CHAPTER 18

When Data Visualization Works—and When It Doesn’t

by Jim Stikeleather

I am uncomfortable with the growing emphasis on big data and its stylist, visualization. Don’t get me wrong—I love infographic representations of large data sets. The value of representing information concisely and effectively dates back to Florence Nightingale, when she developed a new type of pie chart to clearly show that more soldiers were dying from preventable illnesses than from their wounds. On the other hand, I see beautiful exercises in special effects that show off statistical and technical skills, but do not clearly serve an informing purpose. That’s what makes me squirm.

Ultimately, data visualization is about communicating an idea that will drive action. Understanding the criteria for information to provide valuable insights and the reasoning behind constructing data visualizations will help you do that with efficiency and impact.

For information to provide valuable insights, it must be interpretable, relevant, and novel. With so much unstructured data today, it is critical that the data being analyzed generates interpretable information. Collecting lots of data without the associated metadata—such as what is it, where was it collected, when, how, and by whom—reduces the opportunity to play with, interpret, and draw conclusions from the data. It must also be relevant to the people who are looking to gain insights, and to the purpose for which the information is being examined (see the sidebar “Understand Your Audience”). Finally, it must be original, or shed new light on an area. If the information fails any one of these criteria, then no visualization can make it valuable. That means that only a tiny slice of the data we can bring to life visually will actually be worth the effort.

Once we’ve narrowed the universe of data down to that which satisfies these three requirements, we must also understand the legitimate reasons to construct data visualizations, and recognize what factors affect the quality of data visualizations. There are three broad reasons for visualizing data:

• Confirmation: If we already have a set of assumptions about how the system we are interested in operates—for example, a market, customers, or competitors—visualizations can help us check those assumptions. They can also enable us to observe whether the underlying system has deviated from the model we had and assess the risk of the actions we are about to undertake based on those assumptions. You see this approach in some enterprise dashboards.
• Education: There are two forms of education that visualization offers. One is simply reporting: here is how we measure the underlying system of interest, and here are the values of those measures in some comparative form—for instance, over time, or against other systems or models. The other is to develop intuition and new insights on the behavior of a known system as it evolves and changes over time, so that humans can get an experiential feel of the system in an extremely compressed time frame. You often see this model in the “gamification” of training and development.
• Exploration: When we have large sets of data about a system we are interested in and the goal is to provide optimal human-machine inter actions (HMI) to that data to tease out relationships, processes, models, etc., we can use visualization to help build a model to allow us to predict and better manage the system. The practice of using visual discovery in lieu of statistics is called exploratory data analysis (EDA), and too few businesses make use of it.

Assuming the visualization creator has gotten it all right—a well-defined purpose, the necessary and sufficient amount of data and metadata to make the visualization interpretable, enabling relevant and original insights for the business—what gives us confidence that these findings are now worthy of action? Our ability to understand and to a degree control three areas of risk can define the visualization’s resulting value to the business:

• Data quality: The quality of the underlying data is crucial to the value of visualization. How complete and reliable is it? As with all analytical processes, putting garbage in means getting garbage out.
• Context: The point of visualization is to make large amounts of data approachable so we can apply our evolutionarily honed pattern detection computer—our brain—to draw insights from it. To do so, we need to access all of the potential relationships of the data elements. This context is the source of insight. To leave out any contextual information or metadata (or more appropriately, “metacontent”) is to risk hampering our understanding.
• Biases: The creator of the visualization may influence the visualization’s semantics and the syntax of the elements through color choices, positioning, and visual tricks (such as unnecessary 3D, or 2D when 3D is more informative)—any of which can challenge the interpretation of the data. This also creates the risk of pre-specifying discoverable features and results via the embedded algorithms used by the creator (something EDA is intended to overcome). These in turn can significantly influence how viewers under stand the visualization, and what insight they will gather from it.

Ignoring these requirements and risks can undermine the visualization’s purpose and confuse rather than enlighten.

__________

Jim Stikeleather, DBA, is a serial entrepreneur and was formerly Chief Innovation Officer at Dell. He teaches innovation, business models, strategy, governance, and change management at the graduate level at the University of South Florida and The Innovation Academy at Trinity College Dublin. He is also a senior executive coach.

Adapted from content posted on hbr.org, March 27, 2013 (product #H00ADJ).