Abstract

Most work on supervised learning is undertaken on static problems. However, in many real world classification problems, the environment in which the classifiers operate is dynamic i.e. the descriptions of classes change with time. In this paper, the process of generating drifting data is introduced in order to assess two adaptive approaches that deal with dynamically changing data. These approaches are: retraining on evolving data set and incremental learning. The empirical evaluation has shown that both these approaches improve the performance compared to the non-adaptive mode though a number of outstanding research issues remain.