df_init1["Length of time-series"] = 500 144 | 145 | 146 | with open( 147 | r"pickled_data/9_nonlinear_" + plot_type + "_1000.pickle", "rb" 148 | ) as input_file: 149 | df_init2 = pickle.load(input_file) 150 | 151 | df_init2["Length of time-series"] = 1000 152 | 153 | with open( 154 | r"pickled_data/9_nonlinear_" + plot_type + "_1500.pickle", "rb" 155 | ) as input_file: 156 | df_init3 = pickle.load(input_file) 157 | 158 | df_init3["Length of time-series"] = 1500 159 | 160 | with open( 161 | r"pickled_data/9_nonlinear_" + plot_type + "_2000.pickle", "rb" 162 | ) as input_file: 163 | df_init4 = pickle.load(input_file) 164 | 165 | df_init4["Length of time-series"] = 2000 166 | df4 = pd.concat([df_init1, df_init2, df_init3, df_init4]).reset_index(drop=True) 167 | df4 = df4.melt("Length of time-series", var_name="Model", value_name="Val") 168 | 169 | with open( 170 | r"pickled_data/11_nonlinear_" + plot_type + "_500.pickle", "rb" 171 | ) as input_file: 172 | df_init1 = pickle.load(input_file) 173 | 174 | df_init1["Length of time-series"] = 500 175 | 176 | 177 | with open( 178 | r"pickled_data/11_nonlinear_" + plot_type + "_1000.pickle", "rb" 179 | ) as input_file: 180 | df_init2 = pickle.load(input_file) 181 | 182 | df_init2["Length of time-series"] = 1000 183 | 184 | with open( 185 | r"pickled_data/11_nonlinear_" + plot_type + "_1500.pickle", "rb" 186 | ) as input_file: 187 | df_init3 = pickle.load(input_file) 188 | 189 | df_init3["Length of time-series"] = 1500 190 | 191 | with open( 192 | r"pickled_data/11_nonlinear_" + plot_type + "_2000.pickle", "rb" 193 | ) as input_file: 194 | df_init4 = pickle.load(input_file) 195 | 196 | df_init4["Length of time-series"] = 2000 197 | df5 = pd.concat([df_init1, df_init2, df_init3, df_init4]).reset_index(drop=True) 198 | df5 = df5.melt("Length of time-series", var_name="Model", value_name="Val") 199 | 200 | with open(r"pickled_data/34_zachary1_" + plot_type + "_500.pickle", "rb") as input_file: 201 | df_init1 = pickle.load(input_file) 202 | 203 | df_init1["Length of time-series"] = 500 204 | 205 | 206 | with open( 207 | r"pickled_data/34_zachary1_" + plot_type + "_1000.pickle", "rb" 208 | ) as input_file: 209 | df_init2 = pickle.load(input_file) 210 | 211 | df_init2["Length of time-series"] = 1000 212 | 213 | with open( 214 | r"pickled_data/34_zachary1_" + plot_type + "_1500.pickle", "rb" 215 | ) as input_file: 216 | df_init3 = pickle.load(input_file) 217 | 218 | df_init3["Length of time-series"] = 1500 219 | 220 | with open( 221 | r"pickled_data/34_zachary1_" + plot_type + "_2000.pickle", "rb" 222 | ) as input_file: 223 | df_init4 = pickle.load(input_file) 224 | 225 | df_init4["Length of time-series"] = 2000 226 | df6 = pd.concat([df_init1, df_init2, df_init3, df_init4]).reset_index(drop=True) 227 | df6 = df6.melt("Length of time-series", var_name="Model", value_name="Val") 228 | 229 | with open(r"pickled_data/34_zachary2_" + plot_type + "_500.pickle", "rb") as input_file: 230 | df_init1 = pickle.load(input_file) 231 | 232 | df_init1["Length of time-series"] = 500 233 | 234 | 235 | with open( 236 | r"pickled_data/34_zachary2_" + plot_type + "_1000.pickle", "rb" 237 | ) as input_file: 238 | df_init2 = pickle.load(input_file) 239 | 240 | df_init2["Length of time-series"] = 1000 241 | 242 | with open( 243 | r"pickled_data/34_zachary2_" + plot_type + "_1500.pickle", "rb" 244 | ) as input_file: 245 | df_init3 = pickle.load(input_file) 246 | 247 | df_init3["Length of time-series"] = 1500 248 | 249 | with open( 250 | r"pickled_data/34_zachary2_" + plot_type + "_2000.pickle", "rb" 251 | ) as input_file: 252 | df_init4 = pickle.load(input_file) 253 | 254 | df_init4["Length of time-series"] = 2000 255 | df7 = pd.concat([df_init1, df_init2, df_init3, df_init4]).reset_index(drop=True) 256 | df7 = df7.melt("Length of time-series", var_name="Model", value_name="Val") 257 | 258 | title_font = { 259 | "size": "32", 260 | "color": "black", 261 | "weight": "bold", 262 | "verticalalignment": "bottom", 263 | } 264 | title_font2 = { 265 | "size": "52", 266 | "color": "black", 267 | "weight": "bold", 268 | "verticalalignment": "bottom", 269 | } 270 | f, axes = plt.subplots(1, 7, sharex=False, sharey=True, figsize=(30, 14)) 271 | f.text(0.5, 1.025, plot_type_cap, ha="center", va="center", **title_font2) 272 | f.text(-0.01, 0.5, plot_type_cap, ha="center", va="center", rotation=90, **title_font) 273 | 274 | sns.boxplot( 275 | x="Model", 276 | y="Val", 277 | hue="Length of time-series", 278 | data=df1, 279 | palette="Set3", 280 | ax=axes[0], 281 | linewidth=2, 282 | saturation=1, 283 | width=0.7, 284 | color="black", 285 | notch=False, 286 | medianprops=dict(color="black", linewidth=10, alpha=1, solid_capstyle="butt"), 287 | whis=(0, 100), 288 | showfliers=False, 289 | bootstrap=10000, 290 | whiskerprops=dict(color="black", alpha=1), 291 | boxprops=dict(edgecolor="black", alpha=1), 292 | capprops=dict(color="black", alpha=1), 293 | ) 294 | sns.boxplot( 295 | x="Model", 296 | y="Val", 297 | hue="Length of time-series", 298 | data=df2, 299 | palette="Set3", 300 | ax=axes[1], 301 | linewidth=2, 302 | saturation=1, 303 | width=0.7, 304 | color="black", 305 | notch=False, 306 | medianprops=dict(color="black", linewidth=10, alpha=1, solid_capstyle="butt"), 307 | whis=(0, 100), 308 | showfliers=False, 309 | bootstrap=10000, 310 | whiskerprops=dict(color="black", alpha=1), 311 | boxprops=dict(edgecolor="black", alpha=1), 312 | capprops=dict(color="black", alpha=1), 313 | ) 314 | sns.boxplot( 315 | x="Model", 316 | y="Val", 317 | hue="Length of time-series", 318 | data=df3, 319 | palette="Set3", 320 | ax=axes[2], 321 | linewidth=2, 322 | saturation=1, 323 | width=0.7, 324 | color="black", 325 | notch=False, 326 | medianprops=dict(color="black", linewidth=10, alpha=1, solid_capstyle="butt"), 327 | whis=(0, 100), 328 | showfliers=False, 329 | bootstrap=10000, 330 | whiskerprops=dict(color="black", alpha=1), 331 | boxprops=dict(edgecolor="black", alpha=1), 332 | capprops=dict(color="black", alpha=1), 333 | ) 334 | sns.boxplot( 335 | x="Model", 336 | y="Val", 337 | hue="Length of time-series", 338 | data=df4, 339 | palette="Set3", 340 | ax=axes[3], 341 | linewidth=2, 342 | saturation=1, 343 | width=0.7, 344 | color="black", 345 | notch=False, 346 | medianprops=dict(color="black", linewidth=10, alpha=1, solid_capstyle="butt"), 347 | whis=(0, 100), 348 | showfliers=False, 349 | bootstrap=10000, 350 | whiskerprops=dict(color="black", alpha=1), 351 | boxprops=dict(edgecolor="black", alpha=1), 352 | capprops=dict(color="black", alpha=1), 353 | ) 354 | sns.boxplot( 355 | x="Model", 356 | y="Val", 357 | hue="Length of time-series", 358 | data=df5, 359 | palette="Set3", 360 | ax=axes[4], 361 | linewidth=2, 362 | saturation=1, 363 | width=0.7, 364 | color="black", 365 | notch=False, 366 | medianprops=dict(color="black", linewidth=10, alpha=1, solid_capstyle="butt"), 367 | whis=(0, 100), 368 | showfliers=False, 369 | bootstrap=10000, 370 | whiskerprops=dict(color="black", alpha=1), 371 | boxprops=dict(edgecolor="black", alpha=1), 372 | capprops=dict(color="black", alpha=1), 373 | ) 374 | sns.boxplot( 375 | x="Model", 376 | y="Val", 377 | hue="Length of time-series", 378 | data=df6, 379 | palette="Set3", 380 | ax=axes[5], 381 | linewidth=2, 382 | saturation=1, 383 | width=0.7, 384 | color="black", 385 | notch=False, 386 | medianprops=dict(color="black", linewidth=10, alpha=1, solid_capstyle="butt"), 387 | whis=(0, 100), 388 | showfliers=False, 389 | bootstrap=10000, 390 | whiskerprops=dict(color="black", alpha=1), 391 | boxprops=dict(edgecolor="black", alpha=1), 392 | capprops=dict(color="black", alpha=1), 393 | ) 394 | sns.boxplot( 395 | x="Model", 396 | y="Val", 397 | hue="Length of time-series", 398 | data=df7, 399 | palette="Set3", 400 | ax=axes[6], 401 | linewidth=2, 402 | saturation=1, 403 | width=0.7, 404 | color="black", 405 | notch=False, 406 | medianprops=dict(color="black", linewidth=10, alpha=1, solid_capstyle="butt"), 407 | whis=(0, 100), 408 | showfliers=False, 409 | bootstrap=10000, 410 | whiskerprops=dict(color="black", alpha=1), 411 | boxprops=dict(edgecolor="black", alpha=1), 412 | capprops=dict(color="black", alpha=1), 413 | ) 414 | 415 | axes[0].set_title("5-linear", fontweight="bold", fontsize=48) 416 | axes[1].set_title("5-nonlin", fontweight="bold", fontsize=48) 417 | axes[2].set_title("7-nonlin", fontweight="bold", fontsize=48) 418 | axes[3].set_title("9-nonlin", fontweight="bold", fontsize=48) 419 | axes[4].set_title("11-nonlin", fontweight="bold", fontsize=48) 420 | axes[5].set_title("Zachary1", fontweight="bold", fontsize=48) 421 | axes[6].set_title("Zachary2", fontweight="bold", fontsize=48) 422 | 423 | for i in range(7): 424 | if i < 6: 425 | axes[i].get_legend().remove() 426 | else: 427 | axes[i].legend( 428 | prop=dict(size=26, weight="bold"), 429 | title="$\\bf{Num. obs.}$", 430 | title_fontsize=26, 431 | ) 432 | x_axis = axes[i].axes.get_xaxis() 433 | x_label = x_axis.get_label() 434 | x_label.set_visible(False) 435 | y_axis = axes[i].axes.get_yaxis() 436 | y_label = y_axis.get_label() 437 | y_label.set_visible(False) 438 | if i == 0: 439 | start, end = axes[i].get_ylim() 440 | new_start = float(truncate(deepcopy(start), 1)) 441 | if plot_type.lower() == "accuracy" or plot_type.lower() == "accuracy2": 442 | new_start = abs(round(max(new_start - 0.1, 0), 1)) 443 | else: 444 | new_start = abs(round(max(new_start, 0), 1)) 445 | axes[i].yaxis.set_ticks(np.arange(new_start, end, 0.1)) 446 | axes[i].yaxis.set_major_formatter(ticker.FormatStrFormatter("%0.1f")) 447 | if i == 0: 448 | axes[i].tick_params(width=5, size=10) 449 | else: 450 | axes[i].tick_params( 451 | top=False, bottom=True, left=False, right=False, width=5, size=10 452 | ) 453 | axes[i].tick_params(axis="x", rotation=90, labelsize=48) 454 | axes[i].tick_params(axis="y", labelsize=48) 455 | [ 456 | axes[i].axvline(x, color="0.3", linestyle="--", linewidth=3) 457 | for x in [0.5, 1.5, 2.5] 458 | ] 459 | axes[i].grid(visible=True, which="major", axis="y", linewidth=3) 460 | axes[i].set(axisbelow=True) 461 | for axis in ["top", "bottom", "left", "right"]: 462 | axes[i].spines[axis].set_linewidth(5) # change width 463 | if i == 0: 464 | start2, end2 = axes[i].get_ylim() 465 | for axis in ["left", "right"]: 466 | axes[i].spines[axis].set_bounds( 467 | low=start2 - (end2 - start2) * 0.21153846153846154, 468 | high=end2 + (end2 - start2) * 0.06937799043062202, 469 | ) # Extend spine bounds 470 | 471 | f.tight_layout() 472 | f.subplots_adjust(wspace=0, hspace=0) 473 | plt.rcParams["svg.fonttype"] = "none" 474 | plt.savefig( 475 | "plots/" + plot_type.replace("_", "") + "plotmulti.pdf", bbox_inches="tight" 476 | ) 477 | plt.savefig( 478 | "plots/" + plot_type.replace("_", "") + "plotmulti.eps", bbox_inches="tight" 479 | ) 480 | # plt.show() 481 | -------------------------------------------------------------------------------- /LICENSE: -------------------------------------------------------------------------------- 1 | GNU GENERAL PUBLIC LICENSE 2 | Version 3, 29 June 2007 3 | 4 | Copyright (C) 2007 Free Software Foundation, Inc. 5 | Everyone is permitted to copy and distribute verbatim copies 6 | of this license document, but changing it is not allowed. 7 | 8 | Preamble 9 | 10 | The GNU General Public License is a free, copyleft license for 11 | software and other kinds of works. 12 | 13 | The licenses for most software and other practical works are designed 14 | to take away your freedom to share and change the works. By contrast, 15 | the GNU General Public License is intended to guarantee your freedom to 16 | share and change all versions of a program--to make sure it remains free 17 | software for all its users. We, the Free Software Foundation, use the 18 | GNU General Public License for most of our software; it applies also to 19 | any other work released this way by its authors. You can apply it to 20 | your programs, too. 21 | 22 | When we speak of free software, we are referring to freedom, not 23 | price. Our General Public Licenses are designed to make sure that you 24 | have the freedom to distribute copies of free software (and charge for 25 | them if you wish), that you receive source code or can get it if you 26 | want it, that you can change the software or use pieces of it in new 27 | free programs, and that you know you can do these things. 