结合类属特征及因果发现的序列优化分类器链

分类器链是利用标签间相关性实现挖掘特定对象多维标记信息的重要多标签分类方法.面向现有分类器链算法,针对各标签的基学习器均在完整特征空间中训练导致学习特征冗余,以及因标签学习顺序随机且分类器链训练过程单向无反馈导致的标签间相关信息利用不充分等问题,本文提出一种结合类属特征及因果发现的序列优化分类器链.该方法采用类内仿射传播聚类为每个基学习器构建高级结构化特征,减少冗余信息;利用条件熵准则挖掘标签间因果关系,优化学习序列提高对标签间相关信息的利用程度.在多个公开数据集的实验结果表明,序列优化分类器链有效增强了单节点学习效果以及对多标签间关联信息的利用,有效提升了多标签分类效果,实用价值高. 

Sequence Optimization Classifier Chain Based on Label-Specific Features and Causal Discovery

Classifier chain is an important multi-label classification method to mine multi-dimensional label information of specific objects by using the correlation between labels. To solve the problems in the existing classifier chain algorithm,including the redundancy of learning features caused by the base learner training of each label in the complete feature space,and the low efficiency of information utilization among labels caused by the random sequence of label learning and the one-way non-feedback in the training process of classifier chain,a sequence optimization classifier chain based on label-specific features and causal discovery was proposed. In this method,affine propagation clustering was used to construct advanced structured features for each base learner,reducing the difficulty of training single label nodes. At the same time,conditional entropy was used to mine the causal relationship between labels,optimize the learning sequence and improve the utilization density of relevant information between labels. The experimental results on several open datasets show that the sequential optimization classifier chain can effectively enhance the learning effect of single node and the utilization of correlation information between multi-labels,and improve the classification effect of multi-labels,possessing high practical value.