Interactive visualization and exploration of network data with gephi


Published on

Presentation for a workshop given at the Centre for Interdisciplinary Methodologies at Warwick University on May 9 2013. Focuses on conceptual and historical questions. Comments, references, and explanations are in the notes.

Published in: Education, Technology
  • Be the first to comment

No Downloads
Total views
On SlideShare
From Embeds
Number of Embeds
Embeds 0
No embeds

No notes for slide
  • Anatomy of a tweet.
  • Very large scale systems on the one side, but highly concentrated data repositories on the other.The promise of data analysis is, of course, to use that data to make sense of all the complexity.
  • Reduction to the point and line model requires purging of context, etc. but provides considerable computational capacities. No free lunch, as it were.
  • Gitlin, The Dominant Paradigm, 1978Many people argue that we no longer need that grant, we already have the data.
  • Tukey, The Future of Data Analysis, 1962
  • Allows for all kinds of folding, combinations, etc. – Math is not homogeneous, but sprawling!Different forms of reasoning, different modes of aggregation.These are already analytical frameworks, different ways of formalizing.Statistics: atomism, structure is implicit ("hidden forces", "social forces" cf. Durhkeim) => groups are abstractions, constituted by socioeconomic similaritySocial Network Analysis: atomism, structure is explicit ("dyadic forces") => groups are concrete, constituted by social exchange
  • Now we can calculate (in particular via matrix algebra).
  • Handbooks on graph theory are full of exhaustive discussions of basic graph types. Loads of vocabulary and analytical approaches.
  • (Created by WaelGhonim, considered to be a central place for the sparking of the Egyptian Revolution) (tool used for extraction)
  • Classic, simple statistics are often very useful.
  • It's easy and interesting to produce correlations.
  • A very different concept of "relation" – no longer derived and probabilistic (correlation) but explicit.Formalization: here: posts and users, but we could do many other things, e.g. users linked by co-reaction, posts linked by co-reaction, we could look at relationships between words, etc.How do we interpret this: understand the platform, understand the context of the phenomenon, understand the algorithm, etc.
  • How do we interpret something like this?
  • Visualization is, again, one type of analysis.Which properties of the network are "made salient" by an algorithm? behind: spring simulation, simulated annealing (
  • Non force-based layouts can be extremely useful. Gephi can produce those as well
  • Extend word lists (what am I missing?), account for refraction. Rieder & Gerlitz 2013: 2012:
  • Project variables into the graph User diversity = no of unique users of a hashtag divided by hashtag frequency
  • Larger roles of hashtags, not all are issue markers!
  • There is no need to analyze and visualize a graph as a network.Characterize hashtags in relation to a whole. (their role beyond a particular topic sample), better understand our "fishing pole" (the sample technique) and the weight it carries.Tbt: throwback thursday
  • From DMI workshop on anti-Islamism and right-wing extremism.We can also look at interaction patters: activity structure, held together by leaders?
  • From DMI workshop on anti-Islamism and right-wing extremism.Netvizz also allows for looking into interaction patterns in groups..
  • Unique user id allows for large scale analysis.The connectors are often the admins.See:
  • We can also look at interaction patters: activity structure, held together by leaders?
  • Combination of methods is most interesting.The move to posting a larger number of photos is highly successful for this page.What is happening in April 2012?For more details see: diagrams are not so good with time.
  • Simply looking at "images" in quantitative terms is not enough, here the qualitative part begins and netvizz makes it easier to take that step:Extracting photo URLsExtracting commentsStudying most "successful" tropes in depth, etc.For more details see:
  • Network analysis has produced a large number of calculated metrics that take into account the structure of the network."All in all, this process resulted in the specification of nine centrality measures based on three conceptual foundations. Three are based on the degrees of points and are indexes of communication activity. Three are based on the betweenness of points and are indexes of potential for control of communication. And three are based on closeness and are indexes either of independence or efficiency." (Freeman 1979)What concepts are these metricsbased on?
