| Abstrakt: |
In this paper, we propose a sequential decomposition method for multi-class pattern classification problems based on domain knowledge. A novel modular decision tree architecture is used to divide a K-class classification problem into a series of L smaller (binary or multi-class) sub-problems. The set of all K classes c = {c1, c2, ...cK} is divided into smaller subsets (c = {s1, s2, ...sL}) each of which contains classes that are related to each other. A modular approach is then used to solve (1) the binary sub-problems ( ]> ) and (2) the smaller multi-class problem si = {ci1, ci2, ...cin}. Problem decomposition helps in a better understanding of the problem without compromising on the classification accuracy. This is demonstrated using the rules generated by the C4.5 classifier using a monolithic system and the modular system. [ABSTRACT FROM AUTHOR] |
