PaperlayerView exact revision on arXiv

Source-linked AI summary

A Comprehensive Survey on Community Detection with Deep Learning

Xing Su, Shan Xue, Fanzhen Liu, Jia Wu, Jian Yang, Chuan Zhou, Wenbin Hu, Cecile Paris, Surya Nepal, Di Jin, Quan Z. Sheng, Philip S. Yu

arXiv:2105.12584v2cs.SIcs.AIcs.LG

TL;DR

Community detection is important for analyzing network structure, but traditional methods are less applicable to real-world networks with complex topology and high-dimensional information. This survey proposes a taxonomy and comprehensive review of deep learning-based methods, together with resources and applications. It concludes that these methods improve the effectiveness, efficiency, robustness, and applicability of community detection while identifying unresolved challenges.

  • Problem

    Traditional methods are less applicable to real-world networks containing complex topology and high-dimensional nonlinear features, motivating a comprehensive deep learning-focused review.

  • Method

    The survey proposes a taxonomy and reviews deep learning methods, benchmark resources, implementations, applications, and future directions.

  • Results

    The survey reports that deep learning-based methods significantly increase community detection effectiveness, efficiency, robustness, and applicability.

  • Takeaways & Limitations

    The taxonomy and collected resources support researchers and practitioners in understanding, evaluating, and applying deep learning-based community detection.

Abstract

from arXiv · show

A community reveals the features and connections of its members that are different from those in other communities in a network. Detecting communities is of great significance in network analysis. Despite the classical spectral clustering and statistical inference methods, we notice a significant development of deep learning techniques for community detection in recent years with their advantages in handling high dimensional network data. Hence, a comprehensive overview of community detection's latest progress through deep learning is timely to academics and practitioners. This survey devises and proposes a new taxonomy covering different state-of-the-art methods, including deep learning-based models upon deep neural networks, deep nonnegative matrix factorization and deep sparse filtering. The main category, i.e., deep neural networks, is further divided into convolutional networks, graph attention networks, generative adversarial networks and autoencoders. The survey also summarizes the popular benchmark data sets, evaluation metrics, and open-source implementations to address experimentation settings. We then discuss the practical applications of community detection in various domains and point to implementation scenarios. Finally, we outline future directions by suggesting challenging topics in this fast-growing deep learning field.

I. INTRODUCTION

Community detection identifies meaningful groups in networks and has growing practical significance across social, citation, biological, and brain-network applications. This survey addresses the limits of traditional methods by organizing deep learning approaches, resources, applications, and future directions.

  • Motivation: Community detection reveals network structure and supports applications including targeted promotion, research-trend analysis, biological-function analysis, and brain-region studies.Examples span social, citation, metabolic, protein-interaction, and brain networks.
  • Motivation: Traditional spectral clustering and statistical inference methods often target small, simple networks, while real-world networks contain complex topology and high-dimensional nonlinear features.The survey also notes that traditional methods can incur expensive computational costs in such settings.
  • Deep Learning Approach: Deep learning methods learn nonlinear network properties, produce lower-dimensional embeddings that preserve structure, and use varied information for community detection.Their inputs can include topology and attributes represented through adjacency, signed-adjacency, modularity, and node-attribute matrices.
  • Survey Contributions: The survey proposes a systematic taxonomy, reviews and compares representative works, and introduces real-world community-detection applications.The taxonomy covers traditional methods and deep learning categories, with the article organized around detailed reviews of each category.
  • Survey Contributions: The survey collects benchmark datasets, evaluation metrics, technique implementations, and high-impact references to support experimentation by researchers and practitioners.It also identifies supporting materials and implementation resources as part of the survey scope.
  • Future Directions: The survey discusses current limitations, critical challenges, and open opportunities for future research in deep learning-based community detection.Future directions are presented alongside practical applications and implementation scenarios.

IV. A TAXONOMY OF COMMUNITY DETECTION WITH DEEP LEARNING

The survey organizes deep community detection into six model-based categories and reviews CNN-based approaches that preprocess graphs into image-formatted data or recover representations from incomplete topology.

  • The taxonomy contains six categories: CNNs and GCNs, GATs, GANs, AEs, DNMF, and DSF.Autoencoder methods include stacked, sparse, denoising, graph convolutional, graph attention, and variational autoencoders.
  • CNN-based Community Detection: CNN-based methods preprocess node or edge graphs into image data before convolutional mapping and classification.Node-focused models predict k community classes, whereas edge-focused models classify edges as inner or inter.
  • CNN-based Community Detection: Supervised CNNs for topologically incomplete networks use convolution and pooling layers to recover latent features from rudimentary inputs.The final fully connected layer updates each node’s community assignment.
  • CNN-based Community Detection: CNN experiments on topologically incomplete networks achieve around 80% community detection accuracy with 10% labeled nodes.The reported result associates multi-hop, high-order neighborhood representations with improved accuracy.
  • CNN-based Community Detection: Sparse CNN variants target large-scale social networks by operating on nonzero adjacency elements to reduce computational operations.ComNet-R instead classifies edges, removes inter-community edges, and merges preliminary communities using local modularity.

