Link Fraction Mixed Membership Reveals Community Diversity in Aggregated Social Networks

Community detection is a critical tool for understanding the mesoscopic structure of large-scale networks. However, when applied to aggregated or coarse-grained social networks, disjoint community partitions cannot capture the diverse composition of community memberships within aggregated nodes. While existing mixed membership methods alleviate this issue, they may detect communities that are highly sensitive to the aggregation resolution, not reliably reflecting the community structure of the underlying individual-level network. This paper presents the Link Fraction Mixed Membership (LFMM) method, which computes the mixed memberships of nodes in aggregated networks. Unlike existing mixed membership methods, LFMM is consistent under aggregation. Specifically, we show that it conserves community membership sums at different scales. The method is utilized to study a population-scale social network of the Netherlands, aggregated at different resolutions. Experiments reveal variation in community membership across different geographical regions and evolution over the last decade. In particular, we show how our method identifies large urban hubs that act as the melting pots of diverse, spatially remote communities.

💡 Research Summary

This paper tackles a fundamental problem in the analysis of aggregated (coarse‑grained) social networks: traditional community detection methods produce disjoint partitions that cannot capture the overlapping nature of individual‑level community memberships, and existing mixed‑membership models (e.g., MMSBM, overlapping SBM) are highly sensitive to the resolution of aggregation, often violating the ecological fallacy. To address these issues, the authors introduce the Link Fraction Mixed Membership (LFMM) method.

LFMM defines a node’s (or aggregate set’s) membership in community k as the fraction of total link weight that connects the node to nodes belonging to that community, excluding self‑loops. Formally, for an original weighted undirected graph G with adjacency matrix W, the unnormalized membership is
(M_i(k)=\sum_{j\in C_k} w_{ij}(1-\delta_{ij})).
The normalized vector (m_i(k)=M_i(k)/\sum_k M_i(k)) represents a probability‑like distribution over communities. Crucially, because the definition is linear in the edge weights, LFMM is provably consistent under any aggregation: the sum of the M‑vectors of all individual nodes inside an aggregated set S_x equals the M‑vector computed directly on the aggregated graph G′ (Equation 4). This “aggregation invariance” is not shared by non‑linear mixed‑membership formulations.

The methodology consists of two stages. First, an arbitrary disjoint community detection algorithm (the authors use the Leiden‑Potts implementation) partitions the aggregated graph into communities C_k. Second, LFMM vectors are computed via a single matrix multiplication of the aggregated adjacency matrix with a community indicator matrix, followed by optional normalization. The authors also define a community‑diversity index (entropy of the normalized LFMM vector) and a spatial null model based on a gravity formulation to assess statistical significance of observed patterns.

Synthetic experiments use stochastic block models (SBM) with two communities, varying intra‑community affinity μ and aggregation mixing probability m. Results confirm perfect correlation (r = 1.0) between LFMM computed on the aggregated graph and the sum of individual‑level LFMM values, validating the theoretical consistency. Normalized values show a near‑perfect correlation (r ≈ 0.999). When community detection is performed on the individual‑level graph instead of the aggregated one, LFMM still approximates the ground truth (r ≈ 0.997) but tends to over‑estimate minority memberships in aggregates. The authors also demonstrate that high μ or high m each increase LFMM values, yet when both are present the effects partially cancel, indicating that LFMM cannot distinguish between a homogeneous high‑affinity aggregate and a heterogeneous mixed‑membership aggregate without additional structural cues.

The real‑world case study analyzes a register‑based social network of the Netherlands, comprising roughly 17 million residents linked through family, work, and school ties over 13 years. The network is aggregated at two spatial resolutions: ~3 000 neighbourhoods and ~400 municipalities. Applying LFMM reveals several key patterns:

1. Urban melting pots – Metropolitan areas (Amsterdam, Rotterdam, The Hague) exhibit the highest average LFMM values, indicating that these locations host a rich mixture of members from many distant communities.
2. Spatial significance – Compared against a gravity null model that accounts for distance and population size, urban LFMM excesses are statistically significant (p < 0.01), confirming that the observed mixing cannot be explained by simple spatial decay alone.
3. Temporal evolution – Over the 2012‑2024 period, LFMM in expanding suburbs and newly built districts (e.g., Almere, Albert’s Pond) rises sharply, while traditional industrial zones (e.g., Rotterdam port area) show a modest decline, reflecting demographic shifts and changing social integration patterns.
4. Normalization effects – Because the normalized LFMM vector m′ is a strength‑weighted sum of constituent nodes, large high‑degree aggregates exert disproportionate influence; the authors propose an optional degree‑normalization to mitigate this bias, which yields similar qualitative results.

The paper discusses limitations: (i) LFMM’s edge‑centric nature may over‑emphasize hubs; (ii) identical LFMM values can arise from either homogeneous high‑affinity aggregates or heterogeneous mixes, requiring supplementary diagnostics; (iii) the current framework treats each snapshot and relationship type separately, suggesting future extensions to multilayer, dynamic mixed‑membership models.

In conclusion, LFMM provides a mathematically sound, computationally efficient tool for extracting mixed‑membership information from aggregated social networks while guaranteeing consistency across aggregation scales. Its application to a nation‑wide dataset demonstrates that urban centers function as “melting pots” of diverse, spatially remote communities and that community composition evolves measurably over time. The method opens avenues for policy‑relevant analyses of social integration, urban planning, and the impact of demographic change on mesoscopic network structure.