A self-adaptive ensemble for user interest drift learning

User interest reflects user preference which plays an important role in commercial decision-making. Learning and predicting user interest has attracted significant attention in recent years, however, user interest will change under uncertain environments as time passes, called user interest drift. This may adversely impact the accuracy of machine learning model prediction and lead to a delay in decision-making. How to detect and adapt to user interest drift in streaming data is an important problem which needs to be addressed. In this paper, we propose a novel method to detect user interest drift, called the topic-based user interest drift detection method (T-IDDM), which can recognize the severity of user interest drift. Then, a self-adaptive ensemble (SA-Ensemble) method with an adaptive weighted voting strategy is proposed to deal with user interest drift and reduce the time decay of the voting process. Next, a dynamic voting strategies selection process is proposed and applied to improve model robustness. Finally, an application study of user interest drift learning is presented to verify the proposed method. Twelve sequential online reviews datasets are collected and tested for the experiment. A comparison of our method with state-of-the-art benchmark methods shows its efficiency.