public boolean update() {
zeroFill(im_grad);
zeroFill(um_grad);
int cnt = 0;
iter.reset();
while (iter.hasNext()) {
cnt++;
Rating r = iter.next();
double rdiff = r.rate - product(userMatrix[r.user], itemMatrix[r.movie]);
for (int f = 0; f < factorNum; f++) {
um_grad[r.user][f] += -rdiff * itemMatrix[r.movie][f] + userReg * userMatrix[r.user][f];
im_grad[r.movie][f] += -rdiff * userMatrix[r.user][f] + itemReg * itemMatrix[r.movie][f];
}
}
while (this.learningRate > 1e-10) {
test_update(um_grad, im_grad);
double next_like = likelihood(um_next, im_next);
System.out.println(cur_like + " " + next_like + " " + learningRate);
if (next_like > cur_like) {
next();
this.learningRate *= 1.25;
this.cur_like = next_like;
return true;
} else {
this.learningRate *= 0.5;
}
}
return false;
}
public double likelihood(double users[][], double items[][]) {
double ret = 0.0;
iter.reset();
while (iter.hasNext()) {
Rating r = iter.next();
double rdiff = r.rate - product(users[r.user], items[r.movie]);
ret += rdiff * rdiff;
}
double reg = userReg * norm(users) + itemReg * norm(items);
return -(ret + reg);
}
