-
Notifications
You must be signed in to change notification settings - Fork 0
/
booktest.m
279 lines (198 loc) · 8.15 KB
/
booktest.m
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
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
274
275
276
277
%% pick dataset
% int0: books
data_raw = int0;
max_value = 10;
%% use movielens dataset
data_raw = movielens(:,[1 2 3])';
max_value = 5;
%% my course_review data
data_raw = reviews(:,[1 2 4])';
max_value = 5;
%% normalize my course_review data
data_raw = normalize_reviews(reviews');
data_raw = data_raw([1 2 4], :);
max_value = 5;
%% make a super dense dataset!
limit = 4000;
user_limit = limit;
item_limit = limit / 2;
R = dense_R(data_raw, item_limit, user_limit, 100, 5);
display(sprintf('%d items, %d users. avg %f reviews per item, %f per user', size(R,1), size(R,2), mean(sum(R > 0,2)), mean(sum(R > 0,1))));
%% INIT DATA
% generate sparse matrix from input data
% just pick the first N points
% NOTE: convert all to double so sparse matrix doesn't complain
% data = double(int0(:,1:20000));
limit = 4000;
user_limit = limit;
item_limit = limit / 2;
under_limit = (data_raw(2,:) < item_limit) & (data_raw(1,:) < user_limit);
data = data_raw(:,under_limit);
num_results = size(data,2);
collision_ratio = num_results / limit;
display(sprintf('found %d results w/ limit %d. collision ratio: %f', num_results, limit, collision_ratio));
% num_nonzeros = sum(data(3,:) > 0);
nonzero_rows = data(3,:) ~= 0;
length(find(nonzero_rows));
data = data(:,nonzero_rows);
% data format: userid, bookid, rating
% want: Rs = IxU matrix
% num_users = length(unique(data(1,:)));
% num_books = length(unique(data(2,:)));
num_users = max(data(1,:));
num_books = max(data(2,:));
% note: max(user) and max(books) are both smaller than 10k
collapse_factor = 1;
I = ceil(num_books / collapse_factor + 1);
U = num_users + 1;
R = zeros(I, U);
for j = 1:size(data,2)
col = data(:,j);
row_i = floor(col(2) / collapse_factor + 1); % item
col_i = col(1) + 1; % user
% note: matlab doesn't like indexing from 0
R(row_i, col_i) = col(3);
end
display(sprintf('avg reviews per item: %f, avg reviews per user: %f', mean(sum(R > 0,2)), mean(sum(R > 0,1))));
%%
%
% P = [1 1 2 2 3 3 4 4; 1 2 1 2 1 2 1 2];
% Q = [1 1 1 1 1 1 1 1 1; 0 1 0 1 0 1 0 1 0];
%
% R = P' * Q;
% max_value = 6;
% f = 2;
%
% indices = 1:numel(R);
%% PREP THE LOOP
k = 10; % for k-fold cross-validation
svd_err = zeros(k,2); % row 1: rms, row 2: mae
base_err = zeros(k,2);
% permute the indices before partitioning for cross-validation
nonzero_indices = find(R(:));
i_count = length(nonzero_indices);
indices = nonzero_indices(randperm(i_count));
%% PREP DATA
gamma = 0.007;
lambda = 0.1;
max_iter = 10;
f = 20; % number of latent dimensions
i = 1;
%% RUN THE LOOP
display(sprintf('starting %d-fold cross-validation.',k));
% for gamma = [0.001, 0.005, 0.007, 0.01, 0.02, 0.05];
for lambda = [0.0005, 0.001, 0.003, 0.009, 0.02, 0.06, 0.1, 0.2, 0.3, 0.5, 1, 2];
% loop for k-fold cross-validation
% for i = 1:k
%% prep the loop
train = R;
low = 1 + floor((i_count * (i-1))/k);
hi = 1 + floor((i_count * i)/k);
test_indices = indices(low:hi);
train(test_indices) = 0; % zero the test indices
%% run & evaluate
R_svd = collab_svd(f, lambda, gamma, max_iter, max_value, train, 0.25, 3);
R_base = baseline(train, 3);
svd_err(i,1) = rms_error(R_svd, R, test_indices);
svd_err(i,2) = mae_error(R_svd, R, test_indices);
base_err(i,1) = rms_error(R_base, R, test_indices);
base_err(i,2) = mae_error(R_base, R, test_indices);
display(sprintf('lambda %d. SVD {rms: %f, mae: %f}, BASE {rms: %f, mae: %f}',lambda,svd_err(i,1),svd_err(i,2),base_err(i,1),base_err(i,2)));
end
%%
i = 5;
mu = mean(train(train ~= 0));
smu = sqrt(mu);
[I,U] = size(train);
noise = 0.5;
Q = rand(f,I) .* noise - noise/2;
P = rand(f,U) .* noise - noise/2;
% for lambda = [0.001, 0.003, 0.009, 0.02, 0.06, 0.1, 0.2, 0.3, 0.5, 1, 2, 4, 8];
lambda = 0.02;
gamma = 0.001;
max_iter = 30;
momentum = 0.4; % adding a momentum allows you to decrease the learning rate (gamma)
%%
display(sprintf('starting loop. max_iter = %d. momentum = %d', max_iter, momentum));
