This is the repository of the KDD 2023 paper "Adaptive Disentangled Trasnformer for Sequential Recommendation"
This repository reimplements the backbone model from SASRec、Bert4Rec and STOSA.
The following libraries and versions are used in the experiments, these requirements are not mandatory and are only for reference.
numpy = 1.22.4
Python = 3.8.10
Pytorch = 1.11.0
scipy = 1.9.3
tqdm = 4.61.2
For each backbone, you can use this command to search the best lambdas for training the model. You can go the source file for more details.
python evolution.py --dataset xxx
After searching process, you will get the best candidate in /res and you can copy the candidate vector and your search space to candidates_to_lambdas.py and run it to get the best lambdas, or copy the value of key rec and ind to get the best lambdas. And then copy the lambdas to function get_lambdas in utils.py and retrain the model to get the final score.
For convenience, we have recorded part of the lambdas used in our experiments in utils.py and hyperparameters in /templates as an example.
python main.py --dataset xxx
Part of the datasets used for SASRec backbone and Bert4Rec backbone are uploaded to the repository, and ml-20m can download from this link. For STOSA backbone, you can download the 5-core dataset from Amazon Dataset or go to the original repository for the details.