Trust Decay-Based Temporal Learning for Dynamic Recommender Systems With Concept Drift Adaptation

Modeling temporal dynamics in recommendation systems is essential for capturing drifts in user preferences over time, particularly under conditions of data sparsity and non-stationarity. In this study, we propose a novel framework, Trust Decay-based Temporal Learning (TDTL), which integrates trust-region optimization, temporal regularization, and drift-aware adaptation to enhance the stability and adaptability of matrix factorization models. A key contribution of TDTL is the incorporation of bias-aware prediction and interaction-specific confidence weighting, which enables the model to capture rating tendencies and filter out noisy updates based on user-item frequency. TDTL also employs a BiGRU-based autoencoder to detect and refine temporal drift in user latent factors. Extensive experiments on nine benchmark datasets from diverse domains, including technology, health, entertainment, and social platforms, demonstrate that TDTL consistently outperforms both classical and deep learning-based baselines. For instance, TDTL reduces the root mean square error (RMSE) from 1.395 to 0.860 on the highly sparse Software dataset compared to TimeSVD++ and from 0.733 to 0.5349 on the Cellphones dataset compared to ALSTM. An ablation study further confirms that bias modeling and confidence weighting significantly contribute to predictive accuracy, particularly in extreme sparsity scenarios. These results highlight the robustness and generalization capability of TDTL as a reliable solution for real-world temporal recommendation tasks.