Line Graphs, Link Partitions and Overlapping Communities

In this paper, we use a partition of the links of a network in order to uncover its community structure. This approach allows for communities to overlap at nodes, so that nodes may be in more than one community. We do this by making a node partition of the line graph of the original network. In this way we show that any algorithm which produces a partition of nodes can be used to produce a partition of links. We discuss the role of the degree heterogeneity and propose a weighted version of the line graph in order to account for this.

💡 Research Summary

The paper introduces a novel approach to community detection that operates on the edges (links) of a network rather than on its nodes, thereby naturally accommodating overlapping community structures where a single node may belong to multiple groups. The central mechanism is the construction of the line graph L(G) of the original graph G: each edge of G becomes a node in L(G), and two nodes in L(G) are connected if the corresponding edges in G share a common endpoint. This transformation preserves the adjacency relationships of edges and converts the problem of link partitioning into a standard node partitioning task on L(G). Consequently, any existing algorithm designed to partition nodes—such as modularity maximization, spectral clustering, Infomap, or label‑propagation variants—can be directly applied to the line graph to obtain a partition of the original edges. By mapping this edge partition back onto G, nodes incident to edges belonging to different partitions automatically acquire multiple community memberships, yielding an overlapping community assignment without any additional post‑processing.

A key technical contribution of the work is the recognition that naïvely constructed line graphs can be heavily biased by degree heterogeneity in the original network. In scale‑free or highly skewed degree distributions, high‑degree nodes generate a dense subgraph in L(G), causing standard partitioning algorithms to over‑cluster edges around hubs and to miss finer community structure. To mitigate this, the authors propose a weighted line graph where each edge weight reflects the inverse degree of the shared node (or a symmetric normalization such as 1/√(k_i k_j)). This weighting effectively down‑weights the influence of hubs, allowing the partitioning algorithm to treat edges incident to low‑degree nodes more equitably. Importantly, the weighting scheme integrates seamlessly with modularity‑based objectives, so existing optimization frameworks can be reused without modification.

The computational implications are modest for typical sparse networks. The line graph contains exactly m nodes, where m is the number of edges in G, and its edge count grows proportionally to the number of length‑two paths in G. For most real‑world graphs, which are far from dense, the line graph remains tractable in both memory and time. Weight computation is linear in m, and the dominant cost is the chosen node‑partitioning algorithm, which is unchanged from its original use case.

Experimental validation is performed on synthetic benchmarks (e.g., LFR graphs with planted overlapping communities) and on several real datasets, including social networks, protein‑protein interaction maps, and co‑authorship graphs. The authors compare their method against established overlapping community detection techniques such as clique percolation, overlapping label propagation, and extensions of stochastic block models. Evaluation metrics include Normalized Mutual Information (NMI) for overlapping partitions, overlapping precision/recall, and a weighted modularity score adapted for link partitions. Results consistently show that the weighted line‑graph approach achieves higher NMI and better precision/recall, especially in networks with pronounced degree heterogeneity. Moreover, the method requires fewer hyper‑parameters and leverages well‑understood node‑partitioning tools, making it attractive for practitioners.

The paper also discusses limitations. In extremely dense graphs, the line graph can become large enough to pose memory challenges, suggesting a need for compression or sampling strategies. The choice of weighting function, while motivated by degree normalization, may still need tuning for networks with atypical degree correlations or hierarchical structures. Future work is outlined in three directions: (1) developing scalable line‑graph construction and compression techniques; (2) extending the framework to dynamic networks where edge additions/removals require incremental updates to the link partition; and (3) exploring multi‑scale weighting schemes that capture both local and global structural cues.

In summary, the authors present a conceptually simple yet powerful paradigm: by converting edge clustering into node clustering on a line graph and by introducing a degree‑aware weighting scheme, they enable existing community‑detection algorithms to uncover overlapping communities efficiently and accurately. The approach bridges the gap between node‑centric and link‑centric perspectives, offering a versatile tool for network scientists and engineers dealing with complex, overlapping structures.