Activity-based effectiveness of fundamental visualizations

Task-based effectiveness of basic visualizations Saket et al., IEEE Transactions on Visualization and Pc Graphics 2019

Up to now this week we’ve seen the right way to create all kinds of improbable interactive visualisations, and brought a take a look at what information analysts truly do after they do ‘exploratory data analysis.’ To spherical off the week right now’s selection is a latest paper on an age-old matter: what visualisation ought to I take advantage of?

No prizes for guessing “it depends!”

…the effectiveness of a visualization relies on a number of components together with job on the hand, and information attributes and datasets visualized.

Is that this the paper to lastly settle the age-old debate surrounding pie-charts??

Saket et al. take a look at 5 of essentially the most fundamental visualisations —bar charts, line charts, pie charts, scatterplots, and tables— and examine their effectiveness when presenting modest quantities of knowledge (lower than 50 visible marks) throughout 10 completely different duties. The duty taxonomy comes from the work of Amar et al., describing a set of ten low-level evaluation duties that describe customers’ actions whereas utilizing visualization instruments.

  1. Discovering anomalies
  2. Discovering clusters (counting the variety of teams with comparable information attribute values)
  3. Discovering correlations (figuring out whether or not or not there’s a correlation between two information attributes)
  4. Computing derived values, for instance, computing an combination worth
  5. Characterising distributions, for instance, determining which share of knowledge factors have a worth over a sure threshold
  6. Discovering extremes (i.e., min and max)
  7. Filtering (discovering information factors that fulfill a situation)
  8. Ordering (rating information factors in line with some metric)
  9. Figuring out a spread (discovering the span of values – fairly easy if you will discover the extremes – #6)
  10. Retrieving a worth (figuring out the values of attributes for given factors)

If solely our customers had been utilizing SQL and never a visualisation software :).


Research design

Two datasets had been chosen such that contributors could be conversant in the that means of the information attributes, however not the precise content material: a automobiles dataset with information on 407 new automobiles, and a films dataset with particulars of 335 films. Each datasets include a mixture on nominal (categorical), ordinal (portions with a particular vary and pure ordering, e.g. film rankings), and numerical information attributes.

Following an preliminary pilot, the examine centered in on visualisations containing not more than 50 visible marks (contributors discovered it too time-consuming above this quantity). Inside this constraint visualisations had been generated throughout all give visualisation varieties utilizing the next pairwise mixtures of all of the attribute varieties: nominal * numerical, ordinal * numerical, and numerical * numerical.

180 Mechanical Turk contributors with earlier expertise of visualisation accomplished the examine. 18 contributors had been assigned to every of the ten job varieties, and contributors for a given job kind answered 30 completely different questions referring to the duty (5 visualisation varieties x 2 datasets x 3 trials). Instance job questions are proven within the determine beneath. After finishing their questions, contributors had been then requested a set of rating questions to find out their most well-liked visualisation kind(s) for performing the duty.


And the winner is…?


Outcomes, aggregated over duties and datasets, present the Bar Chart is the quickest and most correct visualization kind.

However to cease there could be to overlook the purpose. Making an allowance for the accuracy with which customers carried out a job, the time it took them, and their very own said preferences, right here’s a concise abstract of how nicely the completely different visualisation do on a per-task foundation.

Bar charts and tables are the visualisation varieties most most well-liked by contributors. This holds although efficiency when utilizing desk is comparatively sluggish and fewer correct. If you need the total particulars, try the (page-and-a-half-long!) Desk 1 within the paper. Chopping to the chase, right here’s the condensed recommendation on deciding on a visualisation kind:

  • Use bar charts for locating clusters. Efficiency on this job was comparable with each pie charts and bar charts, however customers most well-liked bar charts. Recall that ‘clusters’ right here means answering questions resembling ‘how many different movie genres are shown in the chart?’ – not the form of clustering chances are you’ll be primed to think about from kNN and many others..
  • Use line charts for locating correlations. Line charts and scatterplots are each good for this, however person favor line charts.
  • Use scatterplots for locating anomalies – they’ve excessive accuracy and pace, and are extremely most well-liked by customers on this job
  • Keep away from line charts for duties that require readers to exactly determine the worth of a particular information level – it’s more durable to determine a particular worth utilizing this visualisation kind.
  • Keep away from tables and pie charts for correlation duties, they’re much less correct, slower, and fewer most well-liked for this job.

When you’re utilizing visualisations to accompany a report or article, you may flip that round and say e.g. “use line charts to support a point about correlation” and so forth. When you simply wish to current information for readers to discover on their very own with no particular a priori job or level to assist, bar charts and tables are the go-to visualisation.


The final phrase

So, about these pie-charts…

…all through the historical past of graphical notion analysis, pie charts have been the topic of passionate arguments for and towards their use. Though the present widespread knowledge amongst visualization researchers is to keep away from them, pie charts proceed to be widespread in on a regular basis visualizations. Outcomes of our examine current a extra nuanced view of pie charts.

Must you use a pie-chart? It relies upon… they had been proven to be aggressive on cluster, extremum, filter, vary, and retrieval duties.

Original. Reposted with permission.


About the Author

Leave a Reply

Your email address will not be published. Required fields are marked *