page 1  (17 pages)
2to next section

An Algorithm for Classification by Feature

Partitioning

_Izzet S?irin H. Altay Guvenir

Computer Engineering and Information Science Department
Bilkent University, Ankara 06533 TURKEY
E-mail: fsirin,[email protected]

Abstract

This paper presents a new methodology for learning from examples, called Classification by Feature Partitioning (CFP), which is an inductive, incremental and supervised learning method. Learning in CFP is accomplished by storing the objects separately in each feature dimension as disjoint partitions of values. A partition, which is initially a point in the feature dimension, is expanded through generalization. The CFP algorithm specializes a partition by subdividing it into sub-partitions. It is shown that the CFP algorithm has low sample complexity and training complexity. CFP is also empirically evaluated in three different domains, and the results are compared with Instance-based learning, Nested Generalized Exemplars and Decision Tree techniques.