Graph-Based Competence Model for Concept Drift Detection
Yiping Sun, Yansen Yu, Cheng Jin, Qingyong Zhang, Weixu Liu, Hang Yu · 2023
Concept drift is the fundamental reason for the degradation of accuracy in evolving data stream learning, as the change in data stream distribution leads to an increase in error rates of the learning model. In online learning models or other detection models, it is crucial to detect whether concept drift has occurred for accurate predictions of the learning model. On the other hand, Case-Based Reasoning (CBR) systems utilize solutions to similar problems stored in a case library to address new problems. The system's competence can be measured to assess its effectiveness in solving problems, and this competence can be evaluated by the changes in the probability distribution of cases. Similarly, detection methods for concept drift operate on a similar principle, aiming to measure the distribution changes of cases relative to their competence. Inspired by this, this paper models the original data using a graph structure, establishing discrete regions, and then constructs a competence model to calculate the competence values for each discrete region. Differences are detected by evaluating the temporal changes in competence values for each region. The proposed method is validated through five experiments in three categories, demonstrating the effectiveness of our detection method for concept drift.