├── README.txt
└── deep_mf.py


/README.txt:
--------------------------------------------------------------------------------
1 | Deep-MF
2 | ==========
3 | 
4 | Deep-MF uses a simple 2-layer neural network to learn latent embeddings for users and items, using implicit feedback data. The network takes a user's full history of implicit signals as input and tries to predict their recent history (~2 weeks) as output. I'll add a fuller description and a comparison with other methods (IMF, Logistic-MF, word2vec) later when I have the time.  
5 | 


--------------------------------------------------------------------------------
/deep_mf.py:
--------------------------------------------------------------------------------
 1 | import matplotlib as mp
 2 | import numpy as np
 3 | import random
 4 | from scipy.spatial.distance import cosine
 5 | import tensorflow as tf
 6 | import time
 7 | 
 8 | 
 9 | def init_weights(shape):
10 |     return tf.Variable(tf.random_normal(shape, stddev=0.01))
11 | 
12 | def model(X, w_h1, w_h2, w_o, p_keep_input, p_keep_hidden):
13 |     X = tf.nn.dropout(X, p_keep_input)
14 | 
15 |     h1 = tf.nn.relu(tf.matmul(X, w_h1))
16 |     h1 = tf.nn.dropout(h1, p_keep_hidden)
17 | 
18 |     h2 = tf.nn.relu(tf.matmul(h1, w_h2))
19 |     h2 = tf.nn.dropout(h2, p_keep_hidden)
20 | 
21 |     return tf.matmul(h2, w_o)
22 | 
23 | def train_model(input_counts, output_counts, hidden_units=128, batch_size=100):
24 | 
25 |     num_users, num_items = input_counts.shape
26 |     # Balance pos and neg output weight in cross entropy
27 |     pos_weight = num_users * num_items / np.count_nonzero(input_counts)
28 | 
29 |     # Initialize input + output placeholders
30 |     X = tf.placeholder(tf.float32, [None, num_items])
31 |     Y = tf.placeholder(tf.float32, [None, num_items])
32 | 
33 |     # Initialize weights
34 |     w_h1 = init_weights([num_items, hidden_units])
35 |     w_h2 = init_weights([hidden_units, hidden_units])
36 |     # w_o == item vectors
37 |     w_o = init_weights([hidden_units, num_items])
38 | 
39 |     # Dropout probabilities
40 |     p_keep_input = tf.placeholder("float")
41 |     p_keep_hidden = tf.placeholder("float")
42 | 
43 |     py_x = model(X, w_h1, w_h2, w_o, p_keep_input, p_keep_hidden)
44 | 
45 |     # Define cost and training procedure
46 |     cost = tf.reduce_mean(tf.nn.weighted_cross_entropy_with_logits(py_x, Y, pos_weight))
47 |     train_op = tf.train.AdagradOptimizer(0.1).minimize(cost)
48 |     predict_op = tf.argmax(py_x, 1)
49 | 
50 |     with tf.Session() as sess:
51 |         # you need to initialize all variables
52 |         tf.initialize_all_variables().run()
53 | 
54 |         for i in range(1000):
55 |             avg_loss = 0
56 |             for start, end in zip(range(0, num_users, batch_size), range(batch_size - 1, num_users, batch_size)):
57 |                 _, loss = sess.run([train_op, cost], feed_dict={X: input_counts[start:end],
58 |                                                                 Y: output_counts[start:end],
59 |                                                                 p_keep_input: 0.8,
60 |                                                                 p_keep_hidden: 0.5})
61 |                 avg_loss += loss / (num_users / batch_size)
62 | 
63 |             if i % 10 == 0:
64 |                 print('%i iterations finished' % i)
65 |                 print('cross entropy: %f' % avg_loss)
66 |         vecs =  w_o.eval()
67 | 
68 |     return vecs.T
69 | 
70 | 


--------------------------------------------------------------------------------