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
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
|
/*
Copyright (c) by respective owners including Yahoo!, Microsoft, and
individual contributors. All rights reserved. Released under a BSD (revised)
license as described in the file LICENSE.
*/
#ifdef _WIN32
#include <winsock2.h>
#else
#include <netdb.h>
#endif
#include "reductions.h"
#include "simple_label.h"
#include "gd.h"
#include "rand48.h"
using namespace std;
using namespace LEARNER;
namespace MF {
struct mf {
vector<string> pairs;
size_t rank;
uint32_t increment;
// array to cache w*x, (l^k * x_l) and (r^k * x_r)
// [ w*(1,x_l,x_r) , l^1*x_l, r^1*x_r, l^2*x_l, r^2*x_2, ... ]
v_array<float> sub_predictions;
// array for temp storage of indices during prediction
v_array<unsigned char> predict_indices;
// array for temp storage of indices
v_array<unsigned char> indices;
// array for temp storage of features
v_array<feature> temp_features;
vw* all;
};
template <bool cache_sub_predictions>
void predict(mf& data, base_learner& base, example& ec) {
float prediction = 0;
if (cache_sub_predictions)
data.sub_predictions.resize(2*data.rank+1, true);
// predict from linear terms
base.predict(ec);
// store linear prediction
if (cache_sub_predictions)
data.sub_predictions[0] = ec.partial_prediction;
prediction += ec.partial_prediction;
// store namespace indices
copy_array(data.predict_indices, ec.indices);
// erase indices
ec.indices.erase();
ec.indices.push_back(0);
// add interaction terms to prediction
for (vector<string>::iterator i = data.pairs.begin(); i != data.pairs.end(); i++) {
int left_ns = (int) (*i)[0];
int right_ns = (int) (*i)[1];
if (ec.atomics[left_ns].size() > 0 && ec.atomics[right_ns].size() > 0) {
for (size_t k = 1; k <= data.rank; k++) {
ec.indices[0] = left_ns;
// compute l^k * x_l using base learner
base.predict(ec, k);
float x_dot_l = ec.partial_prediction;
if (cache_sub_predictions)
data.sub_predictions[2*k-1] = x_dot_l;
// set example to right namespace only
ec.indices[0] = right_ns;
// compute r^k * x_r using base learner
base.predict(ec, k + data.rank);
float x_dot_r = ec.partial_prediction;
if (cache_sub_predictions)
data.sub_predictions[2*k] = x_dot_r;
// accumulate prediction
prediction += (x_dot_l * x_dot_r);
}
}
}
// restore namespace indices and label
copy_array(ec.indices, data.predict_indices);
// finalize prediction
ec.partial_prediction = prediction;
ec.pred.scalar = GD::finalize_prediction(data.all->sd, ec.partial_prediction);
}
void learn(mf& data, base_learner& base, example& ec) {
// predict with current weights
predict<true>(data, base, ec);
float predicted = ec.pred.scalar;
// update linear weights
base.update(ec);
ec.pred.scalar = ec.updated_prediction;
// store namespace indices
copy_array(data.indices, ec.indices);
// erase indices
ec.indices.erase();
ec.indices.push_back(0);
// update interaction terms
// looping over all pairs of non-empty namespaces
for (vector<string>::iterator i = data.pairs.begin(); i != data.pairs.end(); i++) {
int left_ns = (int) (*i)[0];
int right_ns = (int) (*i)[1];
if (ec.atomics[left_ns].size() > 0 && ec.atomics[right_ns].size() > 0) {
// set example to left namespace only
ec.indices[0] = left_ns;
// store feature values in left namespace
copy_array(data.temp_features, ec.atomics[left_ns]);
for (size_t k = 1; k <= data.rank; k++) {
// multiply features in left namespace by r^k * x_r
for (feature* f = ec.atomics[left_ns].begin; f != ec.atomics[left_ns].end; f++)
f->x *= data.sub_predictions[2*k];
// update l^k using base learner
base.update(ec, k);
// restore left namespace features (undoing multiply)
copy_array(ec.atomics[left_ns], data.temp_features);
// compute new l_k * x_l scaling factors
// base.predict(ec, k);
// data.sub_predictions[2*k-1] = ec.partial_prediction;
// ec.pred.scalar = ec.updated_prediction;
}
// set example to right namespace only
ec.indices[0] = right_ns;
// store feature values for right namespace
copy_array(data.temp_features, ec.atomics[right_ns]);
for (size_t k = 1; k <= data.rank; k++) {
// multiply features in right namespace by l^k * x_l
for (feature* f = ec.atomics[right_ns].begin; f != ec.atomics[right_ns].end; f++)
f->x *= data.sub_predictions[2*k-1];
// update r^k using base learner
base.update(ec, k + data.rank);
ec.pred.scalar = ec.updated_prediction;
// restore right namespace features
copy_array(ec.atomics[right_ns], data.temp_features);
}
}
}
// restore namespace indices
copy_array(ec.indices, data.indices);
// restore original prediction
ec.pred.scalar = predicted;
}
void finish(mf& o) {
// restore global pairs
o.all->pairs = o.pairs;
// clean up local v_arrays
o.indices.delete_v();
o.sub_predictions.delete_v();
}
base_learner* setup(vw& all) {
new_options(all, "MF options")
("new_mf", po::value<size_t>(), "rank for reduction-based matrix factorization");
if(missing_required(all)) return NULL;
mf& data = calloc_or_die<mf>();
data.all = &all;
data.rank = (uint32_t)all.vm["new_mf"].as<size_t>();
// store global pairs in local data structure and clear global pairs
// for eventual calls to base learner
data.pairs = all.pairs;
all.pairs.clear();
all.random_positive_weights = true;
learner<mf>& l = init_learner(&data, setup_base(all), learn, predict<false>, 2*data.rank+1);
l.set_finish(finish);
return make_base(l);
}
}
|
