基于k-邻域相关性的多标签分类

摘要

Multi-label classification is such a problem that a example belongs to multiple classes, data character and relevance among labels have great influences on classification result. Thus, this paper proposed an algorithm that combined those two factors above. Data were mapped characteristically, then used to get the maximal posteriori probability of every label and added to corresponding data character. Data character with maximal posterior proba-bility trained classifiers the same number as labels. Experiments on classic databases of yeast, scene and emotion denoted the effectiveness of the proposed algorithm.%多标签分类是指数据可以同属于多个类的分类问题，其数据特征和标签间相关性对分类结果存在影响。因此，提出一种融合前述两种因素的多标签分类算法。将数据用核函数进行特征映射，根据训练数据的k-邻域计算得到每个标签的最大后验概率；并将其加入到对应的数据特征中。用加入最大后验概率的数据特征训练分类器。通过在经典的yeast、scene和emotion数据库上实验，证明了算法的有效性。