current page

history

user

Lecture on Oct 21, 2009. (Slides)

Readings

  • Required

    • Graph Visualization and Navigation in Information Visualization: A Survey, Herman, Melancon, and Marshall, IEEE TVCG 2000. (pdf) NOTE: Just skim this to get an overview

    • Dig-cola: Directed graph layout through constrained energy minimization. Dwyer and Koren. InfoVis 2005. (pdf)

    • Visual Exploration of Multivariate Graphs. Wattenberg. CHI 2006. (pdf)

  • Optional

    • Hierarchical Edge Bundles: Visualization of Adjacency Relations in Hierarchical Data. Danny Holten. InfoVis 2006. (pdf)

    • A Comparison of the Readability of Graphs Using Node-Link and Matrix-Based Representations. Ghoniem, Fekete, Castagliola. InfoVis 2004. (pdf)

    • Vizster: Visualizing Online Social Networks. Heer & boyd. InfoVis 2005. (pdf)

    • Interactive Visualization of Genealogical Graphs. Michael J. McGuffin and Ravin Balakrishnan. InfoVis 2005. (pdf)

    • A Focus+Context Technique Based on Hyperbolic Geometry for Visualizing Large Hierarchies. Lamping, Rao, Pirolli. CHI 1995. (html)

Comments

nmarrocc wrote:

So its hard to have an algorithm design a good graph using constraints and then trying to optimize those constraints. In addition there are multiple ways of viewing the same graph and its not clear how one drawing of the same graph would be better than another. Why not build a tool for a graph designer that allows him to view the "design space" of many possible drawings of the same graph. Now a user could select some optimizations and give them more weight, for example he could view a design space of all the graphs with an emphasis on angles, or a space of many graphs that concentrate on the space of a graph.

vagrant wrote:

Frankly, I never gave much thought to trees and graphs beyond a algorithmic perspective--certainly I never considered them as practical visual representations of data, strange as that may seem. The readings for this topic have been somewhat dry, but I found them engaging nonetheless.

The Dig-cola approach seems to be a very enthusiastic approach, and were I to design a graphing schema myself, I would take much inspiration from the Dig-cola treatment of stress concepts. But for all the evidence presented in the paper, I found that when large quantities of data exist in a tree, that almost any result seems overwhelming, and it would take some doing to convince me that one layout is better than another. Still, given the obvious applications that trees have in computing and modern information database design, I feel just as surprised that more exploration has not been done on the subject, as I am at my own lack of ruminations.

As for Pivotgraph, while there is a nice simple and minimalist face to the results, as an uninformed viewer, I would fall under Tufte's category of viewers who become suspicious at an apparently emaciated representation of dense information. I guess that's just another trick to be solved when it comes to tree representations--balancing visual ease with informative depth.

bowenli wrote:

@vagrant Pivotgraph is used on non-tree graphs as well. I agree that there is some simplification but I think for certain types of patterns it actually does quite well. For example in their matrix vs pivot graph example I found the latter to be much more readable than trying to compare different cells. I wish they explained the ordering algorithm a little more since it seems like the order of the nodes can really make or break your understanding of the graph.

vizster: the paper shows that the program is actually pretty complex. You can do a lot more with it than first meets the eye. I thought the bit at the end about how party goers reacted was great. It's always interesting when users attribute some behavior to your tool that you didn't actually include. And since facebook killed off advanced search, I think tools like this will be pretty useful.

jieun5 wrote:

In search of graphs and trees used to visualize music (esp. with recent trends in social-music tagging and online listening), I discovered a blog full of useful and fascinating examples: http://visualizingmusic.com/.

The entry from October 22, 2009 is a HUGE slide show, and contents from p121~182, in particular, show many examples of graphs that have been created recently to visualize relationship between artists, songs, genres, and listeners. It's helpful to see the slide show in full-screen.

Many of these examples, however, experience the common problems found in designing graphs (as discussed in lecture), such as overcrowded nodes and connections. Because of this, I particularly liked examples of graphs that try to focus on a small set of information: data collected on one's own listening-trend through last.fm.

First awesome example is a stream graph by Lee Byron.This example, though not strictly a "graph" with nodes and edges, quite effectively shows relative volumes of artists and genres listened-to across time, and their interconnections across time.

The second example is Tube Tags that represents a single user's listening history in the style of the London Tube, where each colored line represents a genre / social tag. The choice of the London Tube "map" as the form is more catchy than is practical, in that there is little interconnection relationship between nodes (subway stops) other than possibly that the nodes have been arranged top-to-bottom in order of listening frequency within a time frame. But because the data is not cluttered, it is nevertheless an effective and definitely a fun way of visualizing the relative order of genres--along with top artists in the genres-- one listens to over time.

rnarayan wrote:

jieun5 - good links/examples - seems like the subject of the today's class lecture could be a course in itself...