28 | 29 | To protect your rights, we need to prevent others from denying you 30 | these rights or asking you to surrender the rights. Therefore, you have 31 | certain responsibilities if you distribute copies of the software, or if 32 | you modify it: responsibilities to respect the freedom of others. 33 | 34 | For example, if you distribute copies of such a program, whether 35 | gratis or for a fee, you must pass on to the recipients the same 36 | freedoms that you received. You must make sure that they, too, receive 37 | or can get the source code. And you must show them these terms so they 38 | know their rights. 39 | 40 | Developers that use the GNU GPL protect your rights with two steps: 41 | (1) assert copyright on the software, and (2) offer you this License 42 | giving you legal permission to copy, distribute and/or modify it. 43 | 44 | For the developers' and authors' protection, the GPL clearly explains 45 | that there is no warranty for this free software. For both users' and 46 | authors' sake, the GPL requires that modified versions be marked as 47 | changed, so that their problems will not be attributed erroneously to 48 | authors of previous versions. 49 | 50 | Some devices are designed to deny users access to install or run 51 | modified versions of the software inside them, although the manufacturer 52 | can do so. This is fundamentally incompatible with the aim of 53 | protecting users' freedom to change the software. The systematic 54 | pattern of such abuse occurs in the area of products for individuals to 55 | use, which is precisely where it is most unacceptable. Therefore, we 56 | have designed this version of the GPL to prohibit the practice for those 57 | products. If such problems arise substantially in other domains, we 58 | stand ready to extend this provision to those domains in future versions 59 | of the GPL, as needed to protect the freedom of users. 60 | 61 | Finally, every program is threatened constantly by software patents. 62 | States should not allow patents to restrict development and use of 63 | software on general-purpose computers, but in those that do, we wish to 64 | avoid the special danger that patents applied to a free program could 65 | make it effectively proprietary. To prevent this, the GPL assures that 66 | patents cannot be used to render the program non-free. 67 | 68 | The precise terms and conditions for copying, distribution and 69 | modification follow. 70 | 71 | TERMS AND CONDITIONS 72 | 73 | 0. Definitions. 74 | 75 | "This License" refers to version 3 of the GNU General Public License. 76 | 77 | "Copyright" also means copyright-like laws that apply to other kinds of 78 | works, such as semiconductor masks. 79 | 80 | "The Program" refers to any copyrightable work licensed under this 81 | License. Each licensee is addressed as "you". "Licensees" and 82 | "recipients" may be individuals or organizations. 83 | 84 | To "modify" a work means to copy from or adapt all or part of the work 85 | in a fashion requiring copyright permission, other than the making of an 86 | exact copy. The resulting work is called a "modified version" of the 87 | earlier work or a work "based on" the earlier work. 88 | 89 | A "covered work" means either the unmodified Program or a work based 90 | on the Program. 91 | 92 | To "propagate" a work means to do anything with it that, without 93 | permission, would make you directly or secondarily liable for 94 | infringement under applicable copyright law, except executing it on a 95 | computer or modifying a private copy. Propagation includes copying, 96 | distribution (with or without modification), making available to the 97 | public, and in some countries other activities as well. 98 | 99 | To "convey" a work means any kind of propagation that enables other 100 | parties to make or receive copies. Mere interaction with a user through 101 | a computer network, with no transfer of a copy, is not conveying. 102 | 103 | An interactive user interface displays "Appropriate Legal Notices" 104 | to the extent that it includes a convenient and prominently visible 105 | feature that (1) displays an appropriate copyright notice, and (2) 106 | tells the user that there is no warranty for the work (except to the 107 | extent that warranties are provided), that licensees may convey the 108 | work under this License, and how to view a copy of this License. If 109 | the interface presents a list of user commands or options, such as a 110 | menu, a prominent item in the list meets this criterion. 111 | 112 | 1. Source Code. 113 | 114 | The "source code" for a work means the preferred form of the work 115 | for making modifications to it. "Object code" means any non-source 116 | form of a work. 117 | 118 | A "Standard Interface" means an interface that either is an official 119 | standard defined by a recognized standards body, or, in the case of 120 | interfaces specified for a particular programming language, one that 121 | is widely used among developers working in that language. 122 | 123 | The "System Libraries" of an executable work include anything, other 124 | than the work as a whole, that (a) is included in the normal form of 125 | packaging a Major Component, but which is not part of that Major 126 | Component, and (b) serves only to enable use of the work with that 127 | Major Component, or to implement a Standard Interface for which an 128 | implementation is available to the public in source code form. A 129 | "Major Component", in this context, means a major essential component 130 | (kernel, window system, and so on) of the specific operating system 131 | (if any) on which the executable work runs, or a compiler used to 132 | produce the work, or an object code interpreter used to run it. 133 | 134 | The "Corresponding Source" for a work in object code form means all 135 | the source code needed to generate, install, and (for an executable 136 | work) run the object code and to modify the work, including scripts to 137 | control those activities. However, it does not include the work's 138 | System Libraries, or general-purpose tools or generally available free 139 | programs which are used unmodified in performing those activities but 140 | which are not part of the work. For example, Corresponding Source 141 | includes interface definition files associated with source files for 142 | the work, and the source code for shared libraries and dynamically 143 | linked subprograms that the work is specifically designed to require, 144 | such as by intimate data communication or control flow between those 145 | subprograms and other parts of the work. 146 | 147 | The Corresponding Source need not include anything that users 148 | can regenerate automatically from other parts of the Corresponding 149 | Source. 150 | 151 | The Corresponding Source for a work in source code form is that 152 | same work. 153 | 154 | 2. Basic Permissions. 155 | 156 | All rights granted under this License are granted for the term of 157 | copyright on the Program, and are irrevocable provided the stated 158 | conditions are met. This License explicitly affirms your unlimited 159 | permission to run the unmodified Program. The output from running a 160 | covered work is covered by this License only if the output, given its 161 | content, constitutes a covered work. This License acknowledges your 162 | rights of fair use or other equivalent, as provided by copyright law. 163 | 164 | You may make, run and propagate covered works that you do not 165 | convey, without conditions so long as your license otherwise remains 166 | in force. You may convey covered works to others for the sole purpose 167 | of having them make modifications exclusively for you, or provide you 168 | with facilities for running those works, provided that you comply with 169 | the terms of this License in conveying all material for which you do 170 | not control copyright. Those thus making or running the covered works 171 | for you must do so exclusively on your behalf, under your direction 172 | and control, on terms that prohibit them from making any copies of 173 | your copyrighted material outside their relationship with you. 174 | 175 | Conveying under any other circumstances is permitted solely under 176 | the conditions stated below. Sublicensing is not allowed; section 10 177 | makes it unnecessary. 178 | 179 | 3. Protecting Users' Legal Rights From Anti-Circumvention Law. 180 | 181 | No covered work shall be deemed part of an effective technological 182 | measure under any applicable law fulfilling obligations under article 183 | 11 of the WIPO copyright treaty adopted on 20 December 1996, or 184 | similar laws prohibiting or restricting circumvention of such 185 | measures. 186 | 187 | When you convey a covered work, you waive any legal power to forbid 188 | circumvention of technological measures to the extent such circumvention 189 | is effected by exercising rights under this License with respect to 190 | the covered work, and you disclaim any intention to limit operation or 191 | modification of the work as a means of enforcing, against the work's 192 | users, your or third parties' legal rights to forbid circumvention of 193 | technological measures. 194 | 195 | 4. Conveying Verbatim Copies. 196 | 197 | You may convey verbatim copies of the Program's source code as you 198 | receive it, in any medium, provided that you conspicuously and 199 | appropriately publish on each copy an appropriate copyright notice; 200 | keep intact all notices stating that this License and any 201 | non-permissive terms added in accord with section 7 apply to the code; 202 | keep intact all notices of the absence of any warranty; and give all 203 | recipients a copy of this License along with the Program. 204 | 205 | You may charge any price or no price for each copy that you convey, 206 | and you may offer support or warranty protection for a fee. 207 | 208 | 5. Conveying Modified Source Versions. 209 | 210 | You may convey a work based on the Program, or the modifications to 211 | produce it from the Program, in the form of source code under the 212 | terms of section 4, provided that you also meet all of these conditions: 213 | 214 | a) The work must carry prominent notices stating that you modified 215 | it, and giving a relevant date. 216 | 217 | b) The work must carry prominent notices stating that it is 218 | released under this License and any conditions added under section 219 | 7. This requirement modifies the requirement in section 4 to 220 | "keep intact all notices". 221 | 222 | c) You must license the entire work, as a whole, under this 223 | License to anyone who comes into possession of a copy. This 224 | License will therefore apply, along with any applicable section 7 225 | additional terms, to the whole of the work, and all its parts, 226 | regardless of how they are packaged. This License gives no 227 | permission to license the work in any other way, but it does not 228 | invalidate such permission if you have separately received it. 229 | 230 | d) If the work has interactive user interfaces, each must display 231 | Appropriate Legal Notices; however, if the Program has interactive 232 | interfaces that do not display Appropriate Legal Notices, your 233 | work need not make them do so. 234 | 235 | A compilation of a covered work with other separate and independent 236 | works, which are not by their nature extensions of the covered work, 237 | and which are not combined with it such as to form a larger program, 238 | in or on a volume of a storage or distribution medium, is called an 239 | "aggregate" if the compilation and its resulting copyright are not 240 | used to limit the access or legal rights of the compilation's users 241 | beyond what the individual works permit. Inclusion of a covered work 242 | in an aggregate does not cause this License to apply to the other 243 | parts of the aggregate. 