Well-Connected Clusters in Community Detection

• Well-Connected Clusters (WCC) are clusters with strong internal connectivity that resist fragmentation by small edge cuts.
• WCC methods leverage min-cut thresholds, spectral connectivity measures, and post-processing steps to ensure superior community cohesion.
• These approaches improve clustering accuracy and scalability in applications such as stochastic block models and large-scale network analytics.

Searching arXiv for the cited WCC-related papers to ground the article in current records. Well-Connected Clusters (WCC) denotes a family of concepts in which a cluster is required to satisfy a stronger internal cohesion condition than mere edge density. In graph clustering and community detection, the central idea is that a cluster should not only contain many internal edges, but also should not be separable by deleting only a small number of edges. This viewpoint appears in several distinct but related strands of the literature: local spectral clustering based on internal connectivity (Zhu et al., 2013), connectivity guarantees in Leiden-style community detection (Traag et al., 2018), explicit min-cut-based post-processing criteria for communities in empirical networks (Park et al., 2023), and cut-based repair of stochastic block model outputs (Park et al., 2 Feb 2025). The acronym is also used differently in other domains, including Weighted Community Clustering, Weakly Cool-Core galaxy clusters, and rare connected solution regions in the binary perceptron (Saltz et al., 2014, Hernández-Martínez et al., 21 Jul 2025, Abbe et al., 2021).

1. WCC as a graph-theoretic community criterion

In the graph-clustering literature, WCC is motivated by the claim that a community should be “well connected,” not merely “dense.” The distinction is explicit in “Well-Connected Communities in Real-World and Synthetic Networks” (Park et al., 2023), which treats well-connectedness as stronger than ordinary dense-community structure: a community should not just have many internal edges, it should also not be separable by a small edge cut. The same paper operationalizes this with a minimum-edge-cut threshold. For a cluster $C$ with $n$ nodes, the default rule is that $C$ is well connected only if its minimum edge cut is strictly greater than

$f(n)=\log_{10} n.$

A cluster is poorly connected if its minimum cut is at most $f(n)$ (Park et al., 2023).

This threshold is presented as a deliberately mild criterion. The same source compares it to a bound associated with CPM-optimal clusterings: $|X| \ge r \times |A| \times |B|,$ and therefore

$\mincut(C) \ge r \times (n-1).$

The threshold $f(n)=\log_{10}n$ is described as slower-growing than that guarantee, hence a comparatively weak but meaningful standard (Park et al., 2023).

A related formulation appears in the local graph clustering literature, where well-connectedness is captured through internal connectivity rather than a direct min-cut threshold. “Local Graph Clustering Beyond Cheeger’s Inequality” introduces an internal connectivity parameter $\mathsf{Conn}(A)\in[0,1]$, defined as the reciprocal of the mixing time of the random walk over the induced subgraph on $A$ (Zhu et al., 2013). The conductance guarantee for a local random-walk-based algorithm improves from the previously known $n$0 to

$n$1

Using $n$2, where $n$3 is the second eigenvalue of the Laplacian of the induced subgraph on $n$4, the guarantee can be as good as

$n$5

This places WCC-like structure in a spectral framework: the better the internal connectivity, the stronger the local clustering guarantee (Zhu et al., 2013).

A plausible implication is that the literature converges on two closely related but non-identical notions of WCC. One is cut-based, defined directly through minimum edge cuts; the other is spectral or random-walk-based, defined through fast internal mixing or high internal expansion.

2. Connectivity guarantees and the Leiden line of work

A major turning point in the study of well-connected communities is “From Louvain to Leiden: guaranteeing well-connected communities” (Traag et al., 2018). That work argues that the Louvain algorithm may yield arbitrarily badly connected communities and, in the worst case, communities may even be disconnected, especially when running the algorithm iteratively. The paper introduces the Leiden algorithm to address this defect and proves that Leiden yields communities that are guaranteed to be connected (Traag et al., 2018).

