Graph-Enhanced Deep Autoencoder with SHAP-Based Interpretability for Multi-Criteria Recommender Systems

Abstract

The problem of delivering customized recommendations when users face too much information has led scientists to develop multi-criteria and context-aware recommender systems. The Deep Autoencoder-Based Multi-Criteria Recommender System (DAE-MCRS) serves as the foundation for this research which develops a new hybrid system that unites deep autoencoders with Graph Neural Networks (GNNs) and Reinforcement Learning (RL). The method solves previous model restrictions because it handles complex user-item-context interactions while generating flexible sequential recommendations which update based on changing user preferences. The proposed methodology combines GNN-based representations with the multi-criteria autoencoder pipeline through RL optimization which uses user feedback to learn recommendations over time. The research used Yahoo! and other public multi-aspect datasets to perform extensive experiments. The extended model achieves superior results than baseline DAE-MCRS and state-of-the-art methods in both Movie and TripAdvisor datasets through its improved RMSE and MAE and NDCG performance. The system achieves better explainability because SHAP analysis reveals how both criteria and context factors affect the system which leads users to trust the system more deeply. The research creates a base for intelligent recommender systems which deliver precise results through explainable outputs while handling increasing interaction data types at large scale. This research is presented as a primary method for creating future personalized information systems. These systems are designed to provide precise results by using outputs that can be easily understood.