An advanced context-aware method for detecting user interest drift: Methodology, statistical analysis, and empirical evaluation

Modern data-driven systems increasingly operate on continuous data streams. In such environments, the statistical properties of data change with external events, contextual factors and system dynamics. This phenomenon, known as concept drift, is lately recognized as a hard-core challenge in learning systems—especially in real-world applications such as recommender systems, and online personalization. User interest drift, an underexplored form of concept drift, is critical in user-centric and decision-critical applications, where delayed or inaccurate adaptation to evolving user behavior might lead to substantial performance degradation. The lack of a statistically principled, context-aware framework in non-stationary environments leaves a significant gap in the literature. Therefore, this thesis introduces C-IDDM (Context-Aware Interest Drift Detection Method), a novel and statistically grounded skeleton for the detection of user interest drift through the explicit integration of contextual information such as time-of-day. The proposed method models user behavior derived from contextual text streams and detects preference changes using distribution-free, permutation-based statistical testing. A real-world Last.fm music listening dataset and synthetic streaming datasets with controlled drift scenarios are utilized to evaluate comparative experiments, which indicate that C-IDDM outperforms a context-free baseline and classical streaming drift detectors, particularly under context-dependent drift, achieving lower false positive rates and better context purity.