Abstract:

As a decomposing framework, error-correcting output codes (ECOC) can effectively reduce the multiclass to the binary and attract much attention, in which the construction of coding matrix based on data is the key to use ECOC to solve multiclass problems. An approach of hierarchical errorcorrecting output codes (HECOC) based on support vector domain description (SVDD) and Fisher theory is presented. Firstly, the SVDD is used to measure the class separabilty quantitatively. Then the interclass separability matrix is got gradually. The binary tree is built based on the matrixes from the bottom to the top. Then, each node of the binary tree is encoded by the level to get the final HECOC. The separability of base classifiers trained by different class partition is compared in experiments. The results show that the HECOC can promote the diversity of the base classifiers and the error-correcting ability of codewords as well as enhance the classification accuracy.