13 |
14 |
--------------------------------------------------------------------------------
/.travis.yml:
--------------------------------------------------------------------------------
1 | language: R
2 | sudo: false
3 | cache: packages
4 | dist: trusty
5 |
6 | before_install:
7 | - sudo apt-get install --yes udunits-bin libproj-dev libgeos-dev libgdal-dev libgdal1-dev libudunits2-dev
8 |
9 | env:
10 | global:
11 | - R_CHECK_ARGS="--timings"
12 |
13 | notifications:
14 | email: false
15 |
16 | r_packages:
17 | - archivist
18 | - DALEX
19 | - ggplot2
20 | - covr
21 |
22 | after_success:
23 | - Rscript -e 'library(covr); codecov()'
24 |
--------------------------------------------------------------------------------
/EIX.Rproj:
--------------------------------------------------------------------------------
1 | Version: 1.0
2 |
3 | RestoreWorkspace: No
4 | SaveWorkspace: No
5 | AlwaysSaveHistory: Default
6 |
7 | EnableCodeIndexing: Yes
8 | UseSpacesForTab: Yes
9 | NumSpacesForTab: 2
10 | Encoding: UTF-8
11 |
12 | RnwWeave: Sweave
13 | LaTeX: pdfLaTeX
14 |
15 | AutoAppendNewline: Yes
16 | StripTrailingWhitespace: Yes
17 |
18 | BuildType: Package
19 | PackageUseDevtools: Yes
20 | PackageInstallArgs: --no-multiarch --with-keep.source
21 | PackageRoxygenize: rd,collate,namespace
22 |
--------------------------------------------------------------------------------
/man/tableOfTrees.Rd:
--------------------------------------------------------------------------------
1 | % Generated by roxygen2: do not edit by hand
2 | % Please edit documentation in R/calculateGain.R
3 | \name{tableOfTrees}
4 | \alias{tableOfTrees}
5 | \title{tableOfTrees}
6 | \usage{
7 | tableOfTrees(model, data)
8 | }
9 | \arguments{
10 | \item{model}{a xgboost or lightgbm model}
11 |
12 | \item{data}{a data table with data used to train the model}
13 | }
14 | \value{
15 | a data table
16 | }
17 | \description{
18 | tableOfTrees
19 | }
20 | \keyword{internal}
21 |
--------------------------------------------------------------------------------
/man/calculateGain.Rd:
--------------------------------------------------------------------------------
1 | % Generated by roxygen2: do not edit by hand
2 | % Please edit documentation in R/calculateGain.R
3 | \name{calculateGain}
4 | \alias{calculateGain}
5 | \title{calculateGain}
6 | \usage{
7 | calculateGain(xgb.model, data)
8 | }
9 | \arguments{
10 | \item{xgb.model}{a xgboost or lightgbm model}
11 |
12 | \item{data}{a data table with data used to train the model}
13 | }
14 | \value{
15 | a list
16 | }
17 | \description{
18 | List of trees with pairs of variable and other needed fields
19 | }
20 | \keyword{internal}
21 |
--------------------------------------------------------------------------------
/R/package.R:
--------------------------------------------------------------------------------
1 | #' EIX package
2 | #'
3 | #' Structure mining from 'XGBoost' and 'LightGBM' models.
4 | #' Key functionalities of this package cover: visualisation of tree-based ensembles models,
5 | #' identification of interactions, measuring of variable importance,
6 | #' measuring of interaction importance, explanation of single prediction
7 | #' with break down plots (based on 'xgboostExplainer' and 'iBreakDown' packages).
8 | #' To download the 'LightGBM' use the following link:
39 |
40 |
95 |
96 |
128 |
129 |
130 |
131 |
132 |
--------------------------------------------------------------------------------
/docs/articles/index.html:
--------------------------------------------------------------------------------
1 |
2 |
3 |
4 |
5 |
6 |
7 |
8 |
9 |
41 |
92 |
93 |
94 |
42 |
91 |
43 |
48 |
49 | 
50 |
51 |
52 |
53 |
90 |
50 |
51 |
52 |
97 |
115 |
116 |
117 |
127 |
98 |
113 |
114 |
99 |
101 | Authors
100 |-
102 |
-
103 |
Ewelina Karbowiak. Author, maintainer. 104 |
105 |
106 | -
107 |
Przemyslaw Biecek. Author, thesis advisor. 108 |
109 |
110 |
39 |
40 |
95 |
96 |
125 |
126 |
127 |
128 |
129 |
--------------------------------------------------------------------------------
/docs/pkgdown.css:
--------------------------------------------------------------------------------
1 | /* Sticker footer */
2 | body > .container {
3 | display: flex;
4 | padding-top: 60px;
5 | min-height: calc(100vh);
6 | flex-direction: column;
7 | }
8 |
9 | body > .container .row {
10 | flex: 1;
11 | }
12 |
13 | footer {
14 | margin-top: 45px;
15 | padding: 35px 0 36px;
16 | border-top: 1px solid #e5e5e5;
17 | color: #666;
18 | display: flex;
19 | }
20 | footer p {
21 | margin-bottom: 0;
22 | }
23 | footer div {
24 | flex: 1;
25 | }
26 | footer .pkgdown {
27 | text-align: right;
28 | }
29 | footer p {
30 | margin-bottom: 0;
31 | }
32 |
33 | img.icon {
34 | float: right;
35 | }
36 |
37 | img {
38 | max-width: 100%;
39 | }
40 |
41 | /* Section anchors ---------------------------------*/
42 |
43 | a.anchor {
44 | margin-left: -30px;
45 | display:inline-block;
46 | width: 30px;
47 | height: 30px;
48 | visibility: hidden;
49 |
50 | background-image: url(./link.svg);
51 | background-repeat: no-repeat;
52 | background-size: 20px 20px;
53 | background-position: center center;
54 | }
55 |
56 | .hasAnchor:hover a.anchor {
57 | visibility: visible;
58 | }
59 |
60 | @media (max-width: 767px) {
61 | .hasAnchor:hover a.anchor {
62 | visibility: hidden;
63 | }
64 | }
65 |
66 |
67 | /* Fixes for fixed navbar --------------------------*/
68 |
69 | .contents h1, .contents h2, .contents h3, .contents h4 {
70 | padding-top: 60px;
71 | margin-top: -60px;
72 | }
73 |
74 | /* Static header placement on mobile devices */
75 | @media (max-width: 767px) {
76 | .navbar-fixed-top {
77 | position: absolute;
78 | }
79 | .navbar {
80 | padding: 0;
81 | }
82 | }
83 |
84 |
85 | /* Sidebar --------------------------*/
86 |
87 | #sidebar {
88 | margin-top: 30px;
89 | }
90 | #sidebar h2 {
91 | font-size: 1.5em;
92 | margin-top: 1em;
93 | }
94 |
95 | #sidebar h2:first-child {
96 | margin-top: 0;
97 | }
98 |
99 | #sidebar .list-unstyled li {
100 | margin-bottom: 0.5em;
101 | }
102 |
103 | /* Reference index & topics ----------------------------------------------- */
104 |
105 | .ref-index th {font-weight: normal;}
106 | .ref-index h2 {font-size: 20px;}
107 |
108 | .ref-index td {vertical-align: top;}
109 | .ref-index .alias {width: 40%;}
110 | .ref-index .title {width: 60%;}
111 |
112 | .ref-index .alias {width: 40%;}
113 | .ref-index .title {width: 60%;}
114 |
115 | .ref-arguments th {text-align: right; padding-right: 10px;}
116 | .ref-arguments th, .ref-arguments td {vertical-align: top;}
117 | .ref-arguments .name {width: 20%;}
118 | .ref-arguments .desc {width: 80%;}
119 |
120 | /* Nice scrolling for wide elements --------------------------------------- */
121 |
122 | table {
123 | display: block;
124 | overflow: auto;
125 | }
126 |
127 | /* Syntax highlighting ---------------------------------------------------- */
128 |
129 | pre {
130 | word-wrap: normal;
131 | word-break: normal;
132 | border: 1px solid #eee;
133 | }
134 |
135 | pre, code {
136 | background-color: #f8f8f8;
137 | color: #333;
138 | }
139 |
140 | pre .img {
141 | margin: 5px 0;
142 | }
143 |
144 | pre .img img {
145 | background-color: #fff;
146 | display: block;
147 | height: auto;
148 | }
149 |
150 | code a, pre a {
151 | color: #375f84;
152 | }
153 | table {
154 | display: block;
155 | overflow: auto;
156 | width: 100% !important;
157 | }
158 |
159 | .fl {color: #1514b5;}
160 | .fu {color: #000000;} /* function */
161 | .ch,.st {color: #036a07;} /* string */
162 | .kw {color: #264D66;} /* keyword */
163 | .co {color: #888888;} /* comment */
164 |
165 | .message { color: black; font-weight: bolder;}
166 | .error { color: orange; font-weight: bolder;}
167 | .warning { color: #6A0366; font-weight: bolder;}
168 |
169 | .navbar-mi2logo {
170 | float: left;
171 | margin-right: 15px;
172 | margin-top: 2px;
173 | }
174 | .navbar-mi2 {
175 | background-color: #4a3c89;
176 | color: #fff !important;
177 | margin-right: 0px;
178 | }
179 | .navbar-mi2 > li > a {
180 | color: #fff !important;
181 | }
182 | .navbar-mi2 > .active > a{
183 | background-color: #370f54 !important;
184 | }
185 | .navbar-mi2 > .open > a:focus, .nav-pills> .open > a:focus{
186 | background-color: #370f54 !important;
187 | }
188 | .dropdown-menu > .active > a, .dropdown-menu > .active > a:focus{
