history

user

Lecture on Nov 4, 2010. (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

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

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

  • Optional

    • Visual Exploration of Multivariate Graphs. Wattenberg. CHI 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

wulabs wrote:

In "Graph Visualization and Navigation in Information Visualization: a Survey", I honestly do not see what the applications of the complex graph visulization methods such as fisheye grid distortion and space-scale diagrams. The basic spanning and tree based views I can see as being useful, then the secondary level of visualization seems to be a way of focusing or segregating the main data in view.

I have seen the tree map with colours being used successfully in a computer disk space application which clusters directories on one's hard drive as individual rectangles, and the color of the rectangles depends on the type of files contained there-in. It is a very useful application which allows users to quickly see from a birds eye view what directories or files or type of files are taking up the most disk space, and then also allowing users to quickly drill down in one particular area or directory.

wulabs wrote:

In "Graph Visualization and Navigation in Information Visualization: a Survey", I honestly do not see what the applications of the complex graph visulization methods such as fisheye grid distortion and space-scale diagrams. The basic spanning and tree based views I can see as being useful, then the secondary level of visualization seems to be a way of focusing or segregating the main data in view.

I have seen the tree map with colours being used successfully in a computer disk space application which clusters directories on one's hard drive as individual rectangles, and the color of the rectangles depends on the type of files contained there-in. It is a very useful application which allows users to quickly see from a birds eye view what directories or files or type of files are taking up the most disk space, and then also allowing users to quickly drill down in one particular area or directory.

hyatt4 wrote:

I took the note to heart and just skimmed through the Graph Visualization and Navigation paper. I found it interesting, that after looking at all of the graphs, I found figure 1, the tree graph to be the most intuitive to me. The authors go on to explain how the graph fails to meet certain objectives. And I agree that it makes it more difficult to focus in on one part (section 1.2 presents the comprehension issue that arises when looking at the entire graph at once), but I still had an easier time knowing how things related. That is, I knew how things related if there was a particular parent or root node and children nodes. The radial views, cone tree, information cube, and some others did not have the same instant comprehension from an initial look. Perhaps I have been trained or biased from looking at tree graphs on computer systems, or perhaps there is something inherently more perceptually intuitive to this layout.

gdavo wrote:

In "Hierarchical Edge Bundles: Visualization of Adjacency Relations in Hierarchical Data", I was impressed by the results obtained in fig. 15 when the bundling is used on top of a scarified treemap. Contrary to most people in the survey, I think it is more powerful than the radial layout, even though it can seem more "pleasing". But the advantage of the treemap is that it keeps the tree hierarchy visible, whereas for the radial layout it is moved outside of the circle. So I find it easier to visualize the hierarchy in the treemap. But it might be improved by reordering the elements in each rectangle so that the relationship lines are shorter. In general I wonder if it is possible to adapt the tree layout for the bundling algorithm instead of just drawing in top of the layout.

msavva wrote:

I was really impressed by the results of the hierarchical edge bundling paper. It applied the very natural concept of bundling connections together along shared paths on an underlying hierarchical structure in order to effectively reduce clutter and present higher level patterns. In particular, I found the section on interaction with the bundling strength parameter (see Fig. 16) to be extremely interesting. The described approach gives users the ability to focus on higher level or lower level structure on demand, over a continuous axis of complexity. I thought that it would be really nice to tie other parameters for graph visualization in such an interactive way and facilitate graph comprehension at multiple scales. For example, a very elegant way to handle complex graphs (in terms of multi-scale structure) would be to tie the degree of semantic zoom and clustering to the size of selected node groups (or an implicit selection size in the case of panning and zooming). Such animated transitions seem to impart a lot of power when they are used to facilitate interaction in the way that was described and I'm not sure to what extent this has already been explored by research in graph visualization.

selassid wrote:

I think the force-directed edge bundling is a powerful visualization, except I think it might fall short in a few ways. I don't have a sense of what locations nodes in a graph should be placed at in order to highlight the high-level patterns in the bundles. I also am wondering what would need to be done in order to make the visualization apply to a directed graph and highlight differences in direction. Those edge bundling papers have had some of the tastiest looking images thus far, though.

yanzhudu wrote:

It is interesting to see Force-Directed and DiG-Cola algorithms draw their inspiration from physical system.

