Which scatterplot is preferred?

Next week, we’ll be presenting a machine learning (ML) and visualization paper at the IV2019 conference in Paris. The core idea is to tag and display relevant news headlines in a real-time ambient visualization system.

From an ML perspective, the challenge is to use an open source news dataset (e.g. GDELT) where thousands of headlines are available and updated frequently (e.g. every 15 minutes) but the tags provided don’t match the needs. In general, classifying news stories is an ongoing challenge, as new topics and words emerge (e.g. Brexit, tariffs), and topics may change over time (e.g. Clinton, environment, etc.) We provide a classification module where expert users start with a simple text search against the headlines. Then we automatically suggest additional relevant keywords, which the user may explore, add, or remove. Additionally, the user may inspect sample headlines associated with any of the keywords, grouped by similarity. Then the the user can explicitly select target headlines and keywords that match their intended search topic (or not). Finally, we can run a classifier to tag all the headlines with respect to this topic defined by sample keywords and headlines. The user can iterate, modify, reclassify, and so on. Lots more detail on the technical approach is in the paper.

From the visualization perspective, the challenge is how to display a subset of these headlines. We were working with an existing animated ambient visualization which provided either a scatterplot or map upon which we could display headlines and would automatically select headlines at random to popup. Note that a map with point data is essentially the same as a scatterplot: the x and y location of the data points is based on latitude and longitude; plus an underlying image is used to show geographic features, such as land/sea.

We created four scatterplot variations: 1) a scatterplot map (with underlying land/sea image); 2) a scatterplot with explicit axes (recency vs. # stories); 3) a scatterplot based on a multi-dimensional projection (e.g. PCA or t-SNE); and 4) scatterplot based on random layout (i.e. like a wordcloud) – which is visually similar to the multi-dimensional projection:

VisIRML1
VisIRML2
VisIRML3

These visualizations were then reviewed with a small target group of users, in a meeting environment. This is where things get really interesting:

  1. Map: Everyone likes the map. It is immediately understandable.
  2. Scatterplot with explicit axes: Everyone logically understands the representation, but otherwise largely unenthusiastic. A few people are very interested, but given the broad audience for an ambient visualization, there is not enough support to push this into production.
  3. Scatterplot projection: The experts are not expert in multidimensional data projection. Multidimensional data projection is difficult to explain and difficult for people to comprehend. They can’t use what they don’t understand.  Unfortunately, dead on arrival.
  4. Scatterplot random layout (aka cloud): Surprisingly, some people really like word clouds, but this community of experts is not interested. It’s art, not information.

So, the map wins: there is a strong preference for the map over all other scatterplot variants. Why should the map have such a strong preference? Unfortunately, the project didn’t have scope to consider this, and there are various confounding effects going on too: titles are smaller on the map, the scatterplot also uses color-coding of text, the map had leader lines to association headlines with locations, etc.

Here are some hypotheses why there was such a strong preference for maps:

a. Maps are easy to decode. The specific map we used is always global. A global map is easy to decode because people are very familiar with them. A global map has very low cognitive load. Low cognitive load may be very important in an ambient visualization as people don’t want to have to think too hard about what they are seeing. Scatterplots, however, have higher cognitive load. You have to be actively engaged to decode it: you have to reference back and forth between data points and an axis, or you need to understand what a projection means. And a cloud doesn’t have anything to decode positionally, so there’s no information there. From an information standpoint, the map provides more information than a random cloud.

b. Maps automatically engage prior knowledge. A news headline situated in Iowa means that we can bring all of our knowledge about Iowa immediately as context to interpret the visual representation. If the news headline is about corn, that probably matches prior knowledge that Iowa is farmland and may have a lot of corn crops.

c. Maps are visceral. In Norman’s Design of Everyday Things, some objects elicit a visceral response. A visceral responses are immediate and emotive. The map is immediately accessible and engaging. In a ranking of maps vs. scatterplot with axes vs. multidimensional projection scatterplot, my hunch is that the map is most viscerally engaging.

Thoughts? Comments?

 

 

About richardbrath

Richard is a long time visualization designer and researcher. Professionally, I am one of the partners of Uncharted Software Inc. I have recently completed a PhD in data visualization at LSBU. The opinions on this blog are related to my personal interests in data visualization, particularly around research interests related to my PhD work- this blog is about exploratory aspects of data visualization not proven principles.
This entry was posted in Data Visualization, Text Visualization and tagged . Bookmark the permalink.

Leave a Reply

Fill in your details below or click an icon to log in:

WordPress.com Logo

You are commenting using your WordPress.com account. Log Out /  Change )

Google photo

You are commenting using your Google account. Log Out /  Change )

Twitter picture

You are commenting using your Twitter account. Log Out /  Change )

Facebook photo

You are commenting using your Facebook account. Log Out /  Change )

Connecting to %s