Causality-aware Social Recommender System with Network Homophily Informed Multi-treatment Confounders

Area of Science:

• Computer Science
• Machine Learning
• Causal Inference

Background:

• Recommender systems typically predict user preferences from observed ratings, but ignore causality where item exposure influences ratings.
• Existing causal methods often overlook the complexity of multiple items and simultaneous inference in real-world scenarios.
• Social network information, while valuable, confounds user preferences and complicates deconfounding in social recommender systems.

Purpose of the Study:

• To frame recommendation as a multiple causal inference problem for improved accuracy.
• To develop a causality-aware social recommender system that integrates social network structures.
• To mitigate confounding bias in networked observational data for enhanced social recommendations.

Main Methods:

• Framing recommendation as a multiple causal inference problem.
• Incorporating social network structures with matrix factorization for deconfounding.
• Utilizing network homophily within matrix factorization models via regularization.
• Employing a proximal gradient-based optimization framework for efficient model estimation.

Main Results:

• The proposed method effectively mitigates confounding bias by learning network-informed multi-treatment confounders.
• Latent variables capture network structure, leading to improved rating prediction accuracy.
• The proximal gradient optimization framework enhances computational efficiency and incorporates network constraints.

Conclusions:

• Treating recommendation as a multiple causal inference problem is crucial for accurate predictions.
• Integrating social network homophily into matrix factorization improves deconfounding and recommendation quality.
• The developed causality-aware social recommender offers a computationally efficient and effective approach for networked data.