  • Network metrics are highly dependent on individual variables. Here: the same network with PageRank with four different values for the dampening parameter alpha. (red=highest PR value, yellow=second highest, turquoise=third highest)See Rieder 2012:
  • Interactive visualization and exploration of network data with gephi

    1. 1. Interactive visualization and explorationof network data with gephiBernhard RiederUniversiteit van AmsterdamMediastudies Departmentand some conceptual context
    2. 2. ContextTerms like "big data", "computational social science", "digital humanities","digital methods", etc. are receiving a lot of attention.They point to a set of practices of knowledge production: data analysis,visualization, modeling, etc.Instead of a totalizing search for a "logic" of data analysis, we couldinquire into the vocabulary of concepts and analytical gestures thatconstitute the practice of data analysis.A twofold approach to methods:☉ Engagement, development, application => digital methods☉ Conceptual, historical, and political analysis and critique => software studies
    3. 3. This workshopHow do we talk about data? How do we analyze them? What is our frameof thought? How do we go further in terms of imagination, expressivity?☉ Introduction☉ A bit of math☉ Two kinds of mathematics☉ Concepts and techniques from graph theory☉ Working with gephiEngage the theory of knowledge (epistemology) mobilized in data analysis,but through the actual techniques and not generalizing concepts.
    4. 4. Basic ideasWhy?Why do network analysis and visualization? Which arguments are putforward?☉ New media: technical and conceptual structures modeled as networks☉ Calculative capacities: powerful techniques and tools☉ Visualization: the network diagram, "visual analytics"☉ Logistics: data and software are available☉ Methodology: dissatisfaction with statistics (SNA)☉ Society: diversification, problems with demographics / statistics / theory
    5. 5. Platforms like Twitterboost opportunities forconnectivity betweenvarious types of actors.
    6. 6. At the same time, theyproduce detailed datatraces that are highlycentralized and searchable.Much of these data can beanalyzed as graphs.
    7. 7. What styles of reasoning?Hacking (1991) building the concept of "style of reasoning" on A. C.Crombie’s (1994) "styles of scientific thinking":☉ postulation and deduction☉ experiment and empirical research☉ reasoning by analogy☉ ordering by comparison and taxonomy☉ statistical analysis of regularities and probabilities☉ genetic developmentWhat kind of reasoning are we mobilizing in data analysis?Is it one type of reasoning or many?Are we "positivists" when we do data analysis? Reductionists?
    8. 8. Quality / quantity"One of my favorite fantasies is a dialogue between Mills and Lazarsfeld in which the formerreads to the latter the first sentence of The Sociological Imagination: Nowadays men oftenfeel that their private lives are a series of traps. Lazarsfeld immediately replies: How manymen, which men, how long have they felt this way, which aspects of their private livesbother them, do their public lives bother them, when do they feel free rather than trapped,what kinds of traps do they experience, etc., etc., etc. If Mills succumbed, the two of themwould have to apply to the National Institute of Mental Health for a million-dollar grant tocheck out and elaborate that first sentence. They would need a staff of hundreds, and whenfinished they would have written Americans View Their Mental Health rather than TheSociological Imagination, provided that they finished at all, and provided that either of themcared enough at the end to bother writing anything." (Maurice Stein, cit. in Gitlin 1978)Theory vs. empiricism, macro vs. micro, qualitative vs. quantitative, inductive vs.deductive, associative vs. formalistic, etc.The promise of data analysis tools, applied to exhaustive (and cheap) data, is tobridge the gap, to allow zooming, "quali-quanti" (Latour 2010).
    9. 9. Two kinds of mathematicsCan there be data analysis without math? No.Does this imply epistemological commitments? Yes.But there are choices, e.g. between:☉ Confirmatory data analysis => deductive☉ Exploratory data analysis (Tukey 1962) => inductiveThere is a fast growing variety of formal analytical gestures relying onmathematical modeling and computation.
