-
Notifications
You must be signed in to change notification settings - Fork 8
/
sqlrank12.jl
341 lines (318 loc) · 8.84 KB
/
sqlrank12.jl
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
278
279
280
281
282
283
284
285
286
287
288
289
290
291
292
293
294
295
296
297
298
299
300
301
302
303
304
305
306
307
308
309
310
311
312
313
314
315
316
317
318
319
320
321
322
323
324
325
326
327
328
329
330
331
332
333
334
335
336
337
338
339
340
341
# version 12.0:
# 1. Rewrite gradient calculation to reduce time complexity to O(d2_bar)
# 2. Remove line search and calculate average of gradients
# main("ml1m_oc_50_train_ratings.csv", "ml1m_oc_50_test_ratings.csv", 0.2, 0.9, 1, 4, 100, 3)
function main(train, test, learning_rate, decay_rate, T, lambda, r, ratio)
#train = "ml1m_oc_50_train_ratings.csv"
#test = "ml1m_oc_50_test_ratings.csv"
# requires ratio to be integer, usually 3 works best
X = readdlm(train, ',' , Int64);
x = vec(X[:,1]);
y = vec(X[:,2]);
v = vec(X[:,3]);
Y = readdlm(test, ',' , Int64);
xx = vec(Y[:,1]);
yy = vec(Y[:,2]);
vv = vec(Y[:,3]);
n = max(maximum(x), maximum(xx));
msize = max(maximum(y), maximum(yy));
X = sparse(x, y, v, n, msize); # userid by movieid
Y = sparse(xx, yy, vv, n, msize);
# julia column major
# now moveid by userid
X = X';
Y = Y';
rows = rowvals(X);
vals = nonzeros(X);
cols = zeros(Int, size(vals)[1]);
index = zeros(Int, n + 1);
d2, d1 = size(X);
cc = 0;
# need to record new_index based on original index
# such that later, no need to shift each iteration, only need to swap the zero part
new_len = 0;
new_index = zeros(Int, d1 + 1);
new_index[1] = 1;
for i = 1:d1
index[i] = cc + 1;
tmp = nzrange(X, i);
nowlen = size(tmp)[1];
newlen = nowlen * (1 + ratio);
new_len += newlen;
new_index[i + 1] = new_index[i] + newlen;
for j = 1:nowlen
cc += 1;
cols[cc] = i;
end
end
index[d1 + 1] = cc + 1;
# no need to sort for 0/1 data, so we don't need sort_input function in sqlrank10.jl
# ASSUMPTION: new_vals containing all 0's and 1's
# we also need a function to shuffle all 1's and swapping 0's
new_rows = zeros(Int, new_len);
new_cols = zeros(Int, new_len);
new_vals = zeros(Int, new_len);
for i = 1:d1
rows_set = Set{Int}();
for j = index[i]:(index[i + 1] - 1)
push!(rows_set, rows[j]);
end
nowlen = new_index[i + 1] - new_index[i];
nowOnes = div(nowlen, 1 + ratio);
for j = 1:nowOnes
new_rows[new_index[i] + j - 1] = rows[index[i] + j - 1];
new_cols[new_index[i] + j - 1] = i;
new_vals[new_index[i] + j - 1] = vals[index[i] + j - 1];
end
nowStart = new_index[i] + nowOnes;
nowEnd = new_index[i + 1] - 1;
for j = nowStart:nowEnd
while true
row_idx = rand(1:d2);
if !(row_idx in rows_set)
new_rows[j] = row_idx;
new_cols[j] = i;
new_vals[j] = 0.0;
push!(rows_set, row_idx);
break;
end
end
end
end
rows_t = rowvals(Y);
vals_t = nonzeros(Y);
cols_t = zeros(Int, size(vals_t)[1]);
index_t = zeros(Int, n + 1)
cc = 0;
for i = 1:d1
index_t[i] = cc + 1;
tmp = nzrange(Y, i);
