CLEAR-Rec: A Contrastive Long-Term Memory-Enhanced, Explainable, and Adaptive Recommendation Framework

Abstract

Recommender systems frequently face difficulties such as data sparsity, limited personalization, and challenges in interpretability. We introduce CLEAR-Rec, a comprehensive framework for Contrastive Long-Term Memory-Enhanced, Explainable, and Adap- tive Recommendation that addresses these challenges via four essential components. The Episodic Memory Module (EMM) focuses on long-term user-item interactions to improve personalization. Contrastive Semantic Alignment (CSA) aligns collaborative and content embeddings through contrastive learning. Explainability through Memory Trace (EMT) produces natural-language justi- fications for recommendations by leveraging memory cues. Finally, the Reinforced Feedback Loop (RFL) integrates real-time user feedback to dynamically adjust recommendations using reinforcement learning. Experiments on multiple real-world datasets show that CLEAR-Rec consistently achieves superior performance compared to competitive baselines on standard ranking metrics like HR@10 and NDCG@10. Moreover, it provides explanations that are more accurate and aligned with user preferences than those of existing explainable recommenders. Our framework balances accuracy, adaptability, and interpretability, offering a scalable solution for a variety of recommender scenarios.