189 | background-color: #370f54 !important;
190 | }
191 |
192 | .contents-mi2 > li > a:focus, .nav-pills > li > a:focus {
193 | background-color: #4a3c89 !important;
194 | color: #fff;
195 | }
196 | .contents-mi2 > li.active > a, .nav-pills > li.active > a{
197 | background-color: #370f54 !important;
198 | }
199 | .contents-mi2 > li > a, .nav-pills > li > a{
200 | background-color: #4a3c89 !important;
201 | color: #fff;
202 | }
203 |
204 | .sidebar-logo {
205 | display:block;
206 | margin-left:auto;
207 | margin-right:auto;
208 | text-align: justify;
209 | }
--------------------------------------------------------------------------------
/docs/reference/tableOfTrees.html:
--------------------------------------------------------------------------------
1 |
2 |
3 |
4 |
5 |
6 |
7 |
8 |
9 |
41 |
92 |
93 |
94 |
42 |
91 |
43 |
48 |
49 |
50 |
51 |
52 |
53 |
90 |
50 |
51 |
52 |
97 |
113 |
114 |
124 |
98 |
112 |
99 |
101 |
102 | Articles
100 |
103 |
111 | All vignettes
104 | 105 | 106 |-
107 |
- EIX: Explain Interactions in XGBoost 108 |
- EIX: Titanic data 109 |
39 |
40 |
95 |
96 |
149 |
150 |
151 |
152 |
153 |
--------------------------------------------------------------------------------
/docs/reference/calculateGain.html:
--------------------------------------------------------------------------------
1 |
2 |
3 |
4 |
5 |
6 |
7 |
8 |
9 |
41 |
92 |
93 |
94 |
42 |
91 |
43 |
48 |
49 |
50 |
51 | tableOfTrees
52 |
53 |
90 |
50 |
51 | tableOfTrees
52 |
97 |
137 |
138 |
148 |
98 |
110 |
120 |
121 |
127 |
136 |
99 |
101 |
102 |
103 | tableOfTrees
100 |tableOfTrees
104 | 105 | 106 |tableOfTrees(model, data)107 | 108 |
Arguments
109 || model | 113 |a xgboost or lightgbm model |
114 |
|---|---|
| data | 117 |a data table with data used to train the model |
118 |
Value
122 | 123 |a data table
124 | 125 | 126 |
39 |
40 |
95 |
96 |
149 |
150 |
151 |
152 |
153 |
--------------------------------------------------------------------------------
/R/calculateGain.R:
--------------------------------------------------------------------------------
1 | #' calculateGain
2 | #'
3 | #' List of trees with pairs of variable and other needed fields
4 | #'
5 | #' @param xgb.model a xgboost or lightgbm model
6 | #' @param data a data table with data used to train the model
7 | #'
8 | #' @return a list
9 | #'
10 | #' @import data.table
11 | #'
12 | #' @keywords internal
13 | #'
14 |
15 | # @import stats
16 | # @import utils
17 | calculateGain <- function(xgb.model, data) {
18 |
19 | leaf <- Feature <- Yes <- No <- ID <- parentsGain <- Quality <- parentsCover <-
20 | Cover <- name_pair <- childsGain <- depth <- parentsName <- NULL
21 |
22 | trees = tableOfTrees(xgb.model, data)
23 | trees[, leaf := Feature == "Leaf"]
24 | trees$depth <- 0
25 | treeList = split(trees, as.factor(trees$Tree))
26 |
27 | for (tree in treeList) {
28 | num_nodes = nrow(tree)
29 | non_leaf_rows = which(tree[, leaf] == F)
30 | for (r in non_leaf_rows) {
31 | left = tree[r, Yes]
32 | right = tree[r, No]
33 | if (tree[ID == left, leaf] == F) {
34 | # newDepth <- tree[r , depth] + 1
35 | tree[ID == left,`:=`(parentsGain = tree[r, Quality],
36 | parentsCover = tree[r, Cover],
37 | name_pair = paste(tree[r, Feature], tree[ID == left, Feature], sep = ":"),
38 | childsGain = Quality,
39 | depth = tree[r , depth] + 1,
40 | parentsName = tree[r, Feature])]
41 | tree[ID == left, interaction := ((parentsGain < childsGain) & (Feature != parentsName))]
42 | }
43 |
44 | if (tree[ID == right, leaf]==F) {
45 |
46 | #newDepth <- tree[r , depth] + 1
47 | tree[ID == right, `:=`(parentsGain = tree[r, Quality],
48 | parentsCover = tree[r, Cover],
49 | name_pair = paste(tree[r, Feature], tree[ID == right, Feature], sep = ":"),
50 | childsGain = Quality,
51 | depth = tree[r , depth] + 1,
52 | parentsName = tree[r, Feature])]
53 | tree[ID == right, interaction := ((parentsGain < childsGain) & (Feature != parentsName))]
54 | }
55 | }
56 | }
57 |
58 | return(treeList)
59 | }
60 |
61 | #'tableOfTrees
62 | #'
63 | #'tableOfTrees
64 | #'
65 | #' @param model a xgboost or lightgbm model
66 | #' @param data a data table with data used to train the model
67 | #'
68 | #' @return a data table
69 | #'
70 | #' @import data.table
71 | #' @importFrom xgboost xgb.model.dt.tree
72 | #'
73 | #' @keywords internal
74 | #'
75 | # @import lightgbm
76 |
77 |
78 | tableOfTrees <- function(model, data){
79 | count <- split_feature <- leaf_count <- internal_count <-
80 | split_index <- tree_index <- leaf_index <- threshold <-
81 | leaf_value <- split_gain <- flag <- node_parent <- leaf_parent<-
82 | Node <- Feature <- . <- Cover <- Yes <- No <- ID <-
83 | Tree<- Quality <- Missing <-Leaf_old_num<- Split <- NULL
84 |
85 |
86 | if(class(model)[1] == "xgb.Booster") {
87 | return(xgb.model.dt.tree(colnames(data), model)[])
88 | }
89 | if (class(model)[1] == "lgb.Booster") {
90 | lgb.trees <- lightgbm::lgb.model.dt.tree(model)
91 |
92 | lgb.trees <- lgb.trees[, count := ifelse(is.na(split_feature), leaf_count, internal_count)]
93 |
94 | lgb.trees <- lgb.trees[, max := max(split_index, na.rm = TRUE), by = tree_index]
95 |
96 | #UWAGA: nie jest tu istotne rodzaj nierówności, interesuje nas, że ktoś jest rodzicem, a nie, czy idzie w prawo i w lewo, dlatego losowe przypisanie Yes, No
97 |
98 | trees <- lgb.trees[, `:=`(Tree = tree_index,
99 | Node = ifelse(is.na(split_index), max + leaf_index + 1, split_index),
100 | Feature = ifelse(is.na(split_feature), "Leaf", split_feature),
101 | Split = threshold, Missing = NA, Quality = ifelse(is.na(split_feature), leaf_value, split_gain),
102 | Cover = (1/4)*count,
103 | Leaf_old_num=leaf_index)]
104 | trees[, `:=`(ID = paste(Tree, Node, sep = "-"))]
105 |
106 | trees[, flag := FALSE]
107 | treeList = split(trees, as.factor(trees$Tree))
108 |
109 | for (tree in treeList) {
110 | num_nodes = nrow(tree)
111 | for (i in 1:num_nodes) {
112 | if (is.na(tree[i, node_parent]) == FALSE) {
113 | if (tree[Node == tree[i, node_parent] , flag] == FALSE) {
114 | tree[Node == tree[i, node_parent] , Yes := paste(tree[i, Tree], tree[i, Node], sep = "-")]
115 | tree[Node == tree[i, node_parent] , flag := TRUE]
116 | } else{
117 | tree[Node == tree[i, node_parent] , No := paste(tree[i, Tree], tree[i, Node], sep = "-")]
118 | }
119 | }
120 | if (is.na(tree[i, leaf_parent]) == FALSE) {
121 | if (tree[Node == tree[i, leaf_parent] , flag] == FALSE) {
122 | tree[Node == tree[i, leaf_parent] , Yes := paste(tree[i, Tree], tree[i, Node], sep = "-")]
123 | tree[Node == tree[i, leaf_parent] , flag := TRUE]
124 | } else{
125 | tree[Node == tree[i, leaf_parent] , No := paste(tree[i, Tree], tree[i, Node], sep = "-")]
126 | }
127 | }
128 | }
129 | }
130 | trees <- rbindlist(treeList)
131 | trees[, .(Tree, Node, ID, Feature, Split, Yes, No, Missing, Quality, Cover)]
132 | return(trees[, .(Tree, Node, ID, Feature, Split, Yes, No, Missing, Quality, Cover,Leaf_old_num)][])
133 | }
134 | if (class(model)[1] != "xgb.Booster" || "lgb.Booster") {
135 | return(cat( "You should choose one of two available models: xgboost, lightgbm \n" ))
136 | }
137 | }
138 |
--------------------------------------------------------------------------------
/docs/index.html:
--------------------------------------------------------------------------------
1 |
2 |
3 |
4 |
5 |
6 |
7 |
8 |
41 |
92 |
93 |
94 |
42 |
91 |
43 |
48 |
49 |
50 |
51 | calculateGain
52 |
53 |
90 |
50 |
51 | calculateGain
52 |
97 |
137 |
138 |
148 |
98 |
110 |
120 |
121 |
127 |
136 |
99 |
101 |
102 |
103 | calculateGain
100 |List of trees with pairs of variable and other needed fields
104 | 105 | 106 |calculateGain(xgb.model, data)107 | 108 |
Arguments
109 || xgb.model | 113 |a xgboost or lightgbm model |
114 |
|---|---|
| data | 117 |a data table with data used to train the model |
118 |
Value
122 | 123 |a list
124 | 125 | 126 |
19 |
139 |
140 |
141 |
142 |
--------------------------------------------------------------------------------
/docs/reference/HR_data.html:
--------------------------------------------------------------------------------
1 |
2 |
3 |
4 |
5 |
6 |
7 |
8 |
9 |
20 |
71 |
72 |
73 |
74 |
21 |
69 |
70 |
22 |
27 |
28 |
29 |
30 |
31 |
67 |
68 |
29 |
30 |
75 |
127 |
128 |
129 |
138 |
76 |
77 |
95 |
96 |
78 |
82 | 79 | EIX - Explain Interactions in Xgboost
80 |A set of tools to explain XGBoost and LightGBM models.