The formal framework uses several connectivity notions tied to a resolution parameter $n$6 and a quality function $n$7. A community $n$8 is connected if the induced subgraph on $n$9 is connected. More strongly, a set $C$0 is $C$1-connected if either $C$2, or $C$3 can be split into two sets $C$4 and $C$5 such that

$C$6

and both $C$7 and $C$8 are $C$9-connected recursively. A community is $f(n)=\log_{10} n.$0-connected if $f(n)=\log_{10} n.$1 is $f(n)=\log_{10} n.$2-connected (Traag et al., 2018).

Leiden differs from Louvain by inserting a refinement phase between local moving and aggregation. Nodes are merged only within each community of the partition produced by the local moving phase, and only if the node and candidate community are sufficiently well connected to their parent community. The refinement rules include conditions of the form

$f(n)=\log_{10} n.$3

and

$f(n)=\log_{10} n.$4

This refinement is the mechanism that prevents badly connected substructures from being frozen into the aggregate graph (Traag et al., 2018).

The paper proves that after each Leiden iteration, the resulting partition is $f(n)=\log_{10} n.$5-connected. It also proves stronger properties along the convergence path: after each iteration, communities are $f(n)=\log_{10} n.$6-separated; after a stable iteration, communities are subpartition $f(n)=\log_{10} n.$7-dense and all nodes are node optimal; asymptotically, communities are uniformly $f(n)=\log_{10} n.$8-dense and subset optimal (Traag et al., 2018). The convergence statement is

$f(n)=\log_{10} n.$9

for some $f(n)$0, because Leiden only accepts strictly positive quality improvements (Traag et al., 2018).

Empirically, the paper reports that Louvain can return substantial fractions of badly connected communities. In the first iteration, the fraction is about $f(n)$1 on Amazon, about $f(n)$2 on DBLP, and about $f(n)$3 on Web UK; disconnected communities were usually around $f(n)$4, but more than $f(n)$5 on the Web of Science network (Traag et al., 2018). When iterated, the problem can worsen; on DBLP, disconnected communities reached $f(n)$6, and on Amazon the paper reports $f(n)$7 disconnected but $f(n)$8 badly connected communities after multiple iterations (Traag et al., 2018). The central misconception addressed by this line of work is that connectivity problems are rare pathologies of modularity; the paper argues that the defect is procedural, not reducible to the resolution limit, and can arise even when CPM rather than modularity is used (Traag et al., 2018).

3. Empirical evidence that standard methods often violate WCC

The empirical critique is broadened in “Well-Connected Communities in Real-World and Synthetic Networks” (Park et al., 2023). That paper examines five methods optimizing different criteria: Leiden optimizing the Constant Potts Model, Leiden optimizing modularity, Iterative K-Core Clustering with $f(n)$9, Infomap, and Markov Clustering. Its headline result is that all five methods produce, to varying extents, communities that fail even a mild requirement for well connectedness (Park et al., 2023).

The observed failure modes are method-dependent. For Leiden-CPM, as the resolution parameter $|X| \ge r \times |A| \times |B|,$0 decreases, clusters get larger and node coverage increases, but the number of poorly connected clusters also increases. For Leiden-modularity, the behavior is similar to very low-resolution CPM. On the Open Citations network,

$|X| \ge r \times |A| \times |B|,$1

while still covering $|X| \ge r \times |A| \times |B|,$2 of nodes (Park et al., 2023). IKC is more conservative and yields lower node coverage, but most clusters are well connected. Infomap produces many well-connected clusters on some networks, but also some disconnected clusters; on orkut, up to $|X| \ge r \times |A| \times |B|,$3 of clusters were disconnected. MCL scaled only to the smallest network analyzed and still produced some disconnected clusters (Park et al., 2023).

The central empirical pattern is a tradeoff between node coverage and connectivity. Parameter settings that maximize coverage tend to produce larger clusters, but many of those clusters are weakly connected or have tiny edge cuts. Parameter settings that favor well-connected clusters tend to produce smaller clusters and lower coverage (Park et al., 2023). This tradeoff is especially visible for Leiden: high CPM resolution yields smaller clusters, lower coverage, and fewer small cuts; low CPM resolution or modularity yields higher coverage, but many small min cuts (Park et al., 2023).

The broader interpretation advanced there is that the expectation that communities are well connected is not automatically satisfied by standard community detection methods, and that some networks may not be fully covered by communities that are both large and well connected (Park et al., 2023). This suggests a more restrictive notion of “clusterability,” where only parts of a network exhibit community structure under a well-connectedness requirement.

