Abstract:This paper is concerned with the problem of handwritten digits recognition in high dimensional feature space. The residual feature information may bring undesirable complexity to the underlying probability distribution of the concept label for learning algorithm to capture. The recognition accuracy and efficiency of the so trained learning model are usually depressed. According to this situation, an improved confusion-crossed support vector machine tree is proposed. A feature selection process based on the sensitivity of the margin to a feature is presented for the training step of each support vector machine embedded in each internal node. The experimental results on optical handwritten digits recognition problem in UCI database indicate that the proposed approach achieves competitive or even better recognition accuracy in the condensed feature space. Further, it also obtains lower structure complexity on internal nodes and the whole hybrid learning model than the compared approaches.