Sunday 8 June 2014


Visualizing Dynamic Interactions

Institut National de Recherche en Informatique et Automatique (INRIA) Saclay - ile-de-France




OVERVIEW: Graphs are powerful mathematical structures for modeling and representing many natural phenomena. In trying to explore and make sense of graphs collected in the wild — such as social interactions stored by social network sites or correlations between brain signals obtained using fMRI — visualization is often used. However, traditional visualization techniques are limited to sparse graphs: dense graphs are unreadable. Much progress has been made recently using matrix-based and hybrid visualizations to explore large and dense networks. Although understanding the visualization of the adjacency matrix of a graph is not as immediate as the traditional node-link representation, it does not suffer from most of its drawbacks and only takes a few minutes to grasp, a very reasonable time considering its expressive power. I’ll show how this relatively novel representation can be used to visualize many types of graphs, even dynamic graphs, with no limitation on density and good scalability. I'll show some results on social networks and brain signals.

READINGS:
    Wybrow, M., Elmqvist, N., Fekete, J. D., von Landesberger, T., van Wijk, J. J., & Zimmer, B. (2014). Interaction in the Visualization of Multivariate Networks. In Multivariate Network Visualization (pp. 97-125). Springer International Publishing. 
    Bach, B., Pietriga, E., & Fekete, J. D. (2014, April). 
Visualizing Dynamic Networks with Matrix Cubes. In SICCHI Conference on Human Factors in Computing Systems (CHI).

23 comments:

  1. Pourquoi un graphe qui représente les interactions entre les individus (par exemple, le graphe entre les liens de compte sur LinkedIn) ne doit pas être utilisé pour tirer des données et des conclusions?

    ReplyDelete
    Replies
    1. Un graphe représenté avec des noeuds et des liens devient illisible lorsque sa densité augmente. On peut l'utiliser, mais il est illisible: on ne peut pas répondre à des questions simples comme "A est-il connecté à B ?".
      C'est pourquoi la représentation matricielle est plus efficace, même si elle demande un peu d'apprentissage.

      Delete
  2. Most examples of graph visualizations are usually based on single-typed graph. In the real world (including non-trivial social networks) graphs involve multiple types of links between same vertices. In other words, they are multipgraphs (http://en.wikipedia.org/wiki/Multigraph). I am interested about analysis and visualization techniques for these structures. Are there any?

    ReplyDelete
    Replies
    1. The Sci2 Tool, see http://sci2.cns.iu.edu (taught in the IVMOOC http://ivmooc.cns.iu.edu/) supports bi-modal graph layouts, see http://wiki.cns.iu.edu/display/CISHELL/Bipartite+Network+Graph

      Delete
    2. Thanks! CISHELL is wonderful.

      Delete
    3. We're working on it, but it can become complex. See our work on OntoTrix for visualizing ontologies:
      http://hal.archives-ouvertes.fr/docs/00/56/05/48/PDF/hal.pdf
      There's a more recent and extensive paper publshed in a journal but the pdf is not accessible for free:
      http://www.igi-global.com/article/visualizing-populated-ontologies-with-ontotrix/102706

      Delete
  3. Cette étude présuppose que tous les humains partagent la même facilité ou difficulté à lire les graphiques. Cependant on pourrait penser que la compréhension des schémas est subjectif et que ce qui est simple pour l’un ne l’est pas pour l’autre.

    ReplyDelete
    Replies
    1. Il y a des différences interprersonnelles mais elles ne sont pas importantes. La plupart des expérimentations montrent au contraire que les humains sont très similaires en terme de perception visuelle. La compréhension des schémas n'est pas subjective mais peut demander de l'expérience (lorsque vous montez votre prmier meuble IKEA).
      Mais la visualisation utilise des mécanismes différents des schémas, au moins pour le décodage initial.
      Par exemple. Cleveland & McGill ont fait des expérimentation de perception de formes pour la visualisation des les années 80:
      Graphical Perception: Theory, Experimentation, and Application to the Development of Graphical Methods. W. S. Cleveland and R. McGill (1984). Journal of the American Statistical Association, 79:531-554.

      L'étude a été reproduite en 2010:
      Jeffrey Heer and Michael Bostock. 2010. Crowdsourcing graphical perception: using mechanical turk to assess visualization design. In Proceedings of the SIGCHI Conference on Human Factors in Computing Systems (CHI '10). ACM, New York, NY, USA, 203-212.

      Les résultats sont convergents et reproduisibles.

