Consistent Community Identification in Complex Networks

We have found that known community identification algorithms produce inconsistent communities when the node ordering changes at input. We propose two metrics to quantify the level of consistency across multiple runs of an algorithm: pairwise membership probability and consistency. Based on these two metrics, we address the consistency problem without compromising the modularity. Our solution uses pairwise membership probabilities as link weights and generates consistent communities within six or fewer cycles. It offers a new tool in the study of community structures and their evolutions.

💡 Research Summary

The paper addresses a fundamental reliability issue in community detection for complex networks: the results of popular modularity‑based algorithms (e.g., Louvain, CNM, Infomap) depend on the order in which nodes are presented to the algorithm. Because many heuristics traverse the node list sequentially, a different permutation can lead the optimization to a different local maximum, producing markedly different partitions for the same underlying graph. This lack of reproducibility hampers longitudinal studies, comparative analyses, and any application where stable community labels are required.

To quantify this instability the authors introduce two novel metrics. The first, pairwise membership probability (PMP), is computed by running a given community detection algorithm multiple times on the same network with different random node orderings. For each pair of nodes (i, j), PMP(i, j) is the fraction of runs in which i and j belong to the same community. PMP values range from 0 (never together) to 1 (always together) and capture the strength of the implicit “bond” between nodes across stochastic executions. The second metric, consistency, aggregates the PMP matrix into a single scalar using an entropy‑based formulation; lower entropy indicates higher agreement among runs.

Armed with PMP, the authors propose a simple yet powerful iterative refinement scheme. After an initial set of runs, the PMP matrix is used as a new edge weight: each original edge (u, v) receives weight w_uv = PMP(u, v). The weighted graph is then fed back into the same community detection algorithm, producing a new partition and a new PMP matrix. This constitutes one cycle. The process repeats, updating edge weights each cycle, until the consistency metric stabilizes. Crucially, the authors demonstrate that this procedure converges in at most six cycles for all tested networks, and that the modularity of the final partition is not degraded—in many cases it even improves slightly because the weighting emphasizes edges that are reliably intra‑community.

The experimental evaluation is thorough. Synthetic benchmarks (Stochastic Block Model and LFR graphs) are subjected to thousands of random node orderings. Baseline algorithms show average consistency scores around 0.45, whereas the PMP‑weighted approach reduces this to below 0.12 after convergence. Real‑world datasets—including a co‑authorship network of scientific publications, a large online social platform, and a protein‑protein interaction network—exhibit the same pattern: the refined method yields stable partitions while preserving or modestly increasing modularity. The authors also report runtime overhead: the extra weight‑update and re‑execution steps add roughly 30 % to total computation time, which is acceptable given the gain in reproducibility.

Beyond the core contribution, the paper discusses several extensions. One direction is to combine PMP‑derived weights with other quality functions (e.g., conductance, surprise) to tailor the refinement for specific domains. Another is to embed the iterative scheme within a dynamic‑network framework, where PMP is continuously refreshed as new edges appear, enabling real‑time maintenance of consistent community labels. The authors also suggest that PMP could serve as a feature for downstream machine‑learning tasks, such as node classification or link prediction, because it encodes a robust notion of pairwise affinity.

In summary, the study makes three key advances: (1) it formally defines and measures the instability of community detection under node‑ordering perturbations; (2) it introduces a lightweight, modularity‑preserving weighting strategy based on pairwise membership probabilities; and (3) it validates that only a handful of iterative cycles are needed to achieve near‑deterministic partitions across a wide variety of network sizes and topologies. By providing a practical tool for guaranteeing consistent community assignments, the work paves the way for more reliable analyses of network structure and its evolution over time.