A comparative study of social network analysis tools
1. A comparative study of social network analysistools David Combe, Christine Largeron, Előd Egyed-Zsigmondand Mathias Géry International Workshop on Web Intelligence and Virtual Enterprises 2 (2010)
2. Outline Context: social networks and analysissoftwareExpected functionalities of network analysis softwareBenchmarkConclusion
3. Context Definition (Wikipedia) A social network isa social structure made up of individuals called "nodes," which are tied by one or more specific types of interdependency, such as friendship, commoninterest, etc. Sociologic analysis Sociological works (Moreno 1934, Milgram 1967, Cartwright and Harary, 1977) Web 2.0 : Renewed interest from the Web based social networks websites development. Context
4. Context: Social network in business For the Gartner Institute: “By 2014, social networking services will replace e-mail as the primary vehicle for interpersonal communications for 20 percent of business users.” (Gartner 2008) Social network analysisisgetting mature. Some applications in business: Workflowstudyto adapt management to the real flow in a company; Identifykeyactors, ie. for viral marketing. These applications needadapted software. Context
5. Context: social networks and analysis software Network analysis software A previousstatisticalanalysisorientedsurvey (Huisman& Van Duijn, 2003) Networks and needs are changing Size Complex graphs Necessity to make a new benchmark Context
10. Aim: give a visualrepresentation of the graph, withdifferentapproaches: Fish eye Centered on an actor Force drivenvisualizationlayouts FruchtermanReingold (1984) Iterative algorithm Randomlayout F-R convergence 2. Visualization Expected functionalities of network analysis software
11. 3. Characterization by indicators Global indicatorsat network level by: Number of nodes Number of edges Diameter … Local indicatorsatnodelevel: Number of neighboors degree … Distance Length of the shortestpath Expected functionalities of network analysis software 2 1 Density 2 4 3 5 5 4
12. 3. Characterization by indicators : how to decide if an actoris « central »? Many ways to determine central actors. Ex: Betweenness centrality Which node is the most likely to be an intermediary for a random communication? higher betweennesscentrality Selection depends on what they are needed for. Expected functionalities of network analysis software
13. 4. Communitydetection Community: A set of actorshavingstrong connexions. Communitydetectionalgorithms Newman–Girvan (Newman and Girvan, 2002) Walktrap (Latapy & Pons, 2005) Expected functionalities of network analysis software
15. Benchmark methodology Required points: Asocial network analysis point of view Scalability Free for educational purposes A balance between well established software and newer ones, based on recent development standards (ergonomics, modularity and data portability). Datasets: Zachary’skarate-club, DBLP Benchmark
19. Pajek(Vladimir Batagelj and AndrejMrvar) Developmentstarted in 1996 Data miningoriented Many graph operatorsavailable Fast Exports 3D visualization Macro Supports matrices, adjacencylists and arcs listsoriented input files Benchmark
20. Gephi(Bastian M., Heymann S., Jacomy M.) Benchmark Development started in 2008 Interactive GUI Uses Java Recent scriptability improvements « Photoshop for graphs » with customizable visualization Supports the main file formats for networks Improvable by plugins Community detection still experimental
22. Igraph(Csárdi G., Nepusz T.) For R (a statisticalenvironment) and Python. The lowlevel routines are written in C. GUI available for R. Communitydetectionready. Not custom attributes-friendly Benchmark > g <- graph.ring(10) > degree(g) [1] 2 2 2 2 2 2 2 2 2 2 > g2 <- erdos.renyi.game(1000, 10/1000) > degree.distribution(g2) [1] 0.000 0.000 0.002 0.009 0.020 0.039 0.064 0.107 0.111 0.115 0.118… [21] 0.003 0.001
23. Benchmark How to choose the right tool? ++ Mature functionality - - Not available or weak
28. Bibliography Gartner http://www.gartner.com/it/page.jsp?id=1293114 Gartner Hype Cycle for Social Software, 2008 Fortunato, S. (2009). Community detection in graphs. Physics Reports, 103. Retrieved from http://arxiv.org/abs/0906.0612.Pons, P., & Latapy, M. (2005). Computing communities in large networks using random walks. Computer and Information Sciences-ISCIS 2005. Retrieved from http://www.springerlink.com/index/P312811313637372.pdf. Newman, M., & Girvan, M. (2004). Finding and evaluating community structure in networks. Physical review E. Retrieved from http://link.aps.org/doi/10.1103/PhysRevE.69.026113. Kamada, T., & Kawai, S. (1989). An algorithm for drawing general undirected graphs. Information processing letters, 31(12), 7--15. Retrieved from http://linkinghub.elsevier.com/retrieve/pii/0020019089901026.