81 |
83 |
88 |
93 |
94 | 84 | Installation
85 |Install from GitHub
86 |devtools::install_github("ModelOriented/EIX")
97 |
102 |
126 |
103 |
111 | Links
104 |-
105 |
- Browse source code at
https://github.com/ModelOriented/EIX 106 |
107 | - Report a bug at
https://github.com/ModelOriented/EIX/issues 108 |
109 |
112 |
117 | License
113 |-
114 |
- GPL-2 115 |
118 |
124 |
125 | Developers
119 |-
120 |
- Ewelina Karbowiak
Author, maintainer
121 | - Przemyslaw Biecek
Author, thesis advisor
122 |
39 |
40 |
95 |
96 |
159 |
160 |
161 |
162 |
163 |
--------------------------------------------------------------------------------
/R/plot_importance.R:
--------------------------------------------------------------------------------
1 | #' Plot importance measures
2 | #'
3 | #' This functions plots selected measures of importance for variables and interactions.
4 | #' It is possible to visualise importance table in two ways: radar plot with six measures
5 | #' and scatter plot with two choosen measures.
6 | #'
7 | #' Available measures:
8 | #'\itemize{
9 | #'\item "sumGain" - sum of Gain value in all nodes, in which given variable occurs,
10 | #'\item "sumCover" - sum of Cover value in all nodes, in which given variable occurs; for LightGBM models: number of observation, which pass through the node,
11 | #'\item "mean5Gain" - mean gain from 5 occurrences of given variable with the highest gain,
12 | #'\item "meanGain" - mean Gain value in all nodes, in which given variable occurs,
13 | #'\item "meanCover" - mean Cover value in all nodes, in which given variable occurs; for LightGBM models: mean number of observation, which pass through the node,
14 | #'\item "freqency" - number of occurrences in the nodes for given variable.
15 | #'}
16 | #'
17 | #' Additionally for plots with single variables:
18 | #'\itemize{
19 | #'\item "meanDepth" - mean depth weighted by gain,
20 | #'\item "numberOfRoots" - number of occurrences in the root,
21 | #'\item "weightedRoot" - mean number of occurrences in the root, which is weighted by gain.
22 | #'}
23 | #'
24 | #' @param x a result from the \code{importance} function.
25 | #' @param top number of positions on the plot or NULL for all variable. Default 10.
26 | #' @param radar TRUE/FALSE. If TRUE the plot shows
27 | #' six measures of variables' or interactions' importance in the model.
28 | #' If FALSE the plot containing two chosen measures
29 | #' of variables' or interactions' importance in the model.
30 | #' @param text_start_point place, where the names of the particular feature start. Available for `radar=TRUE`. Range from 0 to 1. Default 0.5.
31 | #' @param text_size size of the text on the plot. Default 3.5.
32 | #' @param xmeasure measure on the x-axis.Available for `radar=FALSE`. Default "sumCover".
33 | #' @param ymeasure measure on the y-axis. Available for `radar=FALSE`. Default "sumGain".
34 | #' @param ... other parameters.
35 | #'
36 | #' @return a ggplot object
37 | #'
38 | #' @import ggplot2
39 | #' @import data.table
40 | #' @importFrom DALEX theme_drwhy
41 | #' @importFrom ggrepel geom_label_repel
42 | #' @importFrom ggiraphExtra coord_radar
43 | #'
44 | #' @examples
45 | #' library("EIX")
46 | #' library("Matrix")
47 | #' sm <- sparse.model.matrix(left ~ . - 1, data = HR_data)
48 | #'
49 | #' library("xgboost")
50 | #' param <- list(objective = "binary:logistic", max_depth = 2)
51 | #' xgb_model <- xgboost(sm, params = param, label = HR_data[, left] == 1, nrounds = 25, verbose=0)
52 | #'
53 | #' imp <- importance(xgb_model, sm, option = "both")
54 | #' imp
55 | #' plot(imp, top = 10)
56 | #'
57 | #' imp <- importance(xgb_model, sm, option = "variables")
58 | #' imp
59 | #' plot(imp, top = nrow(imp))
60 | #'
61 | #' imp <- importance(xgb_model, sm, option = "interactions")
62 | #' imp
63 | #' plot(imp, top = nrow(imp))
64 | #'
65 | #' imp <- importance(xgb_model, sm, option = "variables")
66 | #' imp
67 | #' plot(imp, top = NULL, radar = FALSE, xmeasure = "sumCover", ymeasure = "sumGain")
68 | #'
69 | #'\donttest{
70 | #'library(lightgbm)
71 | #'train_data <- lgb.Dataset(sm, label = HR_data[, left] == 1)
72 | #'params <- list(objective = "binary", max_depth = 2)
73 | #'lgb_model <- lgb.train(params, train_data, 25)
74 | #'
75 | #' imp <- importance(lgb_model, sm, option = "both")
76 | #' imp
77 | #' plot(imp, top = nrow(imp))
78 | #'
79 | #' imp <- importance(lgb_model, sm, option = "variables")
80 | #' imp
81 | #' plot(imp, top = NULL, radar = FALSE, xmeasure = "sumCover", ymeasure = "sumGain")
82 | #'
83 | #'}
84 | #'
85 | #' @export
86 |
87 |
88 | plot.importance <- function(x, ..., top = 10, radar = TRUE, text_start_point = 0.5, text_size=3.5,
89 | xmeasure = "sumCover", ymeasure = "sumGain"){
90 |
91 | Feature <- sumGain <- sumCover <- meanGain <- meanCover <-
92 | mean5Gain <- . <- value <- variable <- hjust <- NULL
93 |
94 | if (is.null(top))
95 | top <- nrow(x)
96 |
97 |
98 | if (radar == FALSE) {
99 | ggplot(data.frame(x[1:top, ]),
100 | aes_string(x = xmeasure, y = ymeasure, label = "Feature")) +
101 | geom_point() +
102 | scale_size() + geom_label_repel() + theme_drwhy()
103 |
104 | }else{
105 | import <- as.data.table(x[1:top, ])
106 | import <- import[1:top, .(Feature,
107 | sumGain = sumGain / max(import[, sumGain]),
108 | sumCover = sumCover / max(import[, sumCover]),
109 | meanGain = meanGain / max(import[, meanGain]),
110 | meanCover = meanCover / max(import[, meanCover]),
111 | mean5Gain = mean5Gain / max(import[, mean5Gain]),
112 | frequency = frequency / max(import[, frequency]))]
113 | data<-import[,Feature:= ifelse(nchar(import[,Feature])>20, gsub(":", ": :",import[,Feature]),Feature)]
114 |
115 | import$Feature <- factor(import$Feature, levels = import$Feature[order(import$sumGain, decreasing = TRUE)])
116 |
117 | #angles and hjust of labels
118 | numberOfBars=nrow(import)
119 | angle= 90-360*(row(import)[,1]-0.5)/numberOfBars
120 |
121 | import$hjust<-ifelse( angle < -90, 1, 0)
122 | import$angle<-ifelse(angle < -90, angle+180, angle)
123 |
124 | data_to_plot <- melt(import, id = c(1,8,9), measures = 2:6, value.factor = FALSE)
125 | data<-data_to_plot[,.(hjust=mean(hjust),angle=mean(angle)), by=Feature]
126 |
127 | ggplot(data.frame(data_to_plot),
128 | aes(x = Feature, y = value, colour = variable, group = variable)) +
129 | geom_line(size = 1.5) +
130 | geom_point(size = 2.5) +
131 | theme_drwhy() +
132 | theme(axis.title.x = element_blank(),
133 | axis.title.y = element_blank(),
134 | legend.position = "bottom",
135 | panel.grid.major.y = element_line(colour = "gray68", linetype = "dashed", size = 0.4),
136 | axis.line = element_blank(),
137 | axis.text.x=element_blank(),) +
138 | labs(fill = "Measures")+
139 | coord_radar() +
140 | geom_text(data=data, aes(x=Feature, y= rep(text_start_point,top), label=lapply(strwrap(data[,Feature], width = 10, simplify = FALSE), paste, collapse="\n"), hjust=hjust), color="#371ea3", fontface="bold",alpha=0.6, size=text_size, angle= data$angle, inherit.aes = FALSE )
141 |
142 | }
143 | }
144 |
--------------------------------------------------------------------------------
/docs/reference/index.html:
--------------------------------------------------------------------------------
1 |
2 |
3 |
4 |
5 |
6 |
7 |
8 |
9 |
41 |
92 |
93 |
94 |
42 |
91 |
43 |
48 |
49 |
50 |
51 | Why are our best and most experienced employees leaving prematurely?
