Concept drift detector based on centroid distance analysis
Jakub Klikowski · 2022 International Joint Conference on Neural Networks (IJCNN) · 2022
The interest in data stream mining is continuously growing due to the increasing volume of data arriving at high speed produced by various systems. Processing and classification to gain valuable information from this kind of data require special treatment. The main problem is the non-stationary nature of data streams, which reveals in concept drifts - changes in sample characteristics describing classes. There are many approaches to mitigate the adverse effects of this phenomenon. However, one of the most popular techniques is drift detection. This work will propose a new drift detection approach based on the distance analysis between the subsequent data chunk centroids. The original solution in this methodology is to gather and preserve information about the data characteristics starting from the previous drift detection or data stream beginning. Two variants of the proposed method were compared with state-of-the-art concept drift detection algorithms. The evaluation was made on a comprehensive set of synthetic streams with different concept drift types and selected real streams. The evaluated results subjected to statistical analysis showed that the proposed concept drift detection algorithm could improve the classification of tested classifiers with a statistically significant advantage over the compared drift detection methods.