29. Bibliography (2) Brin, S., & Page, L. (1998). The anatomy of a large-scale hypertextual Web search engine* 1. Computer networks and ISDN systems. Retrieved from http://linkinghub.elsevier.com/retrieve/pii/S016975529800110X. Fruchterman, T. M., & Reingold, E. M. (1991). Graph Drawing by Force-directed Placement. Huisman, M., & Van Duijn, M. (2003). Software for social network analysis. In Models and methods in social network analysis (p. 270–316). Freeman, L. (1979). Centrality in Social Networks Conceptual Clarification. Social Networks.
Notas del editor
24 minutesGood morning,mynameis DC. I come fromLaHC , University of St-Etienne. The aim of mythesis I amhere to presentyou an article titled « A comparative study of social network analysistools » writtenwith Christine Largeron, Előd Egyed-Zsigmond and Mathias Géry.
(for the Gartner …)Many experts insist about social network new behavioursappearing in the enterprise and necessity to takethemintoaccount.Gartner Institute predicts #CITATION#.It alsoconsidersthat SNA isgetting mature.So progresscanbe made in business use of SNA for workflow study intended to permit adaptation of the management to the real communication flow in a company;And Identify key actors in networks, ie. for viral marketing aware campains
A previoussurvey by Huisman & Van Duijn has been published.Havingpreviousbench, werealizedthatitshouldbeupdated, as networks and needs have changed.Main Changes in networks concerns the sizes (networks are far largerthanbefore) and the type of information (as networks tend to getricher, withcontextual data).
There are 4 main domains of functionalities :the Representations of the network as datafilesVisualizationCharacterization by quantitativeindicatorsCommunitydetectioncapabilities
There are different sorts of indicators, which help discribe the network or just one actor.First, among global indicatorsyoucanfindsomegeneral values such as the number of nodes of the graph, the number of edges, the diameter (longestshortest path, how far at the maximum are two nodes in the graph). You can compose these values and deduce the density of the graph (whichdepends on the maximum number of links youcan put in the graph ie. the complete graph). An highdensity (high links ratio) defines a dense graph. A graph with a lowdensity value issaid to besparse.The second type of indicatorsis local indicators. You cancalculateitwithonly a small part of the graph information. An exampleis the number of neighboors (called the degree of a node).You canalso use distances between 2 nodes (or evenedges). The shortestpathis one of them.
Nowlet’sadress the benchmark itself.
The methodology for the selection of the tools takes into account the following points:First, in our vision blocking criteria for selection are that:The tools must provide the basic metricsused in the Social Network Analysisdomain ;and the networks size can reach tens of thousands of nodes, as well as smallerones.The toolsstudied in depthwerechoosenbecausetheyrepresentlegacy, wellestablished software, or new developments base on recent standards such as modularization and ergonomics.The datasetsused for experimentation are the Zachary’skarate-club dataset and the DBLP database.
Seven areas of functionalities were evaluated:Input/output formatsCustom attribute handlingBipartite graphs specific functionsLongitudinal analysis (ietimestamped events)IndicatorsVisualizationClustering capabilities
Hereis the list of studied software.Several of them are vizualisationoriented: Pajek, Gephi, GraphViz, GUESS, Tulip.Others claim to permit graph manipulation: igraph, NetworkX, and JUNG.Bothapproaches have led for manytools in incorporating SNA metrics.
The retained softwareis split in twocategories: Stand-alone softwares and libraries.The Stand-alone software are Pajek and Gephi.The libraries are igraph and NetworkX.Some of the software are emblematic of an approach (Gephi has somecommonalitieswithTulip for example).Nowlet’sdetailtheseones.
This table summarizes how well each tool performs in the seven areas.Pajek performs quite well at the seven domains.If you need more than a simple visualization of a small network, you should avoid NetworkX.Igraph is very good at providing indicator, visualization and community detection.
Hereis an otherrepresentation of the results of the benchmark.Weseeclearlywiththis radar viewthatNetworkXisvery good at custom attributeshandling and providesmanyindicators.No software performsfairlywellattemporality and bipartite graphs management.
Inorder to concludethispresentation, wecansaythere aredifferentscientificdomainsconcerned by SN, and thisexplainwhythere are somuch approches and toolsavailable.The advancesshould come in short or midtermfrom smart modularisation of user interface, algorithms and data, with the goal of allowing the reuse of methodsbetween software. Contextualanalysisneedtools for longitudinal analysis and attributesstudy.Hierarchical graph visualizationisnow an important subjecttoo for smart manipulation of large graphs.