B. GCN-based Community Detection

GCN-based community detection aggregates neighborhood information for classification or unsupervised clustering, while specialized filters and architectures address smoothing, graph signals, and community-specific structure.

  • GCN-based Community Detection: GCN methods comprise supervised or semi-supervised community classification and unsupervised community clustering through network representations.Classification is constrained by limited real-world labels, whereas representations can be optimized through reconstruction and clustering objectives.
  • GCN-based Community Detection: LGNN integrates non-backtracking message passing with belief propagation and SBM ideas to learn node features and classify community membership.Its softmax outputs conditional community probabilities and training minimizes permutation-invariant cross-entropy.
  • GCN-based Community Detection: MRFasGCN adds a network-specific Markov Random Field convolutional layer, while SGCN samples structural centers without prior labels.MRFasGCN refines coarse GCN results; SGCN combines topology and node attributes during membership training.
  • GCN-based Community Detection: GCN frameworks support overlapping-community detection through Bernoulli–Poisson affiliation vectors and threshold-based removal of weak memberships.NOCD minimizes the Bernoulli–Poisson model’s negative log-likelihood using a two-layer GCN.
  • GCN-based Community Detection: Specialized GCN architectures address over-smoothing, redundant relation representations, and global cluster structure in community-oriented embeddings.GCLN uses a ladder-shaped design, IPGDN disentangles independent latent features, and SENet introduces a spectral-clustering loss.
  • GCN-based Community Detection: Adaptive filters smooth graph signals and select neighborhood scales, while CayleyNets uses narrow-band high-order filtering because low frequencies contain community information.AGE applies similarity measurement and stacked Laplacian smoothing; larger k improves AGC filtering performance.
  • GAT-based Community Detection: GAT-based detection learns neighborhood attention coefficients so similar-node correlations and multiple relation types inform community representations.The approach supports attributed, multiplex, and heterogeneous networks by filtering spatial relations or integrating metapaths.

2 KD−1 2 H(3)), (8)

Deep learning community-detection methods use graph attention, infomax, heterogeneous-network modeling, and adversarial training to learn community-aware representations from topology, attributes, and relations.

  • Graph Attention Networks: Graph attention aggregates neighborhood features with trainable weights, adaptively emphasizing nodes and relation types relevant to community membership.These mechanisms can filter less significant relations and integrate metapaths in attributed, multiplex, and heterogeneous networks.
  • Graph Attention Networks: DMGI embeds multiplex relation types separately, applies contrastive learning against corrupted networks, and uses consensus regularization to integrate embeddings.
  • Graph Attention Networks: HDMI extends DMGI by capturing intrinsic attribute signals and their interactions, then fusing multiplex-layer embeddings through semantic attention on community memberships.
  • Graph Attention Networks: MAGNN uses multi-informative semantic metapaths and graph-attention layers to distinguish heterogeneous structures and generate node attributes from semantic information.
  • Graph Attention Networks: CP-GNN learns embeddings from non-predefined context paths, using attention to distinguish relationship importance and capture high-order relationships.Meaningful predefined metapaths require substantial domain knowledge.
  • Generative Adversarial Networks: GAN-based methods competitively train generators and discriminators using node embeddings, local topology, or communities as real samples.Clique-level GANs infer membership, node-level approaches learn competitive representations, and discriminator-based flows directly distinguish communities.
  • Generative Adversarial Networks: JANE jointly uses topology and node attributes, encoding embeddings with multi-head self-attention and adding Gaussian noise to fake features for adversarial competition.

VIII. AUTOENCODER-BASED COMMUNITY DETECTION

Autoencoders provide an unsupervised framework for community detection by encoding network structure and attributes into low-dimensional representations and reconstructing the input.

  • Autoencoders are widely used in unsupervised community detection, including stacked, sparse, denoising, convolutional, and variational forms.
  • An encoder maps high-dimensional structure and attributes into latent features, while a decoder reconstructs the network or attributes from those representations.
  • The reconstruction loss maximizes likelihood between source data and decoded data, allowing latent representations to preserve preferred network information.

A. Stacked AE-based Community Detection

Stacked and sparse autoencoders extend community detection to multi-level, dynamic, attributed, signed, transferred, and large-scale sparse networks by combining reconstruction with structural or community constraints.

