PaperlayerView exact revision on arXiv

Source-linked AI summary

Network Community Detection: A Review and Visual Survey

Bisma S. Khan, Muaz A. Niazi

arXiv:1708.00977v1cs.SIcs.CCcs.DLcs.DMcs.DS

TL;DR

Community-detection literature is dispersed across disciplines, and few prior studies show how the field has advanced over time. This paper uses a visual scientometric survey and network analysis to examine its evolution, identifying Liu J as the most important researcher in the author collaboration network and Jiao LC as having the strongest citation burst.

  • Problem

    Community-detection research is spread across related disciplines, and only a few prior studies show how the field advanced over time.

    Source: §1 · pp. 1–2

  • Method

    The paper uses visual scientometric analysis and network techniques to identify emerging trends and examine the field’s structure and evolution.

    Source: §1 · p. 2, §3.2 · p. 11

  • Results

    Liu J is identified as the most important researcher by largest radii, while Jiao LC shows the strongest citation burst in the author collaboration network.

    Source: §4 · p. 12

  • Takeaways & Limitations

    The survey identifies emerging trends, structure, and evolution in community detection through analysis of central, landmark, and bursting nodes.

    Source: §1 · p. 2

    Initial-based author-name disambiguation insufficiently resolves ambiguity, and records with matching surnames and first initials may be merged as one author.

    Source: §4 · p. 13

Abstract

from arXiv · show

Community structure is an important area of research. It has received a considerable attention from the scientific community. Despite its importance, one of the key problems in locating information about community detection is the diverse spread of related articles across various disciplines. To the best of our knowledge, there is no current comprehensive review of recent literature which uses a scientometric analysis using complex networks analysis covering all relevant articles from the Web of Science (WoS). Here we present a visual survey of key literature using CiteSpace. The idea is to identify emerging trends besides using network techniques to examine the evolution of the domain. Towards that end, we identify the most influential, central, as well as active nodes using scientometric analyses. We examine authors, key articles, cited references, core subject categories, key journals, institutions, as well as countries. The exploration of the scientometric literature of the domain reveals that Yong Wang is a pivot node with the highest centrality. Additionally, we have observed that Mark Newman is the most highly cited author in the network. We have also identified that the journal, "Reviews of Modern Physics" has the strongest citation burst. In terms of cited documents, an article by Andrea Lancichinetti has the highest centrality score. We have also discovered that the origin of the key publications in this domain is from the United States. Whereas Scotland has the strongest and longest citation burst. Additionally, we have found that the categories of "Computer Science" and "Engineering" lead other categories based on frequency and centrality respectively.

1. Introduction

The paper addresses the difficulty of finding and understanding community-detection literature dispersed across disciplines, and presents a scientometric visual survey to trace the field’s structure and evolution.

1. Introduction

  • Only a few existing surveys show how community detection advanced over time, while relevant work is dispersed across disciplines, making the field’s history and developments difficult to locate.The introduction also frames community structure as important for understanding complex-network functionality.Source: §1 · pp. 1–2
  • Using Web of Science literature, CiteSpace, and Pajek, the study examines publication performance, temporal evolution, intellectual structure, and bibliographic networks.The authors describe the work as addressing the lack of a recent comprehensive scientometric review of community detection literature from WoS.Source: §1 · p. 2
  • The survey identifies emerging trends and domain evolution by examining central, landmark, and bursting nodes in the scientometric literature.Its visual-survey approach draws on the Cognitive Agent-based Computing framework.Source: §1 · p. 2
  • The analysis covers productive authors, key articles, cited references, journals, institutions, countries, subject categories, and structural patterns in coauthor, cited-document, and cited-journal clusters.The paper also outlines background, methodology, results and discussion, and conclusions.Source: §1 · p. 2

2. Background

The background introduces network and community concepts, then surveys modularity measures and families of community-detection techniques across static, overlapping, and dynamic settings.

