Implementation based on pytorch for DIN recommendation algorithm
- For convenience, referring to authors tensorflow implementation, feature-embedding dimension is identical.
- Without any L1/L2 normalization or dropout strategy, it's supposed to choose suitable model according to the evaluation stage manually.
file name | description |
---|---|
main.ipynb | Session for training and evaluation |
model.py | Defination of target models |
DataLoader.py | Self-defined data loader |
environment.yml | Conda envrionment yaml |
Deep Interest Network for Click-Through Rate Prediction
Deep Interest Evolution Network for Click-Through Rate Prediction
Preprocessed data wrapped within data.tar.gz
came from mouna99/dien
- DIN
- AUGRU
- DICE activation layer
- Auxialary loss with neg_sample