Interest-aware Message-Passing GCN for Recommendation

April 6, 2026 · View on GitHub

IMP-GCN improves GCN-based recommendation by restricting high-order message passing to interest-consistent subgraphs, effectively alleviating over-smoothing and noisy propagation.

Authors

Fan Liu1, Zhiyong Cheng2*, Lei Zhu3, Zan Gao2, Liqiang Nie1*

1 Shandong University, China
2 Shandong Artificial Intelligence Institute, China
3 Shandong Normal University, China
* Corresponding author

Updates

  • [02/2021] Initial release
  • [02/2021] Release paper / arXiv version
  • [02/2021] Release code

Introduction

本项目是论文 Interest-aware Message-Passing GCN for Recommendation 的官方实现。

请在这里简要说明：

  • 论文要解决的问题

    • 现有基于 GCN 的推荐方法（如 NGCF、LightGCN）虽然能够利用高阶邻居信息提升表示能力，但存在过平滑（over-smoothing）问题：随着层数加深，不同用户的表示逐渐趋同，导致推荐性能下降。

  • 方法的核心思想是什么：

    • 只利用“兴趣一致”的高阶邻居信息进行表示学习

  • 与现有方法相比有什么特点：

    • 减少噪声传播
    • 缓解过平滑问题
    • 支持更深层模型

  • 本仓库提供了：

    • 训练代码
    • 训练集
    • 测试集

Description

We present IMP-GCN, a framework for collaborative filtering recommendation.
Our method addresses the over-smoothing and noisy message propagation problem by introducing an interest-aware subgraph-based message passing mechanism, restricting high-order propagation within groups of users sharing similar interests.

Highlights

  • 支持 <CF-based recommendation Task>
  • 提供 <training / inference / evaluation> 脚本
  • 提供 <dataset>
  • 适合用于 <论文复现 / 后续研究>

Method / Framework

Framework Figure

Figure 1. Overall framework of IMP-GCN.

Usage

Environment Settings

  • Tensorflow-gpu version: 1.3.0

Example to run the codes.

Training

gowalla

Run IMP_GCN.py

python IMP_GCN.py --dataset gowalla  --regs [1e-4] --embed_size 64 --layer_size [64,64,64,64,64,64] --lr 0.001 --batch_size 2048 --epoch 2000 --groups 3 --Ks [20,10] --gpu_id 0

Example Results

Citation

@inproceedings{10.1145/3442381.3449986,
author = {Liu, Fan and Cheng, Zhiyong and Zhu, Lei and Gao, Zan and Nie, Liqiang},
title = {Interest-aware Message-Passing GCN for Recommendation},
year = {2021},
publisher = {ACM},
booktitle = {Proceedings of the Web Conference 2021},
pages = {1296–1305}

Acknowledgement

  • Thanks to our supervisor and collaborators for valuable support.
  • Thanks to the open-source community for providing useful baselines and tools.

License

This project is released under the Apache License 2.0.