2. Background

  • Communities are groups of nodes more densely connected internally than to the rest of a network, but no universal definition exists and large-scale detection is computationally intractable.Networks are represented as graphs with node and edge sets; adjacency and degree notions describe their connections.Source: §2 · pp. 2–3

2.2.2. Modularity Density

  • Modularity density D measures community structure using subgraph density, with higher D indicating a better partition; optimizing it is NP-hard.The measure is defined in terms of within-subset and between-subset links.Source: §2.2.2 · p. 5

2.3. Community Detection Techniques

  • The survey groups community-detection techniques across static, dynamic, disjoint, and overlapping settings, against a background of computational difficulty and no universally accepted solution.It presents six state-of-the-art groups and points to existing comprehensive reviews for further coverage.Source: §2.3 · p. 6

2.3.1.2. Hierarchical clustering

  • Hierarchical methods form multilevel communities without requiring a predefined community count; agglomerative methods merge clusters, while divisive methods split them.The section also describes graph partitioning, partitional and spectral clustering, modularity optimization, and greedy, simulated-annealing, extremal, spectral, and evolutionary approaches.Source: §2.3.1.2 · pp. 6–7, §2.3 · p. 6

2.3.3. Overlapping community detection techniques

  • Traditional detection methods can miss nodes belonging to multiple communities; clique percolation identifies overlapping groups through connected series of adjacent k-cliques.Two k-cliques are adjacent when they share k −1 nodes.Source: §2.3.3 · p. 8

2.3.4. Dynamic Community Detection Algorithms

  • Dynamic community-detection methods update community assignments for changed or new vertices as the network changes over time.The background introduces Potts-model, random-walk, diffusion, and synchronization perspectives on community structure.Source: §2.3.4 · p. 8, §2.3.4.1 · p. 8, §2.3.4.2 · pp. 8–9

2.3.4.1. Potts model

2.3.4.2. Random walk

  • Random-walk methods rely on walkers spending longer within dense communities, where multiple paths keep them inside; examples include WalkTrap and Infomap.The section also describes diffusion communities through shared propagation and synchronization-based identification through earlier synchronization within partitions.Source: §2.3.4.2 · pp. 8–9

3. Methodology

The study gathers Web of Science records and uses CiteSpace and Pajek to trace publication and citation trends and analyze collaboration, co-citation, and category networks. It also describes the visualization tools’ capabilities and the network measures used to interpret them.

3. Methodology

  • The analysis covers author collaboration, author/journal/document co-citation, institution and country collaboration, and category co-occurrence.The workflow also identifies the largest clusters in the collaboration and co-citation networks.Source: §3 · p. 9
  • The WoS search used ((communit*) AND (detect*) AND (network*)) across all years and four databases, yielding 3,168 English-language journal-article records.The search was run on November 11, 2016, with full records and cited references selected.Source: §3 · p. 9
  • 485 articles appeared in 2015 and 480 in 2016; citations rose from 100 in 1998 to 13,182 in 2015 and approximately 10,900 in 2016.The article counts describe the publication trend, while the citation figures describe citation history.Source: §3 · pp. 9–10, Figure 2 · p. 10, Figure 3 · p. 10

3.2. Network Analysis and Visualisation Tools

  • CiteSpace directly processes WoS data and visualizes citation networks by betweenness centrality, citation frequency, or citation burstness.The study supplements CiteSpace with Pajek for structural analysis because CiteSpace is described as relatively weak in analytics.Source: §3.2 · p. 11
  • CiteSpace marks high-centrality nodes with purple trims, while red circles indicate significant citation bursts and node size reflects total citation frequency.Links represent interaction, and line thickness indicates co-authorship strength.Source: §3.2 · p. 11
  • Landmark nodes are highly cited, hubs are widely co-cited, pivots join clusters, and turning points are highly central; the paper distinguishes turning points from pivotal points.The authors use Pajek alongside CiteSpace for structural analysis of bibliographic networks.Source: §3.2 · p. 11

4. Results and Discussion

The results examine author and journal networks, including prominent collaborators and brokerage roles, co-cited authors, network clusters, publication venues, and citation patterns. They also note limitations in author-name disambiguation and introduce journal co-citation analysis.