% for max_iter = [2, 5, 10, 20, 40, 100]
%
fake_test_is = find(train > 0);
fake_test_is(randperm(length(fake_test_is))) = fake_test_is;
fake_test_is = fake_test_is(1:100);
% for gamma = [0.002, 0.004, 0.01] % large gamma => divergence (papers agree)
% for lambda = [0.0005, 0.001, 0.01, 0.1]
%%
for gamma = [0.001, 0.005, 0.01] % large gamma => divergence (papers agree)
% for gamma = [0.005, 0.01, 0.02] % large gamma => divergence (papers agree)
for lambda = [0.0001, 0.001, 0.01, 0.02]
%%
base = baseline(train, 2);
% prepped_train = zeros(size(train,1),size(train,2));
% prepped_train(train ~= 0) = train(train ~= 0) - base(train ~= 0);
% R_svd = rsvd(f, lambda, gamma, max_iter, base, train, Q, P, max_value);
% R_svd = rsvd_absolute(f, lambda, gamma, max_iter, train, 1, 10, bound);
% R_svd = rsvd_momentum(f, lambda, gamma, max_iter, train, 1, 10, bound, momentum, 0.00005);
% R_svd = rsvd_relative(f, lambda, gamma, max_iter, train, 1, 10, noise, momentum, -1);
R_svd = rsvd_standard(f, lambda, lambda/2, gamma, max_iter, train, 1, max_value, noise, momentum, -1);
% R_svd = rsvd_incremental(f, lambda, lambda*2, gamma, max_iter, train, 1, max_value, noise);
% R_svd2 = real(R_svd) ;
svd_err(i,1) = rms_error(R_svd, R, test_indices);
svd_err(i,2) = mae_error(R_svd, R, test_indices);
base_err(i,1) = rms_error(base, R, test_indices);
base_err(i,2) = mae_error(base, R, test_indices);
R_err = rms_error(R_svd, train, fake_test_is);
% svd_2_err = rms_error(R_svd2, R, test_indices);
% svd_2_mae_err = mae_error(R_svd2, R, test_indices);
% display(sprintf('loop %d. lambda %f. gamma %f. SVD {rms: %f, mae: %f}, SVD2 {rms: %f, mae: %f}, BASE {rms: %f, mae: %f}',i,lambda,gamma,svd_err(i,1),svd_err(i,2),svd_2_err, svd_2_mae_err,base_err(i,1),base_err(i,2)));
display(sprintf('loop %d. lambda %f. gamma %f. SVD {rms: %f, mae: %f}, BASE {rms: %f, mae: %f} (overfitting? %d)',i,lambda,gamma,svd_err(i,1),svd_err(i,2),base_err(i,1),base_err(i,2), R_err));
end
end
%% incremental SVD
display(sprintf('starting loop. max_iter = %d', max_iter));
for gamma = [0.005, 0.01, 0.03] % large gamma => divergence (papers agree)
for lambda = [0.001, 0.01, 0.02, 0.04, 0.1]
%%
base = baseline(train, 2);
% R_svd = rsvd_standard(f, lambda, lambda/2, gamma, max_iter, train, 1, max_value, noise, momentum, -1);
R_svd = rsvd_incremental(f, lambda, lambda/2, gamma, max_iter, train, 1, max_value, noise);
svd_err(i,1) = rms_error(R_svd, R, test_indices);
svd_err(i,2) = mae_error(R_svd, R, test_indices);
base_err(i,1) = rms_error(base, R, test_indices);
base_err(i,2) = mae_error(base, R, test_indices);
R_err = rms_error(R_svd, train, fake_test_is);
display(sprintf('loop %d. lambda %f. gamma %f. SVD {rms: %f, mae: %f}, BASE {rms: %f, mae: %f} (overfitting? %d)',i,lambda,gamma,svd_err(i,1),svd_err(i,2),base_err(i,1),base_err(i,2), R_err));
end
end
%% run svd++
train = R;
low = 1 + floor((i_count * (i-1))/k);
hi = 1 + floor((i_count * i)/k);
test_indices = indices(low:hi);
train(test_indices) = 0; % zero the test indices
%%
R_svd_plus = collab_svd_plus(f, 0.005, 0.015, 0.007, max_iter, max_value, train);
R_svd = collab_svd(f, lambda, gamma, max_iter, max_value, train);
R_base = baseline(train, 3);
svd_err(i,1) = rms_error(R_svd, R, test_indices);
svd_err(i,2) = mae_error(R_svd, R, test_indices);
base_err(i,1) = rms_error(R_base, R, test_indices);
base_err(i,2) = mae_error(R_base, R, test_indices);
%% RUN IT
R_svd = collab_svd(f, lambda, gamma, max_iter, max_value, train, mu);
%%
R_base = baseline(train, 3);
'R_svd error:'
rms_error(R_svd,R,test_indices)
mae_error(R_svd,R,test_indices)
'R_base error:'
rms_error(R_base,R,test_indices)
mae_error(R_base,R,test_indices)
% check if it moves in the right/wrong direction
both_up = (R_base > R) .* (R_base >= R_svd);
both_down = (R_base < R) .* (R_base <= R_svd);
right_side = sum(both_up(R > 0)) + sum(both_down(R > 0));
nonzero = R > 0;
total = sum(nonzero(:));
percent_right_side = right_side / total
% NOTEWORTHY: svd does way better than neighbor-based approach from
% published paper here. NOTE: condensed the dataset though, so it's easier