Graph visualizations of many different networks (social, knowledge, semantic, transportation, etc.) have exploded taking more and more complex forms, although Vizster seems to be one of the first such efforts pioneering several nice original features. Found this site that links to over 700 such visualizations: http://www.visualcomplexity.com/vc/ There is also a visual browser to traverse the site in a semantically connected sequence: http://bestiario.org/research/remap/#project_id=586

Statutory warning - enter site at your own risk - it could take hours if not days going thru' the many visuals :)

wchoi25 wrote:

I really liked Wattenberg's pivot graph idea despite all the limitations discussed. It seems to scale well as the data set one's dealing with becomes extremely big, making it almost impossible to get any meaningful relationship without some sort of aggregation. This is the sort of interaction where animation seems very important (as discussed in the paper), in order to make it clear to the viewer what nodes are being aggregated to where from their initial placements in the graph structure.

I wonder if there could be an improvement to be made here. Currently, if the system puts together, say, 10 nodes, it would draw a single "bigger circle" node representing an aggregation of these ten nodes. But why not just leave them as a cluster? I'm thinking the ten small circles clustered to form the big circle. This way, when doing roll-ups and downs, each individual data point will still be visible. It also opens up a possibility for supporting at least one other variable, as these individual nodes could be colored differently based on a third dimension. This would let the viewer quickly see the makeup of a rolled-up node along this third dimension, perhaps mitigating what I see as the biggest problem of pivot graphs - hiding the effect of a certain third variable and thus overemphasizing an apparent relationship between the two dimensions selected to pivot on.

malee wrote:

Echoing wchoi25's comment, I also enjoyed reading about Wattenberg's pivot graph, and I agree with the importance of being able to easily aggregate information to deduce relationships between a multitude of data points.

Reading this and the other dataviz papers, I wonder at what point these visualizations and interactions will shift from being created/used by researchers to being created/used for everyday visualizations by laypeople. There are certainly advances being made in 'everyday' visualizations (NYTimes is one obvious example that comes to mind), but there is always a lag between when the newest viz comes out and when it is absorbed by the so-called masses. Is there a better way to increase adoption of the latest and greatest viz?

joeld wrote:

I really liked the DigCola paper - yet another fascinating example of convex optimization applied to a problem. I was suprised to read that the Majorization routines converge to a global optimum - especially because the many possibilities for symmetrical layouts in graphs make it possible to have combinatorially many layouts with essentially the same value. A visualization of the cost-function of the layouts would be interesting to see as well.

tessaro wrote:

I was very interested in how the many tree layout strategies deal with the focus + context challenge when faced with very large data sets. In particular, those approaches for looking ahead to gage the amount of children nodes in dynamically expanding or collapsing visualizations. Nested folders on a computer are a perfect example of the problem of anticipating parent to child relationships–you simply have to click through the folders to see their contents. The desktop metaphor derived from actual paper office folders entirely ignores an important real world look-ahead; seeing whether a folder is fat or thin, empty or full. (The desktop trash or recycler bin being a limited exception) One attempt at building upon the physical desktop metaphor is the use of scale and dynamic physics in a GUIs like this one from BumpTop3D

http://www.youtube.com/watch?v=M0ODskdEPnQ

It may be interesting to look at the possibility of incorporating the dynamic feedback of gravity or inertial friction in a node based representation. A local part of a tree structure could be toggled to behave like a hanging mobile, heavy nodes could imply their contents - perhaps by 'weighing' nested file sizes. If anyone has seen examples along these lines please share them...

nornaun wrote:

Although there are still limitation in Wattenberg's pivot graph, I think it can be used for a lot of practical work. For example, when I want to create some simple graph from multi-dimension data, I can it as an exploration tool to come up with the format of my final visualization. As I am working on my assignment3, this paper gives me some valuable insight. I have never seen an exclusive tool for exploration of data without producing final visualization package. I think it would be really helpful if I can have a transition tool that I can quickly playing with my data without having to worry much about how it look.

@malee, I agree with you that there should be a way to let the public know and make use of these new visualization methods. Visualization is a powerful tool, but I have never heard or thought about many things mentioned in our class before. One reason that not many people know of these new viz tools and methods may have to do with technology adoption cycle, where researchers in the field are the only ones enthusiastic about it. The public are late adopter who follows the flow once the standard is settled. Plus, the public is more interested in simplified/generalized data to detect the trend. In my opinion, what is lacking here is the channel to the public (or at least people who are interested in applying visualization to their work).

zdevito wrote:

One thing noted in lecture was that trees tend to grow exponentially, but the space on a 2D display grows only quadratically, making it difficult to fit a tree on a display if it grows too large. However, while trees can grow exponentially, it is clear that some do not allowing algorithms like Reingold and Tilford's tidier layout to exploit the less-than exponential growth. If the data actually grows exponentially then most of the tree layout methods are going to break down unless they aggregate information in some way. DoI trees are a nice solution to this problem because they recognize that some trees cannot be laid out without exponential space in the depth of the tree and instead try to quantify which parts are worth showing. One could then make a distinction between algorithms that try to layout an entire tree using minimal space and those that try to layout as much of a tree as possible given a constant amount of space.

alai24 wrote:

I fascinated by emulating physical processes for optimization. Dig-CoLa's energy and tension minimization are reminiscent of atoms rearranging themselves into a lattice.

Due to the exponential growth of graphs, I think the ability to filter out data is one of the most important aspects of a visualization. Using a hyperbolic space helps make the data look comprehensible, but even then it can be very difficult to get a sense of the relationships. I like PivotGraph's take on collapsing dimensions since it seems to make the important edges salient.

Leave a comment