面向不平衡数据的特征加权聚类算法

Feature Weighted Clustering Algorithm for Unbalanced Data

不平衡数据集类别分布严重倾斜,传统的聚类算法由于以提高整体学习性能为目标,往往偏向于聚集多数类,而忽视更有价值的稀有类.本文提出一种基于迭代的特征加权聚类算法,根据当前聚类后簇的特点以及特征重要性度量函数确定特征权值,利用所得权值进行下一轮聚类,直到权值稳定后结束迭代.在多个UCI不平衡数据集上的实验效果表明,本文算法能够较好地识别出重要特征并提高它们的权重,避免聚类算法过度偏向多数类,有效地提高了聚类性能.

The class distribution in imbalanced data sets is serious inclined, as the traditional clustering algorithm mainly designed for improving the overall learning performance, the majority class usually tends to be clustered and the minority class which is more valu- able would be ignored. In This paper, a new weighted clustering algorithm based on iteration is proposed. It updates feature weights according to the current character of clusters and a formula of feature importance measurements. With the calculated weights, cluster is processed iteratively until the weights are stable. Experimental results on the imbalance datasets in UCI shows that the new cluste- ring algorithm can recognize and enhance the weights of important feature, which avoids the clustering algorithm in favor of the ma- jority class and effectively improve the clustering performance.