WinDrift++: A Diversified Detector for Concept Drift
Naureen Naqvi, Sabih ur Rehman, Md Zahidul Islam · IEEE Transactions on Artificial Intelligence · 2025
With the rise of hyperconnectivity, vast amounts of time series data are generated by applications, many of which are business-critical and are used by incremental learning models to make informed decisions. A key requirement for such models is the ability to detect concept drift changes in the statistical characteristics of data over time. A recent method, WinDrift (WD), effectively detects drift when only a single statistical characteristic changes. However, real-world datasets often exhibit changes across multiple characteristics simultaneously, requiring a more robust detection approach. We propose WinDrift++ (WD++), a novel method that extends WD by employing a committee of statistical tests rather than a single test. This enhances robustness and accuracy in detecting diverse types of drift. WD++ retains the efficient windowing strategy of WD while improving reliability across a broader range of scenarios. We evaluate WD++ on 22 synthetic and 8 real-world datasets, where it consistently outperforms WD and other state-of-the-art detectors. Code and datasets are available onhttps://github.com/naureenaqvi/windriftplusplus.