基于概率LS-SVM 的多标签非均衡样本分类算法

针对多标签分类问题,提出了一种面向样本不均衡及类属不确定性的多标签分类算法。首先,结合“一对一”分解策略和贝叶斯理论,将多标签数据集分解为单标签数据子集,并利用Parzen窗方法估计子集样本后验概率,对类标签进行了基于概率的不确定性表示。然后,在融合概率类标签和LS-SVM模型的基础上,利用样本差异信息来调节惩罚参数值,建立了考虑样本不均衡的概率LS-SVM子分类器模型。依据正态分布的3σ原理,设计了子分类器决策阈值确定方法。最后,结合实例对算法进行了性能分析,结果证明了新算法的合理性和有效性。

Multi-label unbalanced classification algorithm based on probability LS-SVM

To tackle the data unbalanced and label uncertain problems in multi label classification, a new multi label classification algorithm was proposed. Firstly, with Bayes theory and "one versus one" decomposition strategy, the multi label datasets were decomposed into single label datasets. By using Parzen window method to estimate data posteriori probabilities, the class label was expanded into probability class label. Then the probability class label was combined with least square support vector machine (LS SVM), and the probability LS SVM sub classifiers model that takes the data unbalanced into account were designed, and the data difference information was used to modify the penalty parameter value. And with the 3σ theory, the classifier decision threshold determination method was put forward. Finally, combined with examples, the algorithm performances were analyzed, which proves the rationality and effectiveness of the algorithm.