کنفرانس مهندسی برق ایران

صفحه اصلی / سی و چهارمین کنفرانس بین المللی مهندسی برق

A Hybrid Explainability Framework for Graph-Based Movie Recommendation Using PinSAGE

نویسندگان :

Elahe Sadat Abdolkarimi¹ Zahra Bayuni²

1- دانشگاه تفرش 2- موسسه آموزش عالی شهاب دانش قم

کلمات کلیدی :

Explainable AI،Pins AGE،Graph Neural Networks،Recommender Systems،Fidelity Analysis

چکیده :

Large-scale movie recommendation systems increasingly use graph models based on graph neural networks; however, the opacity of these models reduces their reliability and limits the ability to explain decisions. In this paper, we present a hybrid explainability framework for movie recommenders that extracts the importance of each movie by combining two complementary sources: model-based importance based on Graph Neural Network (GNN( Explainer and structure-based importance based on graph centrality measures. This framework allows for the generation of explanations that are both consistent with the actual behavior of the model and reflect the structure of user-movie interactions. To assess the quality of explanations, a fidelity measurement protocol is designed that examines the impact of removing or retaining important elements on the model output. Experimental results on the MovieLens-1M dataset show that removing the identified components leads to a 22% reduction in the prediction score, indicating a good fidelity of the explanations. Furthermore, integrating the proposed explainability mechanism with the Pinterest Scalable Graph Embedding (PinSAGE) model improves the recommendation performance (Hit@10: 0.698→0.706, NDCG@10: 0.1349→0.1383). Finally, this study presents an end-to-end explainable version of Pins AGE that combines structural explainability, fidelity measurement, and recommendation quality improvement into an integrated framework.