- Add more models of year 2021.
- Add support for saving pb file, exporting embeddings, exporting feature vocab
- Add DESTINE/MaskNet models
- Refactor the code of layers.EmbeddingLayer
- Add support for loading blocks of h5 data
- Add tests for DIN, FmFM
- Refine the DIN model to support feature concatenation
- Add tutorials on how to use sequence features and pretrained embeddings
- Fix the defect in using padding_idx (no impact on Criteo/Avazu results)
- Fix the defect in loading pretrain embeddings (no impact on Criteo/Avazu results)
Refactor the code and documentation to support reproducing steps on the BARS benchmark.
This is the first release of FuxiCTR, including 28 models: LR, FM, CCPM, FFM, DNN, Wide&Deep, FNN, IPNN, DeepCross, HOFM, DeepFM, NFM, AFM, DCN, FwFM, xDeepFM, DIN, FiGNN, AutoInt+, FiBiNET, FGCNN, HFM+, ONN, AFN+, LorentzFM, InterHAt, FLEN, FmFM. Especially, this version corresponds to the original implementations used for reproducing the experiments in the following paper: Jieming Zhu, Jinyang Liu, Shuai Yang, Qi Zhang, Xiuqiang He, Open Benchmarking for CTR Prediction, CIKM 2021.