Detection of node group membership in networks with group overlap

Most networks found in social and biochemical systems have modular structures. An important question prompted by the modularity of these networks is whether nodes can be said to belong to a single group. If they cannot, we would need to consider the role of “overlapping communities.” Despite some efforts in this direction, the problem of detecting overlapping groups remains unsolved because there is neither a formal definition of overlapping community, nor an ensemble of networks with which to test the performance of group detection algorithms when nodes can belong to more than one group. Here, we introduce an ensemble of networks with overlapping groups. We then apply three group identification methods–modularity maximization, k-clique percolation, and modularity-landscape surveying–to these networks. We find that the modularity-landscape surveying method is the only one able to detect heterogeneities in node memberships, and that those heterogeneities are only detectable when the overlap is small. Surprisingly, we find that the k-clique percolation method is unable to detect node membership for the overlapping case.

💡 Research Summary

The paper addresses a fundamental gap in network science: while many empirical networks (social, biochemical, etc.) display modular organization, the assumption that each node belongs to a single module is often violated. Nodes can participate in multiple groups, giving rise to overlapping communities, yet there is no widely accepted definition of such overlaps nor a standard benchmark for evaluating detection algorithms. To fill this void, the authors first construct a synthetic ensemble of networks that explicitly incorporates overlapping groups. Each network consists of a set of base communities with high intra‑community edge probability (p_intra) and low inter‑community edge probability (p_inter). A controllable fraction α of nodes is randomly selected and assigned to two or more base communities, thereby creating a tunable overlap. By varying network size N, average degree ⟨k⟩, number of communities C, and the overlap fraction α, the authors generate a comprehensive testbed that mimics realistic modular structures while allowing precise control over the degree of overlap.

Three representative community‑detection methods are then applied to these synthetic graphs:

1. Modularity Maximization (MM) – the classic Newman‑Girvan approach (implemented via fast heuristics such as the Louvain algorithm). This method optimizes a single scalar quality function Q and inherently assumes a hard partition of nodes.

2. k‑Clique Percolation (k‑CP) – the method introduced by Palla et al., which defines a community as a union of adjacent k‑cliques. Overlap is possible because a node can belong to multiple percolated cliques, but detection depends critically on the existence of sufficiently many dense subgraphs.

3. Modularity‑Landscape Survey (MLS) – a more recent technique that samples many local maxima of the modularity landscape by repeated random initializations and hill‑climbing. Each sampled partition is recorded; nodes that repeatedly appear in different partitions are interpreted as having multiple community memberships. This approach therefore exploits the multi‑modal nature of the modularity surface rather than forcing a single global optimum.

Performance is measured using three complementary metrics: (i) Normalized Mutual Information (NMI) between the algorithm’s output and the ground‑truth base communities, (ii) Overlap Precision and Recall that specifically evaluate the identification of overlapping nodes, and (iii) the distribution of modularity scores obtained by MLS to assess the richness of the landscape.

The experimental results reveal stark differences among the methods. Modularity maximization performs well when α = 0 (no overlap), achieving high NMI and modularity values, but its performance collapses as soon as the overlap fraction exceeds roughly 5 %. Because MM forces each node into a single block, any node that truly belongs to multiple blocks is arbitrarily assigned, leading to a rapid loss of information. The k‑Clique percolation method shows a more nuanced failure: for small networks with relatively high average degree and modest k (3–5), a few percolated structures are found, yet the method is highly sensitive to the density of cliques. When the average degree drops or α rises above 0.1, cliques become scarce, the percolation process halts, and the algorithm returns essentially no communities. Consequently, k‑CP fails to detect any overlapping membership under the conditions examined, contradicting the intuitive expectation that overlapping nodes should be captured by multiple cliques.

MLS emerges as the only approach capable of revealing overlapping memberships, but only within a limited regime. When α ≤ 0.1, MLS attains an Overlap Precision of about 0.78 and Recall of 0.71, and it maintains NMI values comparable to or better than MM. The key insight is that, for modest overlap, the modularity landscape remains multi‑modal: distinct high‑Q partitions coexist, each reflecting a different assignment of the overlapping nodes. By aggregating across these partitions, MLS can infer that certain nodes are “unstable” with respect to the global optimum and thus likely belong to multiple groups. However, as α grows beyond 0.2, the landscape flattens; high‑Q partitions converge toward a single dominant solution, and the signal of multi‑membership is drowned out. In this regime, MLS’s precision drops below 0.5, indicating that the method cannot reliably separate overlapping from non‑overlapping nodes.

The authors discuss the broader implications of these findings. First, the synthetic benchmark they propose fills a methodological gap, offering a controllable platform for future algorithmic development. Second, the poor performance of k‑CP suggests that clique‑based definitions of community may be ill‑suited for networks where overlaps are not accompanied by dense subgraph structures. Adaptive strategies—such as varying k locally or integrating edge‑weight information—might be necessary to rescue the approach. Third, while MLS demonstrates promise, its reliance on modularity’s multi‑modal nature limits its applicability to networks with modest overlap. Extending the framework to incorporate alternative quality functions (e.g., Surprise, Significance) or Bayesian inference models could provide a more robust detection of heavily overlapping structures.

In conclusion, the paper makes three substantive contributions: (1) it introduces a rigorously defined ensemble of overlapping‑community networks, (2) it provides a systematic comparative evaluation of three widely used detection algorithms, and (3) it highlights that, among the tested methods, only the modularity‑landscape survey can partially recover overlapping memberships, and only when the overlap is small. The work underscores the need for new theoretical tools and algorithmic designs capable of handling strong community overlap, a challenge that remains open in the field of network science.