52 |
53 |
90 |
50 |
51 | Why are our best and most experienced employees leaving prematurely?
52 |
97 |
147 |
148 |
158 |
98 |
134 |
146 |
99 |
101 |
102 |
103 | Why are our best and most experienced employees leaving prematurely?
100 |A dataset from Kaggle competition Human Resources Analytics. 104 | https://www.kaggle.com/ludobenistant/hr-analytics/data
105 | 106 | 107 | 108 |Format
109 | 110 |A data table with 14999 rows and 10 variables
111 | 112 |Source
113 | 114 |https://www.kaggle.com/ludobenistant/hr-analytics/data, https://cran.r-project.org/package=breakDown
115 | 116 |Details
117 | 118 |The description of the dataset was copied from the breakDown package.
-
120 |
satisfaction_level Level of satisfaction (0-1)
121 | last_evaluation Time since last performance evaluation (in Years)
122 | number_project Number of projects completed while at work
123 | average_montly_hours Average monthly hours at workplace
124 | time_spend_company Number of years spent in the company
125 | Work_accident Whether the employee had a workplace accident
126 | left Whether the employee left the workplace or not (1 or 0) Factor
127 | promotion_last_5years Whether the employee was promoted in the last five years
128 | sales Department in which they work for
129 | salary Relative level of salary (high)
130 |
39 |
40 |
95 |
96 |
199 |
200 |
201 |
202 |
203 |
--------------------------------------------------------------------------------
/R/importance.R:
--------------------------------------------------------------------------------
1 | #' Importance of variables and interactions in the model
2 | #'
3 | #' This functions calculates a table with selected measures of importance
4 | #' for variables and interactions.
5 | #'
6 | #' Available measures:
7 | #'\itemize{
8 | #'\item "sumGain" - sum of Gain value in all nodes, in which given variable occurs,
9 | #'\item "sumCover" - sum of Cover value in all nodes, in which given variable occurs; for LightGBM models: number of observation, which pass through the node,
10 | #'\item "mean5Gain" - mean gain from 5 occurrences of given variable with the highest gain,
11 | #'\item "meanGain" - mean Gain value in all nodes, in which given variable occurs,
12 | #'\item "meanCover" - mean Cover value in all nodes, in which given variable occurs; for LightGBM models: mean number of observation, which pass through the node,
13 | #'\item "freqency" - number of occurrences in the nodes for given variable.
14 | #'}
15 | #'
16 | #' Additionally for table with single variables:
17 | #'\itemize{
18 | #'\item "meanDepth" - mean depth weighted by gain,
19 | #'\item "numberOfRoots" - number of occurrences in the root,
20 | #'\item "weightedRoot" - mean number of occurrences in the root, which is weighted by gain.
21 | #'}
22 | #'
23 | #' @param xgb_model a xgboost or lightgbm model.
24 | #' @param data a data table with data used to train the model.
25 | #' @param option if "variables" then table includes only single variables,
26 | #' if "interactions", then only interactions
27 | #' if "both", then both single variable and interactions.
28 | #' Default "both".
29 | #' @param digits number of significant digits that shall be returned. Will be passed to the signif() functions.
30 | #'
31 | #' @return a data table
32 | #'
33 | #' @import data.table
34 | #' @importFrom stats frequency
35 | #' @importFrom stats weighted.mean
36 | #'
37 | #' @examples
38 | #' library("EIX")
39 | #' library("Matrix")
40 | #' sm <- sparse.model.matrix(left ~ . - 1, data = HR_data)
41 | #'
42 | #' library("xgboost")
43 | #' param <- list(objective = "binary:logistic", max_depth = 2)
44 | #' xgb_model <- xgboost(sm, params = param, label = HR_data[, left] == 1, nrounds = 25, verbose=0)
45 | #'
46 | #' imp <- importance(xgb_model, sm, option = "both")
47 | #' imp
48 | #' plot(imp, top = 10)
49 | #'
50 | #' imp <- importance(xgb_model, sm, option = "variables")
51 | #' imp
52 | #' plot(imp, top = nrow(imp))
53 | #'
54 | #' imp <- importance(xgb_model, sm, option = "interactions")
55 | #' imp
56 | #' plot(imp, top = nrow(imp))
57 | #'
58 | #' imp <- importance(xgb_model, sm, option = "variables")
59 | #' imp
60 | #' plot(imp, top = NULL, radar = FALSE, xmeasure = "sumCover", ymeasure = "sumGain")
61 | #'
62 | #'\donttest{
63 | #'library(lightgbm)
64 | #'train_data <- lgb.Dataset(sm, label = HR_data[, left] == 1)
65 | #'params <- list(objective = "binary", max_depth = 2)
66 | #'lgb_model <- lgb.train(params, train_data, 25)
67 | #'
68 | #' imp <- importance(lgb_model, sm, option = "both")
69 | #' imp
70 | #' plot(imp, top = nrow(imp))
71 | #'
72 | #' imp <- importance(lgb_model, sm, option = "variables")
73 | #' imp
74 | #' plot(imp, top = NULL, radar = FALSE, xmeasure = "sumCover", ymeasure = "sumGain")
75 | #'
76 | #'}
77 | #'
78 | #' @export
79 |
80 | importance <- function(xgb_model, data, option = "both", digits = 4){
81 | importance <- NULL
82 |
83 | if (option == "both") {
84 | importance <- importanceTableMixed(xgb_model, data)
85 | }
86 | if (option == "variables") {
87 | importance <- importanceSingleVariable(xgb_model, data)
88 | }
89 | if (option == "interactions") {
90 | importance <- importanceInteraction(xgb_model, data)
91 | }
92 |
93 | importance <- cbind(importance[, 1], signif(importance[, -1], digits = digits))
94 | #importance <- unlist(importance)
95 |
96 | class(importance) <- c("importance", "data.table")
97 |
98 | return(importance[])
99 |
100 | }
101 |
102 | importanceTableMixed <- function(xgb_model, data){
103 | parentsGain <- childsGain <- name_pair <- Cover <- Feature <-
104 | Gain <- indx <- . <- Quality <- NULL
105 |
106 | trees <- noLeavesGainTable(xgb_model, data)
107 |
108 | #single variables
109 | importanceSingle <-
110 | trees[(interaction == FALSE) | (is.na(interaction)), .(Feature, Gain = Quality, Cover)]
111 |
112 | #interactions
113 | interactions <- trees[interaction == TRUE]
114 | importanceInter <- interactions[, .(Feature = name_pair, Gain = childsGain, Cover)]
115 | importance <- rbind(importanceSingle, importanceInter)
116 |
117 | importance4 <-
118 | merge(importance[, .(sumGain = sum(Gain),
119 | sumCover = sum(Cover),
120 | meanGain = mean(Gain),
121 | meanCover = mean(Cover),
122 | frequency = .N), by = Feature],
123 | mean5gain(importance), by = "Feature")
124 |
125 | setorderv(importance4, "sumGain", -1)
126 |
127 | return(importance4[])
128 |
129 | }
130 |
131 |
132 | importanceInteraction <- function(xgb_model, data) {
133 | parentsGain <- childsGain <- name_pair <- Cover <- . <- Feature <- Gain <- indx <- NULL
134 |
135 | trees <- noLeavesGainTable(xgb_model, data)
136 | trees <- trees[interaction == TRUE]
137 | tress <- trees[, `:=`(Feature = name_pair, Gain = childsGain)]
138 | tress <- trees[, .(Feature, Gain, Cover)]
139 | importance <- merge(trees[, .(sumGain = sum(Gain),
140 | sumCover = sum(Cover),
141 | meanGain = mean(Gain),
142 | meanCover = mean(Cover),
143 | frequency = .N), by = Feature],
144 | mean5gain(trees), by = "Feature")
145 |
146 | setorderv(importance, "sumGain", -1)
147 |
148 | return(importance[])
149 | }
150 |
151 |
152 | importanceSingleVariable <- function(xgb_model, data) {
153 | Feature <- Gain <- Quality <- Cover <- indx <- . <- NULL
154 |
155 | trees <- noLeavesGainTable(xgb_model, data)
156 | trees[, Gain := Quality]
157 |
158 | importance1 <- merge(countRoots(trees),calculateWeightedDepth(trees), by = "Feature", all = TRUE)[, -"sumGain"]
159 |
160 | trees <- trees[, .(Feature, Gain, Cover)]
161 |
162 | importance2 <- merge(trees[,.(sumGain=sum(Gain),
163 | sumCover=sum(Cover),
164 | meanGain=mean(Gain),
165 | meanCover=mean(Cover),
166 | frequency=.N),,by=Feature],
167 | mean5gain(trees), by="Feature")
168 | importance <- merge(importance1, importance2, by = "Feature")[, -"count"]
169 |
170 | setorderv(importance, "sumGain", -1)
171 | importance[is.na(importance)] <- 0
172 |
173 | return(importance[])
174 |
175 | }
176 |
177 | #Table with number of roots and weighedRoot
178 | #counts how many times each variable is in the root of the tree and calculates the weighedRoot-number of occurrences in root weighed by Gain.
