DEASRec

Purify and Generalize: Efficient Dual-End Adapters for Sequential Recommendation (KDD 2026).

You can find our paper on [ACM Digital Library].

This repository contains an implementation of DEASRec built on the RecBole framework.

Requirements

• Python >= 3.7
• PyTorch >= 1.10.0
• See requirements.txt for complete dependencies

Quick Start

Dataset

This project evaluates the DEASRec model on the following benchmark datasets:

• Amazon_ratings:
  • Beauty
  • Sports & Outdoors
  • Video Games
  • Toys & Games

You can download the processed datasets from the RecBole Library "Google Drive" and place the files in ./dataset/.

Basic Usage

Then run:

python run_recbole.py --model=DEASRec --dataset=beauty

Optimal Hyperparameters (SASRec Backbone)

The following are the best hyperparameter settings of DEASRec across datasets:

Parameter	Sports	Video	Toys	Beauty
Number of Subspaces $G$	4	8	8	2
KL Regularization $\lambda_{KL}$	0.01	0.01	0.01	0.01
$\ell_2$ Regularization $\lambda_{\ell_2}$	5e-4	1e-4	1e-4	5e-5
Max Noise Scale $\sigma_{\max}$	1.0	1.0	1.0	1.5
Modulation Scale $\epsilon$	0.3	0.5	0.4	0.5

Run Commands with Best Hyperparameters

The following commands use run_recbole.py directly and append dataset-specific best hyperparameters:

# Sports
python run_recbole.py --model=DEASRec --dataset=sports --num_subspaces=4 --lambda_kl=0.01 --l2_reg_weight=5e-4 --max_noise_scale=1.0 --epsilon=0.3

# Video
python run_recbole.py --model=DEASRec --dataset=video --num_subspaces=8 --lambda_kl=0.01 --l2_reg_weight=1e-4 --max_noise_scale=1.0 --epsilon=0.5

# Toys
python run_recbole.py --model=DEASRec --dataset=toys --num_subspaces=8 --lambda_kl=0.01 --l2_reg_weight=1e-4 --max_noise_scale=1.0 --epsilon=0.4

# Beauty
python run_recbole.py --model=DEASRec --dataset=beauty --num_subspaces=2 --lambda_kl=0.01 --l2_reg_weight=5e-5 --max_noise_scale=1.5 --epsilon=0.5

Acknowledgments

This implementation is based on the RecBole recommendation library. We appreciate their outstanding work.

Citation

If you find this work useful, please cite it as:

@inproceedings{10.1145/3770855.3818208,
	author = {Hu, Juntao and Zhou, Wei and Shen, Huayi and Wen, Junhao and Zhang, Hongyu},
	title = {Purify and Generalize: Efficient Dual-End Adapters for Sequential Recommendation},
	year = {2026},
	booktitle = {Proceedings of the 32nd ACM SIGKDD Conference on Knowledge Discovery and Data Mining V.2},
	pages = {1758–1769}
}