[Paper Review] On weakly optimal partitions in modular networks

J. Ignacio Alvarez-Hamelin, Mariano G. Beiró+1 more|arXiv (Cornell University)|Aug 20, 2010

Complex Network Analysis Techniques21 references3 citations

TL;DR

This paper introduces a probabilistic formalism for modularity in weighted networks using a resolution parameter $ t $, and proposes a novel algorithm based on weakly optimal partitions to identify high-quality community structures. The method efficiently scales to large networks and achieves modularity scores comparable or superior to existing state-of-the-art algorithms, including Blondel et al.'s fast algorithm, while offering theoretical insights into modularity's scaling limitations and weak optimality.

ABSTRACT

Modularity was introduced as a measure of goodness for the community structure induced by a partition of the set of vertices in a graph. Then, it also became an objective function used to find good partitions, with high success. Nevertheless, some works have shown a scaling limit and certain instabilities when finding communities with this criterion. Modularity has been studied proposing several formalisms, as hamiltonians in a Potts model or laplacians in spectral partitioning. In this paper we present a new probabilistic formalism to analyze modularity, and from it we derive an algorithm based on weakly optimal partitions. This algorithm obtains good quality partitions and also scales to large graphs.

Motivation & Objective

• To address the scaling and instability issues in modularity-based community detection, particularly the resolution limit problem.
• To formalize a probabilistic framework for modularity in weighted networks using a resolution parameter $ t $.
• To introduce and analyze the concept of weakly optimal partitions as a more robust alternative to strict modularity optimization.
• To develop a fast, scalable algorithm that identifies weakly optimal partitions and evaluates its performance on synthetic and real-world networks.
• To compare the algorithm’s modularity results with existing methods, revealing strong performance even on large-scale networks.

Proposed method

• Formalizes modularity using a signed measure $ \mu_t(S) = m_E(S) - t \cdot m_{VV}(S) $, where $ m_E $ is edge probability and $ m_{VV} $ is the product of marginal probabilities.
• Introduces a resolution parameter $ t $ to control community size, generalizing standard modularity and enabling multi-scale community detection.
• Defines weak optimality of a partition as maximizing $ \mu_t $ over all refinements, providing a more stable criterion than strict modularity maximization.
• Develops a fast, greedy algorithm that iteratively merges communities to achieve weakly optimal partitions, inspired by but distinct from Blondel et al.'s approach.
• Employs a hierarchical merging strategy based on local $ \mu_t $ gain, ensuring scalability to large graphs.
• Validates results using synthetic networks (e.g., binary trees) and real-world networks, comparing modularity scores with multiple benchmarks.

Experimental results

Research questions

• RQ1 How does the introduction of a resolution parameter $ t $ affect the stability and scalability of modularity-based community detection?
• RQ2 What are the theoretical properties of weakly optimal partitions, and how do they relate to standard modularity optimization?
• RQ3 Can a weakly optimal partitioning algorithm achieve competitive modularity scores while scaling efficiently to large networks?
• RQ4 To what extent does weak optimality mitigate the resolution limit problem inherent in standard modularity maximization?
• RQ5 How do the modularity values of weakly optimal partitions compare to those obtained by state-of-the-art algorithms like Blondel et al. on real-world networks?

Key findings

• The proposed algorithm achieves modularity scores comparable to or better than Blondel et al.'s fast algorithm on all tested real-world networks, including large-scale ones like web-google and ir_dimes.
• For binary trees of height $ n $, the algorithm achieves modularity values nearly identical to the theoretical upper bound, with $ Q_1 \approx 0.953 $ at $ n=10 $, indicating strong performance on synthetic benchmarks.
• On the karate club network, the algorithm achieves $ Q_1 = 0.405 $, slightly below Newman’s original result but competitive with other methods.
• For the arXiv co-citation network, the algorithm achieves $ Q_1 = 0.797 $, outperforming Clauset et al. ($ 0.772 $) and Newman’s method ($ 0.770 $).
• On the PGP trust network, the algorithm achieves $ Q_1 = 0.864 $, surpassing Newman’s $ 0.855 $ and Blondel’s $ 0.884 $, indicating strong performance on dense, real-world graphs.
• The evolution of $ Q_t(\mathcal{C}_t) $ shows that optimal modularity is often reached near $ t=1 $, suggesting that $ t=1 $ is a robust default setting for community detection.