Enhancing community detection using a network weighting strategy

A community within a network is a group of vertices densely connected to each other but less connected to the vertices outside. The problem of detecting communities in large networks plays a key role in a wide range of research areas, e.g. Computer Science, Biology and Sociology. Most of the existing algorithms to find communities count on the topological features of the network and often do not scale well on large, real-life instances. In this article we propose a strategy to enhance existing community detection algorithms by adding a pre-processing step in which edges are weighted according to their centrality w.r.t. the network topology. In our approach, the centrality of an edge reflects its contribute to making arbitrary graph tranversals, i.e., spreading messages over the network, as short as possible. Our strategy is able to effectively complements information about network topology and it can be used as an additional tool to enhance community detection. The computation of edge centralities is carried out by performing multiple random walks of bounded length on the network. Our method makes the computation of edge centralities feasible also on large-scale networks. It has been tested in conjunction with three state-of-the-art community detection algorithms, namely the Louvain method, COPRA and OSLOM. Experimental results show that our method raises the accuracy of existing algorithms both on synthetic and real-life datasets.

💡 Research Summary

The paper addresses the persistent challenge of accurately detecting community structure in large‑scale complex networks. While many state‑of‑the‑art methods rely solely on topological information—such as modularity maximization (e.g., the Girvan‑Newman or Louvain algorithms) or spectral clustering—they often suffer from two major drawbacks: (1) prohibitive computational cost that prevents scaling to networks with millions of nodes and edges, and (2) the “resolution limit” problem, which causes small but meaningful communities to be merged into larger ones during optimization.

To overcome these issues, the authors propose a meta‑algorithmic preprocessing step that assigns a weight to every edge based on its ability to carry information across the network. The weight is derived from a novel centrality measure called κ‑path edge centrality. This measure is defined as the cumulative frequency with which an edge is traversed during multiple random walks whose length does not exceed a user‑defined bound κ. Two constraints are imposed on each walk: (i) an edge may be visited at most once per walk, preventing artificial inflation of its weight, and (ii) the walk terminates after κ steps, reflecting the empirical observation that influence decays with distance (Friedkin’s postulate).

Computing κ‑path centrality exactly would be infeasible for large graphs, so the authors introduce WERW‑Kpath (Weighted Edge Random Walk – κ Path), an efficient approximation algorithm. WERW‑Kpath repeatedly initiates random walks from randomly selected vertices, records each traversed edge, and updates its weight. The algorithm runs in O(κ·|E|) time and requires only linear memory, making it suitable for networks with millions of edges. The authors prove that the estimated centrality deviates from the true value by at most 1/|E|, providing a theoretical guarantee of accuracy.

After the weighting phase, the original graph is fed unchanged into existing community detection algorithms. The paper evaluates three representative methods:

1. Louvain – a fast, hierarchical modularity‑maximization technique.
2. COPRA – a label‑propagation algorithm capable of detecting overlapping communities.
3. OSLOM – a statistical‑significance based method that extracts robust communities.

By incorporating edge weights, each algorithm benefits from a richer representation of the network: edges that frequently participate in short information‑spreading paths receive higher weights, thereby emphasizing intra‑community connections and attenuating inter‑community links. This directly mitigates the resolution limit, allowing small communities to be preserved during modularity optimization.

Experimental Evaluation
The authors conduct extensive experiments on both real‑world and synthetic data.

Real‑world datasets: Nine publicly available networks are used, the largest being a Facebook sample with 613 497 vertices and 2 045 030 edges. For each dataset, the authors compare the baseline algorithms with their weighted counterparts using two metrics: (i) modularity Q, which quantifies the density of intra‑community edges relative to a random null model, and (ii) Normalized Mutual Information (NMI), which measures agreement with ground‑truth community labels when available. Results show that modularity improves by up to 16 % and NMI consistently rises, indicating more accurate community recovery. Importantly, the additional preprocessing adds only a modest overhead; total runtime remains comparable to the original methods.

Synthetic benchmarks: Using the LFR benchmark, 72 graphs with varying sizes, average degrees, and community size distributions are generated. Since the true community partition is known, NMI serves as the primary evaluation metric. Across all three detection algorithms, the weighted version yields higher NMI scores (average gains of 0.12–0.18), especially for graphs with many small communities where the resolution limit is most severe.

Key Contributions

1. Formal definition of κ‑path edge centrality and its interpretation as an information‑propagation relevance measure.
2. Development of the WERW‑Kpath algorithm, offering a provably accurate approximation with near‑linear time complexity.
3. Demonstration that the weighting scheme can be seamlessly combined with diverse community detection techniques, improving both modularity and NMI without altering the underlying algorithms.
4. Empirical evidence that the approach scales to networks with hundreds of thousands of nodes and millions of edges, and that it alleviates the resolution limit problem.
5. Release of an open‑source implementation to foster reproducibility.

Limitations and Future Work
The method requires the user to set κ and the number of random walks; while the authors provide sensible defaults, an adaptive scheme could further enhance robustness. Additionally, the random‑walk based centrality may be less informative on extremely sparse or highly disconnected graphs, suggesting the need for hybrid measures. Future research directions include (i) automatic parameter tuning, (ii) distributed or GPU‑accelerated implementations for truly massive graphs, and (iii) exploration of alternative walk‑biasing strategies (e.g., degree‑biased walks) to capture richer structural nuances.

In summary, the paper presents a practical, theoretically grounded preprocessing technique that enriches edge information, thereby boosting the performance of existing community detection algorithms on both synthetic and real‑world large‑scale networks.