H.-J. Zhou’s scientific contributions

It is known that error-correcting output codes (ECOC) is a common way to model multiclass classification problems, in which the research of encoding based on data especially attracts attentions. In this paper, we proposed a method for learning error-correcting output codes with the help of a single layered perception neural network. To achieve this goal, the code elements of ECOC are mapped to the weights of network for the given decoding strategy, and an object function with the constrained weights used as a cost function of network. After the training, we can obtain a coding matrix including lots of subgroups of class. Experimental results on artificial data and UCI with logistic linear classifier (LOGLC) as the binary learner show that our scheme provides better performance of classification with shorter length of coding matrix than other state-of-the-art encoding strategies.

It is an effective way to transform multiclass into binary by using error-correcting output code (ECOC) as a decomposing frame, in which the research of encoding based on data especially attracts attentions. An encoding method for adaptively constructing ECOC is proposed. To achieve this goal, the relativity between each pair of patterns is achieved with the help of confusion matrixes. Then abiding by the Fisher's principle, the most favorable combination of patterns for classifying is obtained, and a binary partition based on the way of pattern combination and a data driven coding matrix are gotten. Experimental results on UCI datasets and three kinds of HRRP datasets show that the proposed scheme provides a better performance and robustness of classification than the classical encoding strategies.