179 | countRoots <- function(trees) {
180 | Node <- Quality <- Feature <- sumGain <- . <-
181 | weightedRoot <- numberOfRoots <- NULL
182 |
183 | roots <- trees[Node == 0, ]
184 | roots <- roots[, .(sumGain = sum(Quality), numberOfRoots = .N), by = Feature]
185 | sumGains <- sum(roots[, sumGain])
186 | roots <- roots[, weightedRoot := round(roots[, sumGain] * roots[, numberOfRoots] / sumGains, 4)]
187 |
188 | return(roots[])
189 |
190 | }
191 |
192 | #Mean form 5 nodes with the highests gain
193 | mean5gain <- function(trees) {
194 | indx <- Gain <- . <- Feature <- NULL
195 |
196 | setorder(setDT(trees), Feature,-Gain)[, indx := seq_len(.N), by = Feature]
197 | importanceTop <- trees[indx <= 5]
198 | importance <- importanceTop[, .(mean5Gain = mean(Gain)), by = Feature]
199 |
200 | return(importance[])
201 | }
202 |
203 | #calculates depth mean for every variable weighted by Gain
204 | calculateWeightedDepth <- function(trees) {
205 | Feature <- depth <- Quality <- . <- NULL
206 |
207 | trees <- trees[, .(meanDepth = weighted.mean(depth, Quality), count = .N), by = Feature]
208 |
209 | return(trees[])
210 | }
211 |
212 |
213 | noLeavesGainTable <- function(xgb_model, data) {
214 | parentsName <- Feature <- Tree <- name_pair <- parentsGain <- childsGain <-
215 | . <- Cover <- parentsCover <- interaction <- Node <- Quality <- depth <- NULL
216 |
217 | treeList <- calculateGain(xgb_model, data)
218 | trees <- rbindlist(treeList)
219 | trees <- trees[Feature != "Leaf", .(Tree, Node, name_pair, parentsGain, childsGain, Cover,
220 | parentsCover, Feature, Quality, parentsName, interaction, depth)]
221 |
222 | return(trees[])
223 | }
224 |
--------------------------------------------------------------------------------
/docs/reference/titanic.html:
--------------------------------------------------------------------------------
1 |
2 |
3 |
4 |
5 |
6 |
7 |
8 |
9 |
41 |
92 |
93 |
94 |
42 |
91 |
43 |
48 |
49 |
50 |
51 |
52 |
53 |
90 |
50 |
51 |
52 |
97 |
187 |
188 |
198 |
98 |
179 |
180 |
186 |
99 |
104 |
105 | 100 | Reference 101 | version 1.0 102 |
103 |
106 |
178 |
39 |
40 |
90 |
91 |
175 |
176 |
177 |
178 |
179 |
--------------------------------------------------------------------------------
/docs/reference/titanic_data.html:
--------------------------------------------------------------------------------
1 |
2 |
3 |
4 |
5 |
6 |
7 |
8 |
9 |
41 |
87 |
88 |
89 |
42 |
86 |
43 |
48 |
49 |
50 |
51 | Passengers and Crew on the RMS Titanic
52 |
53 |
85 |
50 |
51 | Passengers and Crew on the RMS Titanic
52 |
92 |
163 |
164 |
174 |
93 |
148 |
162 |
94 |
96 |
97 |
98 | Passengers and Crew on the RMS Titanic
95 |The titanic data is a complete list of passengers and crew members on the RMS Titanic.
99 | It includes a variable indicating whether a person did survive the sinking of the RMS
100 | Titanic on April 15, 1912.
data(titanic)104 | 105 |
Format
106 | 107 |a data frame with 2207 rows and 11 columns
108 | 109 |Source
110 | 111 |The description of dataset was copied from the DALEX package.
112 | This dataset was copied from the stablelearner package and went through few variable
113 | transformations. The complete list of persons on the RMS titanic was downloaded from
114 | https://www.encyclopedia-titanica.org on April 5, 2016. The information given
115 | in sibsp and parch was adopoted from a data set obtained from http://biostat.mc.vanderbilt.edu/DataSets.
Details
118 | 119 |The description of the dataset was copied from the DALEX package.
This dataset was copied from the stablelearner package and went through few variable
121 | transformations. Levels in embarked was replaced with full names, sibsp, parch and fare
122 | were converted to numerical variables and values for crew were replaced with 0.
123 | If you use this dataset please cite the original package.
From stablelearner: The website https://www.encyclopedia-titanica.org offers detailed information about passengers and crew
125 | members on the RMS Titanic. According to the website 1317 passengers and 890 crew member were abord.
126 | 8 musicians and 9 employees of the shipyard company are listed as passengers, but travelled with a
127 | free ticket, which is why they have NA values in fare. In addition to that, fare
128 | is truely missing for a few regular passengers.
-
130 |
gender a factor with levels
maleandfemale.
131 | age a numeric value with the persons age on the day of the sinking.
132 | class a factor specifying the class for passengers or the type of service aboard for crew members.
133 | embarked a factor with the persons place of of embarkment (Belfast/Cherbourg/Queenstown/Southampton).
134 | country a factor with the persons home country.
135 | fare a numeric value with the ticket price (
0for crew members, musicians and employees of the shipyard company).
136 | sibsp an ordered factor specifying the number if siblings/spouses aboard; adopted from Vanderbild data set (see below).
137 | parch an ordered factor specifying the number of parents/children aboard; adopted from Vanderbild data set (see below).
138 | survived a factor with two levels (
noandyes) specifying whether the person has survived the sinking.
139 |
References
142 | 143 |https://www.encyclopedia-titanica.org, http://biostat.mc.vanderbilt.edu/DataSets, 144 | https://CRAN.R-project.org/package=stablelearner, https://cran.r-project.org/web/packages/DALEX/index.html.
145 | 146 | 147 |
39 |
40 |
95 |
96 |
180 |
181 |
182 |
183 |
184 |
--------------------------------------------------------------------------------
/docs/reference/countPairs.html:
--------------------------------------------------------------------------------
1 |
2 |
3 |
4 |
5 |
6 |
7 |
8 |
9 |
41 |
92 |
93 |
94 |
42 |
91 |
43 |
48 |
49 |
50 |
51 | Passengers and Crew on the RMS Titanic
52 |
53 |
90 |
50 |
51 | Passengers and Crew on the RMS Titanic
52 |
97 |
168 |
169 |
179 |
98 |
153 |
167 |
99 |
101 |
102 |
103 | Passengers and Crew on the RMS Titanic
100 |The titanic data is a complete list of passengers and crew members on the RMS Titanic.
104 | It includes a variable indicating whether a person did survive the sinking of the RMS
105 | Titanic on April 15, 1912.
data(titanic_data)109 | 110 |
Format
111 | 112 |a data frame with 2207 rows and 11 columns
113 | 114 |Source
115 | 116 |The description of dataset was copied from the DALEX package.
