[Paper Review] How to Measure Significance of Community Structure in Complex Networks

Yanqing Hu, Yiming Ding+2 more|arXiv (Cornell University)|Feb 10, 2010

Complex Network Analysis Techniques14 references3 citations

TL;DR

This paper proposes a novel method to measure the significance of community structure in complex networks without prior knowledge of community partitions, using the stability of eigenvectors from the Laplacian matrix. By analyzing eigenvalue gaps and eigenvector robustness, the method identifies the optimal number of communities and reveals that social networks and *C. elegans* neural networks exhibit high significance, while protein and metabolic networks show lower significance.

ABSTRACT

Community structure analysis is a powerful tool for complex networks, which can simplify their functional analysis considerably. Recently, many approaches were proposed to community structure detection, but few works were focused on the significance of community structure. Since real networks obtained from complex systems always contain error links, and most of the community detection algorithms have random factors, evaluate the significance of community structure is important and urgent. In this paper, we use the eigenvectors' stability to characterize the significance of community structures. By employing the eigenvalues of Laplacian matrix of a given network, we can evaluate the significance of its community structure and obtain the optimal number of communities, which are always hard for community detection algorithms. We apply our method to many real networks. We find that significant community structures exist in many social networks and C.elegans neural network, and that less significant community structures appear in protein-interaction networks and metabolic networks. Our method can be applied to broad clustering problems in data mining due to its solid mathematical basis and efficiency.

Motivation & Objective

• To address the lack of methods that evaluate the significance of community structure independently of community detection algorithms.
• To quantify how robust community structures are against error links and algorithmic randomness, which are common in real-world networks.
• To develop a mathematically grounded index that determines the optimal number of communities before partitioning.
• To apply the method to diverse real-world networks and compare the significance of community structures across network types.

Proposed method

• Uses the Laplacian matrix of a network to compute eigenvalues and eigenvectors, with the smallest eigenvalue always being zero.
• Defines a robustness index R based on the stability of eigenvectors under perturbation, derived from the ratio of the gap between consecutive eigenvalues to the norm of the perturbation.
• Introduces the H-index as a measure of community structure significance, calculated from the eigenvector stability and eigenvalue gaps.
• Applies the method to real networks by varying the number of communities c and identifying the c that minimizes R and maximizes the eigenvalue gap λ_{c+1} - λ_c.
• Employs the LFR benchmark to validate the method’s ability to recover known community structures.
• Uses the H-index to rank networks by significance, with higher values indicating more stable and meaningful community structures.

Experimental results

Research questions

• RQ1 How can the significance of community structure be evaluated without prior knowledge of the community partition?
• RQ2 What mathematical properties of the Laplacian matrix can reliably indicate the presence of a meaningful community structure?
• RQ3 Can eigenvector stability and eigenvalue gaps be used to determine the optimal number of communities in a network?
• RQ4 How does the significance of community structure vary across different types of real-world networks, such as social, biological, and technological networks?
• RQ5 To what extent do error links and algorithmic randomness affect the reliability of detected community structures?

Key findings

• The *C. elegans* neural network exhibits a high H-index of 0.57, indicating a highly significant and stable community structure.
• The H-index for the *C. elegans* metabolic network is 0.62, also indicating strong community structure, which contrasts with previous findings due to improved community number estimation.
• Social networks such as the College football network (H = 0.79) and Jazz network (H = 0.47) show high significance, while the Political blogs network has a low H-index of 0.22.
• Protein interaction networks like *E. coli* (H = 0.14), *H. Sapiens* (H = 0.21), and Yeast (H = 0.40) show moderate to low significance, suggesting less distinct community structures.
• Metabolic networks such as *Aquifex aeolicus*, *Helicobacter pylori*, and *Yersinia pestis* all have H-indices of 0.36, indicating consistent but moderate significance.
• The method successfully identifies the true number of communities in benchmark networks: 2 for Zachary’s karate club, 12 for College football, and 3 for Political books, as confirmed by minimal R values and maximal eigenvalue gaps.