Reacting to different types of concept drift with adaptive and incremental one-class classifiers

摘要

Modern computer systems generate massive amounts of data in real-time. We have come to the age of big data, where the amount of information exceeds the perceptive abilities of any human being. Frequently the massive data collections arrive over time, in the form of a data stream. Not only the volume and velocity of data poses a challenge for machine learning systems, but also its variability. Such an environment may have non-stationary properties, i.e. change its characteristic over time. This phenomenon is known as concept drift, and is considered as one of the main challenges for moder learning systems. In this paper, we propose to investigate different methods for handling concept drift with adaptive soft one-class classifiers. One-class classification is a promising direction in data stream analytics, as it allows for a novelty detection, data description and learning with limited access to class labels. We describe an adaptive model of Weighted One-Class Support Vector Machine, augmented with mechanisms for incremental learning and forgetting. These allow for our models to swiftly adapt to changes in data, without any need for a dedicated drift detector. We carry out an experimental analysis of the behavior of our method with different forgetting rates for various types of concept drift. Additionally, we compare our classifier with state-of-the-art one-class methods for streaming data. We observe, that our adaptive soft one-class model can efficiently handle different types of concept drifts, while delivering a highly satisfactory accuracy for streaming data.