117 | This dataset was copied from the stablelearner package and went through few variable
118 | transformations. The complete list of persons on the RMS titanic was downloaded from
119 | https://www.encyclopedia-titanica.org on April 5, 2016. The information given
120 | in sibsp and parch was adopoted from a data set obtained from http://biostat.mc.vanderbilt.edu/DataSets.
Details
123 | 124 |The description of the dataset was copied from the DALEX package.
This dataset was copied from the stablelearner package and went through few variable
126 | transformations. Levels in embarked was replaced with full names, sibsp, parch and fare
127 | were converted to numerical variables and values for crew were replaced with 0.
128 | If you use this dataset please cite the original package.
From stablelearner: The website https://www.encyclopedia-titanica.org offers detailed information about passengers and crew
130 | members on the RMS Titanic. According to the website 1317 passengers and 890 crew member were abord.
131 | 8 musicians and 9 employees of the shipyard company are listed as passengers, but travelled with a
132 | free ticket, which is why they have NA values in fare. In addition to that, fare
133 | is truely missing for a few regular passengers.
-
135 |
gender a factor with levels
maleandfemale.
136 | age a numeric value with the persons age on the day of the sinking.
137 | class a factor specifying the class for passengers or the type of service aboard for crew members.
138 | embarked a factor with the persons place of of embarkment (Belfast/Cherbourg/Queenstown/Southampton).
139 | country a factor with the persons home country.
140 | fare a numeric value with the ticket price (
0for crew members, musicians and employees of the shipyard company).
141 | sibsp an ordered factor specifying the number if siblings/spouses aboard; adopted from Vanderbild data set (see below).
142 | parch an ordered factor specifying the number of parents/children aboard; adopted from Vanderbild data set (see below).
143 | survived a factor with two levels (
noandyes) specifying whether the person has survived the sinking.
144 |
References
147 | 148 |https://www.encyclopedia-titanica.org, http://biostat.mc.vanderbilt.edu/DataSets, 149 | https://CRAN.R-project.org/package=stablelearner, https://cran.r-project.org/package=DALEX.
150 | 151 | 152 |
63 |
64 |
99 |
100 |
101 |
102 |
103 |
172 |
173 |
174 |
175 |
176 |
177 |
178 |
--------------------------------------------------------------------------------
/docs/reference/lollipop.html:
--------------------------------------------------------------------------------
1 |
2 |
3 |
4 |
5 |
6 |
7 |
8 |
9 |
104 |
161 |
162 |
171 |
105 |
121 |
131 |
132 |
149 |
160 |
106 |
109 |
110 |
111 | Table of occurancess number
107 | 108 |countPairs.Rd
112 |
113 |
116 |
117 | Table containing occurancess number of variables' pairs in the model.
114 | 115 |countPairs(xgb.model, data)118 | 119 |
Arguments
120 || xgb.model | 124 |a xgboost or lightgbm model |
125 |
|---|---|
| data | 128 |a data table with data used to train the model |
129 |
Value
133 | 134 |a data table
135 | 136 | 137 |Examples
138 |148 |#> Warning: pakiet 'Matrix' został zbudowany w wersji R 3.4.4library("data.table")#> Warning: pakiet 'data.table' został zbudowany w wersji R 3.4.4library("xgboost")#> Warning: pakiet 'xgboost' został zbudowany w wersji R 3.4.4140 | dt_HR <- data.table(HR_data) 141 | sm <- sparse.model.matrix(left ~ . - 1, data = dt_HR) 142 | 143 | param <- list(objective = "binary:logistic", base_score = 0.5, max_depth = 2) 144 | xgb.model <- xgboost( param = param, data = sm, label = dt_HR[, left] == 1, nrounds = 50, verbose = FALSE) 145 | 146 | countPairs(xgb.model, sm)#> Error in countPairs(xgb.model, sm): nie udało się znaleźć funkcji 'countPairs'147 |
39 |
40 |
95 |
96 |
174 |
175 |
176 |
177 |
178 |
--------------------------------------------------------------------------------
/docs/reference/EIX_lollipop.html:
--------------------------------------------------------------------------------
1 |
2 |
3 |
4 |
5 |
6 |
7 |
8 |
9 |
41 |
92 |
93 |
94 |
42 |
91 |
43 |
48 |
49 |
50 |
51 | Tables needed for lollipop plot
52 |
53 |
90 |
50 |
51 | Tables needed for lollipop plot
52 |
97 |
162 |
163 |
173 |
98 |
113 |
123 |
124 |
150 |
161 |
99 |
101 |
102 |
103 | Tables needed for lollipop plot
100 |This function calculates two tables needed to generate lollipop plot, which visualise the model. 104 | The first table contains information about all nodes in the trees forming a model. 105 | It includes gain value, depth and ID of each nodes. 106 | The second table contains similarly information about roots in the trees.
107 | 108 | 109 |lollipop(xgb_model, data)110 | 111 |
Arguments
112 || xgb_model | 116 |a xgboost or lightgbm model. |
117 |
|---|---|
| data | 120 |a data table with data used to train the model. |
121 |
Value
125 | 126 |an object of the lollipop class
127 | 128 | 129 |Examples
130 |149 |library("EIX") 131 | library("Matrix") 132 | sm <- sparse.model.matrix(left ~ . - 1, data = HR_data) 133 | 134 | library("xgboost") 135 | param <- list(objective = "binary:logistic", max_depth = 2) 136 | xgb_model <- xgboost(sm, params = param, label = HR_data[, left] == 1, nrounds = 25, verbose = 0) 137 | 138 | lolli <- lollipop(xgb_model, sm) 139 | plot(lolli, labels = "topAll", log_scale = TRUE)140 |library(lightgbm) 141 | train_data <- lgb.Dataset(sm, label = HR_data[, left] == 1) 142 | params <- list(objective = "binary", max_depth = 2) 143 | lgb_model <- lgb.train(params, train_data, 25) 144 | 145 | lolli <- lollipop(lgb_model, sm) 146 | plot(lolli, labels = "topAll", log_scale = TRUE)147 |148 |
39 |
40 |
75 |
76 |
192 |
193 |
194 |
195 |
196 |
--------------------------------------------------------------------------------
/vignettes/EIX.Rmd:
--------------------------------------------------------------------------------
1 | ---
2 | title: "EIX: Explain Interactions in XGBoost"
3 | author: "Ewelina Karbowiak"
4 | date: "2018-12-07"
5 | output: rmarkdown::html_vignette
6 | vignette: >
7 | %\VignetteIndexEntry{EIX Explain Interactions in XGBoost}
8 | %\VignetteEngine{knitr::rmarkdown}
9 | %\VignetteEncoding{UTF-8}
10 | ---
11 |
12 | ```{r, echo=FALSE}
13 | knitr::opts_chunk$set(fig.height = 6, fig.width = 6, fig.align = "center")
14 |
15 | ```
16 |
17 | Package `EIX` is the set of tools to explore the structure of XGBoost and lightGBM models. It includes functions finding strong interactions and also checking importance of single variables and interactions by usage different measures. `EIX` consists several functions to visualize results.
18 |
19 | Almost all `EIX` functions require only two parameters: a XGBoost or LightGBM model and data table used as training dataset. The exceptions are the `waterfall` function and its plot. The first one requires parameters: a XGBoost model and observation, which prediction has to be explained). These two functions support only XGBoost models. All plots are created with package `ggplot2`. Most of them use plot theme `theme_mi2` from `DALEX`.
20 |
21 | ## Data Info
22 |
23 | This vignette shows usage of `EIX` package. It lets to explain XGBoost prediction model concerning departures of employees from company using HR_data. Dataset was taken from kaggle and consists 14999 observations and 10 variables. The dataset is also available in package `EIX` and there it is described more precisely.
24 |
25 | ```{r}
26 | #devtools :: install_github("ModelOriented/EIX")
27 | library("EIX")
28 | set.seed(4)
29 | knitr::kable(head(HR_data))
30 | ```
31 |
32 | To create correct XGBoost model, remember to change categorical features to factors and next change the data frame to sparse matrix. The categorical features are one-hot encoded.
33 |
34 | ```{r, warning=FALSE, message=FALSE}
35 | library("Matrix")
36 | sparse_matrix <- sparse.model.matrix(left ~ . - 1, data = HR_data)
37 | head(sparse_matrix)
38 | ```
39 |
40 |
41 | ## Xgboost model creation
42 |
43 | Package `EIX` uses table, which was generated by `xgboost::xgb.model.dt.tree` with information about trees, their nodes and leaves.
44 |
45 | ```{r}
46 | library("xgboost")
47 | param <- list(objective = "binary:logistic", max_depth = 2)
48 | xgb_model <- xgboost(sparse_matrix, params = param, label = HR_data[, left] == 1, nrounds = 50, verbose = FALSE)
49 | knitr::kable(head(xgboost::xgb.model.dt.tree(colnames(sparse_matrix),xgb_model)))
50 | ```
51 |
52 | Function `xgboost::xgb.importance` shows importance of single variables. `EIX` adds new measures of variables’ importance and shows also importance of interactions.
