A Lightweight HITS-Assisted Matrix Factorization Framework for Sparse Recommendation Systems

Abstract

Recommendation systems have been assisting a variety of businesses using data analysis. Information from users is a strong basis to develop such systems, but obtaining sufficient responses is often challenging. Matrix factorization (MF) helps recommendation systems deal with this issue. However, sparse user–item interaction settings degrade the performance. The Hyperlink-Induced Topic Search (HITS) algorithm captures the structural importance within users and items. This model cannot investigate deeper preferences. This paper aims to provide a lightweight and interpretable framework that incorporates the merits of HITS as a preprocessing step to latent factor learning supported by MF. The MovieLens dataset was used to evaluate the proposed work. A three-stage evaluation protocol was employed, with a standard MF as baseline, the HITS model for recommendation, and then the sequential HITS followed by the MF framework. The proposed HITS→MF framework reduces RMSE by approximately 11.3% and MAE by 12.6% compared to baseline MF. Consistent gains in ranking are observed through Precision@10 and NDCG@10. The results demonstrate that HITS alone improves top-N ranking performance, and HITS followed by MF provides a better balance with respect to the evaluation metrics. The proposed framework improves robustness without introducing additional model complexity or auxiliary data, making it suitable for practical recommendation environments. The collaboration between structural graph information and latent factor learning can reduce sparsity without loss of performance. This collaborative approach is an efficient alternative to complex hybrid models, making it a good choice for practical recommendation system applications.