    10. 10. Two kinds of mathematicsStatisticsObserved: objects and propertiesInferred: social forcesData representation: the tableVisual representation: quantity chartsGrouping: "class" (similar properties)Graph-theoryObserved: objects and relationsInferred: structureData representation: the matrixVisual representation: network diagramsGrouping: "clique" (dense relations)
    11. 11. Graph theoryLeonhard Euler, "Seven Bridges of Königsberg", 1735Introducing the "point and line" model
    12. 12. Graph theoryDevelops over the 20th century, in particular the second half.Integrates branches of mathematics (topology, geometry, statistics, etc.).Graph theory is "the mathematics of structure" (Harary 1965), "amathematical model for any system involving a binary relation" (Harary1969); it makes relational structure calculable."Perhaps even more than to the contact between mankind and nature, graph theory owes tothe contact of human beings between each other." (König 1936)
    13. 13. Basic ideasMoreno 1934Graph theory developed inexchange with sociometry,small-group research and(later) social exchangetheory.Starting point:"the sociometric test"(experimental definition of"relation")
    14. 14. Basic ideas
    15. 15. Forsythe and Katz, 1946, "adjacency matrix"
    16. 16. Harary, Graph Theory, 1969
    17. 17. Basic ideasThe late 1990sThe network "singularity":☉ The network imaginary, a "new science of networks" (Watts 2005)☉ Computational capacities (memory, speed, interfaces, etc.)☉ New platforms and datasets☉ Packaged toolsDifferent traditions conflate to form network analysis:☉ Social network analysis and sociometrics☉ Scientometrics / science and technology studies☉ Mathematics / physics / computer science☉ Information and data visualization☉ Digital sociology / new media studies
    18. 18. Basic ideasAdamic and Glance, "Divided They Blog", 2005
    19. 19. Formalization"As we have seen, the basic terms of digraph theory are point and line. Thus, if anappropriate coordination is made so that each entity of an empirical system is identifiedwith a point and each relationship is identified with a line, then for all true statementsabout structural properties of the obtained digraph there are corresponding true statementsabout structural properties of the empirical system." (Harary et al. 1965)There is always an epistemological commitment!=> What can "carry" the reductionism and formalization?=> What types of analytical gestures?
    20. 20. Facebook Page "ElShaheeed", June 2010 – June 2011, (Poell / Rieder, forthcoming)7K posts, 700K users, 3.6M comments, 10M likes (tool: netvizz), work in progress!
    21. 21. Facebook Page "ElShaheeed", June 2010 – June 2011:comment timescatter, log10 y scale, likes on
    22. 22. Facebook Page "ElShaheeed", June 2010 – June 2011:scatterplot comments / likes, per post type
    23. 23. Facebook Page "ElShaheeed"700K nodes, 11M connectionsColor: type
    24. 24. Facebook Page "ElShaheeed"700K nodes, 11M connectionsColor: outdegree
    25. 25. Basic ideasWhat Kind of Phenomena/Data?Interactive networks (Watts 2004): link encodes tangible interaction☉ social network☉ citation networks☉ hypertext networksSymbolic networks (Watts 2004): link is conceptual☉ co-presence (Tracker Tracker, IMDB, etc.)☉ co-word☉ any kind of "structure" that can be as point and line=> do all kinds of analysis (SNA, transportation, text mining, etc.)=> analyze structure in various ways
    26. 26. Basic ideasWhat is a graph?An abstract representation of nodes connected by links.Two ways of dealing with graphs:☉ mathematical analysis (graph statistics, structural measures, etc.)☉ visualization (network diagram, matrix, arc diagram, etc.)
    27. 27. Three different force-based layouts of my FB profileOpenOrd, ForceAtlas, Fruchterman-Reingold
    28. 28. Non force-based layoutsCircle diagram, parallel bubble lines, arc diagram
    29. 29. Network statisticsbetweenness centralitydegreeRelational elements of graphs canbe represented as tables (nodeshave properties) and analyzedthrough statistics.Network statistics bridge the gapbetween individual units and thestructural forms they areembedded in.This is currently an extremelyprolific field of research.
