Fast Heuristic Algorithm for Multi-scale Hierarchical Community Detection

TL;DR

The paper introduces HAMUHI-CODE, a fast heuristic algorithm for multi-scale hierarchical community detection in complex networks, utilizing a new similarity measure and an efficient agglomerative clustering approach.

Contribution

It proposes a novel, scalable algorithm that improves community detection speed and accuracy in large complex networks using a new similarity measure and a flexible merging criterion.

Findings

Achieves linear time complexity per iteration.

Performs well on real-world and synthetic networks.

Outperforms several state-of-the-art algorithms.

Abstract

Complex networks constitute the backbones of many complex systems such as social networks. Detecting the community structure in a complex network is both a challenging and a computationally expensive task. In this paper, we present the HAMUHI-CODE, a novel fast heuristic algorithm for multi-scale hierarchical community detection inspired on an agglomerative hierarchical clustering technique. We define a new structural similarity of vertices based on the classical cosine similarity by removing some vertices in order to increase the probability of identifying inter-cluster edges. Then we use the proposed structural similarity in a new agglomerative hierarchical algorithm that does not merge only clusters with maximal similarity as in the classical approach, but merges any cluster that does not meet a parameterized community definition with its most similar adjacent cluster. The algorithm computes all the similar clusters at the same time is checking if each cluster meets the parameterized community definition. It is done in linear time complexity in terms of the number of cluster in the iteration. Since a complex network is a sparse graph, our approach HAMUHI-CODE has a super-linear time complexity with respect to the size of the input in the worst-case scenario (if the clusters merge in pairs), making it suitable to be applied on large-scale complex networks. To test the properties and the efficiency of our algorithm we have conducted extensive experiments on real world and synthetic benchmark networks by comparing it to several baseline state-of-the-art algorithms.