4. Results and Discussion

  • Liu J has the largest author-network node, Jiao LC the strongest citation burst, and J Liu the highest joint-publication frequency.J Liu published 18 papers; Licheng Jiao published 15, while Hui-Jia Li, Bo Yang, and X Liu each published 12.Source: §4 · pp. 12–13, Table 4 · p. 12
  • WoS name ambiguity leads the analysis to combine authors sharing surnames and first-name initials, and initial-based disambiguation can misidentify authors.For example, several Liu name instances are grouped as J Liu or X Liu, and several Wang instances as Y Wang.Source: §4 · p. 13

4.1.1. Identification of Largest Clusters in Author Collaborative Network

  • The largest author-network component has Watts-Strogatz clustering coefficient 0.778 and network clustering coefficient 0.595; Yong Wang is its pivot, connecting clusters #0, #1, #2, and #31.Its degree, betweenness, and closeness centralization are 0.124, 0.233, and 0.182, respectively. The collaboration network is described as small-world, with average distance 4.44526 between reachable pairs.Source: §4.1.1 · p. 14

4.2. Author Co-Citation Network Analysis

  • Mark J. Newman is the landmark and most-cited author, Barabasi AL is the most central co-cited author, and Albert R has the strongest author citation burst.Newman has 1,184 citations in the dataset; Barabasi AL shares the highest co-cited-author centrality score of 0.04 with Perona P, Grossberg S, and Duda RO.Source: §4.1.1 · pp. 15–17, Table 10 · p. 17

4.2.1. Identification of Largest Clusters in Author Co-Citation Network

  • The largest author co-citation component has seven clusters, but the largest cluster, #0, is outside that component.Cluster #0 has 85 members and is labelled digital radiography by LLR and neuroscience by MI; cluster #2 is labelled complex network by LLR and dimensionality reduction by MI.Source: §4.1.1 · pp. 18–19

4.3. Journal Co-Citation Network Analysis

  • 321 co-cited journals and 1,186 links make up the journal co-citation network constructed for the 1991–2016 timespan.The analysis uses one-year slices and g-index node selection with scaling factor k = 5.Source: §4.1.1 · p. 19, Figure 8 · p. 20

4.1.1. Identification of Largest Clusters in Author Collaborative Network

  • For network community detection bibliographic data spanning 1991 to 2016, Reviews of Modern Physics had the strongest citation burst, 32.64, in 2004; Science had the highest 2016 impact factor.Source: Table 12 · p. 20, §4.1.1 · p. 20
  • 0.29 was Science’s betweenness centrality, the highest among the journals listed.Nature followed at 0.15, and Physical Review E at 0.13.Source: Table 13 · p. 21, §4.1.1 · p. 21
  • 1,677 publications made PNAS the leading journal by frequency in network community detection.Figure 8 also identifies PNAS as the worldwide leader by publication frequency; the next four were Nature (1,427), Physical Review E (1,364), Science (1,036), and JSTAT (802).Source: §4.1.1 · pp. 20–21, Table 14 · p. 21

J STAT MECH-THEORY E

The analyses map influential journals and documents, leading countries and institutions, collaboration structure, category centrality and frequency, and citation-burst patterns in community-detection literature.

4.3.1. Identification of The Largest Cluster in Journal Co-Citation Network

  • Journal cluster #0 is the major research area, while cluster #1 is the second major and most active area, containing many highly cited, central, and burst-active journals.Cluster #0 contains 82 journals (15.471% of the network) with a 0.786 mean silhouette score; cluster #1 contains 77 (14.528%) with a 0.845 score.Source: §4.3.1 · p. 22

4.4. Document Co-Citation Network Analysis

4.4.1. Identification of Largest Network in Document Co-Citation Network

  • The United States is the most central country, with betweenness centrality 0.4, degree centrality 45, and closeness centrality 0.70.The largest country component contains 77 vertices (91.667% of the network), and its mean clustering coefficient is 0.690.Source: §4.4.1 · pp. 27–28, Table 19 · p. 28
  • Scotland has the strongest and longest country citation burst, scoring 4.7932 over 11 years from 1999 to 2009.England has the lowest burst, while Italy has the shortest, lasting two years.Source: §4.4.1 · p. 29, Figure 13 · p. 29
  • MIT is the most central institution, while the Chinese Academy of Sciences is the most productive, with 92 publications.MIT’s betweenness centrality is 0.09; its degree centrality is 36 and closeness centrality is 0.4883. The largest institution component has 147 vertices (42.857% of the network) and a 0.744 average clustering coefficient.Source: §4.4.1 · pp. 29–30, Table 21 · p. 30, Table 22 · p. 31
  • Imperial College London has the strongest institutional citation burst, 6.0668 over 2010–2014, while NASA’s burst lasts longest, from 1994 to 2008.The institution network’s largest component contains 147 vertices (42.857% of the network) and has a 0.558 transitivity.Source: §4.4.1 · pp. 29, 31, Figure 15 · p. 31

