Abstract: A minimax probability machine, called DCMPM, with the consensus regularization between data distributions is proposed for data classification in which the data contain labeled and unlabeled samples in this paper. In the proposed machine, labeled and unlabeled samples be mapped to the space of decision hyperplane and then the decision hyperplane is revised by minimizing the difference of the probability distributions between labeled and unlabeled samples such that the revised decision hyperplane is more close to the real classification hyperplane. Experimental results indicate the power of the proposed method.

Key words: data distributions, minimax probability machine, decision hyperplane

摘要： 针对既包含有标记样本又包含未标记样本的分类数据，提出数据分布一致性原理，并将其融入到最小最大概率机中。把有标记样本和无标记样本映射到决策超平面所在空间（简称超空间），通过最小化有标记样本和无标记样本在超空间的概率分布差异，充分利用无标签样本来修正最小最大概率机的误差，使得修正后的决策超平面更接近于真正的分类超平面。实验证明，数据分布一致性最小最大概率机（DCMPM）比最小最大概率机（MPM）具有更好的分类性能。

关键词: 数据分布一致性, 最小最大概率机, 决策超平面