Looks Good To Me: Visualizations As Sanity Checks
It is also part of a larger project of examining Black Hat Visualization: can visualizations designed unthinkingly or maliciously hide important information from analysts, or otherwise mislead people? We lay out this concept in a position paper for the DECISIVe 2017 workshop.
Our InfoVis paper focuses on univariate visualizations as "sanity checks": that is, visualizations that are meant to be glanced at to confirm that a given data set is reasonably free from data quality issues or other "badness." Our contention was that the adversarial design of these visualizations could result in visualizations that look plausible (that is, they seem to show an error-free "good" distribution), but make the flaw difficult to see, or indistinguishable from ordinary noise or sampling error.
This Attack folder contains code for generating adversarial univariate visualizations: that is, given a dataset into which I have injected one or more data quality issues (extraneous modes, noise, mean shifts, etc.), what is the visualization of the "bad" data that I can generate that will look the most similar to my visualization of "good" data?
The Study folder contains our experimental apparatus, analysis script, and data tables, including a data dictionary. We performed a lineup task, where participants had to identify one instance of "bad" data amongst a set of "good" datasets. More details about our methodology are available in the paper.