Detection of data drift in a two-dimensional stream using the Kolmogorov-Smirnov test
Piotr Porwik, Benjamin Mensah Dadzie
Abstract
In recent years, there has been an increasing amount of streaming information coming from time series. Learning from data appearing in real time is quite a call, due in part to the speed at which new data appears. Hidden data changes that are not previously known to learning algorithms are referred to in the literature as data or concept drift. In classical machine learning, a classifier analyzes new data using past training instances of the data stream. However, the accuracy of the classifier deteriorates due to data drift, which occurs in non-stationary data. In such situations, the classifier must detect a significant change in the data adapt its prediction over time. The motivation of this paper is to show a method for drift detection without knowledge of instance labels. Labels are sometimes not available or periodically missing, making it difficult to apply methods where knowledge of them is required.