I'm looking for a software for visualizing main trends in some field. Specifically, I want it to build a graph of the field with vertexes representing papers and edges representing how close papers are (probably, based on number of co-citations). It seems to be very useful in order to find important subfields (clusters in the graph) and important papers (centers of the clusters).
$\begingroup$ Have you seen this post? $\endgroup$– Sadeq DoustiJan 22, 2011 at 12:50
$\begingroup$ I haven't, thanks. I have just tried both academic search by microsoft and dblp visualiser, but they do not show specific papers, unfortunately. $\endgroup$– Tatiana StarikovskayaJan 22, 2011 at 13:19
$\begingroup$ I can see the resulting web application that allows you to browse your field like that. Good question! I don't know a tool, but you might find something related to semantic web. $\endgroup$– RaphaelJan 22, 2011 at 13:23
$\begingroup$ Not exactly what you are looking for (since it operates on the level of journals, not individual papers) but eigenfactor.org provides some great visualizations of how different fields of science relate and are clustered. The team behind it also have a few good papers on how to identify clusters or sub-fields relatively robustly: eigenfactor.org/papers.htm $\endgroup$– Artem Kaznatcheev ♦Jan 22, 2011 at 16:53
$\begingroup$ Sadeq, Artem: your comments might very well be answers. $\endgroup$– Suresh VenkatJan 22, 2011 at 17:20
To elaborate a little on my comment (which is slightly tangential).
There is a fun site called eigenfactor that provides nice visualizations of how various parts of science cite each other, and their relative size by publication/citation volume. Unfortunately their primary focus seems to be calculating impact factors for journals and so they operate on the level of journals, not individual papers/authors. However, they have a pretty interesting way of identifying subfields.
They use the number of citations from a journal A to a journal B as a weight on the edge from A to B. They then consider a random walk on the resulting graph. Such a walk tends to linger in topic clusters, and thus when they try to create a shortest description of the walk they tend to name the clusters and use a second index inside the cluster (kind of how you would have unique city names, and unique street names in each city, but redundant across cities). I thought it was a pretty nice way of finding clusters, and produces clusters that mimic how most people would divide the topics in science (with the occasional hilarious classification like putting Phys. Rev. Letters in Chemistry instead of Physics) You can read more in their paper or on the mapequation site.
$\begingroup$ Nice thing to play with it, thank you. Would be great to have the same thing for papers! $\endgroup$ Jan 22, 2011 at 20:53
This is actually a non-trivial problem and active area of research.
Plugging citeseer data into a graph visualizer is very doable (ex. Pajek), and will give you a decent way to browse if you size things by PageRank and maybe include a cutoff threshold (though missing obscure papers that would be useful to you is still an issue). Unfortunately, knowing why one paper cites another is often difficult. Computer science papers often have huge lists of citations, most of which are not vital to intellectual comprehension of a particular paper. Whether papers are being cited because they represent a classic result, are being improved upon, are being contrasted with, or actually helped create the ideas of the current paper can't be known without some textual analysis (though other methods might offer some help).
Surfing through a visualizer for "prominent papers" would be more aesthetically pleasant but not much more useful than surfing google scholar; mega-papers with hundreds of citations, but not necessarily related to the subfield you desire, would dominate the landscape. Also, certain interdisciplinary papers could "pollute" otherwise neat topic partitions.
I'm actually exaggerating the problem some. Using a visualizer on citation data the way you described, if well-implemented (mainly trimming the results so they're not too cluttered but so that it isn't reduced to only big-name results), could be very very helpful, especially with awareness of its limitations. All I'm really saying is that a good survey paper is probably the best possible option to examine a subfield, if one exists.
$\begingroup$ Thank for you answer! Actually, after posting a question I have found this link cs.umd.edu/hcil/InfovisRepository/contest-2004/17/unzip/…, where the authors do exactly what I want, but only for a small set of data. $\endgroup$ Jan 28, 2011 at 8:52
$\begingroup$ For a good visualization that can handle a large amount of data and not be an eyesore (in contrast to the otherwise nifty link you give), you probably want the following things: 1) nodes sized proportional to some metric (e.g. pageRank) 2) zoom feature that hides small/less relevant nodes at a distance 3) hierarchical clustering that shows you high-level clusters at long zoom (with a best-guess label) and low-level clusters at close zoom. I wish I could point you to some software that would do this, but I don't know if there is or isn't. $\endgroup$ Jan 28, 2011 at 20:58
$\begingroup$ Think Tulip can tulip.labri.fr. Sorry not to answer earlier. $\endgroup$ Feb 16, 2011 at 9:22