The Research of Ternary Error-Correcting Output Codes Based on Genetic Programming

Error-Correcting Output Codes (ECOC) provides an effective solution for the multiclass classification problem by decomposing a multiclass problem into a set of binary class problems. In an ECOC algorithm, the design of coding matrix is the key to its performance. In this paper, we propose a Genetic Programming (GP) based ECOC algorithm, aiming to produce optimal coding matrices through the evolutionary process. In our GP, each terminal node denotes a column in the coding matrix, and each nonterminal node represents an operator, which combines the colunms represented by its terminal nodes. In this way, an individual is interpreted as a coding matrix, and a set of operators are proposed to exchange information between column pairs, so as to produce new colunms. Feature selection methods are also integrated into the terminal nodes, so that individuals are dynamically assigned to optimal feature subspaces for diverse classification problems. With evolutionary operators, offspring with high discriminant capability would be produced in the evolution. Our experiments compare our algorithm with other 7 classic ECOC algorithms with the deployment of diverse basic classifiers based on a set of UCI data sets, and results prove the superiority and robustness of our algorithm.