Classification model with subspace data-dependent balls

Abstract

Data-Dependent Ball (DDB) is a pre-processing algorithm that transforms quantitative into binary data by mapping them into a set of balls. In datasets with large number of dimensions, data-dependent balls are less significant due to the distance calculation in the mapping process. To reduce number of ball dimensions, this paper proposes a method for subspace data-dependent balls (SDDB) generation. SDDB starts by ranking features using information gain, and then eliminating input features based on ratio r. Subspace data-dependent balls are then created and filtered out with respect to their size and purity. Finally, a C4.5 decision tree classification model is constructed using subspace data-dependent balls as features. Experimental results from 8 TICI datasets show that the accuracy from a combination of SDDB and C4.5 is better than the combination of DDB and C4.5 in terms of accuracy.