Dynamic Hybrid Model for Sequential and Explainable Recommendation Using GCN, Memory, and Attention

Abstract

Recommendation systems have become essential components of modern digital platforms, helping users discover relevant items within increasingly large collections of content. Despite significant progress in recommendation research, traditional approaches still face several challenges, including data sparsity, cold-start issues, limited modeling of longterm user preferences, and insuffcient explainability of generated recommendations. This dissertation proposes DynamicHybridRec, a dynamic hybrid recommendation architecture that combines graph-based collaborative learning, sequential behavior modeling, external memory mechanisms, and explainable artificial intelligence techniques within a unified framework. The proposed model integrates LightGCN to capture collaborative user-item relationships, a Transformer-based sequential encoder to model short-term behavioral dynamics, and an External Memory Module to preserve long-term preference patterns that extend beyond the fixed sequence window. These heterogeneous representations are fused through a Cross-Attention Fusion layer to produce a comprehensive user representation for recommendation generation. To improve transparency and user trust, a hybrid explainability framework is incorporated by combining attention weight analysis, GradientSHAP attribution, and component contribution estimation. We implemented and tested the proposed architecture on rank standard metrics (HR@10, NDCG@10, MRR@10) in the MovieLens 1M Dataset. Experiments show that the hybrid model obtains reliably good recommendation performance as well as interpretability while successfully discovering proximate user preferences and long term user preferences