/
regsvd_sgd.py
150 lines (99 loc) · 4.23 KB
/
regsvd_sgd.py
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
import numpy as np
import pandas as pd
import datetime as dt
import math
import evaluation as eva
from multiprocessing import Pool
reload(eva)
def assign_id(data):
items = pd.unique(data['itemid'])
vks = pd.unique(data['vk'])
itemid = {items[i]:i for i in range(items.shape[0])}
vkid = {vks[i]:i for i in range(vks.shape[0])}
return itemid, vkid
def split_data(data, split_time):
training_data = data.loc[data['real_time'] <= split_time]
test_data = data.loc[data['real_time'] > split_time]
training_data.index = range(training_data.shape[0])
test_data.index = range(test_data.shape[0])
return training_data, test_data
def user_bias(data, itemid, vkid):
avg = np.mean(data['action'])
vk_avg = np.zeros(len(vkid))
vk_count = np.zeros(len(vkid))
item_avg = np.zeros(len(itemid))
item_count = np.zeros(len(itemid))
#update the bias for user and item
for idx, value in data[['vk','itemid','action']].iterrows():
vk_avg[vkid[value['vk']]] = vk_avg[vkid[value['vk']]] + value['action']
vk_count[vkid[value['vk']]] = vk_count[vkid[value['vk']]] + 1
item_avg[itemid[value['itemid']]] = item_avg[itemid[value['itemid']]] + value['action']
item_count[itemid[value['itemid']]] = item_count[itemid[value['itemid']]] + 1
item_avg = item_avg * 1.0 / item_count - 0.5 * avg
vk_avg = vk_avg * 1.0 / vk_count - 0.5 * avg
return avg, item_avg, vk_avg
def sgd(data, avg, vk_avg, item_avg, u, v, itemid, vkid, gamma, lda):
"""The stochastic gradient descent for Matrix factorization"""
sum_err = 0
for idx, value in data[['vk','itemid','action']].iterrows():
vk_num = vkid[value['vk']]
item_num = itemid[value['itemid']]
err = value['action'] - np.dot(u[vk_num, :], v[item_num, :])
vt = v[item_num, :]
ut = u[vk_num, :]
u[vk_num, :] = ut + gamma * (err * vt - lda * ut)
v[item_num, :] = vt + gamma * (err * ut - lda * vt)
sum_err += (err ** 2)
sum_err = math.sqrt(sum_err / data.shape[0])
return u, v, sum_err
def predict(test_data, avg, vk_avg, item_avg, u, v, itemid, vkid):
prediction = np.zeros(test_data.shape[0])
for idx, value in test_data[['vk', 'itemid', 'action']].iterrows():
if value['vk'] not in vkid.keys():
continue
if value['itemid'] not in itemid.keys():
continue
prediction[idx] == (np.dot(u[vkid[value['vk']]], v[itemid[value['itemid']]]))
return prediction
def predict_all(test_data, u, v, itemid, vkid):
prediction_all = np.dot(u, v.T)
return prediction_all
def avg_prediction(test_data, avg, vk_avg, item_avg, itemid, vkid):
prediction = np.zeros((len(vkid), len(itemid)))
for iid in itemid.keys():
for vid in vkid.keys():
itemnum = itemid[iid]
vknum = vkid[vid]
prediction[vknum, itemnum] = avg + vk_avg[vknum] + item_avg[itemnum]
return prediction
def mf(data, split_time, k):
training_data, test_data = split_data(data, split_time)
tr_itemid, tr_vkid = assign_id(training_data)
tst_itemid, tst_vkid = assign_id(test_data)
#initialize the bias and offect
avg, item_avg, vk_avg = user_bias(training_data, tr_itemid, tr_vkid)
u = np.random.rand(len(tr_vkid), k)
v = np.random.rand(len(tr_itemid), k)
#step size and reguarlization
gamma = 0.05
lda = 0.1
beta = 0.99
old_err = 1e10
err = 1e5
while(old_err - err > 0.001):
old_err = err
u, v, err = sgd(training_data, avg, vk_avg, item_avg, u, v, tr_itemid, tr_vkid, gamma, lda)
gamma = gamma * beta
print '----------Error------------'
print err
avg_pd = avg_prediction(test_data, avg, vk_avg, item_avg, tr_itemid, tr_vkid)
prediction_all = predict_all(test_data, u, v, tr_itemid, tr_vkid)
# rmse, avg_rmse, new_vk, new_item = eva.rmse_test(test_data, u, v, tr_itemid, tr_vkid, avg, item_avg, vk_avg)
# print 'MF RMSE:', rmse
# print 'AVG RMSE:', avg_rmse
# print 'Number of New user:%d, Num of total user:%d' % (new_vk, len(tst_vkid))
# print 'Number of New Items:%d, Num of total item:%d' % (new_item, len(tst_itemid))
avg_ndcg = eva.ndcg_test(test_data, avg_pd, tr_itemid, tr_vkid, 25)
ndcg = eva.ndcg_test(test_data, prediction_all, tr_itemid, tr_vkid, 25)
print 'MF NDCG:', ndcg
print 'AVG NDCG:', avg_ndcg