SCKGD

This repository contains source codes and datasets for our DASFAA'26 paper:

• Spurious Correlation Knowledge Graph Disentanglement for Multi-Behavior Recommendation

Framework

The overall architecture of SCKGD. i) Spurious correlation knowledge graph and time-projected knowledge encoder capture the spurious correlations between items and learn SCKG embeddings, where $f_{TP}(\cdot)$ denotes TP-TransR or TP-TATEC. ii) Behavior-specific embedding propagation learns multi-behavior embeddings. iii) Knowledge-empowered multi-behavior learning disentangles spurious correlations and transfers genuine correlated semantics from auxiliary into the target behavior.

Usage

Train & Test

First, run the kg_generate.py file in data/xxx to construct the spurious correlation knowledge graph.

• Training SCKGD on IJCAI15:

python main.py --dataset=IJCAI_15

• Training SCKGD on Tmall:

python main.py --dataset=Tmall

• Training SCKGD on Retail:

python main.py --dataset=retailrocket

• Testing SCKGD using a saved model file:

ipython evaluation.ipynb

Citation

If you find this work useful, please cite our paper:

@inproceedings{SCKGD,
    title={Spurious Correlation Knowledge Graph Disentanglement for Multi-behavior Recommendation},
    author={Xu, Tongxin and Bin, Chenzhong and Xiao, Cihan and Zeng, Zhixin and Li, Yunhui},
    booktitle={International Conference on Database Systems for Advanced Applications},
    pages={356--371},
    year={2026},
}