Towards Prediction Error Reduction in Matrix Factorization Recommenders Using Semantic Similarity Metrics

Recommender systems leverage historical data to provide personalized suggestions, often using Matrix Factorization (MF) techniques like SVD. However, data sparsity remains a key challenge to high precision. Accurate recommendation predictions are crucial for platform success, as significant errors can severely impact user experience. When recommendations consistently miss the mark, users lose trust in the system, leading to decreased engagement and eventual platform abandonment. Our research addresses this critical challenge by developing a novel hybrid recommendation approach that reduces prediction error. By combining matrix factorization with DBpedia-derived semantic similarities, we achieve more precise personalization that better aligns with user preferences. Our approach minimizes sparsity by inferring missing ratings for semantically similar items. We therefore propose a hybrid framework combining similarity-based imputation with MF, merging content-based and collaborative filtering advantages, targeting improving prediction rates through strategic rating pre-filling.