Hi all, I want to ask about the experience to pick up a suitable loss function for unsupervised problems. Here is a description of my current problem.
Assume I divide the samples into positive and negative, then get two groups of scores when passing them into my model, the dummy codes like:
pos_scores = model(pos_samples) # (sample_num, 1) neg_socres = model(neg_samples) # (sample_num, 1) pos_x = pos_scores.mean() neg_x = neg_scores.mean() loss = LossFunction(pos_x, neg_x)
Where I want to maximize the pos_scores while minimizing the neg_scores at the same time, there is no limitation of their values. Note that it is an unsupervised problem, thus no ground-truth label as a reference, only a calculated score for each sample.
One loss function I have tried is:
loss = neg_x - pos_x
by minimizing this loss function, the code tends to minimize neg_x while maximizing pos_x. But the resulted scores are not very brilliant. Thus, are there any other useful functions that could help to handle with this problem? Thanks all!