      Delete
  4. These tools are very impressive! Why has Dr. Fekete chosen to use squares matrices instead of a triangle. It appears to me that all the data is represented twice (if we take a line from the top left to bottom right corner, it mirrors the data on the other side).

    Do people understand the data better when presented as a square?

    ReplyDelete
    Replies
    1. This comment has been removed by the author.

      Delete
    2. In the case of undirected graphs (where if node A links to node B, node B also links to node A) the adjacency matrix will have data represented twice when it is in the form of a square. Not all data will be an undirected graph, however-- for instance, feedforward neural networks are directed acyclic graphs. The adjacency matrices will not have the symmetry that undirected graphs do.

      Delete
  5. Very interesting techniques and software. Most of the software is available for download as Java files or Java WS. I wonder, what are the licences of this software (did not find this information on web pages)? Specifically, what is the software licence of Cubix?

    ReplyDelete
  6. Really fascinating talk. It seems to me that the techniques presented by Professor Fekete might be extremely useful in cognitive science to model the behavior of neural networks over time and analyze the connectivity of the different units. My question for Professor Fekete is whether the technique of unfolding dynamic networks has already been applied to artificial neural networks.

    ReplyDelete
  7. Here's a cool link describing the use of virtual reality with head tracking (Oculus Rift) to view 3D visualizations of data.

    http://www.ebremer.com/haylyn/2013-08-22/semwebthroughoculusrift

    Adding a device like the myo controller (https://www.thalmic.com/en/myo/) would allow for the motor control Professor Harnad spoke of.

    ReplyDelete
  8. Ces nouvelles approches pour visualiser les données toute en simplifiant leur lecture et leur compréhension sont très intéressants, il suffit juste de prendre le temps de les maîtriser. je trouve le cubix formidable.

    ReplyDelete
    Replies
    1. I totally agree with Eltaani Redha. I found great the cubix. Mr JEAN-DANIEL FEKETE, Could you share a web site where I can know more about it? Thank you

      Delete
    2. You can find all the published papers and most of the programs on our web site: http://aviz.fr/

      Delete
  9. Fascinating talk. It seems to me that prof. Feteke and his colleagues must have some valuable knowledge about how researchers use graphs. Has it been published? Is it available in a form or another?

    ReplyDelete
  10. It's worth mentioning that the Cubix tool is based on the dominant model of data mining in Business Intelligence (BI) : the OLAP Cube. http://en.wikipedia.org/wiki/OLAP_cube

    ReplyDelete
  11. I found the conversation around visualization literacy very interesting. It seems that this could be corrected in our education system. We're taught some types of visualizations, like the 2d data plot, bar graph, and pie chart, but so much more could be done to help improve our ability to represent data in increasingly abstract and creative ways.

    ReplyDelete
  12. Dear Jean-Daniel, Thank you very much for sharing your works ! I have one question : I’m always very interested by visual analytics: I’m working myself on graphs and searching the most intelligible form for what I’m model or what’I induce. I would like to be provocative with Saint-Exepury who says, certainly in an other context: « On ne voit bien qu'avec le cœur. L'essentiel est invisible pour les yeux. » The occidental philosophy and science is very marked by the search of visual proofs as the etymology of evidence testify. My question is what’s your opinion about the limits of the effects or power of visualizations. Don’t we have alternatives comparable and as much powerful?

    ReplyDelete
    Replies
    1. Actually, for exploring data, visualization seems more effective than other methods. Statistics are useful, but they describe properties of the data, not the "shape" of the data. Just like providing your height, weight, eye color, hair color, will provide a profile of you, but not your picture.
      Take a look at Anscombe's quartet for a simple example:
      http://en.wikipedia.org/wiki/Anscombe's_quartet

      Antoine de Saint Exupery was a writer, but also an aviator, and he kept his eyes wide open while piloting. Our visual system has evolved to better grasp and make sense of the nature around us, avoiding obstacles and predators. Deciding not to use it for exploring abstract worlds would be, to me, just like we blindfolding our eyes: giving-up on a very useful sense.

      Delete
  13. I thought the ability to convert clusters into matrices and then link those matriced clusters with edges using Node Trix to be very clever. It really cleans things up without having to represent the whole network in one matrix, which I found somewhat less easy to navigate when there were lots of nodes. Is there a way to automate this clustering process using a set of parameters? I think that would be useful in that, it would remove any biases that may be introduced by having a researcher select the clusters and thus displaying the data is such a way that would lead readers to arrive at certain conclusions about the structure of a network. A classic example of biased data visualization is through the manipulation of scale - user selected clustering seems akin to that.

    ReplyDelete