Digitally Training Graph Viewers against Misleading Bar Charts
Bar charts are common visual tools used to convey statistical information. Even though bar charts are effective in making abstract concepts more accessible, poorly-designed bar charts - whether designed intentionally or unintentionally - can easily mislead the viewer. For example, a poorly-designed bar chart may only present part of the effective range on the vertical axis. This exaggerates the contrast among bars, leading an unsuspecting graph viewer to wrong conclusions. More broadly, misrepresentation in data visualization is becoming an increasing societal problem contributing to daily misinformation. This paper presents a computational and cognitive solution to this problem. Our idea is to train viewers by showing them a few dozen carefully designed bar charts that are misleading, together with guidance on why these bar charts are misleading. We then test whether the viewers identify similarly misleading bar charts in the future. Importantly, we use neural networks and cognitive models to optimize the training (i.e., the design of those few dozen bar charts). Our experiment shows that perceptual trainings can help viewers not be fooled by similar misleading graphs in the future.