Abstract
Content-Based Recommender Systems (CBRS) are widely used owing to their independence from extensive interaction histories. However, they often suffer from over-specialization, producing overly homogeneous recommendations that hinder user discovery. Existing enhancements, including TF-IDF modeling, ONMTF-based similarity extraction, and evolutionary techniques such as genetic algorithms, have partially improved novelty or diversity but have not jointly balanced relevance, novelty, diversity, and unexpectedness to induce serendipity. To address this gap, this study proposes ImprovedCBRS-PSO, a Particle Swarm Optimization (PSO) enhanced re-ranking framework designed to inject serendipity while preserving its relevance. The model integrates a discreet PSO formulation, roulette-wheel-based exploratory search, and a multi-criteria fitness function that combines precision, relevance, novelty, diversity, and serendipity. Experiments on MovieLens-1M, MovieLens-10M, and AmazonBooks demonstrated that the proposed method consistently outperformed CBRS+RRSGA (Revolutionary Recommender System using a Genetic Algorithm) and CBRS+EEDPSO (Elite-Evolutionary Discrete Particle Swarm Optimization), achieving up to 92% gains in Serendipity while maintaining a strong precision. Superior performance on the sparse AmazonBooks dataset further validates the effectiveness of PSO in navigating long-tail environments. Moreover, stable results across K = 1–10 indicating a robust balance between exploration and relevance in the results. Overall, the findings confirm PSO as an effective exploratory mechanism for mitigating over-specialization in the CBRS.

This publication has 20 references indexed in Scilit: