A Comparative Study of Hybrid Recommender Systems: Integrating Collaborative Filtering and Transformer-Based Models for Cold-Start and Popularity Bias Mitigation
Esin Seçil Yılmaz, İsmail Duru · 2025
Recommender systems play a crucial role in delivering personalized content; however, they face challenges such as cold-start problems (CSP), data sparsity, and popularity bias (PB). This study shows a mixed method that combines collaborative filtering (CF) methods such as singular value decomposition++ (SVD++) and autoencoder-based model (AEM) with transformer-based models (TBM) such as E5-Large to improve recommendations. The hybrid model performs better on Amazon Gift Card Dataset (AGCD) than the standalone CF and Content-Based Filtering (CBF) methods in terms of precision (0.500) and recall (0.556). Although standalone SVD++ achieves the lowest RMSE (0.3645) and MAE (0.1468) due to its focus on interaction patterns, the hybrid model balances these metrics (RMSE=0.4258, MAE=0.1585) with its ability to mitigate CSP and PB through semantic embeddings and fusion of latent features. The results show that the hybrid framework is strong enough to deal with sparsity and bias while still being very accurate. This shows how important it is to combine deep learning with traditional recommendation.