面向分组类别概率问题的模糊SVM分类算法

A Fuzzy SVM Classification Algorithm from Group Probabilities Problem

分组类别概率问题(Q-GP)给定样本的群组统计信息或类别概率分布,寻求每个个体样本的实际类标签,有着广泛的实际应用,但目前相应的研究仍较少。Q-GP问题求解的关键是如何利用已知的样本群组信息来获取单个样本的分类信息。文中通过比较二分类Q-GP问题与有监督及半监督二分类问题的异同,提出利用模糊分类的思想,根据已知的各群组类别概率分布近似获取个体样本的类隶属度,以此构造有监督样本进行学习。具体方法是:首先使用fuzzy层次分类构造各群组的等价类,并利用等价类将二分类Q-GP问题变换成多个带模糊隶属度的有监督二分类子问题;然后实施fuzzy SVM训练子分类器;最后整合多个子分类器的结果即得到每个样本的类标签估计。

The problem of estimation from group probabilities （Q-GP） is to find the actual labels of the individual samples given the label proportions in each group,which has a wide application,but lacking of the existing study. The Q-GP solution is critical to use the known information of group probabilities to obtain the classification information for single sample. In this paper, present a fuzzy classifica- tion method based on fuzzy support vector machine （SVM） to solve this problem by comparing the binary Q-GP with the supervised and semi-supervised binary classification in difference. Firstly, introduce the fuzzy hierarchical classification to find the relationships between objects in a group,so as to decompose the binary Q-GP into supervised sub-problems with fuzzy memberships. Then A fuzzy SVM is trained for each sub-problem. At last combine multiple sub-classifiers to get the final labels of all individual samples.