Concept Drift Detection in Streaming Classification of Mobile Application Traffic
Oleg I. Sheluhin, S. A. Sekretarev · Automatic Control and Computer Sciences · 2021
Abstract An important aspect of the classification of applications under conditions of a priori uncertainty is the operation of algorithms in streaming mode, with the continuous receipt of measurement data. A distinctive feature of the classification of data in streaming mode is the concept drift. A concept drift occurs when the phenomenon being studied, for which data have been collected, changes over time. Under the conditions of non-stationary data flows, the classification of mobile applications should be paired with a concept drift detector (CDD). The paper proposes a two-stage algorithm for detecting a change in the concept in the observed data stream. The algorithm is based on the statistical characteristics of the attributes analyzed using two sliding windows that control the change in the current statistical characteristics of the attributes of mobile applications. At the first stage, key statistics are applied in accordance with the Fisher criterion. At the second stage, the Page–Hinckley test is applied. As a result of the experiments, using an artificial data set, dependencies were obtained that allow one to evaluate the performance of the proposed two-stage algorithm for detecting the concept drift. It is shown that CDD allows reducing the probability of classification error with each change of concept by about 5%.