Recursive Self Organizing Maps with Hybrid Clustering

Abstract

We introduce the concept of a neural network based recursive clustering which creates an ensemble of clusters by recursive decomposition of data. The work involves a hybrid combination of a global clustering algorithm followed by a corresponding local clustering algorithm. Evolutionary self organizing maps are used to create clusters. A set of core patterns is isolated and separately trained using a SOM. The process is recursively applied to the remaining patterns to create an ensemble of clusters. The partition of each recursion is integrated with the partition of the previous recursion. The correlation of the clusters with ground truth information (in the form of class labels) is used to measure algorithm robustness. The paper shows that a hybrid combination of evolutionary algorithms and neural network based clustering techniques is efficient in finding good partitions of clusters and in finding suitable resultant cluster shapes. The recursive self organizing map proposed aims to improve the clustering accuracy of the self organizing map. Empirical studies show that superior results are obtained when clustering artificially generated data as well as real world problems such as the Iris, Glass and Wine datasets