Concept Drift Detection by Tracking Weighted Prediction Confidence of Incremental Learning
Pingfan Wang, Wai Lok Woo, Nanlin Jin, Duncan Davies · 2022
Data stream mining is great significant in many real-world scenarios, especially in the big data area. However, conventional machine learning algorithms are incapable to process because of its two characteristics (1) potential unlimited number of data is generated in real-time way, it is impossible to store all the data (2) evolving over time, namely, concept drift, will influence the performance of predictor trained on previous data. Concept drift detection method could detect and locate the concept drift in data stream. However, existing methods only utilize the prediction result as indicator. In this article, we propose a weighted concept drift indicator based on incremental ensemble learning to detect the concept. The indicator not only considers the prediction result, but the change of prediction stability of predictor with occurs of concept drift. Also, an incremental ensemble learning based on vote mechanism is especially used to get constantly updated value of indicator. Based on the experiment result on both benchmark and real-world dataset, our method could effectively detect concept drift and outperform other existing methods.