Unified Graph-Temporal Recommendation Model Based on Preference Drift and Prediction

Modern recommender systems operate in highly dynamic environments where user preferences evolve over time. This phenomenon, commonly referred to as preference drift, leads to performance degradation when models assume stationary behavior. In this work, we propose a drift-aware recommendation framework that models and predicts user preference evolution using graph neural networks combined with temporal sequence modeling. We represent user–item interactions as a sequence of time-dependent bipartite graphs and learn user embeddings via GraphSAGE. Temporal dynamics are captured using a recurrent neural network that predicts future user embeddings. We further incorporate Monte Carlo dropout to estimate predictive uncertainty and quantify drift magnitude as geometric displacement in latent space. Experiments on MovieLens and Yambda datasets demonstrate that the proposed method improves embedding prediction accuracy, enhances recommendation robustness under high-drift conditions, and enables reliable detection of anomalous behavioral changes. The results show that drift-aware modeling significantly mitigates performance degradation in non-stationary environments while providing actionable uncertainty signals.