4. Post-processing methods: CM and WCC

A practical response to poorly connected communities is min-cut-based post-processing. “Well-Connected Communities in Real-World and Synthetic Networks” introduces the Connectivity Modifier (CM), a heuristic remediation pipeline rather than a hard constraint built into the original objective (Park et al., 2023). CM transforms an existing clustering so that all retained clusters satisfy two conditions: size at least $|X| \ge r \times |A| \times |B|,$4, and minimum cut size greater than $|X| \ge r \times |A| \times |B|,$5, with default parameters

$|X| \ge r \times |A| \times |B|,$6

Its four-stage process is: cluster the input network with a chosen method; filter out small clusters of size $|X| \ge r \times |A| \times |B|,$7 and trees; for each remaining cluster, remove nodes of degree at most $|X| \ge r \times |A| \times |B|,$8 until none remain, compute a minimum cut using VieCut, split if the min cut is $|X| \ge r \times |A| \times |B|,$9, re-cluster the components, and recurse; then post-filter resulting clusters of size $\mincut(C) \ge r \times (n-1).$0 (Park et al., 2023).

A simplified variant appears in “Improved Community Detection using Stochastic Block Models” (Park et al., 2 Feb 2025), where WCC is presented as a simple post-processing technique for SBM outputs. The motivation is that SBM inference optimizes description length rather than internal edge-connectivity, so SBM clusters can be weakly connected or internally disconnected (Park et al., 2 Feb 2025). WCC processes each cluster independently, computes its minimum edge cut using VieCut, compares that cut to $\mincut(C) \ge r \times (n-1).$1, and if the cluster is not well connected, removes the minimum edge cut, thereby splitting the cluster into two smaller clusters; the same procedure is then applied recursively until every cluster is well connected (Park et al., 2 Feb 2025). In contrast to CM, WCC does not re-cluster after splitting and performs no removal of small clusters (Park et al., 2 Feb 2025).

The relationship among the three repair strategies discussed in the SBM literature is concise.

Method	Operation	Distinctive feature
CC	Replace each internally disconnected cluster by its connected components	Fixes only disconnected clusters
WCC	Recursively remove minimum edge cuts until every cluster passes the threshold	No re-clustering after splitting
CM	Recursively remove small cuts and then re-run the original clustering method on each part	More aggressive reclustering

This placement of WCC between CC and CM is explicit in the SBM study: WCC is more refined than CC because it also fixes poorly connected but still connected clusters, and less aggressive than CM because it never reclusters after splitting (Park et al., 2 Feb 2025).

5. WCC in stochastic block model clustering

The SBM literature gives WCC a specific operational role: repairing outputs whose statistical optimality under description length does not imply internal connectivity. “Improved Community Detection using Stochastic Block Models” reports that on 120 real-world networks, selected SBM clusterings frequently produced disconnected clusters, and that on the studied large networks, most of the clusters were internally disconnected for selected SBM outputs (Park et al., 2 Feb 2025). The paper attributes this to the description-length objective

$\mincut(C) \ge r \times (n-1).$2

arguing that the $\mincut(C) \ge r \times (n-1).$3 term strongly penalizes increasing the number of blocks, so the optimizer can prefer a clustering with disconnected pieces bundled together rather than splitting them into separate connected clusters (Park et al., 2 Feb 2025).

On synthetic networks with ground truth, that study reports clustering accuracy using NMI, ARI, AGRI, and RMI, and concludes that WCC is generally the best post-processing treatment for SBM. Across the heatmaps and comparisons, WCC is mostly neutral or beneficial for NMI and AGRI, and neutral to beneficial for ARI and RMI; CC shows similar but usually weaker trends, while CM is mixed (Park et al., 2 Feb 2025). WCC also remains computationally cheap relative to SBM inference: on the 13.99-million-node CEN network, the degree-corrected SBM took 38.7 hours, while WCC post-processing took 1.4 hours (Park et al., 2 Feb 2025).

