Hyper-Heuristic Recommender System to Balance Accuracy, Diversity, and Fairness

Abstract Personalized recommendation systems have gained prominence and are traditionally evaluated based on accuracy. However, a growing interest is in incorporating popularity bias into a comprehensive evaluation. Achieving a balance among accuracy, diversity, and ranking metrics remains challenging. We propose a Hyper-heuristic recommendation framework that integrates five models: the Neural Collaborative Filtering (NCF) model, Singular Value Decomposition (SVD) model, K-Nearest Neighbors (KNN) model, Light Factorization Machines model (Light-FM), and Non-negative Matrix Factorization (NMF) model, enhanced with genetic algorithms. By exploring various configurations and models and leverag-ing genetic algorithms, we aim to achieve an optimal balance across met-rics in the final recommendation list. Our approach indicates balanced evaluation results, including popularity bias, diversity, hit rate, precision, recall, and NDCG. It enhances recommendation diversity and balances popularity bias and ranking accuracy, ultimately improving the user experience. The results offer users more satisfying recommendations by providing less biased and more diverse content while maintaining relevant ranking quality.