An Online Competence-Based Concept Drift Detection Algorithm.

The ability to adapt to new learning environments is a vital feature of contemporary case-based reasoning system. It is imperative that decision makers know when and how to discard outdated cases and apply new cases to perform smart maintenance operations. Competence-based empirical distance has been recently proposed as a measurement that can estimate the difference between case sample sets without knowing the actual case distributions. It is reportedly one of the most accurate drift detection algorithms in both synthetic and real-world data sets. However, as the construction of competence models have to retain every case in memory, it is not suitable for online drift detection. In addition, the high computational complexity O((n^{2})) also limits its practical application, especially when dealing with large scale data sets with time constrains. In this paper, therefore, we propose a space-based online case grouping strategy, and a new case group enhanced competence distance (CGCD), to address these issues. The experiment results show that the proposed strategy and related algorithms significantly improve the efficiency of the current leading competence-based drift detection algorithm.