Incremental Learning and Forgetting in One-Class Classifiers for Data Streams

摘要

One-class classification and novelty detection is an important task in processing data streams. Standard algorithms used for this task cannot efficiently handle the changing environment to which they are applied. In this paper we present a modification of Weighted One-Class Support Vector Machine that is able to swiftly adapt to changes in data. This was achieved by extending this classifier by adding incremental learning and forgetting procedures. Both addition of new incoming data and removal of outdated objects is carried out on the basis of modifying weights assigned to each observation.We propose two methods for assigning weights to incoming data and two methods for removing the old objects. These approaches work gradually, therefore preserving useful characteristic of the examined dataset from previous iterations. Our approach was tested on two real-life dynamic datasets and the results prove the quality of our proposal.