Toward temporal stability in explainable AI: A framework for measuring explanation drift

Adaptive recommender systems are updated as new user interactions occur. These updates can change both the recommendations and explanations produced by the system. However, most explainable artificial intelligence methods evaluate explanations at a single point in time and do not measure how they change during continual updates. This study investigates explanation drift in dynamic graph-based recommender systems. We propose a temporal framework that combines graph representation learning, streaming embedding updates, Shapley-based feature attribution, and drift-aware optimization. The framework uses three measures: Temporal Attribution Divergence (TAD), Explanation Stability Score (ESS), and Explanation Half-Life (EHL). We evaluated the approach on MovieLens-1M and an Amazon Product Reviews subset under continual update settings. The proposed framework achieved an F1@10 of 0.837 on MovieLens-1M, compared with 0.821 for the strongest reported baseline, and 0.810 on Amazon, compared with 0.789. It also reduced the TAD by approximately 42.2% compared with the Static SHAP and increased the ESS from 0.612 to 0.831. These results show that explanation changes can be reduced while maintaining the recommendation performance in the tested settings. The evaluation is limited to offline benchmark datasets and feature-attribution stability and does not measure how users understand or judge the explanations. These results support the use of temporal explanation stability as an additional evaluation measure for adaptive recommender systems.