244 | 245 | 6. Conveying Non-Source Forms. 246 | 247 | You may convey a covered work in object code form under the terms 248 | of sections 4 and 5, provided that you also convey the 249 | machine-readable Corresponding Source under the terms of this License, 250 | in one of these ways: 251 | 252 | a) Convey the object code in, or embodied in, a physical product 253 | (including a physical distribution medium), accompanied by the 254 | Corresponding Source fixed on a durable physical medium 255 | customarily used for software interchange. 256 | 257 | b) Convey the object code in, or embodied in, a physical product 258 | (including a physical distribution medium), accompanied by a 259 | written offer, valid for at least three years and valid for as 260 | long as you offer spare parts or customer support for that product 261 | model, to give anyone who possesses the object code either (1) a 262 | copy of the Corresponding Source for all the software in the 263 | product that is covered by this License, on a durable physical 264 | medium customarily used for software interchange, for a price no 265 | more than your reasonable cost of physically performing this 266 | conveying of source, or (2) access to copy the 267 | Corresponding Source from a network server at no charge. 268 | 269 | c) Convey individual copies of the object code with a copy of the 270 | written offer to provide the Corresponding Source. This 271 | alternative is allowed only occasionally and noncommercially, and 272 | only if you received the object code with such an offer, in accord 273 | with subsection 6b. 274 | 275 | d) Convey the object code by offering access from a designated 276 | place (gratis or for a charge), and offer equivalent access to the 277 | Corresponding Source in the same way through the same place at no 278 | further charge. You need not require recipients to copy the 279 | Corresponding Source along with the object code. If the place to 280 | copy the object code is a network server, the Corresponding Source 281 | may be on a different server (operated by you or a third party) 282 | that supports equivalent copying facilities, provided you maintain 283 | clear directions next to the object code saying where to find the 284 | Corresponding Source. Regardless of what server hosts the 285 | Corresponding Source, you remain obligated to ensure that it is 286 | available for as long as needed to satisfy these requirements. 287 | 288 | e) Convey the object code using peer-to-peer transmission, provided 289 | you inform other peers where the object code and Corresponding 290 | Source of the work are being offered to the general public at no 291 | charge under subsection 6d. 292 | 293 | A separable portion of the object code, whose source code is excluded 294 | from the Corresponding Source as a System Library, need not be 295 | included in conveying the object code work. 296 | 297 | A "User Product" is either (1) a "consumer product", which means any 298 | tangible personal property which is normally used for personal, family, 299 | or household purposes, or (2) anything designed or sold for incorporation 300 | into a dwelling. In determining whether a product is a consumer product, 301 | doubtful cases shall be resolved in favor of coverage. For a particular 302 | product received by a particular user, "normally used" refers to a 303 | typical or common use of that class of product, regardless of the status 304 | of the particular user or of the way in which the particular user 305 | actually uses, or expects or is expected to use, the product. A product 306 | is a consumer product regardless of whether the product has substantial 307 | commercial, industrial or non-consumer uses, unless such uses represent 308 | the only significant mode of use of the product. 309 | 310 | "Installation Information" for a User Product means any methods, 311 | procedures, authorization keys, or other information required to install 312 | and execute modified versions of a covered work in that User Product from 313 | a modified version of its Corresponding Source. The information must 314 | suffice to ensure that the continued functioning of the modified object 315 | code is in no case prevented or interfered with solely because 316 | modification has been made. 317 | 318 | If you convey an object code work under this section in, or with, or 319 | specifically for use in, a User Product, and the conveying occurs as 320 | part of a transaction in which the right of possession and use of the 321 | User Product is transferred to the recipient in perpetuity or for a 322 | fixed term (regardless of how the transaction is characterized), the 323 | Corresponding Source conveyed under this section must be accompanied 324 | by the Installation Information. But this requirement does not apply 325 | if neither you nor any third party retains the ability to install 326 | modified object code on the User Product (for example, the work has 327 | been installed in ROM). 328 | 329 | The requirement to provide Installation Information does not include a 330 | requirement to continue to provide support service, warranty, or updates 331 | for a work that has been modified or installed by the recipient, or for 332 | the User Product in which it has been modified or installed. Access to a 333 | network may be denied when the modification itself materially and 334 | adversely affects the operation of the network or violates the rules and 335 | protocols for communication across the network. 336 | 337 | Corresponding Source conveyed, and Installation Information provided, 338 | in accord with this section must be in a format that is publicly 339 | documented (and with an implementation available to the public in 340 | source code form), and must require no special password or key for 341 | unpacking, reading or copying. 342 | 343 | 7. Additional Terms. 344 | 345 | "Additional permissions" are terms that supplement the terms of this 346 | License by making exceptions from one or more of its conditions. 347 | Additional permissions that are applicable to the entire Program shall 348 | be treated as though they were included in this License, to the extent 349 | that they are valid under applicable law. If additional permissions 350 | apply only to part of the Program, that part may be used separately 351 | under those permissions, but the entire Program remains governed by 352 | this License without regard to the additional permissions. 353 | 354 | When you convey a copy of a covered work, you may at your option 355 | remove any additional permissions from that copy, or from any part of 356 | it. (Additional permissions may be written to require their own 357 | removal in certain cases when you modify the work.) You may place 358 | additional permissions on material, added by you to a covered work, 359 | for which you have or can give appropriate copyright permission. 360 | 361 | Notwithstanding any other provision of this License, for material you 362 | add to a covered work, you may (if authorized by the copyright holders of 363 | that material) supplement the terms of this License with terms: 364 | 365 | a) Disclaiming warranty or limiting liability differently from the 366 | terms of sections 15 and 16 of this License; or 367 | 368 | b) Requiring preservation of specified reasonable legal notices or 369 | author attributions in that material or in the Appropriate Legal 370 | Notices displayed by works containing it; or 371 | 372 | c) Prohibiting misrepresentation of the origin of that material, or 373 | requiring that modified versions of such material be marked in 374 | reasonable ways as different from the original version; or 375 | 376 | d) Limiting the use for publicity purposes of names of licensors or 377 | authors of the material; or 378 | 379 | e) Declining to grant rights under trademark law for use of some 380 | trade names, trademarks, or service marks; or 381 | 382 | f) Requiring indemnification of licensors and authors of that 383 | material by anyone who conveys the material (or modified versions of 384 | it) with contractual assumptions of liability to the recipient, for 385 | any liability that these contractual assumptions directly impose on 386 | those licensors and authors. 387 | 388 | All other non-permissive additional terms are considered "further 389 | restrictions" within the meaning of section 10. If the Program as you 390 | received it, or any part of it, contains a notice stating that it is 391 | governed by this License along with a term that is a further 392 | restriction, you may remove that term. If a license document contains 393 | a further restriction but permits relicensing or conveying under this 394 | License, you may add to a covered work material governed by the terms 395 | of that license document, provided that the further restriction does 396 | not survive such relicensing or conveying. 397 | 398 | If you add terms to a covered work in accord with this section, you 399 | must place, in the relevant source files, a statement of the 400 | additional terms that apply to those files, or a notice indicating 401 | where to find the applicable terms. 402 | 403 | Additional terms, permissive or non-permissive, may be stated in the 404 | form of a separately written license, or stated as exceptions; 405 | the above requirements apply either way. 406 | 407 | 8. Termination. 408 | 409 | You may not propagate or modify a covered work except as expressly 410 | provided under this License. Any attempt otherwise to propagate or 411 | modify it is void, and will automatically terminate your rights under 412 | this License (including any patent licenses granted under the third 413 | paragraph of section 11). 414 | 415 | However, if you cease all violation of this License, then your 416 | license from a particular copyright holder is reinstated (a) 417 | provisionally, unless and until the copyright holder explicitly and 418 | finally terminates your license, and (b) permanently, if the copyright 419 | holder fails to notify you of the violation by some reasonable means 420 | prior to 60 days after the cessation. 421 | 422 | Moreover, your license from a particular copyright holder is 423 | reinstated permanently if the copyright holder notifies you of the 424 | violation by some reasonable means, this is the first time you have 425 | received notice of violation of this License (for any work) from that 426 | copyright holder, and you cure the violation prior to 30 days after 427 | your receipt of the notice. 428 | 429 | Termination of your rights under this section does not terminate the 430 | licenses of parties who have received copies or rights from you under 431 | this License. If your rights have been terminated and not permanently 432 | reinstated, you do not qualify to receive new licenses for the same 433 | material under section 10. 434 | 435 | 9. Acceptance Not Required for Having Copies. 