loss formulations

[lihuiliullh]

May I know which loss formulations in the paper these two images correspond to?

[yuh-yang]

Hi, @lihuiliullh !

reg_loss computes the L2 regularization for batch user and item embeddings.

loss corresponds to the BPR loss for recommendation task.

[lihuiliullh]

@yuh-yang May I know why you use L2 regularization for batch user and item embeddings? Does this trick can improve performance?

[lihuiliullh]

I also notice that for BPR loss, the formulation in the paper is

But in your code, it is softplus(-(pos_scores - neg_scores)).

Are these two the same?

[yuh-yang]

1. Generally L2 regularization is effective against overfitting. This batch-wise usage follows NGCF and LightGCN.

2. Using softplus instead of logsigmoid in BPR loss is a common practice to avoid NaN loss when, on some cases during training, the model fails to perform well by scoring negative samples very high.

Referring to this: https://github.com/xiangwang1223/neural_graph_collaborative_filtering/issues/17

[yuh-yang]

Closed for being inactive.