MBID

This repository contains source codes and datasets for our CIKM'25 paper:

• Multi-Behavior Intent Disentanglement for Recommendation via Information Bottleneck Principle

Note: Extended version accepted by Neurocomputing.

Framework

The overall architecture of MBID from the extended journal version. i) Behavior-specific embedding propagation learns multi-behavior embeddings, and projection-based intent extraction method decomposes auxiliary behavior embeddings into both SCI and GCI embeddings. ii) The spurious correlation coefficient quantifies the spurious correlation intents between behaviors. iii) IB-based multi-intent learning personally disentangles spurious correlation intents and transfers genuine ones from auxiliary into the target behavior.

Usage

Train & Test

• Training MBID on IJCAI15:

python main.py --dataset=IJCAI_15

• Training MBID on Tmall:

python main.py --dataset=Tmall

• Training MBID on Retail:

python main.py --dataset=retailrocket

• Testing MBID using a saved model file:

ipython evaluation.ipynb

Citation

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

Conference Version (CIKM'25)

@inproceedings{MBID,
    title = {Multi-Behavior Intent Disentanglement for Recommendation via Information Bottleneck Principle},
    author = {Xu, Tongxin and Bin, Chenzhong and Xiao, Cihan and Li, Yunhui and Gu, Tianlong},
    booktitle = {Proceedings of the 34th ACM International Conference on Information and Knowledge Management},
    year = {2025}
}

Extended Journal Version (Neurocomputing)

@article{MBID_Ext,
    title = {User Intent Disentanglement for Multi-Behavior Recommendation via Information Bottleneck Principle},
    author = {Bin, Chenzhong and Xu, Tongxin and Zhang, Feng},
    journal = {Neurocomputing},
    volume = {656},
    pages = {131454},
    year = {2025},
}