基于粗糙集理论的协同训练算法

为了提升风险决策环境下协同训练的效果, 提出了一种基于粗糙子空间的协同决策算法。首先利用粗糙集属性约简的概念, 将部分标记数据属性空间分解为两差异性较大的粗糙子空间; 在各子空间上训练分类器, 并依据各分类器决策风险代价及隶属度将无标记数据划分为可信、噪声和待定样本。综合两分类器的分类结果, 标注少量可信无标记样本后重复协同训练。从理论上分析了算法性能提升的区间界, 并在UCI数据集上进行实验, 验证了模型的有效性及效率。

Novel co-training algorithm based on rough sets

In order to improve the performance of co-training in the context of decision with risk, this paper proposed a rough subspace-based co-training algorithm. Based on the concept of attribute reduction in rough sets, the algorithm first splitted all condition attributes of partially labeled data into two diverse rough subspaces. Then the two classifiers trained from derived rough subspaces and classified the unlabeled data into confident, noise and uncertain samples with consideration of the classification risk and membership of decision class. Finally, it labeled a few of confident samples for the two classifiers to learn from each other in iterative manner. It theoretically analyzed the performance of proposed algorithm, and empirical results on selected UCI data sets also show its effectiveness.