自适应属性加权近邻传播聚类算法

Affinity Propagation Clustering Algorithm Based on Adaptive Feature Weight

针对多维数据属性对聚类分析结果有不同重要程度影响的问题,提出一种基于自适应属性加权的近邻传播聚类算法。该方法通过考虑多维数据属性权值的重要度,在近邻传播聚类过程中引入属性加权相似性矩阵计算,并根据当前数据聚类划分的结果来分析目标评价函数,计算各个属性对当前聚类的贡献程度。随后根据贡献程度的计算结果自适应地更新属性权值,并通过属性加权相似性矩阵来重新计算近邻传播算法中的两种竞争信息,进而提高聚类结果的质量。数值实验结果表明,新方法能够有效实现属性权值的自适应调整,提高近邻传播算法的聚类效果,与其他传统聚类算法相比新方法具有更好的聚类质量。

Due to the different effects of different features on multidimensional data during the clustering,this paper proposes an affinity propagation clustering algorithm based on adaptive feature weight. This methodintroduces a feature weight similarity measure to affinity propagation by considering different importance ofdifferent features. During the clustering the new method evaluates the objective function according to current datapartition, and calculates the contribution of each features made to current clustering. Then the feature weights willbe self-adaptive updated according to the contribution, and the two messages passing between data points inaffinity propagation will be recalculated on the basis of feature weight similarity matrix to enhance the clusteringresults. The experimental results show that feature updating works well on the new algorithm, and compared to theaffinity propagation and other traditional clustering algorithm the new algorithm can obtain better results.