search

rixwew / pytorch-fm

Factorization Machine models in PyTorch
MIT License
1.04k stars 225 forks source link

Add the implementation of AFN and Add residual part in Autoint #13

Closed yfreedomliTHU closed 4 years ago

yfreedomliTHU commented 4 years ago

The pytorch-fm is quite awesome, and it help me a lot in my research.To make it better, I did a little update as follows: 1.I did a little change to improve the performance of Autoint. According to the paper of Autoint and it's source code, the model should contain the residual part in multihead_attention, which can improve the performance of AutoInt. Also,there maybe a little mistake in the implementation of AutoInt, according to the Author's source code of tensorflow, the embedding size is 16 but the embedding size of MultiheadAttention(the tf source code called block) is 64, they are different actually. refer to: https://github.com/DeepGraphLearning/RecommenderSystems/blob/master/featureRec/autoint/model.py#L48

2.I Add the implementation of AFN, which is the AAAI'20 paper about Deep CTR. Cheng W, et al. Adaptive Factorization Network: Learning Adaptive-Order Feature Interactions.AAAI'20

Thanks!

rixwew commented 4 years ago

Thank you, I'll check it. @yfreedomliTHU

rixwew commented 4 years ago

Your PR looks great to me. I merged and released your PR at pypi. (https://pypi.org/project/torchfm/0.7.0) Thank you very much @yfreedomliTHU