# Graph Convolution Machine for Context-aware Recommender System

**Authors:** Jiancan Wu, Xiangnan He, Xiang Wang, Qifan Wang, Weijian Chen, Jianxun Lian, Xing Xie, Yongdong Zhang  
**arXiv:** [2001.11402](https://arxiv.org/abs/2001.11402)  
**Date:** 2020-01-30  
**Summary page:** https://www.summarizepaper.com/en/paper/2001.11402v1/

## Key points

- Importance of leveraging multiple interactions in Context-Aware Recommender Systems (CARS)
- Introduction of Graph Convolution Machine (GCM) to capture interactions among user behaviors using graph neural networks
- Effectiveness of GCM demonstrated through experiments on public datasets, showing improvements over existing Collaborative Filtering (CF) and CARS baselines
- Emphasis on organizing user behaviors with contextual information in graphs to build strong representations for users and items
- Acknowledgment of limitations in GCM's approach, such as neglecting dynamic characteristics like time and failing to capture dynamic user preferences effectively
- Future research directions include building dynamic graphs based on contextual information and developing a dynamic graph neural network
- Proposal to incorporate rich side information to explain diverse intents behind user behaviors by modeling user-item relationships at a granular level of user intents

## Summary

The study "Graph Convolution Machine for Context-Aware Recommender System" highlights the importance of leveraging multiple interactions in Context-Aware Recommender Systems (CARS). The research focuses on converting user, item, and contextual features into an attributed graph that incorporates contexts as edges between user and item nodes. A novel model called Graph Convolution Machine (GCM) is introduced to capture interactions among various user behaviors using graph neural networks and model individual behavior features through factorization machines. The effectiveness of GCM is demonstrated through experiments conducted on three public datasets, showcasing significant improvements over existing Collaborative Filtering (CF) and CARS baselines. <br>
<br>
The study emphasizes the importance of organizing user behaviors with contextual information in graphs to build strong representations for users and items. However, it acknowledges limitations in GCM's approach of unifying all context features as an edge, neglecting dynamic characteristics like time and failing to capture dynamic user preferences effectively. To address these limitations, future research directions include building dynamic graphs based on contextual information instead of static ones and developing a dynamic graph neural network.<br>
<br>
Additionally, incorporating rich side information to explain diverse intents behind user behaviors is proposed by modeling user-item relationships at a granular level of user intents to generate disentangled representations. Overall, the study underscores the potential of utilizing graph convolutions for context-aware recommendation systems and highlights avenues for further research to enhance the representation learning process and improve recommendation accuracy.