Enhancing Cold-Start Recommendations with Content-Based Profiles and Latent Factor Models

Abstract

Recommendation systems have become an important tool for enhancing personalized recommendations across various domains. However, these systems face challenges, including the cold start problem, data sparsity, etc. In this paper, we present a novel recommendation model that integrates content-based and collaborative approaches to overcome these challenges. The proposed model uses TF-IDF vectorization over multiple item attributes to compute content similarity scores, and the SVD collaborative model captures latent user-item interactions. To further strengthen user preferences, a time-aware exponential decay function is used to acquire the most recent user preferences during the construction of user profiles for content-based prediction. Finally, the rating prediction is generated through a weighted fusion of content and collaborative models. Compared to benchmark models, our approach reduces RMSE by 3.06% and MAE by 3.23%, demonstrating an improvement in prediction accuracy. Furthermore, our method shows stable performance, with only a slight increase in prediction error (MAE with 8% and RMSE with 1.5% with a hybrid weight of 0.5) under cold-start conditions, indicating that the proposed method maintains strong stability and robustness even in data sparsity scenarios.