hidasib
commented
7 years ago You are right in the sense, that the biases are not updated during training. But this is not the reason why BPR-MF results are so bad. The reason for the bad performance is that BPR-MF is not suitable for the session-based task in which each test session is basically a user cold-start scenario. It is easy to see that this is the case, as BPR-MF, like any other matrix factorization, works best if the user/item set is fairly static thus it can use the representations it learned during training. It can't directly handle new items or users. In session-based recommendations every new session is basically a new "user" for which BPR doesn't have a model, thus can't recommend anything. Therefore we used only the item feature vectors from BPR to do the recommendations: we represent the session as the average of the feature vectors of the items in the session so far and we recommend the item(s) with the most similar feature vectors to this representation. In this setting, using biases doesn't really make sense. They can't be used during similarity computations. One could argue that they capture some aspects of the data allowing the item feature vectors to be better representations. We tried this, but the results were worse, so we removed this from the final version. Unfortunately it seems that some of the earlier version made it into the public repo. This is the reason while you see the initialization of the bias terms, but not their update. I'll completely remove them eventually with the next bigger update.
Side note: earlier experiments with BPR-MF on standard personalized recommendations (on implicit data) showed that using the bias term is not always beneficial even during the classic recommendation scenario. On 2 dataset of 5 we found that disabling the bias resulted in better accuracy. Since the examination of BPR-MF was not the focus of that project we didn't investigate this further, but my assumption is that sometimes item biases can learn the popularity bias of the training data too well, which is not beneficial if the trendiness of the items cahnge more rapidly. User biases don't mater during ranking, so their role is only to help learning better representations, which generally have lesser effect that the item biases.
voladorlu
In the update function of BPR baseline, you forget to update the bias parameters. That's very important since the defined relevance function includes this term. Please double check your code. Maybe it's the reason why the performance of BPR shown in your paper is so bad.