Hypothesis Assessments as Guidance for Incremental and Meta-learning

In this paper, a new decision tree learning algorithm (INC-DT) is proposed: an incremental one based on hypotheses assessments. INC-DT uses only a xed amount of instance memory during the whole learning process. It bases its hypotheses exclusively on the last one computed , its assessment, and a xed number of seen examples. So, it is able to deal with initial knowledge and concept drift. Additionally, the hypothesis assessments can be used directly for model selection and combination. Moreover, they make compentence and quality assessment of the algorithm explicit which supports meta-learning algorithms. First, the necessity of incremental learning algorithms is motivated. Work, both theoretical and practical one, regarding incremental learners is discussed. The algorithm INC-DT is described and evaluated empirically. The innuence of certain parameters like the buuer size, the base algorithm and the data noise are discussed.