  • Stacked AE-based Community Detection: Stacked autoencoders represent multi-level and dynamic information by assigning different input types to individual encoders within deep hidden layers.
  • Stacked AE-based Community Detection: Semi-DRN combines modularity-based nonlinear representations, k-means clustering, and pairwise constraints indicating whether node pairs share a community.
  • Stacked AE-based Community Detection: Evolutionary autoencoding adds temporal smoothness so previous community structures guide detection at the current time step.
  • Stacked AE-based Community Detection: DANE uses two autoencoder branches to learn nonlinear network structure and node attributes collaboratively through topology- and attribute-based proximity.
  • Stacked AE-based Community Detection: Transfer-CDDTA balances source and target stacked autoencoders with KL divergence and reconstruction losses to learn embeddings across domains.
  • Sparse AE-based Community Detection: Sparse autoencoders add hidden-layer sparsity penalties to address computational difficulties in sparse real-world networks.GraphEncoder improves clustering efficiency for large-scale networks, while WCD extends sparse autoencoding to weighted networks.
  • Sparse AE-based Community Detection: DFuzzy applies stacked sparse autoencoders to head nodes for overlapping and disjoint communities, reporting 63% higher Q and 34% higher CON than non-deep-learning baselines.

C. Denoising AE-based Community Detection

Denoising and variational autoencoders improve community representations by handling corrupted inputs, uncertainty, sparsity, and higher-order structure, while graph-convolutional combinations preserve structural information.

  • Denoising AE-based Community Detection: Denoising autoencoders improve robustness by minimizing reconstruction loss between original inputs and features learned from corrupted inputs.
  • Denoising AE-based Community Detection: DNGR uses stacked denoising encoders with probabilistic co-occurrence and shifted positive-PMI matrices from random walks to capture local structural information.
  • Denoising AE-based Community Detection: MGAE denoises graph attributes and structures by repeatedly corrupting attributes through random removal and optimizing a regularized objective.
  • Graph-Convolutional and Attention-Based AE: GCN-AE models combine neighborhood aggregation with autoencoder reconstruction, helping clarify community boundaries while incorporating structural information.Examples include SDCN, GUCD, and multi-view O2MAC architectures.
  • Graph-Convolutional and Attention-Based AE: GAT-based autoencoders rank neighborhood-node importance and support high-order, multi-view, and cross-graph community clustering.
  • VAE-based Community Detection: VAEs represent graph features as latent distributions, enabling uncertainty-aware community inference and deeper nonlinear relationship modeling.
  • VAE-based Community Detection: VGAECD uses a Gaussian mixture model to model community assignments and higher-order structural features, with ELBO combining reconstruction and clustering terms.
  • VAE-based Community Detection: VGAECD-OPT addresses VGAECD’s sub-optimal detection caused by reconstruction loss receiving greater training emphasis than community loss.

IX. DEEP NONNEGATIVE MATRIX FACTORIZATION-BASED COMMUNITY DETECTION

Deep nonnegative matrix factorization methods represent community membership through layered, nonnegative network mappings, while deep sparse filtering learns sparse community features from graph representations. The surveyed methods incorporate reconstruction, graph structure, modularity, similarity, or sparsity objectives to improve community discovery.

  • NMF factorizes adjacency matrix A into nonnegative matrices, with U mapping nodes to community membership space and P representing node-community probabilities.
  • DANMF extends conventional NMF with an autoencoder framework for network reconstruction through hierarchical mappings.
  • Graph regularization combines reconstruction losses with topology-based similarity, while an added sparsity constraint further restricts DNMF-based community detection.
  • MDNMF addresses the relatively high computational cost of matrix factorizations by applying modularity directly within a multilayer deep learning structure.
  • DSFCD uses network representation, community feature mapping, and community discovery phases, selecting among adjacency, modularity, and similarity-based inputs.
  • DSFCD discovers communities with higher accuracy than SF, while KL divergence models similarity to the most similar node and sparse penalties optimize representations.

C. Open-source Implementations

The survey compiles experimentation resources and illustrates community detection applications across recommendation, biochemistry, social networks, deception, and community search. It also identifies unresolved challenges involving unlabeled data, topology, hierarchy, and multilayer networks.

  • Open-source Implementations: Open-source implementations are organized by tool, category, repository link, implementation tool, and deep learning model, with most using Python 3.x frameworks.
  • Practical Applications: Community detection supports graph-based recommendation systems because community members may share interests and preferences.
  • Practical Applications: Biochemical applications identify protein complexes, functional compounds, disease-related factors, and relationships between tumor communities and survival rates.
  • Practical Applications: Online social-network applications identify communities, support preference-based personalization, address privacy, and identify criminals from social behaviors.
  • Future Directions: Future directions include detecting an unknown number of communities, preserving community hierarchy, and learning from multilayer network information.