4.7. Category Co-Occurrence Network Analysis

  • Engineering is the most central category (0.37), followed by Public, Environmental & Occupational Health (0.35).The next three are Environmental Sciences (0.27), Psychiatry (0.16), and Mathematics, Interdisciplinary Applications (0.14). The merged network contains 173 categories and 595 links.Source: §4.4.1 · pp. 31–32, Table 23 · p. 32
  • Computer Science leads category frequency with 724 occurrences (22.853% of the dataset).The next four are Physics (530; 16.729%), Engineering (344; 10.858%), Computer Science, Information Systems (337; 10.637%), and Environmental Sciences & Ecology (317; 10.006%).Source: §4.4.1 · p. 32, Table 24 · p. 33
  • Public, Environmental & Occupational Health has the strongest citation burst (13.9655), lasting 13 years from 1991 to 2004.Astronomy & Astrophysics and Medicine, General & Internal have the longest bursts, each lasting 17 years from 1993 to 2010.Source: Figure 17 · p. 33, §4.4.1 · p. 33

5. Summary of the Results

CiteSpace analysis maps the domain’s authors, journals, documents, countries, institutions, and subject categories, identifying influential, central, active, and connecting nodes across these networks.

Author collaboration

  • Wang Y connects four author clusters, while Zhang Y joins the “Potts Model; Time; Resolution” and “non-overlapping community” clusters; “modularity” is the largest cluster with 52 vertices.The modularity cluster contains 3.6775% of total vertices in the network.Source: §5 · p. 33

Author co-citation

  • Mark J. Newman is the landmark author in co-citation analysis; Réka Albert has the strongest citation burst, and Barabási, Perona, Grossberg, and Duda are pivot nodes.Source: §5 · pp. 33–34

Journal co-citation

  • Reviews of Modern Physics has the strongest journal citation burst; PNAS is the most productive node, while Science is the most influential journal.The Complex Network; Modularity; Community Structure cluster is the most influential, productive, and active journal research area.Source: §5 · p. 34

Document co-citation

  • Andrea Lancichinetti (2008) is the most central document, while Forunato (2010) is the most cited; cluster #0 is a major and active research area.Okasonen J (2001) is the turning point joining clusters #6 and #11.Source: §5 · p. 34

Collaborative countries

  • The US has the highest country frequency and betweenness centrality, while Scotland has the strongest citation burst.The analysis selected the top 30 countries per one-year time slice from 1991–2016.Source: §5 · p. 34

Institutions and subject categories

  • MIT has the highest institutional betweenness centrality, the Chinese Academy of Sciences has the most publications, and Engineering leads category centrality at 0.37 while Computer Science leads frequency at 724.The University of London Imperial College of Science and Technology and Medicine has the strongest institutional citation burst.Source: §5 · p. 34

6. Conclusion

Using Web of Science publications from 1991–2016, the study applies CiteSpace to trace network community detection research and identify productivity and emerging trends. It reports growth after 2004 and findings across authors, journals, countries, institutions, and subject categories, while calling for further validation and subfield analysis.

6. Conclusion

  • CiteSpace analysis of Web of Science journal articles from 1991–2016 traces the field’s advances and assesses research productivity and emerging trends.The authors report that interest accelerated after Newman’s 2004 article.Source: §6 · p. 34
  • Santo Fortunato is reported as the most highly cited author, while Réka Albert’s publication count grew rapidly during the study period.Source: §6 · p. 34
  • PNAS contributed 52.935% of publications; the United States was the most productive and influential country, with the largest publication count and highest centrality.Source: §6 · p. 34
  • Most contributions came from the Chinese Academy of Sciences, Imperial College London remained active, and Computer Science led the subject categories.The passage names the institution as the “University of London Imperial College of Science and Technology.”Source: §6 · p. 34
  • The authors identify detailed visual analysis of subfields and validation against Pajek and other databases, such as Scopus, as future work.Source: §6 · p. 34

Related papers

Loading 1708.00977v1…