Constructive inductive learning based on meta-attributes

Constructive Inductive Learning, CIL, aims at learning more accurate or comprehensive concept descriptions by generating new features from the basic features initially given. Most of the existing CIL systems restrict the kinds of functions that can be applied to construct new features, because the search space of feature candidates can be very large. However, so far, no constraint has been applied to combining the basic features. This leads to generating many new but meaningless features. To avoid generating such meaningless features, in this paper, we introduce meta-attributes into CIL, which represent domain knowledge about basic features and allow to eliminate meaningless features. We also propose a Constructive Inductive learning system using Meta-Attributes, CIMA, and experimentally show it can significantly reduce the number of feature candidates.