Alex Kale, Matthew Kay, and Jessica Hullman

Copilot can do a better job at visualizing uncertainty. Specifically when it comes to:

1. significance (or lack thereof) based on number of impressions

2. dealing with bad/uncertain data (e.g. ""unknown"" placement type)

3. predicting future results- “how likely are you to hit your goal”, “how likely are you to deliver in full”

The last point specifically sparked some discussion- maybe something similar to the NYTimes needle visualizations (link) could be used.

Our problem: uncertainty is a hard topic for our hands on keyboard users. We should expect some pushback from users when it comes to visualizing this as real advertiser money is on the line. Traders tend to anchor to a single number and are less comfortable with ranges. How can we could better manage user expectations and build trust while incorporating uncertainty?

• Introduce users to new chart types, color scales, etc. that studies have shown to be more effective?Meet them where they are visualizing data with chart types they are used to?

• We landed on a gradual introduction of new visualization techniques as well as always providing information on why certain changes were made.

Sharon Lin, Julie Fortuna, Chinmay Kulkarni, Jeffrey Heer from Stanford University; Maureen Stone from Tableau Software

This paper focuses on categorical color assignment for data visualizations and explores how linguistic information about the terms defining the data can be used to generate semantically meaningful colors.

The study also explores the idea of colorability – or “how well a given categorical value set maps to semantically-resonant colors”. In our discussions about this concept, we found that our Product Analytics team always uses blue to represent TTD, Green for DV360, and Orange for APN – which reduces the cognitive load on our traders who are well aware of these colors, and their associations with the corresponding DSPs.

We also always use the copilot blue to compare copilot vs non-copilot line items in our A/B visualizations.

While having an algorithm to make these choices for us sounds like a great idea, we think we also need to be mindful to not let it reinforce gender and racial stereotypes (for instance, using blue for “boy” and pink for “girl”) .

Richard Brath and Ebad Banissi from London South Bank University, UK

In the data viz world, there exist a lot of visualizations about words and text, but not so many cases of data encoded by the properties of words and sentences. It was cool to see examples of both in this paper.

This paper explored the design space around typography + data viz, and one of our main takeaways is that it can be extremely fruitful to explore domains outside of data visualization to discover novel, creative solutions to old problems.

We are excited to try our hand at using font attributes to encode data; a couple of applications that immediately came to mind:

Using font obliqueness or opacity to encode uncertainty

Using font family to include hierarchy level (Ad Group vs. Campaign)

Michelle A. Borkin, Zoya Bylinskii, Nam Wook Kim, Constance May Bainbridge, Chelsea S. Yeh, Daniel Borkin, Hanspeter Pfister, and Aude Oliva

This study first explores what aspects of a visualization make it memorable, and goes further to assess what makes a visualization effective.

Visualizing data in novel ways often helps in memorability and effectiveness, as opposed to utilizing basic chart types like bar charts and line charts. We have also seen evidence of that in our visualization of a clustering-based algorithm.

Titles and text are other important factors that impact chart effectiveness – In our new Insights visualizations, we have explored this idea to include a title in the form of a question, along with a subtitle that goes into more specific details. We have also include an info box that elaborately explains each visualization.

While this research is limited to static visualizations, we’re excited to explore how animations and interactivity make visualizations more effective.