    30. 30. Basic ideas
    31. 31. Basic ideasWhat is a graph?Vertices and edges!Nodes and lines!Two main types:Directed (e.g. Twitter)Undirected (e.g. Facebook)Properties of nodes:degree, centrality, etc.Properties of edges:weight, direction, etc.Properties of the graph:averages, diameter, communities, etc.
    32. 32. Basic ideas
    33. 33. Basic ideasWikipedia: Glossary of graph theoryTools are easy, concepts are hard
    34. 34. Basic ideasInteractive visual analyticsBringing structure to the surface (gephi panel: "layout")☉ different spatializations (force, geometry, etc.)Projecting variables into the diagram (gephi panel: "ranking")☉ Size (nodes, edges, labels, etc.)☉ Color (nodes, edges, labels, etc.)Deriving measures (gephi panel: "statistics")☉ Properties of nodes, edges, structure => new variablesAnalysis: e.g. correlation between spatial layout and variables?
    35. 35. Basic ideas
    36. 36. Basic ideas
    37. 37. Basic ideasTwitter #ows dataset, co-hashtag analysisStrong topic clustering
    38. 38. Twitter 1% sample, co-hashtag analysis227,029 unique hashtags, 1627 displayed (freq >= 50)Size: frequencyColor: modularity
    39. 39. Size: frequencyColor: user diversityTwitter 1% sample, co-hashtag analysis227,029 unique hashtags, 1627 displayed (freq >= 50)
    40. 40. Size: frequencyColor: degreeTwitter 1% sample, co-hashtag analysis227,029 unique hashtags, 1627 displayed (freq >= 50)
    41. 41. Twitter 1% sampleCo-hashtag analysisDegree vs.wordFrequency
    42. 42. Degree vs. userDiversityTwitter 1% sampleCo-hashtag analysis
    43. 43. FB group "Islam is dangerous"Friendship network, color: betweenness centrality2.339 membersAverage degree of 39.6981.7% have at least one friend in the group55.4% five or more37.2% have 20 or morefounder and admin has 609 friends
    44. 44. FB group "Islam is dangerous"Friendship network, color: Interface languageen_us, de, en_uk, it dominate
    45. 45. Mapping European ExtremismFriendship relations of 18 extreme-right groups
    46. 46. FB page "Educate children about the evils of Islam"Links have more comments, photos more likes.
    47. 47. FB page "Stop the Islamizationof the World"Number of posts and reactions
    48. 48. FB page "Stop theIslamization of the World"
    49. 49. Basic ideasInteractive visual analyticsBringing structure to the surface (gephi panel: "layout")☉ different spatializations (force, geometry, etc.)Projecting variables into the diagram (gephi panel: "ranking")☉ Size (nodes, edges, labels, etc.)☉ Color (nodes, edges, labels, etc.)Deriving measures (gephi panel: "statistics")☉ Properties of nodes, edges, structure => new variablesAnalysis: e.g. correlation between spatial layout and variables?
    50. 50. Basic ideas
    51. 51. Nine measures of centrality (Freeman 1979)
    52. 52. Label PR α=0.85 PR α=0.7 PR α=0.55 PR α=0.4 In-Degree Out-Degree Degreen34 0.0944 0.0743 0.0584 0.0460 4 1 5n1 0.0867 0.0617 0.0450 0.0345 1 2 3n17 0.0668 0.0521 0.0423 0.0355 2 1 3n39 0.0663 0.0541 0.0453 0.0388 5 1 6n22 0.0619 0.0506 0.0441 0.0393 5 1 6n27 0.0591 0.0451 0.0371 0.0318 1 0 1n38 0.0522 0.0561 0.0542 0.0486 6 0 6n11 0.0492 0.0372 0.0306 0.0274 3 1 4
    53. 53. Basic ideasUS Airports
    54. 54. Thank Yourieder@uva.nlhttps://www.digitalmethods.net"Far better an approximate answer to the right question,which is often vague, than an exact answer to the wrongquestion, which can always be made precise. Dataanalysis must progress by approximate answers, at best,since its knowledge of what the problem really is will atbest be approximate." (Tukey 1962)