436 | 437 | You are not required to accept this License in order to receive or 438 | run a copy of the Program. Ancillary propagation of a covered work 439 | occurring solely as a consequence of using peer-to-peer transmission 440 | to receive a copy likewise does not require acceptance. However, 441 | nothing other than this License grants you permission to propagate or 442 | modify any covered work. These actions infringe copyright if you do 443 | not accept this License. Therefore, by modifying or propagating a 444 | covered work, you indicate your acceptance of this License to do so. 445 | 446 | 10. Automatic Licensing of Downstream Recipients. 447 | 448 | Each time you convey a covered work, the recipient automatically 449 | receives a license from the original licensors, to run, modify and 450 | propagate that work, subject to this License. You are not responsible 451 | for enforcing compliance by third parties with this License. 452 | 453 | An "entity transaction" is a transaction transferring control of an 454 | organization, or substantially all assets of one, or subdividing an 455 | organization, or merging organizations. If propagation of a covered 456 | work results from an entity transaction, each party to that 457 | transaction who receives a copy of the work also receives whatever 458 | licenses to the work the party's predecessor in interest had or could 459 | give under the previous paragraph, plus a right to possession of the 460 | Corresponding Source of the work from the predecessor in interest, if 461 | the predecessor has it or can get it with reasonable efforts. 462 | 463 | You may not impose any further restrictions on the exercise of the 464 | rights granted or affirmed under this License. For example, you may 465 | not impose a license fee, royalty, or other charge for exercise of 466 | rights granted under this License, and you may not initiate litigation 467 | (including a cross-claim or counterclaim in a lawsuit) alleging that 468 | any patent claim is infringed by making, using, selling, offering for 469 | sale, or importing the Program or any portion of it. 470 | 471 | 11. Patents. 472 | 473 | A "contributor" is a copyright holder who authorizes use under this 474 | License of the Program or a work on which the Program is based. The 475 | work thus licensed is called the contributor's "contributor version". 476 | 477 | A contributor's "essential patent claims" are all patent claims 478 | owned or controlled by the contributor, whether already acquired or 479 | hereafter acquired, that would be infringed by some manner, permitted 480 | by this License, of making, using, or selling its contributor version, 481 | but do not include claims that would be infringed only as a 482 | consequence of further modification of the contributor version. For 483 | purposes of this definition, "control" includes the right to grant 484 | patent sublicenses in a manner consistent with the requirements of 485 | this License. 486 | 487 | Each contributor grants you a non-exclusive, worldwide, royalty-free 488 | patent license under the contributor's essential patent claims, to 489 | make, use, sell, offer for sale, import and otherwise run, modify and 490 | propagate the contents of its contributor version. 491 | 492 | In the following three paragraphs, a "patent license" is any express 493 | agreement or commitment, however denominated, not to enforce a patent 494 | (such as an express permission to practice a patent or covenant not to 495 | sue for patent infringement). To "grant" such a patent license to a 496 | party means to make such an agreement or commitment not to enforce a 497 | patent against the party. 498 | 499 | If you convey a covered work, knowingly relying on a patent license, 500 | and the Corresponding Source of the work is not available for anyone 501 | to copy, free of charge and under the terms of this License, through a 502 | publicly available network server or other readily accessible means, 503 | then you must either (1) cause the Corresponding Source to be so 504 | available, or (2) arrange to deprive yourself of the benefit of the 505 | patent license for this particular work, or (3) arrange, in a manner 506 | consistent with the requirements of this License, to extend the patent 507 | license to downstream recipients. "Knowingly relying" means you have 508 | actual knowledge that, but for the patent license, your conveying the 509 | covered work in a country, or your recipient's use of the covered work 510 | in a country, would infringe one or more identifiable patents in that 511 | country that you have reason to believe are valid. 512 | 513 | If, pursuant to or in connection with a single transaction or 514 | arrangement, you convey, or propagate by procuring conveyance of, a 515 | covered work, and grant a patent license to some of the parties 516 | receiving the covered work authorizing them to use, propagate, modify 517 | or convey a specific copy of the covered work, then the patent license 518 | you grant is automatically extended to all recipients of the covered 519 | work and works based on it. 520 | 521 | A patent license is "discriminatory" if it does not include within 522 | the scope of its coverage, prohibits the exercise of, or is 523 | conditioned on the non-exercise of one or more of the rights that are 524 | specifically granted under this License. You may not convey a covered 525 | work if you are a party to an arrangement with a third party that is 526 | in the business of distributing software, under which you make payment 527 | to the third party based on the extent of your activity of conveying 528 | the work, and under which the third party grants, to any of the 529 | parties who would receive the covered work from you, a discriminatory 530 | patent license (a) in connection with copies of the covered work 531 | conveyed by you (or copies made from those copies), or (b) primarily 532 | for and in connection with specific products or compilations that 533 | contain the covered work, unless you entered into that arrangement, 534 | or that patent license was granted, prior to 28 March 2007. 535 | 536 | Nothing in this License shall be construed as excluding or limiting 537 | any implied license or other defenses to infringement that may 538 | otherwise be available to you under applicable patent law. 539 | 540 | 12. No Surrender of Others' Freedom. 541 | 542 | If conditions are imposed on you (whether by court order, agreement or 543 | otherwise) that contradict the conditions of this License, they do not 544 | excuse you from the conditions of this License. If you cannot convey a 545 | covered work so as to satisfy simultaneously your obligations under this 546 | License and any other pertinent obligations, then as a consequence you may 547 | not convey it at all. For example, if you agree to terms that obligate you 548 | to collect a royalty for further conveying from those to whom you convey 549 | the Program, the only way you could satisfy both those terms and this 550 | License would be to refrain entirely from conveying the Program. 551 | 552 | 13. Use with the GNU Affero General Public License. 553 | 554 | Notwithstanding any other provision of this License, you have 555 | permission to link or combine any covered work with a work licensed 556 | under version 3 of the GNU Affero General Public License into a single 557 | combined work, and to convey the resulting work. The terms of this 558 | License will continue to apply to the part which is the covered work, 559 | but the special requirements of the GNU Affero General Public License, 560 | section 13, concerning interaction through a network will apply to the 561 | combination as such. 562 | 563 | 14. Revised Versions of this License. 564 | 565 | The Free Software Foundation may publish revised and/or new versions of 566 | the GNU General Public License from time to time. Such new versions will 567 | be similar in spirit to the present version, but may differ in detail to 568 | address new problems or concerns. 569 | 570 | Each version is given a distinguishing version number. If the 571 | Program specifies that a certain numbered version of the GNU General 572 | Public License "or any later version" applies to it, you have the 573 | option of following the terms and conditions either of that numbered 574 | version or of any later version published by the Free Software 575 | Foundation. If the Program does not specify a version number of the 576 | GNU General Public License, you may choose any version ever published 577 | by the Free Software Foundation. 578 | 579 | If the Program specifies that a proxy can decide which future 580 | versions of the GNU General Public License can be used, that proxy's 581 | public statement of acceptance of a version permanently authorizes you 582 | to choose that version for the Program. 583 | 584 | Later license versions may give you additional or different 585 | permissions. However, no additional obligations are imposed on any 586 | author or copyright holder as a result of your choosing to follow a 587 | later version. 588 | 589 | 15. Disclaimer of Warranty. 590 | 591 | THERE IS NO WARRANTY FOR THE PROGRAM, TO THE EXTENT PERMITTED BY 592 | APPLICABLE LAW. EXCEPT WHEN OTHERWISE STATED IN WRITING THE COPYRIGHT 593 | HOLDERS AND/OR OTHER PARTIES PROVIDE THE PROGRAM "AS IS" WITHOUT WARRANTY 594 | OF ANY KIND, EITHER EXPRESSED OR IMPLIED, INCLUDING, BUT NOT LIMITED TO, 595 | THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR 596 | PURPOSE. THE ENTIRE RISK AS TO THE QUALITY AND PERFORMANCE OF THE PROGRAM 597 | IS WITH YOU. SHOULD THE PROGRAM PROVE DEFECTIVE, YOU ASSUME THE COST OF 598 | ALL NECESSARY SERVICING, REPAIR OR CORRECTION. 599 | 600 | 16. Limitation of Liability. 601 | 602 | IN NO EVENT UNLESS REQUIRED BY APPLICABLE LAW OR AGREED TO IN WRITING 603 | WILL ANY COPYRIGHT HOLDER, OR ANY OTHER PARTY WHO MODIFIES AND/OR CONVEYS 604 | THE PROGRAM AS PERMITTED ABOVE, BE LIABLE TO YOU FOR DAMAGES, INCLUDING ANY 605 | GENERAL, SPECIAL, INCIDENTAL OR CONSEQUENTIAL DAMAGES ARISING OUT OF THE 606 | USE OR INABILITY TO USE THE PROGRAM (INCLUDING BUT NOT LIMITED TO LOSS OF 607 | DATA OR DATA BEING RENDERED INACCURATE OR LOSSES SUSTAINED BY YOU OR THIRD 608 | PARTIES OR A FAILURE OF THE PROGRAM TO OPERATE WITH ANY OTHER PROGRAMS), 609 | EVEN IF SUCH HOLDER OR OTHER PARTY HAS BEEN ADVISED OF THE POSSIBILITY OF 610 | SUCH DAMAGES. 611 | 612 | 17. Interpretation of Sections 15 and 16. 613 | 614 | If the disclaimer of warranty and limitation of liability provided 615 | above cannot be given local legal effect according to their terms, 616 | reviewing courts shall apply local law that most closely approximates 617 | an absolute waiver of all civil liability in connection with the 618 | Program, unless a warranty or assumption of liability accompanies a 619 | copy of the Program in return for a fee. 620 | 621 | END OF TERMS AND CONDITIONS 622 | 623 | How to Apply These Terms to Your New Programs 624 | 625 | If you develop a new program, and you want it to be of the greatest 626 | possible use to the public, the best way to achieve this is to make it 627 | free software which everyone can redistribute and change under these terms. 628 | 629 | To do so, attach the following notices to the program. It is safest 630 | to attach them to the start of each source file to most effectively 631 | state the exclusion of warranty; and each file should have at least 632 | the "copyright" line and a pointer to where the full notice is found. 633 | 634 | 635 | Copyright (C) 636 | 637 | This program is free software: you can redistribute it and/or modify 638 | it under the terms of the GNU General Public License as published by 639 | the Free Software Foundation, either version 3 of the License, or 640 | (at your option) any later version. 641 | 642 | This program is distributed in the hope that it will be useful, 643 | but WITHOUT ANY WARRANTY; without even the implied warranty of 644 | MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the 645 | GNU General Public License for more details. 