Drift detection in ADWIN using the Kolmogorov-Smirnov (KS) test
Hezal Lopes, Prashant Nitnaware · 2025
Data stream mining is used to find meaningful patterns or gain some insights from continuously flowing data streams. Change in input data pattern is considered as concept drift. It is crucial to handle and manage concept drift in machine learning applications. This paper discusses various methods and algorithms to detect the change in data streams, especially focusing on real-time IoT applications. Most researchers have done evaluation and analysis on static, synthetic or predesigned datasets. It is essential to consider analysis of real-world dynamics and high velocity data with the drift factor. In this work, the K-S (Kolmogorov-Smirnov) test and z test are used to detect changes in current distribution and reference distribution models. These statistical tests are used to calculate the drift score and identify how specific features influence drift. If the drift score is less than 0.05, then the drift signal gets activated. Further, it also provides a comparative analysis of current and reference models for handling concept drift. This work is used to provide early detection of the presence of drift before its severe impact on classification or regression. If drift is handled properly, then it is easy to take real-time decisions in critical domains such as financial trading, healthcare monitoring, and network security.