关键词: 数据流, 基于约束学习, 直推学习, 最大熵, 分布式组合分类, data streams, constraint-based learning, transductive learning, maximum entropy, distributed ensemble classification

Abstract: The existing strategy of combining decisions for ensemble classification method requires common labeled training samples across these ensemble classifiers. To resolve combining classifiers decisions among ensemble classification over data streams without labeled examples, a transductive constraint-based learning strategy was proposed. It satisfied the constraints measured by each local classifier based on transductive learning theory while choosing decision on test samples; thereby guaranteed the feasibility of the constraints. It solved the problems of transductive extension of maximum entropy for aggregation in distributed classification. Experimental examples prove that the proposed method can achieve higher classifying accuracy over the existing transductive approach and can be applied to ensemble classification fusing for data streams.