On another note, as mentioned by wulabs, there are programs in Linux that shows disk usage using tree map. I have used it to find what is taking up my disk space. The advantage is that it is really easy to spot large files/directories. The disadvantage is that the tree map can get quite cluttered. When there is no particularly large files, the information displayed on tree map is generally the same size. The amount of information can easily overwhelm user.

jtamayo wrote:

Large graphs seem like a perfect candidate for interactive visualization, as the general trends are usually just as interesting as the particular details of each node and edge. One good example of an interactive graph is this one by the Wall Street Journal. It shows the relation between the 50 most popular websites and online tracking companies.

http://blogs.wsj.com/wtk/

Even though it is a relatively simple graph, it's interesting how the full graph is hard to understand. Making it interactive makes it much more compelling and useful.

gneokleo wrote:

I always thought that the representation of a network graph is something hard to do especially when the graph has many nodes with high clustering and high degrees. I think this was also demonstrated in class today. The Holten paper does a really good job in showcasing the strengths of edge bundling and I especially like the edge bundling radial representation. Even though there's a few disadvantages in bundling too much (for example you can't track edges easily) i think where this benefits a lot is when you want to get a general idea on which nodes are more important and have high degrees.

sholbert wrote:

I thought there are some great insights with edge bundling, espcially if it integrates some interactivity with toggling the beta value.

I thought it was also interesting how participants in vizster liked their friend graph to me moving because it felt more "alive." What need is this "alive" factor really getting at; why do users prefer it? Personally, I think that it suggests interactivity which, for better or for worse, will be amenable to users

ankitak wrote:

I found a really cool visualization tool, and I didn't know where else to post it, so here it is: http://www.gapminder.org/

andreaz wrote:

Hierarchical edge bundling really reminds me of the visualization software used to view DTI or DSI scans of fiber tracts in the brain. The interactivity of the edge bundling software is also similar; in DSI software, users can create a selection of an area of the scan and view only the tracts that go through that area, and the edge bundling software provides a similar function to allow users to filter out certain connections.

It's not surprising to me that participants in the edge bundling study preferred the radial layout over the other layouts because the I think that layout allows for the clearest comparison of relationships between categories since all cross-category relationships pass through the same space. I really like that layout since it is easy to take in the hierarchical attributes of nodes, and I like how connections between nodes in the same category are visualized differently (they don't pass through the center) than connections between nodes that don't belong in the same category (which do pass through the center). I also like the use of gradients to encode the directionality of the relationship between nodes. I think this technique is really successful at giving users an accurate overall picture of the relationship patterns in data and I'm hoping to potentially use this technique for the final project.

ankitak wrote:

With respect to the comment by selassid, it is true that force directed layout typically does not use direction or absolute location as encoding mechanisms. But i think there might be a way to constrain the movement of nodes in a particular dimension, and that might be useful in some cases. For example, we could probably use a radial layout with force directed layout, where (some / all) nodes can be constrained in one dimension, say by the radius, but free to move in the other dimension (angular). Such combinations might help us in coming up with interesting visualizations.

asindhu wrote:

I found the Dwyer et al. paper on DiG-CoLa very interesting. From my perspective, one of the most unique features of this layout algorithm was the concept of horizontal "bands" of variable size, each of which denotes one level in the hierarchy. They show various advantages of this approach, including the fact that cycles are displayed much more clearly by putting them all on one, wide band. There are certainly many advantages to this approach. However, I wonder how it compares to the much simpler layouts where hierarchy is encoded as the y position in terms of how easy it is for the viewer to identify the hierarchical structure of the tree. The bands are not normally shown in a DiG-CoLa graph, so the hierarchy isn't actually as directly clear. I think that as always, this ends up coming down to what exactly you're trying to get across in your visualization. If hierarchy is the most important thing, it may be better to go with a simpler layout at the expense of other things.

anomikos wrote:

The methodology presented in Holten's paper seemed to produce by far the most impressive results from all the examples presented in class, especially if we take into account the ability to easily add filters through interactive elements. However as it is briefly mentioned in the paper results where more impressive in a planar and especially radial setting. The collinearity problems of the algorithm seems to indicate that there is no generally applicable solution and the visualization setting plays an important role in the overall performance.

rakasaka wrote:

I was very impressed by the bundling suggested by Holten in his paper, since it is remarkably effective in reducing visual clutter while maintaining salient aspects of node connections. It very much gives the impression of similarity to the roots in a tree, highlighting the "aesthetically pleasing" aspect. In my own work I have come across fairly frustrating problems with space and clutter, and certainly this method works well for the radial layout- though I would only wish some similar form could work for the rooted-tree form. The paper also failed to mention how interactivity was/was not used in the user-based study (though they mention zooming as a "future work" item) - it would be nice to be able to move a cursor over a strand and see its actual connection.

jbastien wrote:

One thing that frustrates me often with graphs is how hard it is, as a user, to explore the graph without losing context, and how hard it is to reconfigure the graph (reorganize it so that exploration is made easier). These things are possible but often very hard to get right.

esegel wrote:

The examples from Thursday's lecture emphasized that different quantities of details are desirable for different levels of analysis. For example, when do you want to see the entire tree structure? When do you want to only see subtrees? When do you want to see leaf labels? Etc. The different graphs from lecture showed different ways of handling this problem.

In regards to treemaps: when treemaps contain several hierarchal levels. it might be wisest use color distinctions for only two levels deep (since, according to Jeff, this is really the max depth people can meaningfully perceive in a treemap). For instance, consider the "map of the market" treemap, and coloring for "up" days and "down" days. In this map, you might get the following hierarchy: market -> sector -> sub-sector -> individual company. In such a map, perhaps, you'd only want color done on the "sector" or "sub-sector" level when getting an overview of the map, and then ONLY color the individual companies when drilling into a particular sector. Is this sort of "zoom detailing" done? i.e. matching the coloring to the likely level of analysis? Just a thought.

mariasan wrote:

I found the hyperbolic tree visualization interesting. I wonder if it could be extended to get around some of its shortcomings, like by adding colored paths to highlight what you haven't seen (or the other way around), using some clever algorithm.

adh15 wrote:

The Reingold-Tilford Algorithm's result as presented in lecture might be improved by allowing more horizontal space between nodes of equal depth that are less closely related. This might facilitate easier visual grouping of more highly related nodes and prevent inadvertent grouping of trivially-related nodes.

arievans wrote:

In lecture we were exposed to a tremendous variety of ways to display network and tree views, and to be honest I'm impressed with how much time and effort has been put into this area of visualization. I particularly agree with @esegel in his comment above. Namely, It is most important for us to figure out how much info is too much, or perhaps, what specific question is being answered and how can we answer it most effectively. I like where @esegel was going with the idea of using color to draw attention to sort of the minimal set of information that the user needs at whatever level they are viewing the data. I'd like to see that in practice.

ericruth wrote:

For me, the most interesting visualization technique from this lecture was the tree layout in hyperbolic space. I thought this was a really clever application of hyperbolic geometry - especially because there was a recurring theme of cluttered space when visualizing trees. I can just picture somebody being like "I wish our space could just grow exponentially ... oh, wait, it can!" Anyway, the thing that impressed me the most about the hyperbolic layout was the ease with which we could explore a really large tree that would become far too cluttered in a regular geometry. In addition, the shortcomings we talked about didn't seem that insurmountable with some visual band-aids. However, while I think the hyperbolic layout is an effective (and fun) layout for exploring a huge tree, I agree with the sentiment of many of the previous comments that we usually want to evaluate whether a tree of that size is actually necessary. Deciding what information to leave out in a given situation seems to be an important trend in making effective visualizations...

nikil wrote:

I also noticed that there were several of the earlier design principles that come into play when displaying networks. One of the primary ones seemed to be color. Representing the edges and nodes in different colors to represent different features of the features of the data was greatly impacted by the decisions on the number and choice of colors. For example in our network analysis class, a graph was displayed with a whole spectrum of colors and the ordering was consequentially not apparent which was evidenced by someone asking what the difference in the green and yellow nodes were (because the scale didn't indicate intuitively). I think that these basic principles are important to keep in mind always.

sklesser wrote:

I was really intrigued by the edge tension bundling algorithm. I wonder if this technique can be enhanced by leveraging color such as shifting the color of the edges throughout the data set to make it a little easier to track a group of edges after they come out of the bundle. Currently it's hard to tell where edges go after they combine in a large bundle.

Leave a comment