The later paper “Using Stochastic Block Models for Community Detection: The issue of edge-connectivity” extends this analysis to additional SBM software and nested SBMs (Vu-Le et al., 5 Aug 2025). It reports that disconnected clusters are common across graph-tool, PySBM, flat SBMs, nested SBMs, and multiple inference algorithms (Vu-Le et al., 5 Aug 2025). On 74 non-bipartite real-world networks, mean fractions of disconnected clusters are reported as $\mincut(C) \ge r \times (n-1).$4 for NDC-Flat, $\mincut(C) \ge r \times (n-1).$5 for DC-Flat, $\mincut(C) \ge r \times (n-1).$6 for PP-Flat, $\mincut(C) \ge r \times (n-1).$7 for NDC-Nested, and $\mincut(C) \ge r \times (n-1).$8 for DC-Nested (Vu-Le et al., 5 Aug 2025). The paper further states that WCC often improves clustering accuracy, especially for dense ground-truth clusters, though it can reduce accuracy when the true community structure contains many sparse large clusters (Vu-Le et al., 5 Aug 2025).

A plausible implication is that WCC has become, in the SBM setting, an explicit correction for a mismatch between probabilistic compression objectives and graph-theoretic cohesion criteria.

6. Scalability, optimization, and terminological ambiguity

The post-processing view of WCC has recently been pushed toward web-scale graph analytics. “On the Optimization of Methods for Establishing Well-Connected Communities” presents optimized parallel implementations of WCC and CM in HPE Chapel and integrates them into Arkouda/Arachne (Dindoost et al., 29 Aug 2025). In that formulation, WCC has two major stages: Connected Component Refinement and recursive well-connectedness testing. For each cluster $\mincut(C) \ge r \times (n-1).$9, the induced subgraph is

$f(n)=\log_{10}n$0

The recursive test computes a global minimum cut and compares it to a user-defined criterion $f(n)=\log_{10}n$1, with common choices

$f(n)=\log_{10}n$2

and experiments using $f(n)=\log_{10}n$3 (Dindoost et al., 29 Aug 2025).

That optimization paper emphasizes that WCC only performs recursive subdivision and never merging, and that the new implementations enable well-connected community detection on massive graphs with more than 2 billion edges (Dindoost et al., 29 Aug 2025). Its reported hardware setup uses 128 cores and 512 GB RAM, with $f(n)=\log_{10}n$4 and Leiden-CPM as the community detection algorithm for CM (Dindoost et al., 29 Aug 2025). Example runtimes include Bitcoin, Livejournal, and CEN, where WCC-Chapel is substantially faster than the baseline, and the Chapel implementation succeeds on Open-Alex and Open-Citations-v2 where the baseline often failed due to OOM or segmentation fault (Dindoost et al., 29 Aug 2025).

The acronym WCC is, however, overloaded. In “Shaping Communities out of Triangles” and “Distributed Community Detection with the WCC Metric,” WCC stands for Weighted Community Clustering, a triangle-based community-quality metric rather than a min-cut-based post-processor (Prat-Pérez et al., 2012, Saltz et al., 2014). Its vertex-level definition is

$f(n)=\log_{10}n$5

and the partition objective is

$f(n)=\log_{10}n$6

(Saltz et al., 2014). In that tradition, WCC is a triangle-based measure designed to reward communities that are rich in triangles internally and sparse in triangle participation outside.

Outside graph mining, the same acronym can denote Weakly Cool-Core clusters in galaxy-cluster thermodynamics (Hernández-Martínez et al., 21 Jul 2025). The phrase “well-connected cluster” also appears in the binary perceptron literature to describe a subdominant connected component of the solution space under Hamming-distance-1 adjacency, with large diameter and algorithmic accessibility (Abbe et al., 2021). These usages are conceptually unrelated to community detection, and the overlap is terminological rather than methodological.

Taken together, the literature presents WCC in the graph-clustering sense as an increasingly explicit cohesion requirement. The common theme is that a meaningful cluster should resist fragmentation by small cuts, and that standard optimization objectives do not in themselves guarantee this property. The strongest recurring conclusion is therefore not that there is a single canonical WCC definition, but that well-connectedness has become a distinct evaluative and algorithmic axis alongside density, conductance, modularity, and description length (Zhu et al., 2013, Traag et al., 2018, Park et al., 2023, Park et al., 2 Feb 2025).