646 | 647 | You should have received a copy of the GNU General Public License 648 | along with this program. If not, see . 649 | 650 | Also add information on how to contact you by electronic and paper mail. 651 | 652 | If the program does terminal interaction, make it output a short 653 | notice like this when it starts in an interactive mode: 654 | 655 | Copyright (C) 656 | This program comes with ABSOLUTELY NO WARRANTY; for details type `show w'. 657 | This is free software, and you are welcome to redistribute it 658 | under certain conditions; type `show c' for details. 659 | 660 | The hypothetical commands `show w' and `show c' should show the appropriate 661 | parts of the General Public License. Of course, your program's commands 662 | might be different; for a GUI interface, you would use an "about box". 663 | 664 | You should also get your employer (if you work as a programmer) or school, 665 | if any, to sign a "copyright disclaimer" for the program, if necessary. 666 | For more information on this, and how to apply and follow the GNU GPL, see 667 | . 668 | 669 | The GNU General Public License does not permit incorporating your program 670 | into proprietary programs. If your program is a subroutine library, you 671 | may consider it more useful to permit linking proprietary applications with 672 | the library. If this is what you want to do, use the GNU Lesser General 673 | Public License instead of this License. But first, please read 674 | . 675 | -------------------------------------------------------------------------------- /run_all_models.py: -------------------------------------------------------------------------------- 1 | # Get time samples and seed from arguments passed when this script is called 2 | import sys 3 | 4 | network_to_check = sys.argv[1] 5 | n_time_samples_str = sys.argv[2] 6 | n_time_samples_int = int(n_time_samples_str) 7 | seed = int(sys.argv[3]) 8 | 9 | import os 10 | 11 | # Set environment variables useful for multiprocessing. 12 | # This will only work if your BLAS is MKL or openblas. 13 | # If using other BLAS implementations, these have no effect 14 | os.environ["MKL_NUM_THREADS"] = "1" 15 | os.environ["OPENBLAS_NUM_THREADS"] = "1" 16 | 17 | os.makedirs("plots", exist_ok=True) 18 | os.makedirs("pickled_data", exist_ok=True) 19 | 20 | import warnings 21 | 22 | warnings.simplefilter(action="ignore", category=FutureWarning) 23 | warnings.simplefilter(action="ignore", category=UserWarning) 24 | 25 | import numpy as np 26 | 27 | np.seterr(all="raise") 28 | 29 | import pandas as pd 30 | 31 | import cao_min_embedding_dimension 32 | import mlcausality 33 | 34 | from utils_mod import lsNGC as granger 35 | from recovery_performance_mod import recovery_performance 36 | 37 | from brier_loss import brier_loss_func 38 | 39 | import causal_ccm 40 | 41 | from tigramite import data_processing as pcmci_data_processing 42 | from tigramite.pcmci import PCMCI 43 | from tigramite.independence_tests.parcorr import ParCorr 44 | 45 | import tqdm 46 | 47 | import matplotlib.pyplot as plt 48 | 49 | from copy import deepcopy 50 | import itertools 51 | from pprint import pprint 52 | import pickle 53 | 54 | import networkx as nx 55 | 56 | from sklearn.metrics import ( 57 | confusion_matrix, 58 | accuracy_score, 59 | balanced_accuracy_score, 60 | f1_score, 61 | ) 62 | 63 | # Get number of cpu cores 64 | import psutil 65 | 66 | num_cores = psutil.cpu_count(logical=True) 67 | # Import multriprocessing plugins from tqdm 68 | # These rely on concurrent.futures 69 | from tqdm.contrib.concurrent import ensure_lock, process_map 70 | 71 | # Set seed. 72 | np.random.seed(seed) 73 | 74 | if network_to_check == "5_linear": 75 | # from spectral_connectivity.simulate import simulate_MVAR 76 | def simulate_MVAR( 77 | coefficients, 78 | noise_covariance=None, 79 | n_time_samples=100, 80 | n_trials=1, 81 | n_burnin_samples=100, 82 | ): 83 | """ 84 | Simulate multivariate autoregressive (MVAR) process. 85 | Parameters 86 | ---------- 87 | coefficients : array, shape (n_time_samples, n_lags, n_signals, n_signals) 88 | noise_covariance : array, shape (n_signals, n_signals) 89 | Returns 90 | ------- 91 | time_series : array, shape (n_time_samples - n_burnin_samples, 92 | n_trials, n_signals) 93 | """ 94 | n_lags, n_signals, _ = coefficients.shape 95 | if noise_covariance is None: 96 | noise_covariance = np.eye(n_signals) 97 | time_series = np.random.multivariate_normal( 98 | np.zeros((n_signals,)), 99 | noise_covariance, 100 | size=(n_time_samples + n_burnin_samples, n_trials), 101 | ) 102 | 103 | for time_ind in np.arange(n_lags, n_time_samples + n_burnin_samples): 104 | for lag_ind in np.arange(n_lags): 105 | time_series[time_ind, ...] += np.matmul( 106 | coefficients[np.newaxis, np.newaxis, lag_ind, ...], 107 | time_series[time_ind - (lag_ind + 1), ..., np.newaxis], 108 | ).squeeze() 109 | return time_series[n_burnin_samples:, ...] 110 | 111 | # Generate data using the spectral_connectivity package 112 | 113 | def baccala_example3(): 114 | """Baccalá, L.A., and Sameshima, K. (2001). Partial directed coherence: 115 | a new concept in neural structure determination. Biological 116 | Cybernetics 84, 463–474. 117 | """ 118 | np.random.seed(seed) 119 | n_time_samples, n_lags, n_signals = n_time_samples_int, 3, 5 120 | coefficients = np.zeros((n_lags, n_signals, n_signals)) 121 | 122 | coefficients[0, 0, 0] = 0.95 * np.sqrt(2) 123 | coefficients[1, 0, 0] = -0.9025 124 | 125 | coefficients[1, 1, 0] = 0.50 126 | coefficients[2, 2, 0] = -0.40 127 | 128 | coefficients[1, 3, 0] = -0.5 129 | coefficients[0, 3, 3] = 0.5 * np.sqrt(2) 130 | coefficients[0, 3, 4] = 0.25 * np.sqrt(2) 131 | 132 | coefficients[0, 4, 3] = -0.5 * np.sqrt(2) 133 | coefficients[0, 4, 4] = 0.5 * np.sqrt(2) 134 | 135 | noise_covariance = None 136 | 137 | return ( 138 | simulate_MVAR( 139 | coefficients, 140 | noise_covariance=noise_covariance, 141 | n_time_samples=n_time_samples, 142 | n_trials=50, 143 | n_burnin_samples=500, 144 | ), 145 | coefficients, 146 | ) 147 | 148 | data, coefficients = baccala_example3() 149 | 150 | data_list = [data[:, i, :] for i in range(data.shape[1])] 151 | 152 | # Find the adjacency matrix (which are essentially the true labels) 153 | adjacency_matrix = np.zeros((coefficients.shape[1], coefficients.shape[2])) 154 | for i in range(coefficients.shape[0]): 155 | adjacency_matrix += np.ceil(np.abs(coefficients[i])) 156 | 157 | adjacency_matrix = np.clip(adjacency_matrix, 0, 1) 158 | np.fill_diagonal(adjacency_matrix, 0) 159 | adjacency_matrix = adjacency_matrix.T.astype(int) 160 | true_labels = adjacency_matrix 161 | elif network_to_check == "5_nonlinear": 162 | # 5-node nonlinear network 163 | true_labels = np.array( 164 | [ 165 | [0, 1, 1, 1, 0], 166 | [0, 0, 0, 0, 0], 167 | [0, 0, 0, 0, 0], 168 | [0, 0, 0, 0, 1], 169 | [0, 0, 0, 1, 0], 170 | ] 171 | ) 172 | n_burnin_samples = 500 173 | n_trials = 50 174 | n_lags = 3 175 | data_list = [] 176 | while True: 177 | np.random.seed(seed) 178 | seed += 1 179 | data = np.random.normal(0, 1, [n_lags, 5]) 180 | try: 181 | for j in range(n_lags, n_time_samples_int + n_burnin_samples): 182 | x1 = ( 183 | 0.95 * np.sqrt(2) * data[-1, 0] 184 | - 0.9025 * data[-2, 0] 185 | + np.random.normal(0, 1) 186 | ) 187 | x2 = 0.5 * data[-2, 0] ** 2 + np.random.normal(0, 1) 188 | x3 = -0.4 * data[-3, 0] + np.random.normal(0, 1) 189 | x4 = ( 190 | -0.5 * data[-2, 0] ** 2 191 | + 0.5 * np.sqrt(2) * data[-1, 3] 192 | + 0.25 * np.sqrt(2) * data[-1, 4] 193 | + np.random.normal(0, 1) 194 | ) 195 | x5 = ( 196 | -0.5 * np.sqrt(2) * data[-1, 3] 197 | + 0.5 * np.sqrt(2) * data[-1, 4] 198 | + np.random.normal(0, 1) 199 | ) 200 | data = np.vstack([data, np.array([x1, x2, x3, x4, x5])]) 201 | except Exception: 202 | continue 203 | data = np.array(data[n_burnin_samples:]) 204 | data_list.append(data) 205 | if len(data_list) == n_trials: 206 | break 207 | elif network_to_check == "7_nonlinear": 208 | # 7-node nonlinear network 209 | true_labels = np.array( 210 | [ 211 | [0, 1, 0, 0, 0, 1, 1], 212 | [0, 0, 1, 0, 0, 0, 0], 213 | [0, 0, 0, 1, 0, 1, 0], 214 | [0, 0, 0, 0, 0, 0, 0], 215 | [0, 0, 0, 0, 0, 0, 0], 216 | [0, 0, 0, 0, 1, 0, 1], 217 | [0, 0, 0, 1, 0, 0, 0], 218 | ] 219 | ) 220 | n_burnin_samples = 500 221 | n_trials = 50 222 | n_lags = 3 223 | data_list = [] 224 | while True: 225 | np.random.seed(seed) 226 | seed += 1 227 | data = np.random.normal(0, 1, [n_lags, 7]) 228 | try: 229 | for j in range(n_lags, n_time_samples_int + n_burnin_samples): 230 | x1 = ( 231 | 0.95 * np.sqrt(2) * data[-1, 0] 232 | - 0.9025 * data[-2, 0] 233 | + np.random.normal(0, 1) 234 | ) 235 | x2 = ( 236 | -0.04 * data[-3, 0] ** 3 237 | + 0.04 * data[-1, 0] ** 3 238 | + np.random.normal(0, 1) 239 | ) 240 | x3 = ( 241 | -0.04 * np.sqrt(2) * data[-1, 1] ** 3 242 | + 0.04 * np.sqrt(2) * data[-2, 1] ** 3 243 | + np.random.normal(0, 1) 244 | ) 245 | x4 = ( 246 | np.log1p(np.abs(data[-1, 2])) * np.sign(data[-1, 2]) 247 | + 0.001 * data[-2, 6] ** 3 248 | - 0.001 * data[-3, 6] ** 3 249 | + np.random.normal(0, 1) 250 | ) 251 | x5 = np.clip(np.random.normal(0, 1), -1, 1) * 0.04 * data[ 252 | -2, 5 253 | ] ** 5 + np.random.normal(0, 1) 254 | x6 = ( 255 | 0.04 * data[-2, 0] ** 3 256 | + 0.04 * data[-1, 2] ** 3 257 | + np.random.normal(0, 1) 258 | ) 259 | x7 = np.clip(np.random.normal(0, 1), -0.5, 0.5) * ( 260 | 0.04 * data[-2, 0] ** 3 261 | + 0.1 * data[-1, 5] ** 2 262 | - 0.1 * data[-2, 5] ** 2 263 | ) + np.random.normal(0, 1) 264 | data = np.vstack([data, np.array([x1, x2, x3, x4, x5, x6, x7])]) 265 | except Exception: 266 | continue 267 | data = np.array(data[n_burnin_samples:]) 268 | data_list.append(data) 269 | if len(data_list) == n_trials: 270 | break 271 | elif network_to_check == "9_nonlinear": 272 | # 9-node nonlinear network 273 | true_labels = np.array( 274 | [ 275 | [0, 1, 1, 1, 0, 0, 0, 1, 1], 276 | [0, 0, 0, 0, 0, 0, 0, 0, 0], 277 | [0, 0, 0, 0, 0, 0, 0, 1, 0], 278 | [0, 0, 0, 0, 1, 1, 0, 0, 0], 279 | [0, 0, 0, 1, 0, 0, 0, 0, 0], 280 | [0, 0, 0, 0, 0, 0, 1, 0, 0], 281 | [0, 0, 0, 0, 0, 0, 0, 0, 0], 282 | [0, 0, 0, 0, 0, 0, 0, 0, 1], 283 | [0, 0, 0, 0, 0, 0, 0, 0, 0], 284 | ] 285 | ) 286 | n_burnin_samples = 500 287 | n_trials = 50 288 | n_lags = 3 289 | data_list = [] 290 | while True: 291 | np.random.seed(seed) 292 | seed += 1 293 | data = np.random.normal(0, 1, [n_lags, 9]) 294 | try: 295 | for j in range(n_lags, n_time_samples_int + n_burnin_samples): 296 | x1 = ( 297 | 0.95 * np.sqrt(2) * data[-1, 0] 298 | - 0.9025 * data[-2, 0] 299 | + np.random.normal(0, 1) 300 | ) 301 | x2 = ( 302 | 0.5 * data[-2, 0] ** 2 303 | + 0.5 * data[-1, 1] 304 | - 0.4 * data[-2, 1] 305 | + np.random.normal(0, 1) 306 | ) 307 | x3 = ( 308 | -0.4 * data[-3, 0] 309 | + 0.5 * data[-1, 2] 310 | - 0.4 * data[-2, 2] 311 | + np.random.normal(0, 1) 312 | ) 313 | x4 = ( 314 | -0.5 * data[-2, 0] ** 2 315 | + 0.5 * data[-1, 3] 316 | - 0.4 * data[-2, 3] 317 | + 0.5 * np.sqrt(2) * data[-1, 3] 318 | + 0.25 * np.sqrt(2) * data[-1, 4] 319 | + np.random.normal(0, 1) 320 | ) 321 | x5 = ( 322 | -0.5 * np.sqrt(2) * data[-1, 3] 323 | + 0.5 * np.sqrt(2) * data[-1, 4] 324 | + np.random.normal(0, 1) 325 | ) 326 | x6 = ( 327 | np.log1p(np.abs(data[-1, 3])) * np.sign(data[-1, 3]) 328 | + 0.5 * data[-1, 5] 329 | - 0.4 * data[-2, 5] 330 | + np.random.normal(0, 1) 331 | ) 332 | x7 = ( 333 | np.clip(np.random.normal(0, 1), -1, 1) * 0.04 * data[-2, 5] ** 5 334 | + 0.5 * data[-1, 6] 335 | - 0.4 * data[-2, 6] 336 | + np.random.normal(0, 1) 337 | ) 338 | x8 = ( 339 | 0.4 * data[-2, 0] 340 | + 0.25 * data[-1, 2] ** 3 341 | + 0.5 * data[-1, 7] 342 | - 0.4 * data[-2, 7] 343 | + np.random.normal(0, 1) 344 | ) 345 | x9 = ( 346 | np.clip(np.random.normal(0, 1), -0.5, 0.5) 347 | * ( 348 | 0.2 * data[-2, 0] 349 | + 0.1 * data[-1, 7] ** 2 350 | - 0.1 * data[-2, 7] ** 2 351 | ) 352 | + 0.5 * data[-1, 8] 353 | - 0.4 * data[-2, 8] 354 | + np.random.normal(0, 