Conductance-Based Community Search (CCS)

• CCS is a graph mining approach that identifies connected subgraphs containing a query vertex by minimizing conductance, balancing internal cohesion and external sparsity.
• It redefines community search by ensuring communities are not only dense internally but also well separated from the surrounding graph, addressing limitations in prior methods.
• The SCCS heuristic employs local sampling, clique-based seeding, and iterative expansion-verification to offer scalable and effective community detection even on billion-scale graphs.

Conductance-Based Community Search (CCS) is a query-centric graph mining problem in which, given an undirected graph $G(V,E)$ and a query vertex $q$, the task is to find a connected subgraph containing $q$ whose conductance is minimum among all connected subgraphs containing $q$. In this formulation, community quality is defined jointly by internal cohesiveness and external sparsity, rather than by internal density alone. The named CCS problem was introduced together with a scalable heuristic, SCCS, in work that also proves CCS is NP-hard and reports effectiveness, efficiency, and scalability on real-world and synthetic datasets, including a one billion-scale graph (Lin et al., 2 Aug 2025).

1. Historical placement within community search

CCS emerged against a community-search literature that had largely emphasized internal cohesiveness. The dominant formulations surveyed in earlier work were based on $k$-core, $k$-truss, $k$-clique and its variants, $k$-ECC, and several “other metrics” such as local modularity, query-biased density, PageRank, and neighborhood expansion. That survey does not define conductance, does not introduce a conductance-based objective, and does not present a CCS algorithm, index, or framework (Fang et al., 2019).

The motivation for CCS is the claim that many prior community-search methods ensure the returned subgraph is “tight” inside but often ignore whether the community is well separated from the rest of the graph. In the CCS formulation, this omission can lead to communities with many external edges, larger-than-desired sizes, or unstable parameter dependence. CCS therefore redefines the search target from “a dense connected subgraph containing $q$” to “a connected, query-containing subgraph that is both internally cohesive and externally sparse” (Lin et al., 2 Aug 2025).

This suggests CCS is best understood not as a replacement for the whole community-search literature, but as a later reformulation that makes boundary quality central to the query problem. Its distinguishing feature is that the output must simultaneously satisfy seed inclusion, connectedness, and minimum conductance.

2. Conductance as the governing objective

For a vertex set $S \subseteq V$, CCS uses conductance

$q$0

where $q$1 is the set of edges with one endpoint in $q$2 and the other in $q$3, $q$4, and $q$5. The formulation also notes the symmetry $q$6 (Lin et al., 2 Aug 2025).

The optimization problem is

$q$7

The feasibility conditions are explicit: $q$8, $q$9 is connected, and $q$0 is minimized among all feasible connected subgraphs containing $q$1 (Lin et al., 2 Aug 2025).

On a much smaller sampled subgraph $q$2, the denominator is often simplified in practice to $q$3, giving the subgraph conductance

$q$4

Within SCCS, local optimization is driven by the quality score

$q$5

with $q$6 and $q$7. This quantity is described as monotonic with conductance in the sense that increasing $q$8 corresponds to improving community quality through more internal density and less external leakage (Lin et al., 2 Aug 2025).

Conductance-oriented local community detection predates the formal naming of CCS. Earlier work treated conductance as “a popular objective function used in many algorithms for local community detection,” studied a continuous relaxation of conductance, and introduced $q$9-conductance together with the EMc and PGDc local optimizers (Laarhoven et al., 2016). A generative-model approach later showed that, in the limit, one local approximation under a degree-corrected stochastic block model corresponds to conductance, giving a probabilistic interpretation of a criterion often treated as an ad hoc graph-cut objective (Laarhoven, 2018). These results place CCS within a broader conductance-based local-search lineage, while the formal CCS problem adds the explicit minimum-conductance query objective.

3. Computational hardness and exact optimization

CCS is NP-hard. The hardness result is established by reduction from the $q$0-clique problem, which is NP-complete (Lin et al., 2 Aug 2025).

The reduction is motivated by the relationship between internal edge volume and boundary size. A $q$1-clique maximizes internal edges, with $q$2. For a connected set $q$3, conductance becomes smallest when the cut size $q$4 is minimized and the internal edge volume is maximized. Under the constructed decision setting, an optimal CCS solution corresponds to a $q$5-clique if and only if one exists, and conversely a $q$6-clique containing $q$7 yields a candidate with minimum conductance. Exact CCS would therefore solve $q$8-clique (Lin et al., 2 Aug 2025).

The hardness result has two immediate algorithmic consequences. First, it rules out an expectation that the exact minimum-conductance community around a query vertex should be efficiently computable in general. Second, it explains why practical CCS work concentrates on heuristics that restrict the search to a local region and then apply local optimization there. SCCS is the principal realization of that strategy.

4. SCCS: subgraph-conductance-based community search

SCCS is a four-stage heuristic designed to make CCS practical at large scale. The first stage is graph reduction via local sampling. Starting from $q$9, SCCS performs BFS-like exploration, but stops when the sampled subgraph is at least a minimum size $k$0 and depth exceeds $k$1, or when the sampled subgraph reaches a maximum size $k$2. The resulting sampled subgraph $k$3 is intended to capture the local neighborhood in which the target community is likely to lie (Lin et al., 2 Aug 2025).

The second stage is seeding. Within the sampled subgraph, SCCS chooses the largest maximal clique containing $k$4 as the initial community. The rationale is explicit: a clique provides very strong internal cohesiveness, so expansion starts from a tight core rather than from a loose neighborhood (Lin et al., 2 Aug 2025).

