Learning Curve in Concept Drift While Using Active Learning Paradigm

摘要

Classification of evolving data stream requires adaptation during exploitation of algorithms to follow the changes in data. One of the approaches to provide the classifier the ability to adapt changes is usage of sliding window - learning on the basis of the newest data samples. Active learning is the paradigm in which algorithm decides on its own which data will be used as training samples; labels of only these samples need to be obtained and delivered as the learning material. This paper will investigate the error of classic sliding window algorithm and its active version, as well as its learning curve after sudden drift occurs. Two novel performance measures will be introduces and some of their features will be highlighted.