1) 355 | ) 356 | data = np.vstack([data, np.array([x1, x2, x3, x4, x5, x6, x7, x8, x9])]) 357 | except Exception: 358 | continue 359 | data = np.array(data[n_burnin_samples:]) 360 | data_list.append(data) 361 | if len(data_list) == n_trials: 362 | break 363 | elif network_to_check == "11_nonlinear": 364 | # 11-node nonlinear network 365 | true_labels = np.array( 366 | [ 367 | [0, 1, 0, 0, 0, 0, 0, 1, 1, 1, 0], 368 | [0, 0, 1, 1, 0, 0, 0, 0, 0, 1, 1], 369 | [0, 0, 0, 0, 0, 0, 0, 1, 0, 1, 0], 370 | [0, 0, 0, 0, 1, 1, 0, 0, 0, 0, 0], 371 | [0, 0, 0, 1, 0, 0, 0, 0, 0, 0, 0], 372 | [0, 0, 0, 0, 0, 0, 1, 0, 0, 0, 1], 373 | [0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0], 374 | [0, 0, 0, 0, 0, 0, 0, 0, 1, 0, 0], 375 | [0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0], 376 | [0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0], 377 | [0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0], 378 | ] 379 | ) 380 | n_burnin_samples = 500 381 | n_trials = 50 382 | n_lags = 3 383 | data_list = [] 384 | while True: 385 | np.random.seed(seed) 386 | seed += 1 387 | data = np.random.normal(0, 1, [n_lags, 11]) 388 | try: 389 | for j in range(n_lags, n_time_samples_int + n_burnin_samples): 390 | x1 = ( 391 | 0.25 * data[-1, 0] ** 2 392 | - 0.25 * data[-2, 0] ** 2 393 | + np.random.normal(0, 1) 394 | ) 395 | x2 = np.log1p(np.abs(data[-2, 0])) * np.sign( 396 | data[-2, 0] 397 | ) + np.random.normal(0, 1) 398 | x3 = -0.1 * data[-3, 1] ** 3 + np.random.normal(0, 1) 399 | x4 = ( 400 | -0.5 * data[-2, 1] ** 2 401 | + 0.5 * np.sqrt(2) * data[-1, 3] 402 | + 0.25 * np.sqrt(2) * data[-1, 4] 403 | + np.random.normal(0, 1) 404 | ) 405 | x5 = ( 406 | -0.5 * np.sqrt(2) * data[-1, 3] 407 | + 0.5 * np.sqrt(2) * data[-1, 4] 408 | + np.random.normal(0, 1) 409 | ) 410 | x6 = np.log1p(np.abs(data[-1, 3])) * np.sign( 411 | data[-1, 3] 412 | ) + np.random.normal(0, 1) 413 | x7 = np.clip(np.random.normal(0, 1), -1, 1) * 0.04 * data[ 414 | -2, 5 415 | ] ** 5 + np.random.normal(0, 1) 416 | x8 = ( 417 | 0.4 * data[-2, 0] + 0.25 * data[-1, 2] ** 3 + np.random.normal(0, 1) 418 | ) 419 | x9 = np.clip(np.random.normal(0, 1), -0.5, 0.5) * ( 420 | 0.2 * data[-2, 0] + 0.1 * data[-1, 7] ** 2 - 0.1 * data[-2, 7] ** 2 421 | ) + np.random.normal(0, 1) 422 | x10 = ( 423 | 0.25 * data[-3, 0] ** 2 424 | - 0.01 * data[-3, 1] ** 2 425 | + 0.15 * data[-3, 2] ** 3 426 | + np.random.normal(0, 1) 427 | ) 428 | x11 = ( 429 | 0.1 * data[-1, 1] ** 4 430 | - 0.1 * data[-2, 1] ** 4 431 | + 0.1 * data[-3, 5] ** 3 432 | + np.random.normal(0, 1) 433 | ) 434 | data = np.vstack( 435 | [data, np.array([x1, x2, x3, x4, x5, x6, x7, x8, x9, x10, x11])] 436 | ) 437 | except Exception: 438 | continue 439 | data = np.array(data[n_burnin_samples:]) 440 | data_list.append(data) 441 | if len(data_list) == n_trials: 442 | break 443 | elif network_to_check == "34_zachary1": 444 | # Generate data 445 | n_burnin_samples = 500 446 | n_trials = 50 447 | n_lags = 1 448 | true_labels_list = [] 449 | data_list = [] 450 | while True: 451 | np.random.seed(seed) 452 | seed += 1 453 | # Load the Zachary karate club dataset 454 | G_orig = nx.karate_club_graph() 455 | adjacency_matrix = nx.adjacency_matrix(G_orig).todense() 456 | 457 | true_labels = np.clip(adjacency_matrix, 0, 1).astype(int) 458 | true_labels_list.append(true_labels) 459 | 460 | data = 0.01 * np.random.normal(0, 1, [n_lags, true_labels.shape[0]]) 461 | try: 462 | for t in range(n_lags, n_time_samples_int + n_burnin_samples): 463 | data_new = np.zeros_like(data[[-1]]) 464 | for i in range(data_new.shape[1]): 465 | data_new[0, i] = ( 466 | (1 - 0.025 * true_labels[:, i].sum()) 467 | * (1 - 1.8 * data[-1, i] ** 2) 468 | + np.matmul( 469 | 0.025 * true_labels[:, i], (1 - 1.8 * np.square(data[-1])) 470 | ) 471 | + 0.01 * np.random.normal(0, 1) 472 | ) 473 | data = np.vstack([data, data_new]) 474 | except Exception: 475 | continue 476 | data = np.array(data[n_burnin_samples:]) 477 | data_list.append(data) 478 | if len(data_list) == n_trials: 479 | break 480 | elif network_to_check == "34_zachary2": 481 | # Generate data 482 | n_burnin_samples = 500 483 | n_trials = 50 484 | n_lags = 1 485 | true_labels_list = [] 486 | data_list = [] 487 | while True: 488 | np.random.seed(seed) 489 | seed += 1 490 | # Load the Zachary karate club dataset 491 | G_orig = nx.karate_club_graph() 492 | adjacency_matrix = nx.adjacency_matrix(G_orig).todense() 493 | comb_list = deepcopy( 494 | list(itertools.combinations(range(adjacency_matrix.shape[0]), 2)) 495 | ) 496 | list_of_edges = [(i, j) for (i, j) in comb_list if adjacency_matrix[i, j] != 0] 497 | # Randomly select 5 edges to keep as bidirectional 498 | idx_bidirectional = np.random.choice(range(len(list_of_edges)), 5) 499 | edges_nonbidirectional = [ 500 | list_of_edges[i] 501 | for i in range(len(list_of_edges)) 502 | if i not in idx_bidirectional 503 | ] 504 | # Randomly assign a direction 505 | for a, b in edges_nonbidirectional: 506 | if np.random.random() < 0.5: 507 | adjacency_matrix[a, b] = 0 508 | else: 509 | adjacency_matrix[b, a] = 0 510 | 511 | true_labels = np.clip(adjacency_matrix, 0, 1).astype(int) 512 | true_labels_list.append(true_labels) 513 | 514 | data = 0.01 * np.random.normal(0, 1, [n_lags, true_labels.shape[0]]) 515 | try: 516 | for t in range(n_lags, n_time_samples_int + n_burnin_samples): 517 | data_new = np.zeros_like(data[[-1]]) 518 | for i in range(data_new.shape[1]): 519 | data_new[0, i] = ( 520 | (1 - 0.05 * true_labels[:, i].sum()) 521 | * (1 - 1.8 * data[-1, i] ** 2) 522 | + np.matmul( 523 | 0.05 * true_labels[:, i], (1 - 1.8 * np.square(data[-1])) 524 | ) 525 | + 0.01 * np.random.normal(0, 1) 526 | ) 527 | data = np.vstack([data, data_new]) 528 | except Exception as e: 529 | print(e) 530 | continue 531 | data = np.array(data[n_burnin_samples:]) 532 | data_list.append(data) 533 | if len(data_list) == n_trials: 534 | break 535 | 536 | 537 | if network_to_check != "34_zachary1" and network_to_check != "34_zachary2": 538 | true_labels_list = [true_labels for i in range(len(data_list))] 539 | 540 | 541 | G = nx.from_numpy_array(true_labels_list[0], create_using=nx.DiGraph) 542 | if network_to_check != "34_zachary1" and network_to_check != "34_zachary2": 543 | G_pos = nx.kamada_kawai_layout(G) 544 | else: 545 | G_pos = nx.circular_layout(G) 546 | 547 | G_d = dict(G.degree) 548 | G_labels = {node: str(node + 1) for node in G_d.keys()} 549 | nx.draw( 550 | G, 551 | G_pos, 552 | labels=G_labels, 553 | with_labels=True, 554 | node_color="white", 555 | node_size=750, 556 | edgecolors="black", 557 | linewidths=2, 558 | arrows=True, 559 | arrowstyle="->", 560 | connectionstyle="arc3, rad = 0.2", 561 | ) 562 | plt.rcParams["svg.fonttype"] = "none" 563 | if n_time_samples_str == "500": 564 | if network_to_check == "5_linear": 565 | plt.savefig("plots/5_linear_nonlinear_network.pdf", bbox_inches="tight") 566 | plt.savefig("plots/5_linear_nonlinear_network.eps", bbox_inches="tight") 567 | elif network_to_check == "5_nonlinear": 568 | pass 569 | else: 570 | plt.savefig("plots/" + network_to_check + "_network.pdf", bbox_inches="tight") 571 | plt.savefig("plots/" + network_to_check + "_network.eps", bbox_inches="tight") 572 | # plt.show(block=False) 573 | # plt.pause(0.001) 574 | 575 | 576 | # Cao's minimum embedding dimension 577 | pprint("### Find Cao's minimum embedding dimension ###") 578 | 579 | 580 | def caoloop(d): 581 | d_list = [[] for i in range(d.shape[1])] 582 | E1_list = [[] for i in range(d.shape[1])] 583 | E2_list = [[] for i in range(d.shape[1])] 584 | embedding_dim_list = [[] for i in range(d.shape[1])] 585 | for c in range(d.shape[1]): 586 | ( 587 | cur_d, 588 | cur_E1, 589 | cur_E2, 590 | embedding_dim, 591 | ) = cao_min_embedding_dimension.cao_min_embedding_dimension( 592 | d[:, c], maxd=20, tau=1, threshold=0.8, max_relative_change=0.3 593 | ) 594 | d_list[c].append(cur_d) 595 | E1_list[c].append(cur_E1) 596 | E2_list[c].append(cur_E2) 597 | embedding_dim_list[c].append(embedding_dim) 598 | return [d_list, E1_list, E2_list, embedding_dim_list] 599 | 600 | 601 | cao_loop_output = process_map(caoloop, data_list, max_workers=num_cores) 602 | # Ensure process_map finished before continuing: 603 | # the following will not pass until the lock can be acquired 604 | with ensure_lock(tqdm.auto.tqdm, lock_name="mp_lock") as lk: 605 | pass 606 | 607 | # Extract output into lists of lists for easier processing downstream 608 | d_list = [[] for i in range(data_list[0].shape[1])] 609 | E1_list = [[] for i in range(data_list[0].shape[1])] 610 | E2_list = [[] for i in range(data_list[0].shape[1])] 611 | embedding_dim_list = [[] for i in range(data_list[0].shape[1])] 612 | for o in range(len(cao_loop_output)): 613 | for v in range(len(cao_loop_output[0][0])): 614 | d_list[v].append(cao_loop_output[o][0][v][0]) 615 | E1_list[v].append(cao_loop_output[o][1][v][0]) 616 | E2_list[v].append(cao_loop_output[o][2][v][0]) 617 | embedding_dim_list[v].append(cao_loop_output[o][3][v][0]) 618 | 619 | 620 | min_embed_dim = round(np.nanmax(np.array(embedding_dim_list).flatten())) 621 | 622 | pprint("Minimum embedding dimension: " + str(min_embed_dim)) 623 | 624 | 625 | # Function to calculate metrics 626 | def calc_metrics(pvalue_matricies, thresh): 627 | # Check sensitivity and specificity 628 | sensitivities = [] 629 | specificities = [] 630 | accuracies = [] 631 | balanced_accuracies = [] 632 | f1wgt = [] 633 | listlen = deepcopy(len(pvalue_matricies)) 634 | for i in range(listlen): 635 | preds = deepcopy(np.array(pvalue_matricies[i])).astype(float) 636 | np.fill_diagonal(preds, np.nan) 637 | preds = preds.flatten() 638 | preds = preds[~np.isnan(preds)] 639 | preds = (preds < thresh).astype(int) 640 | true_labels_curr = deepcopy(true_labels_list[i]).astype(float) 641 | np.fill_diagonal(true_labels_curr, np.nan) 642 | true_labels_curr = true_labels_curr.flatten() 643 | true_labels_curr = true_labels_curr[~np.isnan(true_labels_curr)] 644 | true_labels_curr = true_labels_curr.astype(int) 645 | tn, fp, fn, tp = confusion_matrix( 646 | true_labels_curr.flatten(), preds.flatten() 647 | ).ravel() 648 | sensitivity = tp / (tp + fn) 649 | sensitivities.append(sensitivity) 650 | specificity = tn / (tn + fp) 651 | specificities.append(specificity) 652 | accuracy = accuracy_score(true_labels_curr.flatten(), preds.flatten()) 653 | accuracies.append(accuracy) 654 | balanced_accuracy = balanced_accuracy_score( 655 | true_labels_curr.flatten(), preds.flatten() 656 | ) 657 | balanced_accuracies.append(balanced_accuracy) 658 | f1 = f1_score(true_labels_curr.flatten(), preds.flatten(), average="weighted") 659 | f1wgt.append(f1) 660 | return sensitivities, specificities, accuracies, balanced_accuracies, f1wgt 661 | 662 | 663 | # Run mlcausality: 664 | pprint("### Run mlcausality ###") 665 | # Note: use minimum embedding dimension - 1 for lag if no logdiff is used and 666 | # dimension - 2 for lag if logdiff is used 667 | lag = max(min_embed_dim - 1, 1) 668 | 669 | 670 | def locoloop(d): 671 | mlcausality_output = mlcausality.multiloco_mlcausality( 672 | d, lags=[lag], scaler_init_1="quantile", return_pvalue_matrix_only=True 673 | ) 674 | return mlcausality_output 675 | 676 | 677 | all_mlcausality_output = process_map(locoloop, data_list, max_workers=num_cores) 678 | # Ensure process_map finished before continuing: 679 | # the following will not pass until the lock can be acquired 680 | with ensure_lock(tqdm.auto.tqdm, lock_name="mp_lock") as lk: 681 | pass 682 | 683 | mlcausality_pvalue_matricies = all_mlcausality_output 684 | mlcausality_adjacency_matricies = [1 - i for i in mlcausality_pvalue_matricies] 685 | 686 | mlcausality_recovery_performance = recovery_performance( 687 | mlcausality_adjacency_matricies, true_labels_list 688 | ) 689 | pprint("mlcausality recovery performance:") 690 | pprint("Mean: " + str(mlcausality_recovery_performance.mean())) 