Neighborhood-Based Smoothing of External Cluster Validity Measures

摘要

This paper proposes a methodology for introducing a neighborhood relation of clusters to the conventional cluster validity measures using external criteria, that is, class information. The extended measure evaluates the cluster validity together with connectivity of class distribution based on a neighborhood relation of clusters. A weighting function is introduced for smoothing the basic statistics to set-based measures and to pair wise-based measures. Our method can extend any cluster validity measure based on a set or pairwise of data [joints. In the experiment, we examined the neighbor component of the extended measure1 and revealed an appropriate neighborhood radius and some properties using synthetic and real-world data.