A Comparison of Approaches for Handling Concept Drifts in Data Processed With Machine Learning

In the realm of machine learning models, the pursuit of achieving favorable metrics is undeniably significant. However, these models confront phenomena that can diminish their effectiveness if left unaddressed-notably, the phenomenon of concept drift. Concept drift materializes when unforeseen alterations in the statistical properties of the target variable transpire over time. This article introduces a comprehensive analysis of classical treatment methods, examining the behavior of machine learning models across diverse datasets. Various concept drift detection algorithms are employed, facilitating a holistic assessment. The study encompasses a comparative exploration of different classification algorithms within the scikit-multiflow framework. These algorithms integrate adaptive strategies to contend with concept drifts. Through this generalized analysis, the performance of distinct classification algorithms is contrasted. The overarching aim is to facilitate the selection of optimal classification methods aligned with specific types of concept drift. Ultimately, this study provides a pivotal toolkit for the judicious selection of classification methods, enhancing model adaptability in the presence of concept drifts. In addition to shedding light on the behavior of machine learning models under concept drift, the findings empower practitioners and researchers to make informed decisions to optimize model robustness.