Efficient clustering of large uncertain graphs using neighborhood information

Abstract

This work addresses the problem of clustering large uncertain graphs. The data is represented as a graph where the proposed solution uses the neighborhood information for the purpose of clustering. The proposed approach converts an uncertain graph to a certain graph by predicting about the existence of the edges in the uncertain graph. For the purpose of prediction, a classifier is used. The proposed approach is compared with baseline approaches for clustering graphs having uncertainties over the edges; uncertain k-means (UK-Mean) and Fuzzy-DBSCAN (FDBSCAN). Additionally, the results are also compared with two state-of-the-art approaches namely, CUDAP (clustering algorithm for uncertain data based on approximate backbone) and PEEDR (partially expected edit distance reduction). Experiments are conducted using two natively uncertain and nine synthetically converted uncertain benchmark datasets. The results are compared with the baseline and the state-of-the-art methods using Davies–Bouldin index, Dunn index and Silhouette coefficient, widely used cluster validity indices. The results show that the proposed approach performs better than the other four methods.