Efficient Calculation of Structural Similarity Threshold for the SCAN Network Clustering Algorithm

Community detection algorithms play an important role in discovering knowledge in networks. The Structural Clustering Algorithm for Network (SCAN) is a community detection algorithm which is capable of detecting hubs and outliers, in addition to cluster members. The term hub means node with the ability of collecting and delivering information among clusters while outlier is considered as a noise in the data. Currently, researchers use exhaustive search to determine the structural similarity threshold value (ε) in the SCAN. This paper reports a new approach of using interval ε value to narrow the searching domain for proper ε value for the SCAN. The approach first adopts computational results produced by the Fast Modularity and the Walktrap algorithms to bind the number of clusters of a network and then determine the interval for ε value. For each of our test datasets, the interval prediction reliably finds the true number of clusters. More importantly, the proposed prediction method helps users to eliminate an average of 67.7% of inappropriate ε values used to generate clusters.