For debugging the well-know issue of performance over-estimation addressed in https://www.kdd.org/kdd2020/accepted-papers/view/on-sampled-metrics-for-item-recommendation, SASRec&TiSASRec are still good with high speed and accurate prediction, we are just trying to make it better without negative sampling based evaluation.
This is our TensorFlow implementation for the paper:
Jiacheng Li, Yujie Wang, Julian McAuley (2020). Time Interval Aware Self-Attention for Sequential Recommendation. WSDM'20
We refer to the repo SASRec.
Please cite our paper if you use the code or datasets.
The code is tested under a Linux desktop (w/ GTX 1080 Ti GPU) with TensorFlow.
For Pytorch version of TiSASRec, please refer to repo.
This repo includes ml-1m dataset as an example.
For Amazon dataset, you could download Amazon review data from here..
To train our model on ml-1m
(with default hyper-parameters):
python main.py --dataset=ml-1m --train_dir=default
The implemention of self attention is modified based on this.
If you have any questions, please send me an email (j9li@eng.ucsd.edu).