Data mining via entropy and graph clustering

Abstract

Data analysis often requires the unsupervised partitioning of the data set into clusters. Clustering data is an important but a difficult problem. In the absence of prior knowledge about the shape of the clusters, similarity measures for a clustering technique are hard to specify. In this work, we propose a framework that learns from the structure of the data. Learning is accomplished by applying the K-means algorithm multiple times with varying initial centers on the data via entropy minimization. The result is an expected number of clusters and a new similarity measure matrix that gives the proportion of occurrence between each pair of patterns. Using the expected number of clusters, final clustering of data is obtained by clustering a sparse graph of this matrix.