Abstract

The traditional pattern recognition method assumes that the model that is used does not depend on data timing. This assumption is correct for several practical issues but it is not valid for the case where new data frequently becomes available . We could meet the above-mentioned situation in many practical issues, as spam filtering, intrusion detection/prevention or recognition of client behaviour to enumerate only a few. In these cases, the dependencies between the observation and classes are continually changing. Unfortunately, most pattern recognition methods do not take the so-called concept drift into consideration and they cannot adapt to a new model. Therefore, design classifiers for data streams are currently the focus of intense research. Another important issue related to the recognition of data streams is the problem of data labelling. Traditional machine learning methods use supervised learning algorithms, which could produce a classifier on the basis of a set of labelled examples. In this approach, we should take into consideration the cost of data labelling, which is usually passed over. Let us notice that labels are usually assigned by human experts and therefore they cannot label all new examples if they come too fast. Therefore, methods of classifier design which could produce the recognition system on the basis of a partially labelled set of examples would be an attractive proposition. This article focuses on the problem of the concept drift using active learning approach for the minimal distance classifiers. The potential for adaptation of the proposed method and its quality are evaluated through computer experiments, carried out on several benchmark data sets