Multi-Criteria Analysis of Concept Drift Detection Algorithms: A Decision-Making Approach

Concept Drift is a challenging problem in data streaming, where the underlying data distribution changes over time. Numerous algorithms have been proposed to address this issue, each evaluated using various metrics such as accuracy, runtime, and false alarms. However, a comprehensive evaluation that simultaneously considers all these metrics is lacking. Motivated by this gap, our paper systematically benchmarks eleven leading concept drift detection algorithms using a Multi-Criteria Decision-Making (MCDM) approach to identify the best-performing methods. We employ four datasets and seven performance measures: Average Delay Detection (ADD), Average True Detection (ATD), Average False Alarm (AFA), Average False Negative (AFN), Average Detection Runtime in milliseconds (ARMS), Average Memory Usage in bytes (MUB), and Average Accuracy. Our experimental evaluation and comparison are conducted against eleven existing detectors. The results show that our approach provides a balanced and comprehensive assessment, offering a significant advancement in the evaluation of concept drift detection methods. This paper provides a holistic strategy that integrates multiple performance metrics to enhance timely and efficient detection in various applications.