Combining social relations and interaction data in Recommender System with Graph Convolution Collaborative Filtering

Tin T. Tran, Václav Snásel, Loc Tan NguyenIEEE Access 2023

TL;DR

This paper introduces a novel recommender system that combines user-item interaction data with social network information using Graph Convolutional Networks (GCNs). By effectively integrating these diverse data sources, the proposed model achieves improved recommendation accuracy and efficiency compared to existing methods.

What Problem Does This Paper Address?

Recommender systems are crucial for e-commerce and content platforms, helping users discover relevant items. Traditional collaborative filtering (CF) methods rely on user-item interaction data, but struggle with sparsity and cold-start problems. While incorporating social relationships can enhance recommendations, effectively combining user-item interactions with social network data, especially when dealing with noisy or indirect social connections, remains a challenge. This paper aims to bridge this gap by developing a GCN-based model that leverages both interaction patterns and social ties to provide more accurate and personalized recommendations.

Methods

The core of this work is a Graph Convolutional Network (GCN) model designed to integrate three key sources of information:

  • User-Item Interaction Matrix (A): This matrix captures how users interact with items (e.g., purchases, ratings).
  • User Correlation Matrix (C): This matrix quantifies the similarity between users based on their interaction patterns, calculated using Jaccard similarity to account for shared item interests.
  • Social Friendship Matrix (S): This matrix represents direct social connections between users.

The model preprocesses data by filtering out users and items with minimal interactions (a "10-core setting") to reduce noise and improve efficiency. It then uses GCN layers to propagate and aggregate information from these three matrices. Notably, the authors experiment with and advocate for eliminating feature transformation matrices and non-linear activation functions in the GCN layers, finding that this simplification leads to better performance and faster convergence, similar to the LightGCN approach. The final recommendation score is generated by a prediction layer, and the model is optimized using a Bayesian Personalized Ranking (BPR) loss function.

Datasets

The proposed model was evaluated on several publicly available datasets:

  • Gowalla: A location-based social networking dataset with user check-ins and friendship information.
  • Librarything: A book review website dataset.
  • Ciao: An online shopping site dataset with user ratings and social connections.
  • Epinions: A consumer review site dataset with user ratings and trust relationships.

Key Insights and Ideas

  • Social Data Matters, But How You Use It is Key: While social relationships generally improve recommendations, simply adding them isn't always effective. The paper highlights how models like SocialLGN can perform worse than their non-social counterparts if social signals aren't integrated harmoniously.
  • Simplicity Can Be Powerful: Eliminating feature transformation matrices and non-linear activation functions in GCNs, as seen in LightGCN and adopted here, leads to improved performance and faster convergence. This suggests that complex GCN architectures aren't always necessary for effective recommendation.
  • User Correlation is a Strong Signal: The proposed user correlation matrix (C), derived from Jaccard similarity, plays a crucial role in accelerating the model's learning process and improving its focus on important data.
  • Context is Crucial for Social Data Impact: The influence of social friendship data varies depending on the platform. For geographically-focused social networks like Gowalla, social data has a significant impact. For platforms where social features are secondary (like Librarything), interaction data remains more critical.

Key Results

  • The proposed model consistently outperforms baseline models, including GCN variants like NGCF, LightGCN, and SocialLGN, across multiple datasets in terms of precision, recall, and NDCG@20.
  • The model demonstrates significantly faster convergence speeds compared to other GCN-based models, attributed to the effective use of the user correlation matrix. For instance, on the Gowalla dataset, the proposed model converges in around 440 epochs, while SEPT and LightGCN take over 1000 epochs.
  • When comparing models with and without social data, the proposed model (Our model-all) generally achieves the best results, indicating the successful integration of both interaction and social signals.

Limitations and Future Work

  • Data Noise: The paper acknowledges that social relationship data can be noisy and may not always reflect true influence, especially when relationships change over time or are based on superficial connections (e.g., friends with similar hobbies might not give good advice on unrelated topics).
  • Group Recommendation: The authors suggest exploring group recommendation, which would provide recommendations for a collective of users rather than individuals, a valuable feature for scenarios like group outings or collaborative learning.
  • Item Grouping: The current model treats items individually. Future work could involve clustering or grouping items into catalogs to leverage relationships between items (e.g., items in the same genre or from the same brand) for more comprehensive recommendations.

Summary updated

BibTeX
@inproceedings{tran2023combining,
  author = {Tin T. Tran and Václav Snásel and Loc Tan Nguyen},
  title = {{Combining social relations and interaction data in Recommender System with Graph Convolution Collaborative Filtering}},
  booktitle = {IEEE Access},
  year = {2023},
}
Figure