nowlen = size(tmp)[1];
for j = 1:nowlen
cc += 1
cols_t[cc] = i
end
end
index_t[d1 + 1] = cc + 1;
# again, no need to sort rows_t, vals_t, cols_t under the ASSUMPTION
# we also don't need levels
srand(123456789);
U = 0.1*randn(r, d1);
V = 0.1*randn(r, d2);
#U = rand(r,d1)*(0.1 - (-0.1)) + (-0.1)
#V = rand(r,d2)*(0.1 - (-0.1)) + (-0.1)
m = comp_m(new_rows, new_cols, U, V);
# no need to calculate max_d_bar, since we are using all 1's and 0's appended of ratio 1:ratio
println("rank: ", r, ", ratio of 0 vs 1: ", ratio, ", lambda:", lambda, ", learning_rate: ", learning_rate);
# no need for obtain_R method, but we do need a shuffling method I call it stochasticQueuing
# (that is why I name the method as sqlrank: stochastic queuing listwise ranking algorithm)
println("iter time objective_function precision@K = 1, 5, 10");
obj = objective(new_index, m, new_rows, d1, lambda, U, V);
p1,p2,p3=compute_precision(U, V, X, Y, d1, d2, rows, vals, rows_t, vals_t);
println("[", 0, ",", obj, ", ", p1," ",p2," ",p3, "],");
#println("[", 0, ",", obj, "],");
totaltime = 0.00000;
num_epoch = 121;
num_iterations_per_epoch = 1;
nowobj = obj;
for epoch = 1:num_epoch
tic();
for iter = 1:num_iterations_per_epoch
U, m = obtain_U(new_rows, new_cols, new_index, U, V, learning_rate, d1, r, lambda);
V = obtain_V(new_rows, new_cols, new_index, m, U, V, learning_rate, d1, r, lambda);
end
new_rows = stochasticQueuing(new_rows, new_index, d1, d2, ratio);
totaltime += toq();
#if (epoch - 1) % 3 == 0
# learning_rate = learning_rate * 0.3
#end
#learning_rate = learning_rate * 0.95
if (epoch - 1) % T == 0
learning_rate = learning_rate * decay_rate
p1,p2,p3=compute_precision(U, V, X, Y, d1, d2, rows, vals, rows_t, vals_t);
m = comp_m(new_rows, new_cols, U, V);
nowobj = objective(new_index, m, new_rows, d1, lambda, U, V);
println("[", epoch, ", ", totaltime, ", ", nowobj, ", ", p1,", ",p2,", ",p3, "],");
else
m = comp_m(new_rows, new_cols, U, V);
nowobj = objective(new_index, m, new_rows, d1, lambda, U, V);
println("[", epoch, ", ", totaltime, ", ", nowobj);
end
end
end
function stochasticQueuing(rows, index, d1, d2, ratio)
new_rows = zeros(Int, size(rows)[1]);
for i = 1:d1
nowlen = index[i + 1] - index[i];
nowOnes = div(nowlen, 1 + ratio);
newOrder = shuffle(1:nowOnes);
rows_set = Set{Int}();
for j = 1:nowOnes
oldIdx = index[i] + j - 1;
row_j = rows[oldIdx];
push!(rows_set, row_j);
newIdx = index[i] + newOrder[j] - 1;
new_rows[newIdx] = row_j;
end
nowStart = index[i] + nowOnes;
nowEnd = index[i + 1] - 1;
for j = nowStart:nowEnd
while true
row_idx = rand(1:d2);
if !(row_idx in rows_set)
new_rows[j] = row_idx;
push!(rows_set, row_idx);
break;
end
end
end
end
return new_rows
end
function obtain_U(rows, cols, index, U, V, s, d1, r, lambda)
m = comp_m(rows, cols, U, V);
grad_U = comp_gradient_U(rows, cols, index, m, U, V, s, d1, r, lambda);
U = U - s * grad_U;
m = comp_m(rows, cols, U, V);
return U, m
