Interpretable Recommendations and Parameter-Grounded LLM Explanations with Multigraph Attention

Recommender systems combine ratings, attributes, and network signals to improve accuracy, but the resulting models are increasingly opaque, undermining the trust users need to act on their recommendations. We develop a multigraph attention network (MG-GAT), which instantiates a single design principle: identify the evidence the model uses at prediction time and reuse that same evidence for user-facing explanations rather than approximating it after the fact. MG-GAT exposes two prediction-time signals: a neighbor importance graph identifying which neighboring users or businesses are most influential for a recommendation and feature relevance scores tying that influence to observable features. These signals serve as the inputs to a constrained language-generation procedure that produces parameter-grounded natural-language explanations. We evaluate MG-GAT on Yelp data from two regions against 14 baselines spanning attention-based, heterogeneous network–based, contrastive learning, and knowledge graph–based recommenders. MG-GAT achieves the lowest root mean squared error (RMSE) in both regions, remains competitive on ranking metrics, and substantially reduces contextual misattribution in language model explanations. In a randomized experiment with 1,363 participants, MG-GAT–based explanations significantly increase perceived relevance and future interest relative to all comparison conditions. Among explanation treatments, they also significantly improve user-rated trust, persuasiveness, and satisfaction relative to the strongest explanation baseline. The results suggest a path for explainable recommender design: preserve the evidence used at prediction time and carry it through to the explanation layer, so that natural-language rationales remain tied to the model rather than to post hoc approximations.

History: Ahmed Abbasi, Senior Editor; Jingjing Zhang, Associate Editor.

Funding: Y. Leng is supported by the National Science Foundation [Grant IIS-2153468]. This research was supported in part by the Marketing Science Institute.

Supplemental Material: The online appendices are available at https://doi.org/10.1287/isre.2023.0549.