A case study of incremental concept induction

Abstract

Application of machine induction techniques in complex domains promises to push the computational limits of non-incremental, search intensive induction methods. Learning effectiveness in complex domains requires the development of incremental, cost effective methods. However, discussion of dimensions for comparing the utility of differing incremental methods has been lacking. In this paper we introduce 3 dimensions for characterizing incremental concept induction systems which relate to the cost and quality of learning. The dimensions are used to compare the respective merits of 4 incremental variants of Quinlan's learning from examples program, ID3. This comparison indicates that cost effective induction can be obtained, without significantly detracting from the quality of induced knowledge.