Abstract: An algorithm of confidence evaluation for SVM is presented. Based on the evaluation， decision risk is specified in the context of multi-classification. Evaluation approach combines the theoretic analysis and empirical analysis. The decision with low confidence is refined by a local classifier that is formulated online. The local classifier works in query’s neighborhood，and the neighborhood is developed according to a local metric. Experiments demonstrate the fine performance of the designed algorithm， and its improvement in classification accuracy over the state of the arts.

Key words: SVM, confidence evaluation, decision risk amount, local classifier, local metric

摘要： 设计了SVM置信度在线评估方案，以此确定SVM做多分类时结果的风险程度，对高风险决策结果进行修正。置信度评估采用理论估计和经验估计相结合的方式。多分类决策结果的修正由在线生成的局部分类器完成。局部分类器在待查询数据的邻域内工作，此邻域基于一个局部测度而生成。实验表明，所设计的算法呈现了较好的分类能力，提高了传统同类方法的分类准确率。

关键词: SVM, 置信度评估, 决策风险值, 局部分类器, 局部测度