JC Herz’s talk was originally "Visualization of Social Intelligence", but she felt that "Visualizations of Social Media" was a better description of what she was actually discussing. She started off in an iconoclastic vein, comparing current social visualizations to lava-lamps or snow globes; fascinating, but conveying absolutely no information. Pretty but useless.
To find a better way, she took us on a whistle-stop tour of the history of graphical data, or pictures that tell stories. Snow’s cholera map and Guerry’s suicide map of France were two classic examples. One common theme with all the classic visualizations is that they were snappy graphics for hidebound institutions. Because producing them involved a lot of manual effort and time, they were only created to address pressing problems.
She then moved on to an example of some work she’d done in this area. She had a brief to investigate how you can visualize coordination in a group, using data from the America’s Army game. She created a way of viewing a particular team’s communications over time, with a 3D graph showing communications as circles on a line, with a line for each team member and the distance along the line representing time. Using this tool, she was able to draw out statistical measures that could tell which teams were effective, based purely on their communication patterns.
One thing she thought important to emphasize is the value of time as a parameter in most visualizations. Current social graphs have no awareness of time.
Another important component of a successful visualization is that it should have consequences. She took the example of the study of dating patterns in a mid-west high-school, and asked if anyone would have been interested in the same graph showing SMS patterns in that high-school? Another example that speaks to the same question is a map showing what name people in different parts of the country use for pop or soda. There isn’t a big consequence to this map, but the minor light it sheds on the culture of the US is enough to get people interested.
The Trulia Hindsight visualization combines both something people are very interested in, real estate, and shows it over time. It tells a story, in a very compelling way.
Space is an important dimension for telling stories too, which is why maps showing political fundraising by area are so fascinating. Conflict or drama is the third ingredient to a compelling story, which is why the diagram showing book buying patterns by party affiliations was such a success.
A story isn’t just an automatic result of running data through an algorithm, to get insight, you have to engage in a dialog with the data. If you ask stupid questions, you get stupid answers.
To wrap up, she proposed some principles about what makes a visualization useful. There should be less information, but the right information. Not just a mish-mash of all the data you have, but a focused version that shows a selected subset of interesting or surprising information. All visualizations should tell a story, which requires notions of time, space, and something at stake. This is why so many popular visualizations are political, because people are trying to make important decisions. To be useful, the visualization also has to be sharable. You’re trying to tell a story to affect something in the real world, and the only way of affecting things is by getting other people involved.
A good test for whether a visualization is any good is asking if it has any consequences? It’s such a waste to go to all the effort of producing a diagram if it doesn’t matter. Any artifact you produce must be sharable to be effective.
The first question came from Matt Hurst, the author of the first diagram JC used in her rogues gallery of pretty but useless visualizations! He’s got an online response here too. He wondered if the America’s Army diagrams were any more intelligible? He also brought up the survivorship bias problem; you can’t know if you’ll get something compelling out of your data set before you start attacking it. You never know which question you ask of it will produce something compelling or surprising.
JC agreed with this, and thought the answer was to emphasize the analysis stage, rather than skipping it.
A lady, whose name I didn’t catch, (edit- Marti Hearst, the Berkeley professor, who was in the "Next level discovery" panel, thanks Matt and apologies Marti, you were out of line of sight for me) said that it’s very hard to make visualizations. The reason that the Amazon political preferences example is still used, years after it was created, is that it’s tough to create something that compelling, and it also needs some luck. She agreed with Matt that the army example was confusing.
JC’s response was that you do need a lot of domain knowledge to be effective.
the same lady Marti brought up was that in the high school example, SMS patterns could be incredibly important if it was a forensic investigation after a school shooting.
As support for this idea, JC brought up the example of the correlation between Snickers sales in prison commissaries, and riots. It’s apparently the best method of predicting riots, much better than more obvious metrics like violent incidents, because the inmates know something is brewing, and want to stock up on food.
Mershad Setayesh from Collective Intellect said that visualizations only make sense if you can see a pattern. How do you do that? Is there a methodology to get patterns from your data?
JC suggested various methods, applying calculus, and bundling points up using natural analysis, to find things like cliques.