Decision Tree-Based Learning Using Multi-Attributed Lens

Abstract

Decision tree induction is one of the popular and widely used method for classification. In a general decision tree algorithm, a dataset is split into partitions by the attribute which provides the best information measure value. Then, each partition recursively finds the best attribute until the terminal criteria are met. In this thesis, a new type of decision tree algorithm, multi-attributed lens, is presented. Instead of using one attribute at a time, all attributes are used to create a lens which is further used to split a dataset. A lens is generated from the farthest pair which splits a dataset into two regions. The first region is the outside lens region where its instances will be labeled as the opposite class of the farthest instances. The second region is the inside lens region where its instances need to be further partitioned. In splitting the instances of the inside lens, the best splitting point is computed by measuring the information measure of their projection values onto the core vector which is created from the farthest pair. Then, newly created partitions are recursively called until the termination conditions are met. Empirically, t-tests of the performance measures are gathered on the UCI datasets and the results show that accuracy of this algorithm is more statistically significant than C4.5 decision tree, k nearest neighbor, naïve Bayes, support vector machine and logistic regression in Blood Transfusion Service Center, Breast Cancer Wisconsin (Original), Breast Cancer Wisconsin (Prognostic), Haberman's Survival, ILPD (Indian Liver Patient Dataset), Credit Approval, Hepatitis and Horse Colic datasets.