The third stage is expansion. SCCS adds neighboring vertices that improve the community-quality score. For a set $k$5 of added vertices,

$k$6

The update is written so that only edges involving $k$7 must be changed: $k$8 where

$k$9

and

$k$0

A notable heuristic is that SCCS may add multiple vertices at a time, up to a threshold count, rather than adding only one vertex greedily. This is intended to reduce premature convergence to local optima (Lin et al., 2 Aug 2025).

The fourth stage is verification. Because conductance is non-monotonic, some previously added vertices may later become harmful. SCCS therefore forms the boundary set

$k$1

and tries to remove boundary vertices that increase quality $k$2 while preserving connectivity. The removal gain is

$k$3

with

$k$4

Expansion and verification are repeated until no more changes occur (Lin et al., 2 Aug 2025).

The practical heuristics of SCCS are therefore local sampling, clique-based seeding, multi-vertex expansion, gain-based expansion, boundary pruning subject to connectivity, and an iterative expansion-verification loop. Together they operationalize CCS without attempting exact global optimization.

5. Complexity, scalability, and empirical behavior

The central scalability claim for SCCS is that the expensive operations are confined to the sampled subgraph of size $k$5, rather than to the full graph. The complexity components are: sampling $k$6, maximum clique search on the sampled subgraph with worst-case $k$7, expansion $k$8, and verification $k$9, where $k$0 is the number of boundary vertices. The overall time is

$k$1

where $k$2 is the number of outer expansion-verification iterations. By comparison, the PPRCS baseline has worst-case time

$k$3

and space $k$4 (Lin et al., 2 Aug 2025).

The reported experiments use six real-world graphs—Amazon, DBLP, Youtube, LiveJournal, Orkut, and Friendster—together with synthetic LFR benchmarks. The baselines are CSM, TCP, SM, LM, HK-Relax, and PPRCS. The principal empirical finding is that SCCS achieves the best F1-score overall. It usually returns communities that are smaller, more consistent with ground-truth community sizes, and better separated from the rest of the graph. Cohesive-subgraph methods such as CSM and TCP often output communities that are too large and have poor conductance, while diffusion methods such as HK-Relax and PPRCS improve conductance but still often return communities that are too large or less aligned with ground truth. SCCS is reported to best balance precision and recall (Lin et al., 2 Aug 2025).

The conductance results are more nuanced than a simple minimum-conductance narrative. SCCS generally has low conductance and is often better than most competitors, but its conductance is not always the absolute minimum on every dataset. At the same time, it preserves much better F1-score than diffusion-heavy baselines. This is an important correction to the misconception that the lowest observed conductance automatically yields the best query-community recovery (Lin et al., 2 Aug 2025).

The efficiency results illustrate the method’s scale. On Friendster, with 1.8B edges, PPRCS is about $k$5 s, SCCS about $k$6 s, and many baselines fail to finish within $k$7 hours. Runtime stays stable as graph size grows, consistent with the sampling strategy. On LFR graphs, runtime and F1-score remain stable as $k$8 increases, runtime increases with average degree $k$9, and performance degrades as the mixing parameter $k$0 increases, although SCCS remains stronger than PPRCS. Parameter studies report that larger count lowers conductance but can reduce F1-score after a point by growing communities too much, while larger dp improves sampled-subgraph coverage but increases time and eventually saturates. The recommended settings are count = 2 and dp = 3 (Lin et al., 2 Aug 2025).

6. Extensions, contrasts, and limitations

CCS belongs to a broader family of conductance-oriented local search methods. In streaming graphs, SCDAC formulates streaming local community detection as finding the target community containing a small set of query nodes from a single-pass edge stream under semi-streaming memory limits. It samples a local subgraph around the query nodes and then extracts the target community using approximate conductance

$k$1

where $k$2 comes from exact degree counts recorded during streaming. This is a conductance-based local community search method under one-pass constraints rather than a named CCS framework, but it shows that conductance-based query search had already been adapted to streaming settings (Yang et al., 2021).

Later extensions move beyond purely structural conductance. NCSAC introduces attribute-augmented conductance

$k$3

where $k$4 is topology-based conductance and $k$5 is attribute-based conductance. It first extracts a coarse candidate community using this hybrid objective and then refines it through reinforcement learning. A plausible implication is that CCS can be generalized from structure-only boundary sparsity to hybrid structural-attribute objectives without abandoning conductance as the organizing principle (Lin et al., 5 Nov 2025).

At the same time, not all modern community-search systems are conductance-based. Query Driven-GNN reformulates community search and attributed community search as binary vertex classification, trains with binary cross entropy, and uses constrained BFS for final extraction rather than a cut-based objective (Jiang et al., 2021). In multiplex networks, CS-MLGCN combines query-dependent local structure, global graph embeddings, and attention across views, again treating the task as supervised node classification followed by BFS connectivity enforcement, not as conductance minimization (Behrouz et al., 2022). These contrasts delimit CCS as one methodological family within query-centric community search, rather than the default form of the problem.

A further limitation concerns interpretation of “community quality.” In online social networks, the CHASE framework reports that there is no clear correlation between structural and psychological cohesiveness, and that no evaluated community-search algorithm effectively identifies psychologically cohesive communities in that domain. Conductance is explicitly discussed there as part of the family of structural community-quality criteria, even though the framework’s primary reported structural metrics are diameter, size, and minimum degree. This suggests that CCS should be understood as a rigorous structural model of community quality, not as a universal proxy for human group cohesion (Zhao et al., 28 Apr 2025).

Taken together, these developments define the current scope of CCS. It is a formally specified, conductance-centered query problem with clear optimization constraints, NP-hard exact form, and scalable heuristic solvers. Its strongest setting is structural local community discovery around a query vertex. Its principal limitations arise when community quality depends on signals—attributes, dynamics, or social-behavioral evidence—not captured by structural conductance alone.