基于TCM的多分类算法研究

Multi-classification algorithm research based on TCM

基于算法随机性理论提出的直推式置信机器能够给出预测的可靠性,但其多用于解决两类识别问题。扩展了置信机器,利用了正反类的思想,在识别时比较多个P值来确定测试样本的分类,使其很容易一次性应用于多分类识别问题。为对扩展后的模型性能进行评估,将其应用于经典的模式识别-人脸识别。实验结果表明,扩展后的置信机器具有良好的分类性能,当每类训练集样本增加到6个时,识别率已高于96%。

Transductive confidence machine is method based on random algorithm theory. It can estimate the reliability of a prediction but has mainly been applied binary classification problems. This paper extends the transductive confidence machine using the idea of positive and negative classes. The extended Transductive Confidence Machine（TCM）does classification by comparing multiple P values and can be applied to multi-class problems. The new algorithm is applied to face recognition and achieves a recognition rate of 96%even when each class contains only 6 training samples.