Modeling Concept Drift Detection as Machine Learning Model Using Overlapping Window and Kolmogorov–Smirnov Test

Abstract

Nowadays the large volume of data from different sources especially as streaming data opens us various opportunities for streaming analytics. Concept drift is one of the challenging streaming analytic problems which observes the changes in the distribution of the data over time, and detecting and adapting these attracted many researchers. In this work, we modeled concept drift detection as a machine learning problem. We have followed a semi-supervised learning approach by utilizing a statistical test, the Kolmogorov–Smirnov test which determines the variation of two time series distributions. The core work is to build a classifier that is capable of predicting the given window of data stream holds drift or not. As we have no labels representing drift or not drift for the stream windows, we have explicitly labeled some beginning parts of the stream using the Kolmogorov–Smirnov test and utilized these for building the classifier. Using this classifier, further, we can detect the drift. Also, we have used overlapping windows to avoid information loss. For building the classifier we have applied various classification models like Logistic Regression, Support Vector Machine, K-Nearest Neighbors, Naïve Bayes, Decision Tree, and Random Forest. Among them KNN model has a low false-positive rate and outperformed others with an accuracy 96%.