Feature Transformation and Multivariate Decision Tree Induction

Univariate decision trees (UDT's) have inherent problems of replication, repetition, and fragmentation. Multivariate decision trees (MDT's) have been proposed to overcome some of the problems. Close examination of the conventional ways of building MDT's, however, reveals that the fragmentation problem still persists. A novel approach is suggested to minimize the fragmentation problem by separating hy-perplane search from decision tree building. This is achieved by feature transformation. Let the initial feature vector be x, the new feature vector after feature transformation T is y, i.e., y = T (x). We can obtain an MDT by (1) building a UDT on y; and (2) replacing new features y at each node with the combinations of initial features x. We elaborate on the advantages of this approach, the details of T , and why it is expected to perform well. Experiments are conducted in order to connrm the analysis, and results are compared to those of C4.5, OC1, and CART.