53 |
54 | ```{r}
55 | knitr::kable(head(xgboost::xgb.importance(colnames(sparse_matrix),xgb_model)))
56 | ```
57 |
58 | ## Model visualization
59 |
60 | The `lollipop` plot is used to visualize the model in such way that the most important variables and interactions are visible.
61 |
62 | On the x-axis, there are tree numbers and on the y-axis there is **Gain** measure for each node. One segment is one tree in the model and each point is one node. On the plot there are all nodes, which are not leaves. Shape of points signifies depth of node. All roots on the plot are connected by a red line. If in the same segment there is a variable with a higher depth above the variable with a lower depth, it means that interaction occurs.
63 |
64 | There is opportunity to set a different way of labeling. On the plot we can see the most important variables in roots (horizontal labels), and interactions (vertical labels), this is option `labels = "topAll"` which is default. Moreover, there are two additional options: `labels = "roots"` - for variables in roots only, `labels = "interactions"` for interactions only. The numbers of labels visible on the plot you can change by parametr `threshold` (range from 0 to 1, default 0.1).
65 | The plot is on a logarithmic scale because the initial trees usually are the most important. You can change the scale of the plot by setting the parameter `log_scale = FALSE`.
66 |
67 | ```{r}
68 | lolli<-lollipop(xgb_model,sparse_matrix)
69 | plot(lolli)
70 | #plot(lolli, threshold=0.05)
71 | #plot(lolli, labels="roots")
72 | #plot(lolli, labels="interactions")
73 | #plot(lolli, labels="roots", threshold=0.05)
74 | #plot(lolli, labels="interactions",threshold=0.05)
75 | #plot(lolli, log_scale = FALSE)
76 | ```
77 |
78 |
79 | ## Interactions
80 |
81 | We can consider interactions in two ways. In first approach we can explore all pairs of variable, which occur in the model one above the other. This approach is not the best one, because we cannot distinguish if pair of variables are real interaction or not. In this approach high gain of pair can be a result of high gain of down variable (child). To explore pairs of variables you can generate table with them using function `interactions` with parametr `option = "pairs"`. This table includes **Gain** measure and number of occurrences of pairs. You can also use the function `plot` to visualize **Gain** measure.
82 |
83 | ```{r}
84 | pairs<-interactions(xgb_model, sparse_matrix, option = "pairs")
85 | head(pairs)
86 | plot(pairs)
87 | ```
88 |
89 | The `interactions` plot is a matrix plot with a child from the pair on the x-axis and the parent on the y-axis. The color of the square at the intersection of two variables means value of **sumGain** measure. The darker square, the higher **sumGain** of variable pairs. The range of **sumGain** measure is divided into four equal parts: `very low, low, medium, high`.
90 |
91 | In second approach, to find strong interactions, we can consider only these pairs of variables, where variable on the bottom (child) has higher gain than variable on the top (parent). We can also create ranking of interactions using function `importance` with parameter `option = "interactions"`. More details in the next section.
92 |
93 | ```{r, warning=FALSE, message=FALSE}
94 | interactions<-interactions(xgb_model, sparse_matrix, option = "interactions")
95 | head(interactions)
96 | plot(interactions)
97 | ```
98 |
99 |
100 | ## Variables' and interactions’ importance
101 |
102 | For exploring variables’ and interactions’ importance there are three functions in `EIX` package: `importance`, its `plot` with parameter `radar = TRUE` or `radar = FALSE`.
103 | With `EIX` package we can compare importance of single variables and interactions. The functions `importance` can return three kinds of outputs, depending on the `opt` parameter:
104 |
105 | * `option = "variables"` - it consists only single variables
106 |
107 | * `option = "interactions"`- only interactions
108 |
109 | * `option = "both"`- output shows importance both single variables and interactions.
110 |
111 | NOTE: `option = "both"` is not direct connection `option = "variables"` and `option = "interactions"`, because values of variable importance measure, which were in the interactions, are not included in importance of single variable.
112 |
113 | In `EIX` the following measures are available:
114 |
115 | * **sumGain** - sum of Gain value in all nodes, in which given variable occurs
116 | * **sumCover** - sum of Cover value in all nodes, in which given variable occurs; for LightGBM models: number of observation, which pass through the node
117 | * **mean5Gain** - mean gain from 5 occurrences of given variable with the highest gain
118 | * **meanGain** - mean Gain value in all nodes, in which given variable occurs
119 | * **meanCover** - mean Cover value in all nodes, in which given variable occurs; for LightGBM models: mean number of observation, which pass through the node
120 | * **freqency** - number of occurrences in the nodes for given variable
121 |
122 | `EIX` package gives additionally measures of variables importance for single variable:
123 |
124 | * **counterRoot** - number of occurrences in the root
125 | * **weightedRoot** - mean number of occurrences in the root, which is weighted by gain
126 | * **meanDepth ** - mean depth weighted by gain
127 |
128 | The function `importance` returns a table with all available importance measures for given option.
129 | The table is sorted by descending value of **sumGain**.
130 |
131 | The function `plot` with parameter `radar = FALSE` and a result from the `importance` function as an argument shows two measures of importance, which can be chosen by `xmeasure` and `ymeasure` parameters. By parameter `top` we can decide how many positions will be included in the plot.
132 |
133 | ```{r}
134 | importance<-importance(xgb_model, sparse_matrix, option = "both")
135 | head(importance)
136 | plot(importance, radar=FALSE)
137 | #plot(importance, xmeasure = "mean5Gain", ymeasure = "sumGain", top = 15, radar=FALSE)
138 | ```
139 |
140 | The function `plot` with parameter `radar = TRUE` enables to compare different measures of variables and interactions importance on the radar plot from `ggiraphExtra` package.
141 | Bellow I attach the example of radar plot. On the outside of the circle there are names of variables or interactions. Colored lines represent various measures of importance. The positions on the plot are sorted decreasing. The variable with the highest **sumGain** value is on the right of 12 o'clock. Next the **sumGain** value decreases in a clockwise direction. On the plot it is possible to change place, where the features names start by parameter `text_start_point` (range from 0 to 1, default 0.5), and size of this text by parametrer `text_size`.
142 |
143 | ```{r}
144 | plot(importance)
145 | #plot(importance, text_start_point = 0.3)
146 | #plot(importance, text_size = 4)
147 | #plot(importance, top=15)
148 | ```
149 |
150 | ## Explanation of single prediction including interactions
151 |
152 | For single prediction explaining package `EIX` uses two packages: `xgboostExplainer` i `breakDown`. The package `xgboostExplainer` is a tool to interpreting prediction of xgboost model. The package `EIX` uses its code and modifies it to include interactions. The methodology of plot creation comes from package `breakDown`.
153 |
154 | The function `waterfall` returns table with variables’ impact on the prediction of the model. Depending on the parameter `option`, the table includes interactions (`option = "interactions"`- default) or does not (`option = "variables"`). The function `plot` with `waterfall` object as an argument visualizes this table. On the y-axis there are: intercept (it is the probability that random variable from training dataset will be 1), variables (which have an impact on prediction) and final prognosis of the model. On the x-axis there is log-odds of impact each variables.
155 |
156 | ```{r}
157 | data <- HR_data[9,]
158 | new_observation <- sparse_matrix[9,]
159 | wf<-waterfall(xgb_model, new_observation, data, option = "interactions")
160 | wf
161 | plot(wf)
162 | #wf<-waterfall(xgb_model, new_observation, data, option = "interactions", baseline = "intercept")
163 | #wf
164 | #plot(wf)
165 | ```
166 |
--------------------------------------------------------------------------------
/docs/reference/plot.lollipop.html:
--------------------------------------------------------------------------------
1 |
2 |
3 |
4 |
5 |
6 |
7 |
8 |
9 |
41 |
72 |
73 |
74 |
42 |
71 |
43 |
48 |
49 |
50 |
51 | Tables needed for lollipop plot
52 |
53 |
70 |
50 |
51 | Tables needed for lollipop plot
52 |
77 |
180 |
181 |
191 |
78 |
93 |
103 |
104 |
168 |
179 |
79 |
81 |
82 |
83 | Tables needed for lollipop plot
80 |This function calculates two tables needed to generate lollipop plot, which visualise the model. 84 | The first table contains information about all nodes in the trees forming a model. 85 | It includes gain value, depth and ID of each nodes. 86 | The second table contains similarly information about roots in the trees.
