Interpretable Recommendations and Parameter-Grounded LLM Explanations with Multigraph Attention

Many online platforms now use complex recommender systems to decide which products, restaurants, or services people see. These systems can improve matching, but their recommendations are often hard for users and managers to understand. This article introduces MG-GAT, a recommender system framework that uses multiple networks and attribute data while keeping track of the neighbors and features that influence each recommendation. The same evidence is then used to generate explanations, so the reason shown to a user is tied to the model’s internal decision process rather than added afterward. Across Yelp data from Ontario and Pennsylvania, the method performs competitively with strong deep-learning baselines. In a randomized experiment, explanations based on MG-GAT increased users’ trust, persuasiveness, satisfaction, and engagement relative to similarity-based, social, and SHAP-based explanations. For practice, the results show that platforms do not have to choose between predictive performance and accountable explanations. Recommendation teams can design systems that expose the signals behind predictions, audit generated explanations for unsupported claims, and give users clearer reasons for accepting or questioning automated recommendations.

• 저널Information Systems Research
• 토픽Explainable Artificial Intelligence (XAI) · Artificial Intelligence
• DOI10.1287/isre.2023.0549
• 저자Yan Leng, Xiao Liu, Rodrigo Ruiz