E. Heterogeneous Networks

The survey identifies heterogeneous, signed, incomplete, dynamic, cross-domain, multiview, and large-scale networks as challenging settings for deep community detection. Proposed directions include metapath modeling, heterophily handling, information recovery, adaptive transfer, integrated multiview learning, and scalable computation.

  • Heterogeneous Networks: Heterogeneous networks require models that capture complex structural and semantic information across different entity types and relationships.
  • Heterogeneous Networks: Metapaths aggregate first-order information across node types, but selecting the most meaningful metapaths remains an open problem.
  • Heterophily: Most methods rely on homophily, motivating deep learning approaches that exploit heterophily for community detection.
  • Incomplete Networks: Incomplete topology and isolated subgraphs limit neighborhood-aggregation methods, motivating information-recovery mechanisms for accurate detection.
  • Signed Networks: Signed-network methods must represent both positive and negative ties because existing unsigned-network approaches are not directly applicable.
  • Large-scale Networks: Large-scale networks challenge low-rank dimension reduction and expensive distributed computing, creating demand for frameworks that improve precision and speed.

APPENDIX A SUMMARIZED TECHNIQUES OF DEEP LEARNING-BASED

The appendix organizes surveyed deep learning community detection methods and the benchmark networks used to evaluate them. It covers CNN, GCN, GAT, GAN, and autoencoder methods alongside citation, social, webpage, product, and synthetic network datasets.

  • Summarized Techniques: Tables IV–VIII summarize CNN-, GCN-, GAT-, GAN-, and AE-based community detection methods, including their inputs and ground-truth availability.
  • Benchmark Data Sets: Citation and co-authorship benchmarks include Citeseer, Cora, Pubmed, and DBLP, with publications as nodes, citations or authorship links, and topic or venue communities.
  • Benchmark Data Sets: Online social-network datasets represent users and relationships, with communities defined by groups, profiles, interests, or other platform attributes.
  • Benchmark Data Sets: Traditional social benchmarks include Karate, Dolphin, Football, friendship, email, cellphone-call, and conference-attendee networks.
  • Benchmark Data Sets: Webpage and product co-purchasing datasets encode entities and links such as movies, actors, products, and frequently co-purchased items, with genre or product categories as communities.
  • Benchmark Data Sets: Synthetic benchmarks include GN networks with four equal communities and LFR networks modeling scale-free degree distributions, with extensions for overlapping, directed, and weighted networks.

APPENDIX C DETAILED DESCRIPTION: EVALUATION METRICS

The survey describes metrics for evaluating community detection with and without ground truth, covering agreement, classification, similarity, separability, and density.

  • Ground-truth-based metrics: NMI evaluates detected communities against ground truth using a normalized mutual-information score and has an overlapping-community variant.The survey also defines community-count and node-overlap quantities used in NMI-related calculations.
  • Ground-truth-based metrics: Accuracy measures node-level agreement between detected and ground-truth community labels using the Kronecker delta.The delta equals 1 for matching labels and 0 otherwise.
  • Ground-truth-based metrics: Precision measures detected-community membership against ground truth, while recall measures the percentage of each ground-truth community recovered.F1-score balances precision and recall and evaluates clustering results across the community set.
  • Ground-truth-based metrics: ARI measures the percentage of correctly divided nodes using true-positive, false-positive, true-negative, and false-negative assignments.Its calculation uses detected-community sizes, ground-truth community sizes, and their node intersections.
  • Ground-truth-free metrics: Modularity evaluates community strength without ground truth by comparing within-community edge structure with a degree-matched random-graph null model.Extended modularity evaluates overlapping community detection.
  • Structural metrics: Jaccard measures detected-community similarity to ground truth, conductance measures external-edge separability, and TPR measures within-community triad density.Conductance uses internal and boundary edge counts, whereas TPR focuses on triads contained in a detected community.

APPENDIX D DETAILED DESCRIPTION: OPEN-SOURCE IMPLEMENTATIONS

The survey catalogs open-source implementations across PyTorch, TensorFlow, and Matlab, covering graph neural networks, GANs, autoencoders, and graph clustering methods.

  • PyTorch implementations: PyTorch implementations include LGNN under GNN and GCN settings, NOCD for overlapping communities, and AGE for attributed graph embedding and node clustering.DMGI and HDMI examine unsupervised attributed multiplex-network graph clustering with GAT-related implementations.
  • TensorFlow implementations: TensorFlow implementations include AGC, CayleyNet, CommunityGAN, DIME, and DANE across graph convolution, spectral filtering, GAN, and stacked autoencoder approaches.The listed methods address attributed graphs, varied network types, and representation learning.
  • Matlab implementations: Matlab implementations include semi-DRN, DNE-SBP, and MGAE for stacked autoencoding, signed-network community detection, and marginalized graph autoencoding.These implementations support node-representation-based clustering in different network settings.
Loading 2105.12584v2…