Concept drift region identification via competence-based discrepancy distribution estimation

Real-world data analytics often involves cumulative data. While such data contains valuable information, the pattern or concept underlying these data may change over time and is known as concept drift. When learning under concept drift, it is essential to know when, how and where the context has evolved. Most existing drift detection methods focus only on triggering a signal when drift is detected, and little research has endeavored to explain how and where the data changes. To address this issue, we introduce kernel density estimation into competence-based drift detection method, and invent competence-based discrepancy distribution estimation to identify specific regions in the data feature space where drift has occurred. Two experiments demonstrate that our proposed approach, competence-based discrepancy density estimation, can quantitatively highlight drift regions through data feature space, and produce results that are very close to preset drift regions.