691 | pprint( 692 | pd.DataFrame(mlcausality_recovery_performance).quantile([0, 0.25, 0.5, 0.75, 1]).T 693 | ) 694 | 695 | mlcausality_brier_score_loss = brier_loss_func( 696 | mlcausality_pvalue_matricies, true_labels_list 697 | ) 698 | pprint("mlcausality brier score:") 699 | pprint("Mean: " + str(mlcausality_brier_score_loss.mean())) 700 | pprint(pd.DataFrame(mlcausality_brier_score_loss).quantile([0, 0.25, 0.5, 0.75, 1]).T) 701 | 702 | # Check sensitivity and specificity for the mlcausality model 703 | mlc_sens, mlc_spec, mlc_acc, mlc_bal_acc, mlc_f1wgt = calc_metrics( 704 | mlcausality_pvalue_matricies, 0.05 705 | ) 706 | 707 | 708 | pprint("mlcausality sensitivities:") 709 | pprint(pd.DataFrame(mlc_sens).quantile([0, 0.25, 0.5, 0.75, 1]).T) 710 | pprint("mlcausality specificities:") 711 | pprint(pd.DataFrame(mlc_spec).quantile([0, 0.25, 0.5, 0.75, 1]).T) 712 | pprint("mlcausality (sensitivities*specificities)**0.5:") 713 | pprint( 714 | pd.DataFrame((np.array(mlc_sens) * np.array(mlc_spec)) ** 0.5) 715 | .quantile([0, 0.25, 0.5, 0.75, 1]) 716 | .T 717 | ) 718 | pprint("mlcausality accuracies:") 719 | pprint(pd.DataFrame(mlc_acc).quantile([0, 0.25, 0.5, 0.75, 1]).T) 720 | pprint("mlcausality balanced accuracies:") 721 | pprint(pd.DataFrame(mlc_bal_acc).quantile([0, 0.25, 0.5, 0.75, 1]).T) 722 | pprint("mlcausality f1 scores:") 723 | pprint(pd.DataFrame(mlc_f1wgt).quantile([0, 0.25, 0.5, 0.75, 1]).T) 724 | 725 | optimal_j = 0 726 | best_score = 0 727 | for j in [round(r * 0.01, 2) for r in range(1, 100)]: 728 | mlc_sens2, mlc_spec2, mlc_acc2, mlc_bal_acc2, mlc_f1wgt2 = calc_metrics( 729 | mlcausality_pvalue_matricies, j 730 | ) 731 | best_score_new = np.median((np.array(mlc_sens2) * np.array(mlc_spec2)) ** 0.5) 732 | if best_score_new > best_score: 733 | best_score = deepcopy(best_score_new) 734 | optimal_j = deepcopy(j) 735 | 736 | 737 | mlc_sens2, mlc_spec2, mlc_acc2, mlc_bal_acc2, mlc_f1wgt2 = calc_metrics( 738 | mlcausality_pvalue_matricies, optimal_j 739 | ) 740 | 741 | 742 | pprint("Optimal cutoff: " + str(optimal_j)) 743 | pprint("Optimal mlcausality sensitivities:") 744 | pprint(pd.DataFrame(mlc_sens2).quantile([0, 0.25, 0.5, 0.75, 1]).T) 745 | pprint("Optimal mlcausality specificities:") 746 | pprint(pd.DataFrame(mlc_spec2).quantile([0, 0.25, 0.5, 0.75, 1]).T) 747 | pprint("Optimal mlcausality (sensitivities*specificities)**0.5:") 748 | pprint( 749 | pd.DataFrame((np.array(mlc_sens2) * np.array(mlc_spec2)) ** 0.5) 750 | .quantile([0, 0.25, 0.5, 0.75, 1]) 751 | .T 752 | ) 753 | pprint("Optimal mlcausality accuracies:") 754 | pprint(pd.DataFrame(mlc_acc2).quantile([0, 0.25, 0.5, 0.75, 1]).T) 755 | pprint("Optimal mlcausality balanced accuracies:") 756 | pprint(pd.DataFrame(mlc_bal_acc2).quantile([0, 0.25, 0.5, 0.75, 1]).T) 757 | pprint("Optimal mlcausality f1 scores:") 758 | pprint(pd.DataFrame(mlc_f1wgt2).quantile([0, 0.25, 0.5, 0.75, 1]).T) 759 | 760 | 761 | # Run lsNGC 762 | pprint("### Run lsNGC ###") 763 | 764 | k_f = 25 765 | k_g = 5 766 | ar_order = max(min_embed_dim - 1, 1) 767 | while k_f >= 2: 768 | try: 769 | _ = granger(data_list[0].T, k_f=k_f, k_g=k_g, ar_order=ar_order, normalize=1) 770 | break 771 | except Exception: 772 | k_f -= 1 773 | 774 | pprint( 775 | "lsNGC parameters: k_f=" 776 | + str(k_f) 777 | + ", k_g=" 778 | + str(k_g) 779 | + ", ar_order=" 780 | + str(ar_order) 781 | ) 782 | 783 | 784 | def lsNGCloop(d): 785 | lsNGC_output = granger(d.T, k_f=k_f, k_g=k_g, ar_order=ar_order, normalize=1) 786 | return lsNGC_output 787 | 788 | 789 | all_lsNGC_output = process_map(lsNGCloop, data_list, max_workers=num_cores) 790 | # Ensure process_map finished before continuing: 791 | # the following will not pass until the lock can be acquired 792 | with ensure_lock(tqdm.auto.tqdm, lock_name="mp_lock") as lk: 793 | pass 794 | 795 | lsgc_adjacency_matricies = [] 796 | lsgc_fstat_matricies = [] 797 | lsgc_pvalue_matricies = [] 798 | for out in all_lsNGC_output: 799 | lsgc_adjacency_matricies.append(out[0]) 800 | lsgc_fstat_matricies.append(out[1]) 801 | pvalues_lsgc = out[2] 802 | np.fill_diagonal(pvalues_lsgc, 1) 803 | # lsgc_adjacency_matricies.append(1 - pvalues_lsgc) 804 | lsgc_pvalue_matricies.append(pvalues_lsgc) 805 | 806 | lsgc_recovery_performance = recovery_performance(lsgc_fstat_matricies, true_labels_list) 807 | # lsgc_recovery_performance = recovery_performance(lsgc_adjacency_matricies,true_labels_list) 808 | pprint("lsNGC recovery performance:") 809 | pprint("Mean: " + str(lsgc_recovery_performance.mean())) 810 | pprint(pd.DataFrame(lsgc_recovery_performance).quantile([0, 0.25, 0.5, 0.75, 1]).T) 811 | 812 | lsgc_brier_score_loss = brier_loss_func(lsgc_pvalue_matricies, true_labels_list) 813 | pprint("lsgc brier score:") 814 | pprint("Mean: " + str(lsgc_brier_score_loss.mean())) 815 | pprint(pd.DataFrame(lsgc_brier_score_loss).quantile([0, 0.25, 0.5, 0.75, 1]).T) 816 | 817 | # Check sensitivity and specificity for the lsNGC model 818 | lsgc_sens, lsgc_spec, lsgc_acc, lsgc_bal_acc, lsgc_f1wgt = calc_metrics( 819 | lsgc_pvalue_matricies, 0.05 820 | ) 821 | 822 | 823 | pprint("lsgc sensitivities:") 824 | pprint(pd.DataFrame(lsgc_sens).quantile([0, 0.25, 0.5, 0.75, 1]).T) 825 | pprint("lsgc specificities:") 826 | pprint(pd.DataFrame(lsgc_spec).quantile([0, 0.25, 0.5, 0.75, 1]).T) 827 | pprint("lsgc (sensitivities*specificities)**0.5:") 828 | pprint( 829 | pd.DataFrame((np.array(lsgc_sens) * np.array(lsgc_spec)) ** 0.5) 830 | .quantile([0, 0.25, 0.5, 0.75, 1]) 831 | .T 832 | ) 833 | pprint("lsgc accuracies:") 834 | pprint(pd.DataFrame(lsgc_acc).quantile([0, 0.25, 0.5, 0.75, 1]).T) 835 | pprint("lsgc balanced accuracies:") 836 | pprint(pd.DataFrame(lsgc_bal_acc).quantile([0, 0.25, 0.5, 0.75, 1]).T) 837 | pprint("lsgc f1 scores:") 838 | pprint(pd.DataFrame(lsgc_f1wgt).quantile([0, 0.25, 0.5, 0.75, 1]).T) 839 | 840 | optimal_j = 0 841 | best_score = 0 842 | for j in [round(r * 0.01, 2) for r in range(1, 100)]: 843 | lsgc_sens2, lsgc_spec2, lsgc_acc2, lsgc_bal_acc2, lsgc_f1wgt2 = calc_metrics( 844 | lsgc_pvalue_matricies, j 845 | ) 846 | best_score_new = np.median((np.array(lsgc_sens2) * np.array(lsgc_spec2)) ** 0.5) 847 | if best_score_new > best_score: 848 | best_score = deepcopy(best_score_new) 849 | optimal_j = deepcopy(j) 850 | 851 | 852 | lsgc_sens2, lsgc_spec2, lsgc_acc2, lsgc_bal_acc2, lsgc_f1wgt2 = calc_metrics( 853 | lsgc_pvalue_matricies, optimal_j 854 | ) 855 | 856 | 857 | pprint("Optimal cutoff: " + str(optimal_j)) 858 | pprint("Optimal lsgc sensitivities:") 859 | pprint(pd.DataFrame(lsgc_sens2).quantile([0, 0.25, 0.5, 0.75, 1]).T) 860 | pprint("Optimal lsgc specificities:") 861 | pprint(pd.DataFrame(lsgc_spec2).quantile([0, 0.25, 0.5, 0.75, 1]).T) 862 | pprint("Optimal lsgc (sensitivities*specificities)**0.5:") 863 | pprint( 864 | pd.DataFrame((np.array(lsgc_sens2) * np.array(lsgc_spec2)) ** 0.5) 865 | .quantile([0, 0.25, 0.5, 0.75, 1]) 866 | .T 867 | ) 868 | pprint("Optimal lsgc accuracies:") 869 | pprint(pd.DataFrame(lsgc_acc2).quantile([0, 0.25, 0.5, 0.75, 1]).T) 870 | pprint("Optimal lsgc balanced accuracies:") 871 | pprint(pd.DataFrame(lsgc_bal_acc2).quantile([0, 0.25, 0.5, 0.75, 1]).T) 872 | pprint("Optimal lsgc f1 scores:") 873 | pprint(pd.DataFrame(lsgc_f1wgt2).quantile([0, 0.25, 0.5, 0.75, 1]).T) 874 | 875 | 876 | # Run causal_ccm 877 | pprint("### Run causal_ccm ###") 878 | 879 | 880 | def causal_ccm_loop(d): 881 | causal_ccm_output = causal_ccm.loop_causal_cmm(d.astype(np.float64), tau=1, E=2) 882 | return causal_ccm_output 883 | 884 | 885 | all_causal_ccm_output = process_map(causal_ccm_loop, data_list, max_workers=num_cores) 886 | # Ensure process_map finished before continuing: 887 | # the following will not pass until the lock can be acquired 888 | with ensure_lock(tqdm.auto.tqdm, lock_name="mp_lock") as lk: 889 | pass 890 | 891 | 892 | # Generate pvalue adjacency matricies 893 | permute_lists = list(itertools.permutations(range(data_list[0].shape[1]), 2)) 894 | causal_ccm_pvalue_matricies = [] 895 | for mlout in all_causal_ccm_output: 896 | cur_pvalue_matrix = np.ones([data_list[0].shape[1], data_list[0].shape[1]]) 897 | for X_idx, y_idx in permute_lists: 898 | cur_pvalue_matrix[X_idx, y_idx] = mlout.loc[ 899 | ((mlout.X == X_idx) & (mlout.y == y_idx)), "pvalue" 900 | ].iloc[0] 901 | np.fill_diagonal(cur_pvalue_matrix, 1) 902 | causal_ccm_pvalue_matricies.append(cur_pvalue_matrix) 903 | 904 | 905 | # Generate adjacency matricies and check recovery performance 906 | permute_lists = list(itertools.permutations(range(data_list[0].shape[1]), 2)) 907 | causal_ccm_adjacency_matricies = [] 908 | for mlout in all_causal_ccm_output: 909 | cur_pvalue_adjacency_matrix = np.zeros( 910 | [data_list[0].shape[1], data_list[0].shape[1]] 911 | ) 912 | for X_idx, y_idx in permute_lists: 913 | cur_pvalue_adjacency_matrix[X_idx, y_idx] = mlout.loc[ 914 | ((mlout.X == X_idx) & (mlout.y == y_idx)), "score" 915 | ].iloc[0] 916 | causal_ccm_adjacency_matricies.append(cur_pvalue_adjacency_matrix) 917 | 918 | causal_ccm_recovery_performance = recovery_performance( 919 | causal_ccm_adjacency_matricies, true_labels_list 920 | ) 921 | pprint("causal_ccm recovery performance:") 922 | pprint("Mean: " + str(causal_ccm_recovery_performance.mean())) 923 | pprint( 924 | pd.DataFrame(causal_ccm_recovery_performance).quantile([0, 0.25, 0.5, 0.75, 1]).T 925 | ) 926 | 927 | causal_ccm_brier_score_loss = brier_loss_func( 928 | causal_ccm_pvalue_matricies, true_labels_list 929 | ) 930 | pprint("causal_ccm brier score:") 931 | pprint("Mean: " + str(causal_ccm_brier_score_loss.mean())) 932 | pprint(pd.DataFrame(causal_ccm_brier_score_loss).quantile([0, 0.25, 0.5, 0.75, 1]).T) 933 | 934 | # Check sensitivity and specificity for the causal_ccm model 935 | ( 936 | causal_ccm_sens, 937 | causal_ccm_spec, 938 | causal_ccm_acc, 939 | causal_ccm_bal_acc, 940 | causal_ccm_f1wgt, 941 | ) = calc_metrics(causal_ccm_pvalue_matricies, 0.05) 942 | 943 | 944 | pprint("causal_ccm sensitivities:") 945 | pprint(pd.DataFrame(causal_ccm_sens).quantile([0, 0.25, 0.5, 0.75, 1]).T) 946 | pprint("causal_ccm specificities:") 947 | pprint(pd.DataFrame(causal_ccm_spec).quantile([0, 0.25, 0.5, 0.75, 1]).T) 948 | pprint("causal_ccm (sensitivities*specificities)**0.5:") 949 | pprint( 950 | pd.DataFrame((np.array(causal_ccm_sens) * np.array(causal_ccm_spec)) ** 0.5) 951 | .quantile([0, 0.25, 0.5, 0.75, 1]) 952 | .T 953 | ) 954 | pprint("causal_ccm accuracies:") 955 | pprint(pd.DataFrame(causal_ccm_acc).quantile([0, 0.25, 0.5, 0.75, 1]).T) 956 | pprint("causal_ccm balanced accuracies:") 957 | pprint(pd.DataFrame(causal_ccm_bal_acc).quantile([0, 0.25, 0.5, 0.75, 1]).T) 958 | pprint("causal_ccm f1 scores:") 959 | pprint(pd.DataFrame(causal_ccm_f1wgt).quantile([0, 0.25, 0.5, 0.75, 1]).T) 960 | 961 | optimal_j = 0 962 | best_score = 0 963 | for j in [round(r * 0.01, 2) for r in range(1, 100)]: 964 | ( 965 | causal_ccm_sens2, 966 | causal_ccm_spec2, 967 | causal_ccm_acc2, 968 | causal_ccm_bal_acc2, 969 | causal_ccm_f1wgt2, 970 | ) = calc_metrics(causal_ccm_pvalue_matricies, j) 971 | best_score_new = np.median( 972 | (np.array(causal_ccm_sens2) * np.array(causal_ccm_spec2)) ** 0.5 973 | ) 974 | if best_score_new > best_score: 975 | best_score = deepcopy(best_score_new) 976 | optimal_j = deepcopy(j) 977 | 978 | 979 | ( 980 | causal_ccm_sens2, 981 | causal_ccm_spec2, 982 | causal_ccm_acc2, 983 | causal_ccm_bal_acc2, 984 | causal_ccm_f1wgt2, 985 | ) = calc_metrics(causal_ccm_pvalue_matricies, optimal_j) 986 | 987 | 988 | pprint("Optimal cutoff: " + str(optimal_j)) 989 | pprint("Optimal causal_ccm sensitivities:") 990 | pprint(pd.DataFrame(causal_ccm_sens2).quantile([0, 0.25, 0.5, 0.75, 1]).T) 991 | pprint("Optimal causal_ccm specificities:") 992 | pprint(pd.DataFrame(causal_ccm_spec2).quantile([0, 0.25, 0.5, 0.75, 1]).T) 993 | pprint("Optimal causal_ccm (sensitivities*specificities)**0.5:") 994 | pprint( 995 | pd.DataFrame((np.array(causal_ccm_sens2) * np.array(causal_ccm_spec2)) ** 0.5) 996 | .quantile([0, 0.25, 0.5, 0.75, 1]) 997 | .T 998 | ) 999 | pprint("Optimal causal_ccm accuracies:") 1000 | pprint(pd.DataFrame(causal_ccm_acc2).quantile([0, 0.25, 0.5, 0.75, 1]).T) 1001 | pprint("Optimal causal_ccm balanced accuracies:") 1002 | pprint(pd.DataFrame(causal_ccm_bal_acc2).quantile([0, 0.25, 0.5, 0.75, 1]).T) 1003 | pprint("Optimal causal_ccm f1 scores:") 1004 | pprint(pd.DataFrame(causal_ccm_f1wgt2).quantile([0, 0.25, 0.5, 0.75, 1]).T) 1005 | 1006 | 1007 | # Run pcmci 1008 | pprint("### Run pcmci ###") 1009 | 1010 | 1011 | def pmciloop(d): 1012 | dataframe = pcmci_data_processing.DataFrame(d) 1013 | parcorr = ParCorr() 1014 | pcmci = PCMCI(dataframe=dataframe, cond_ind_test=parcorr, verbosity=0) 1015 | results = pcmci.run_pcmci( 1016 | tau_min=1, tau_max=min_embed_dim - 1, pc_alpha=None, alpha_level=0.05 1017 | ) 1018 | pcmci_pvalues_matrix = results["p_matrix"].min(axis=2) 1019 | np.fill_diagonal(pcmci_pvalues_matrix, 1) 1020 | return pcmci_pvalues_matrix 1021 | 1022 | 1023 | pcmci_pvalue_matricies = process_map(pmciloop, data_list, max_workers=num_cores) 1024 | # Ensure process_map finished before continuing: 1025 | # the following will not pass until the lock can be acquired 1026 | with ensure_lock(tqdm.auto.tqdm, lock_name="mp_lock") as lk: 1027 | pass 1028 | 1029 | pcmci_adjacency_matricies = [] 1030 | for i in range(len(pcmci_pvalue_matricies)): 1031 | pcmci_pvalues_matrix_oneminus = 1 - pcmci_pvalue_matricies[i] 1032 | np.fill_diagonal(pcmci_pvalues_matrix_oneminus, 0) 1033 | pcmci_adjacency_matricies.append(pcmci_pvalues_matrix_oneminus) 1034 | 1035 | pcmci_recovery_performance = recovery_performance( 1036 | pcmci_adjacency_matricies, true_labels_list 1037 | ) 1038 | pprint("pcmci recovery performance:") 1039 | pprint("Mean: " + str(pcmci_recovery_performance.mean())) 1040 | pprint(pd.DataFrame(pcmci_recovery_performance).quantile([0, 0.25, 0.5, 0.75, 1]).T) 1041 | 1042 | pcmci_brier_score_loss = brier_loss_func(pcmci_pvalue_matricies, true_labels_list) 1043 | pprint("pcmci brier score:") 1044 | pprint("Mean: " + str(pcmci_brier_score_loss.mean())) 1045 | pprint(pd.DataFrame(pcmci_brier_score_loss).quantile([0, 0.25, 0.5, 0.75, 1]).T) 1046 | 1047 | # Check sensitivity and specificity for the pcmci model 1048 | pcmci_sens, pcmci_spec, pcmci_acc, pcmci_bal_acc, pcmci_f1wgt = calc_metrics( 1049 | pcmci_pvalue_matricies, 0.05 1050 | ) 1051 | 1052 | 1053 | pprint("pcmci sensitivities:") 1054 | pprint(pd.DataFrame(pcmci_sens).quantile([0, 0.25, 0.5, 0.75, 1]).T) 1055 | pprint("pcmci specificities:") 1056 | pprint(pd.DataFrame(pcmci_spec).quantile([0, 0.25, 0.5, 0.75, 1]).T) 1057 | pprint("pcmci (sensitivities*specificities)**0.5:") 1058 | pprint( 1059 | pd.DataFrame((np.array(pcmci_sens) * np.array(pcmci_spec)) ** 0.5) 1060 | .quantile([0, 0.25, 0.5, 0.75, 1]) 1061 | .T 1062 | ) 1063 | pprint("pcmci accuracies:") 1064 | pprint(pd.DataFrame(pcmci_acc).quantile([0, 0.25, 0.5, 0.75, 1]).T) 1065 | pprint("pcmci balanced accuracies:") 1066 | pprint(pd.DataFrame(pcmci_bal_acc).quantile([0, 0.25, 0.5, 0.75, 1]).T) 1067 | pprint("pcmci f1 scores:") 1068 | pprint(pd.DataFrame(pcmci_f1wgt).quantile([0, 0.25, 0.5, 0.75, 1]).T) 1069 | 1070 | optimal_j = 0 1071 | best_score = 0 1072 | for j in [round(r * 0.01, 2) for r in range(1, 100)]: 1073 | pcmci_sens2, pcmci_spec2, pcmci_acc2, pcmci_bal_acc2, pcmci_f1wgt2 = calc_metrics( 1074 | pcmci_pvalue_matricies, j 1075 | ) 1076 | best_score_new = np.median((np.array(pcmci_sens2) * np.array(pcmci_spec2)) ** 0.5) 1077 | if best_score_new > best_score: 1078 | best_score = deepcopy(best_score_new) 1079 | optimal_j = deepcopy(j) 1080 | 1081 | 1082 | pcmci_sens2, pcmci_spec2, pcmci_acc2, pcmci_bal_acc2, pcmci_f1wgt2 = calc_metrics( 1083 | pcmci_pvalue_matricies, optimal_j 1084 | ) 1085 | 1086 | 1087 | pprint("Optimal cutoff: " + str(optimal_j)) 1088 | pprint("Optimal pcmci sensitivities:") 1089 | pprint(pd.DataFrame(pcmci_sens2).quantile([0, 0.25, 0.5, 0.75, 1]).T) 1090 | pprint("Optimal pcmci specificities:") 1091 | pprint(pd.DataFrame(pcmci_spec2).quantile([0, 0.25, 0.5, 0.75, 1]).T) 1092 | pprint("Optimal pcmci (sensitivities*specificities)**0.5:") 1093 | pprint( 1094 | pd.DataFrame((np.array(pcmci_sens2) * np.array(pcmci_spec2)) ** 0.5) 1095 | .quantile([0, 0.25, 0.5, 0.75, 1]) 1096 | .T 1097 | ) 1098 | pprint("Optimal pcmci accuracies:") 1099 | pprint(pd.DataFrame(pcmci_acc2).quantile([0, 0.25, 0.5, 0.75, 1]).T) 1100 | pprint("Optimal pcmci balanced accuracies:") 1101 | pprint(pd.DataFrame(pcmci_bal_acc2).quantile([0, 0.25, 0.5, 0.75, 1]).T) 1102 | pprint("Optimal pcmci f1 scores:") 1103 | pprint(pd.DataFrame(pcmci_f1wgt2).quantile([0, 0.25, 0.5, 0.75, 1]).T) 1104 | 1105 | 1106 | df_all_recovery_performance = pd.DataFrame( 1107 | np.array( 1108 | [ 1109 | mlcausality_recovery_performance, 1110 | lsgc_recovery_performance, 1111 | causal_ccm_recovery_performance, 1112 | pcmci_recovery_performance, 1113 | ] 1114 | ).T, 1115 | columns=["MLC", "lsNGC", "LM", "PCMCI"], 1116 | ) 1117 | with open( 1118 | "pickled_data/" + network_to_check + "_auc_" + n_time_samples_str + ".pickle", "wb" 1119 | ) as f: # should be 'wb' rather than 'w' 1120 | pickle.dump(df_all_recovery_performance, f) 1121 | 1122 | 1123 | df_all_brier_score_loss = pd.DataFrame( 1124 | np.array( 1125 | [ 1126 | mlcausality_brier_score_loss, 1127 | lsgc_brier_score_loss, 1128 | causal_ccm_brier_score_loss, 1129 | pcmci_brier_score_loss, 1130 | ] 1131 | ).T, 1132 | columns=["MLC", "lsNGC", "LM", "PCMCI"], 1133 | ) 1134 | with open( 1135 | "pickled_data/" + network_to_check + "_brier_" + n_time_samples_str + ".pickle", 1136 | "wb", 1137 | ) as f: # should be 'wb' rather than 'w' 1138 | pickle.dump(df_all_brier_score_loss, f) 1139 | 1140 | 1141 | df_all_sensitivities = pd.DataFrame( 1142 | np.array( 1143 | [ 1144 | mlc_sens, 1145 | lsgc_sens, 1146 | causal_ccm_sens, 1147 | pcmci_sens, 1148 | ] 1149 | ).T, 1150 | columns=["MLC", "lsNGC", "LM", "PCMCI"], 1151 | ) 1152 | with open( 1153 | "pickled_data/" 1154 | + network_to_check 1155 | + "_sensitivity_" 1156 | + n_time_samples_str 1157 | + ".pickle", 1158 | "wb", 1159 | ) as f: # should be 'wb' rather than 'w' 1160 | pickle.dump(df_all_sensitivities, f) 1161 | 1162 | 1163 | df_all_specificities = pd.DataFrame( 1164 | np.array( 1165 | [ 1166 | mlc_spec, 1167 | lsgc_spec, 1168 | causal_ccm_spec, 1169 | pcmci_spec, 1170 | ] 1171 | ).T, 1172 | columns=["MLC", "lsNGC", "LM", "PCMCI"], 1173 | ) 1174 | with open( 1175 | "pickled_data/" 1176 | + network_to_check 1177 | + "_specificity_" 1178 | + n_time_samples_str 1179 | + ".pickle", 1180 | "wb", 1181 | ) as f: # should be 'wb' rather than 'w' 1182 | pickle.dump(df_all_specificities, f) 1183 | 1184 | 1185 | df_all_gmean = pd.DataFrame( 1186 | np.array( 1187 | [ 1188 | list((np.array(mlc_sens) * np.array(mlc_spec)) ** (1 / 2)), 1189 | list((np.array(lsgc_sens) * np.array(lsgc_spec)) ** (1 / 2)), 1190 | list((np.array(causal_ccm_sens) * np.array(causal_ccm_spec)) ** (1 / 2)), 1191 | list((np.array(pcmci_sens) * np.array(pcmci_spec)) ** (1 / 2)), 1192 | ] 1193 | ).T, 1194 | columns=["MLC", "lsNGC", "LM", "PCMCI"], 1195 | ) 1196 | with open( 1197 | "pickled_data/" + network_to_check + "_gmean_" + n_time_samples_str + ".pickle", 1198 | "wb", 1199 | ) as f: # should be 'wb' rather than 'w' 1200 | pickle.dump(df_all_gmean, f) 1201 | 1202 | 1203 | df_all_accuracies = pd.DataFrame( 1204 | np.array([mlc_acc, lsgc_acc, causal_ccm_acc, pcmci_acc]).T, 1205 | columns=["MLC", "lsNGC", "LM", "PCMCI"], 1206 | ) 1207 | with open( 1208 | "pickled_data/" + network_to_check + "_accuracy_" + n_time_samples_str + ".pickle", 1209 | "wb", 1210 | ) as f: # should be 'wb' rather than 'w' 1211 | pickle.dump(df_all_accuracies, f) 1212 | 1213 | 1214 | df_all_balanced_accuracies = pd.DataFrame( 1215 | np.array( 1216 | [ 1217 | mlc_bal_acc, 1218 | lsgc_bal_acc, 1219 | causal_ccm_bal_acc, 1220 | pcmci_bal_acc, 1221 | ] 1222 | ).T, 1223 | columns=["MLC", "lsNGC", "LM", "PCMCI"], 1224 | ) 1225 | with open( 1226 | "pickled_data/" 1227 | + network_to_check 1228 | + "_balanced_accuracy_" 1229 | + n_time_samples_str 1230 | + ".pickle", 1231 | "wb", 1232 | ) as f: # should be 'wb' rather than 'w' 1233 | pickle.dump(df_all_balanced_accuracies, f) 1234 | 1235 | 1236 | df_all_f1wgt = pd.DataFrame( 1237 | np.array([mlc_f1wgt, lsgc_f1wgt, causal_ccm_f1wgt, pcmci_f1wgt]).T, 1238 | columns=["MLC", "lsNGC", "LM", "PCMCI"], 1239 | ) 1240 | with open( 1241 | "pickled_data/" + network_to_check + "_f1wgt_" + n_time_samples_str + ".pickle", 1242 | "wb", 1243 | ) as f: # should be 'wb' rather than 'w' 1244 | pickle.dump(df_all_f1wgt, f) 1245 | 1246 | 1247 | df_all_sensitivities2 = pd.DataFrame( 1248 | np.array( 1249 | [ 1250 | mlc_sens2, 1251 | lsgc_sens2, 1252 | causal_ccm_sens2, 1253 | pcmci_sens2, 1254 | ] 1255 | ).T, 1256 | columns=["MLC", "lsNGC", "LM", "PCMCI"], 1257 | ) 1258 | with open( 1259 | "pickled_data/" 1260 | + network_to_check 1261 | + "_sensitivity2_" 1262 | + n_time_samples_str 1263 | + ".pickle", 1264 | "wb", 1265 | ) as f: # should be 'wb' rather than 'w' 1266 | pickle.dump(df_all_sensitivities2, f) 1267 | 1268 | 1269 | df_all_specificities2 = pd.DataFrame( 1270 | np.array( 1271 | [ 1272 | mlc_spec2, 1273 | lsgc_spec2, 1274 | causal_ccm_spec2, 1275 | pcmci_spec2, 1276 | ] 1277 | ).T, 1278 | columns=["MLC", "lsNGC", "LM", "PCMCI"], 1279 | ) 1280 | with open( 1281 | "pickled_data/" 1282 | + network_to_check 1283 | + "_specificity2_" 1284 | + n_time_samples_str 1285 | + ".pickle", 1286 | "wb", 1287 | ) as f: # should be 'wb' rather than 'w' 1288 | pickle.dump(df_all_specificities2, f) 1289 | 1290 | 1291 | df_all_gmean2 = pd.DataFrame( 1292 | np.array( 1293 | [ 1294 | list((np.array(mlc_sens2) * np.array(mlc_spec2)) ** (1 / 2)), 1295 | list((np.array(lsgc_sens2) * np.array(lsgc_spec2)) ** (1 / 2)), 1296 | list((np.array(causal_ccm_sens2) * np.array(causal_ccm_spec2)) ** (1 / 2)), 1297 | list((np.array(pcmci_sens2) * np.array(pcmci_spec2)) ** (1 / 2)), 1298 | ] 1299 | ).T, 1300 | columns=["MLC", "lsNGC", "LM", "PCMCI"], 1301 | ) 1302 | with open( 1303 | "pickled_data/" + network_to_check + "_gmean2_" + n_time_samples_str + ".pickle", 1304 | "wb", 1305 | ) as f: # should be 'wb' rather than 'w' 1306 | pickle.dump(df_all_gmean2, f) 1307 | 1308 | 1309 | df_all_accuracies2 = pd.DataFrame( 1310 | np.array([mlc_acc2, lsgc_acc2, causal_ccm_acc2, pcmci_acc2]).T, 1311 | columns=["MLC", "lsNGC", "LM", "PCMCI"], 1312 | ) 1313 | with open( 1314 | "pickled_data/" + network_to_check + "_accuracy2_" + n_time_samples_str + ".pickle", 1315 | "wb", 1316 | ) as f: # should be 'wb' rather than 'w' 1317 | pickle.dump(df_all_accuracies2, f) 1318 | 1319 | 1320 | df_all_balanced_accuracies2 = pd.DataFrame( 1321 | np.array( 1322 | [ 1323 | mlc_bal_acc2, 1324 | lsgc_bal_acc2, 1325 | causal_ccm_bal_acc2, 1326 | pcmci_bal_acc2, 1327 | ] 1328 | ).T, 1329 | columns=["MLC", "lsNGC", "LM", "PCMCI"], 1330 | ) 1331 | with open( 1332 | "pickled_data/" 1333 | + network_to_check 1334 | + "_balanced_accuracy2_" 1335 | + n_time_samples_str 1336 | + ".pickle", 1337 | "wb", 1338 | ) as f: # should be 'wb' rather than 'w' 1339 | pickle.dump(df_all_balanced_accuracies2, f) 1340 | 1341 | 1342 | df_all_f1wgt2 = pd.DataFrame( 1343 | np.array([mlc_f1wgt2, lsgc_f1wgt2, causal_ccm_f1wgt2, pcmci_f1wgt2]).T, 1344 | columns=["MLC", "lsNGC", "LM", "PCMCI"], 1345 | ) 1346 | with open( 1347 | "pickled_data/" + network_to_check + "_f1wgt2_" + n_time_samples_str + ".pickle", 1348 | "wb", 1349 | ) as f: # should be 'wb' rather than 'w' 1350 | pickle.dump(df_all_f1wgt2, f) 1351 | 1352 | 1353 | pprint("Program finished running!") 1354 | --------------------------------------------------------------------------------