end
function comp_gradient_U(rows, cols, index, m, U, V, s, d1, r, lambda)
grad_U = zeros(size(U));
for i = 1:d1
d_bar = index[i+1] - index[i];
grad_U[:,i] = comp_gradient_ui(rows, cols, index, d_bar, m, i, V, r);
end
grad_U += lambda * U;
return grad_U
end
function comp_gradient_ui(rows, cols, index, d_bar, m, i, V, r)
cc = zeros(d_bar);
tt = 0.0;
total = 0.0;
for t = d_bar:-1:1
tmp = m[index[i] - 1 + t];
total += exp(tmp);
tt += 1 / total;
end
total = 0.0;
for t = d_bar:-1:1
ttt = m[index[i] - 1 + t];
cc[t] -= ttt * (1 - ttt);
cc[t] += exp(ttt) * ttt * (1 - ttt) * tt;
total += exp(ttt);
tt -= 1 / total;
end
res = zeros(r);
for t = 1:d_bar
res += cc[t] * V[:,rows[index[i] - 1 + t]];
end
return res
end
function obtain_V(rows, cols, index, m, U, V, s, d1, r, lambda)
grad_V = comp_gradient_V(rows, cols, index, m, U, V, s, d1, r, lambda);
V = V - s * grad_V;
return V
end
function comp_gradient_V(rows, cols, index, m, U, V, s, d1, r, lambda)
grad_V = zeros(size(V));
for i = 1:d1
d_bar = index[i+1] - index[i];
cc = zeros(d_bar);
tt = 0.0;
total = 0.0;
for t = d_bar:-1:1
tmp = m[index[i] - 1 + t];
total += exp(tmp);
tt += 1 / total;
end
total = 0.0;
for t = d_bar:-1:1
ttt = m[index[i] - 1 + t];
cc[t] -= ttt * (1 - ttt);
cc[t] += exp(ttt) * ttt * (1 - ttt) * tt;
total += exp(ttt);
tt -= 1 / total;
end
for t = 1:d_bar
j = rows[index[i] - 1 + t]
grad_V[:,j] += cc[t] * U[:,i]
end
end
grad_V += lambda * V;
return grad_V
end
function logit(x)
return 1.0/(1+exp(-x))
end
function comp_m(rows, cols, U, V)
m = zeros(length(rows));
for i = 1:length(rows)
m[i] = logit(dot(U[:,cols[i]], V[:,rows[i]]));
end
return m
end
function objective(index, m, rows, d1, lambda, U, V)
res = 0.0;
for i = 1:d1
tt = 0.0;
d_bar = index[i+1] - index[i];
for t = d_bar:-1:1
# since we will shuffle new_rows, new_cols, and obtain new m
# we don't need to shuffle again for m (ASSUMPTION: we only have 1's and 0's)
tmp = m[index[i] - 1 + t];
tt += exp(m[index[i] - 1 + t]);
res -= tmp;
res += log(tt);
end
end
res += lambda / 2 * (vecnorm(U) ^ 2 +vecnorm(V) ^ 2);
return res
end
function compute_precision(U, V, X, Y, d1, d2, rows, vals, rows_t, vals_t)
K = [1, 5, 10]; # K has to be increasing order
precision = [0, 0, 0];
for i = shuffle(1:d1)[1:1000]
#for i = 1:d1
tmp = nzrange(Y, i);
test = Set{Int64}();
for j in tmp
push!(test, rows_t[j]);
end
#test = Set(rows_t[tmp])
if isempty(test)
continue
end
tmp = nzrange(X, i);
vals_d2_bar = vals[tmp];
train = Set(rows[tmp]);
score = zeros(d2);
ui = U[:, i];
for j = 1:d2
if j in train
score[j] = -10e10;
continue;
end
vj = V[:, j];
score[j] = dot(ui,vj);
end
p = sortperm(score, rev = true);
for c = 1: K[length(K)]
j = p[c];
if score[j] == -10e10
break;
end
if j in test
for k in length(K):-1:1
if c <= K[k]
precision[k] += 1;
else
break;
end
end
end
end
end
#precision = precision./K/d1;
precision = precision./K/1000;
return precision[1], precision[2], precision[3]
end