Abstract: ROCCH theory is mainly used to solve the performance evaluation of two-classifier which are cost sensitive，how to effectively extend it to multi class evaluation is the difficulty for research. With binary-tree and vertical average method，a new cost-sensitive three-classifier evaluation method B-ROCCH was made up. B-ROCCH method was used to transform three-classification problems into two-classification problems with the idea of binary-tree，three types of ROC curves were plotted with the vertical average method，then combinated with the ROCCH thoughts，to determine the potential best classifier and the best classifier in the three-classification problems. The method was implemented on the MBNC platform，the experimental data were compared with the B-AUC evaluation method，the results prove that the B-ROCCH method is feasible，more distinguishable and faster.

Key words: classifier evaluation, cost - sensitive, binary-tree, three-classifier