update on 05/23/2025: thx to Wentworth1028 and Tiny-Snow, we have LayerNorm update, for higher NDCG&HR, and here's the doc👍.
modified based on paper author's tensorflow implementation, switching to PyTorch(v1.6) for simplicity, fixed issues like positional embedding usage etc. (making it harder to overfit, except for that, in recsys, personalization=overfitting sometimes)
code in python folder.
to train:
python main.py --dataset=ml-1m --train_dir=default --maxlen=200 --dropout_rate=0.2 --device=cuda
just inference:
python main.py --device=cuda --dataset=ml-1m --train_dir=default --state_dict_path=[YOUR_CKPT_PATH] --inference_only=true --maxlen=200
output for each run would be slightly random, as negative samples are randomly sampled, here's my output for two consecutive runs:
1st run - test (NDCG@10: 0.5897, HR@10: 0.8190)
2nd run - test (NDCG@10: 0.5918, HR@10: 0.8225)
pls check paper author's repo for detailed intro and more complete README, and here's the paper bib FYI :)
@inproceedings{kang2018self,
title={Self-attentive sequential recommendation},
author={Kang, Wang-Cheng and McAuley, Julian},
booktitle={2018 IEEE International Conference on Data Mining (ICDM)},
pages={197--206},
year={2018},
organization={IEEE}
}
I see a dozen of citations of the repo🫰, pls use the example bib as below if needed.
@online{huang2020sasrec_pytorch,
author = {Zan Huang},
title = {SASRec.pytorch},
year = {2020},
url = {https://github.com/pmixer/SASRec.pytorch}
}
