It is equivalent to exposing the label by consistently output higher score on the right part(positive samples) than the left part(negative samples). Experiments fixed the BN bugs result in lower auc: model_dice_origin(0.8922) and model_dice_fixed(0.8706), which is a lot lower than the result in the paper(0.8871) and minor improvements over PNN(0.8679) and DeepFM(0.8683).
liuhu-bigeye
commented
4 years ago
You didn't share BN parameters in the siamese network: https://github.com/zhougr1993/DeepInterestNetwork/blob/master/din/model_dice.py#L58 .
It is equivalent to exposing the label by consistently output higher score on the right part(positive samples) than the left part(negative samples). Experiments fixed the BN bugs result in lower auc: model_dice_origin(0.8922) and model_dice_fixed(0.8706), which is a lot lower than the result in the paper(0.8871) and minor improvements over PNN(0.8679) and DeepFM(0.8683).