87 | 88 | 89 |EIX_lollipop(xgb.model, data)90 | 91 |
Arguments
92 || xgb.model | 96 |a xgboost or lightgbm model. |
97 |
|---|---|
| data | 100 |a data table with data used to train the model. |
101 |
Value
105 | 106 |an object of the lollipop class
107 | 108 | 109 |Examples
110 |167 |#> Warning: pakiet 'Matrix' został zbudowany w wersji R 3.4.4#> Warning: pakiet 'xgboost' został zbudowany w wersji R 3.4.4param <- list(objective = "binary:logistic", max_depth = 2) 114 | xgb.model <- xgboost(sm, params = param, label = HR_data[, left] == 1, nrounds = 50)#> [1] train-error:0.150077 115 | #> [2] train-error:0.098007 116 | #> [3] train-error:0.098007 117 | #> [4] train-error:0.098007 118 | #> [5] train-error:0.098007 119 | #> [6] train-error:0.098007 120 | #> [7] train-error:0.098007 121 | #> [8] train-error:0.095873 122 | #> [9] train-error:0.095873 123 | #> [10] train-error:0.095606 124 | #> [11] train-error:0.095473 125 | #> [12] train-error:0.093406 126 | #> [13] train-error:0.061271 127 | #> [14] train-error:0.059404 128 | #> [15] train-error:0.055137 129 | #> [16] train-error:0.063271 130 | #> [17] train-error:0.043070 131 | #> [18] train-error:0.042670 132 | #> [19] train-error:0.039203 133 | #> [20] train-error:0.038536 134 | #> [21] train-error:0.037669 135 | #> [22] train-error:0.037869 136 | #> [23] train-error:0.036802 137 | #> [24] train-error:0.037336 138 | #> [25] train-error:0.036602 139 | #> [26] train-error:0.036402 140 | #> [27] train-error:0.036669 141 | #> [28] train-error:0.035802 142 | #> [29] train-error:0.035402 143 | #> [30] train-error:0.032202 144 | #> [31] train-error:0.031869 145 | #> [32] train-error:0.031469 146 | #> [33] train-error:0.030935 147 | #> [34] train-error:0.030602 148 | #> [35] train-error:0.030269 149 | #> [36] train-error:0.029402 150 | #> [37] train-error:0.029269 151 | #> [38] train-error:0.028802 152 | #> [39] train-error:0.028802 153 | #> [40] train-error:0.028535 154 | #> [41] train-error:0.028269 155 | #> [42] train-error:0.028202 156 | #> [43] train-error:0.027935 157 | #> [44] train-error:0.027669 158 | #> [45] train-error:0.027669 159 | #> [46] train-error:0.027402 160 | #> [47] train-error:0.028269 161 | #> [48] train-error:0.027268 162 | #> [49] train-error:0.026668 163 | #> [50] train-error:0.026335#> Warning: Transformation introduced infinite values in continuous x-axis#> Warning: Transformation introduced infinite values in continuous x-axis#> Warning: Transformation introduced infinite values in continuous x-axis#> Warning: Transformation introduced infinite values in continuous x-axis#> Warning: Transformation introduced infinite values in continuous x-axis#> Warning: Transformation introduced infinite values in continuous x-axis#> Warning: Removed 7 rows containing missing values (geom_text_repel).166 |
39 |
40 |
95 |
96 |
188 |
189 |
190 |
191 |
192 |
--------------------------------------------------------------------------------
/docs/reference/waterfall.html:
--------------------------------------------------------------------------------
1 |
2 |
3 |
4 |
5 |
6 |
7 |
8 |
9 |
41 |
92 |
93 |
94 |
42 |
91 |
43 |
48 |
49 |
50 |
51 | Visualiation of the model
52 |
53 |
90 |
50 |
51 | Visualiation of the model
52 |
97 |
176 |
177 |
187 |
98 |
112 |
138 |
139 |
164 |
175 |
99 |
101 |
102 |
103 | Visualiation of the model
100 |The lollipop plots the model with the most important interactions and variables in the roots.
104 | 105 | 106 |# S3 method for lollipop 107 | plot(x, ..., labels = "topAll", log_scale = TRUE, 108 | threshold = 0.1)109 | 110 |
Arguments
111 || x | 115 |a result from the |
116 |
|---|---|
| ... | 119 |other parameters. |
120 |
| labels | 123 |if "topAll" then labels for the most important interactions (vertical label) 124 | and variables in the roots (horizontal label) will be displayed, 125 | if "interactions" then labels for all interactions, 126 | if "roots" then labels for all variables in the root. |
127 |
| log_scale | 130 |TRUE/FALSE logarithmic scale on the plot. Default TRUE. |
131 |
| threshold | 134 |on the plot will occur only labels with Gain higher than `threshold` of the max Gain value in the model. 135 | The lower threshold, the more labels on the plot. Range from 0 to 1. Default 0.1. |
136 |
Value
140 | 141 |a ggplot object
142 | 143 | 144 |Examples
145 |163 |library("EIX") 146 | library("Matrix") 147 | sm <- sparse.model.matrix(left ~ . - 1, data = HR_data) 148 | 149 | library("xgboost") 150 | param <- list(objective = "binary:logistic", max_depth = 2) 151 | xgb_model <- xgboost(sm, params = param, label = HR_data[, left] == 1, nrounds = 25, verbose = 0) 152 | 153 | lolli <- lollipop(xgb_model, sm) 154 | plot(lolli, labels = "topAll", log_scale = TRUE)155 |library(lightgbm) 156 | train_data <- lgb.Dataset(sm, label = HR_data[, left] == 1) 157 | params <- list(objective = "binary", max_depth = 3) 158 | lgb_model <- lgb.train(params, train_data, 25) 159 | 160 | lolli <- lollipop(lgb_model, sm) 161 | plot(lolli, labels = "topAll", log_scale = TRUE)162 |
39 |
40 |
95 |
96 |
215 |
216 |
217 |
218 |
219 |
--------------------------------------------------------------------------------
41 |
92 |
93 |
94 |
42 |
91 |
43 |
48 |
49 |
50 |
51 | Explain prediction of a single observation
52 |
53 |
90 |
50 |
51 | Explain prediction of a single observation
52 |
97 |
203 |
204 |
214 |
98 |
112 |
144 |
145 |
189 |
202 |
99 |
101 |
102 |
103 | Explain prediction of a single observation
100 |This function calculates a table with influence of variables and interactions 104 | on the prediction of a given observation. It supports only xgboost models.
105 | 106 | 107 |waterfall(xgb_model, new_observation, data, type = "binary", 108 | option = "interactions", baseline = 0)109 | 110 |
Arguments
111 || xgb_model | 115 |a xgboost model. |
116 |
|---|---|
| new_observation | 119 |a new observation. |
120 |
| data | 123 |row from the original dataset with the new observation to explain (not one-hot-encoded).
124 | The param above has to be set to merge categorical features.
125 | If you dont wont to merge categorical features, set this parameter the same as |
126 |
| type | 129 |the learning task of the model. Available tasks: "binary" for binary classification or "regression" for linear regression. |
130 |
| option | 133 |if "variables", the plot includes only single variables, 134 | if "interactions", then only interactions. 135 | Default "interaction". |
136 |
| baseline | 139 |a number or a character "Intercept" (for model intercept). 140 | The baseline for the plot, where the rectangles should start. 141 | Default 0. |
142 |
Value
146 | 147 |an object of the broken class
148 | 149 |Details
150 | 151 |The function contains code or pieces of code
152 | from breakDown code created by Przemysław Biecek
153 | and xgboostExplainer code created by David Foster.
Examples
157 |188 |158 |library("EIX") 159 | library("Matrix") 160 | sm <- sparse.model.matrix(left ~ . - 1, data = HR_data) 161 | 162 | library("xgboost") 163 | param <- list(objective = "binary:logistic", max_depth = 2) 164 | xgb_model <- xgboost(sm, params = param, label = HR_data[, left] == 1, nrounds = 25, verbose=0) 165 | 166 | data <- HR_data[9,-7] 167 | new_observation <- sm[9,] 168 | 169 | wf <- waterfall(xgb_model, new_observation, data, option = "interactions") 170 | wf#> contribution 171 | #> xgboost: intercept -1.492 172 | #> xgboost: time_spend_company = 5 1.360 173 | #> xgboost: last_evaluation = 1 1.093 174 | #> xgboost: Work_accident = 0 -0.423 175 | #> xgboost: satisfaction_level = 0.89 -0.390 176 | #> xgboost: last_evaluation:time_spend_company = 1:5 0.297 177 | #> xgboost: last_evaluation:average_montly_hours = 1:224 0.227 178 | #> xgboost: satisfaction_level:time_spend_company = 0.89:5 0.223 179 | #> xgboost: number_project = 5 -0.211 180 | #> xgboost: average_montly_hours:last_evaluation = 224:1 -0.156 181 | #> xgboost: average_montly_hours = 224 -0.096 182 | #> xgboost: time_spend_company:last_evaluation = 5:1 0.095 183 | #> xgboost: salary = 2 0.074 184 | #> xgboost: satisfaction_level:number_project = 0.89:5 -0.003 185 | #> xgboost: prediction 0.597186 | plot(wf)187 |
