Blog post

Big Botched Data

By Svetlana Sicular | July 29, 2014 | 1 Comment

Information Everywheredata paprazziData and Analytics Strategiesdataanalytics

Everybody talks about successes in big data. And everybody is curious about failures.  Today, I want to illustrate some typical causes of big data project failures with real-life examples, no company logos to show, sorry.  I’ll give not necessarily “fail fast” scenarios, but also the uneventful and painful “fail slow.” Let’s start with the amazing success story.

Management inertia. Our client, a household name among early internet travel companies, as well as the early adopter of big data technologies, ran click-stream analysis to find out how people navigate this travel site and how they make purchases. It turned out that the buying patterns were exactly opposite from the sales approach of the company’s upper management. This is the verbatim quote about this rare happy end:

“We’ve had great success with this technology. The insights we’ve had changed the business dramatically. To capitalize on these insights we brought in new management.

How many companies are in a position to get rid of their upper management?

Selecting wrong use cases. Many companies start with advanced use cases that require a better understanding of technologies, which comes with experience. Other companies select the same use cases that they used to implementing on traditional technologies, and, consequently, they don’t see benefits. My blog post The Top Mistake in Evaluating Big Data Initiatives describes this situation.

Asking wrong questions. An automobile manufacturer with thousands of dealerships ran a sentiment analysis project to learn about its customers. Six months and $10M later the findings from big data were distributed to all thousands of dealerships, and all thousands of them were laughing out loud: every one of them knew all along what the big data project was digging out all this time.

Lacking the right skills. Every one of us considers him/herself an expert in human behavior, our native language or our own social life. So are people running big data analytics projects. A financial services company started a project to detect how people’s habits affect their propensity to buy retirement plans. Humans are creatures of habits, and of too many habits. People who ran the project decided (little by little, failing slowly) to narrow down all habits just to smoking / non-smoking. And failed again.  It turned out (from my dialogs with a healthcare company, which coincided with this one) that healthcare professionals instead of a black-and-white “do you smoke?” would have asked, ” how many years did you smoke? How many times did you quit smoking? When was the last time you smoked?” The bottom line: look for professionals who know the field you analyze — healthcare experts, linguists, behavioral psychologists, social anthropologists and others who normally don’t belong to IT.

Unanticipated problems that are wider than just a big data technology. One large retailer ran a big data project in the cloud. The network congestion to stores was a problem that derailed the whole project.  A team member summarized their learning from the failure:

“Supporting any new platforms on a remote site is more than a technology problem.  It must factor in personnel, training, upgrades, maintenance and real estate.”

Disagreement on the enterprise strategy. There are many trains of thought in a large company. Here is an eloquent quote from a client, an information architect:

“We see information as the heart. Others believe cloud is the heart of our strategy.”

As a result, there is no enterprise-wide strategy, but a lot of unrelated initiatives, big data being rather small.

Siloed big data negates the whole idea of having it. This reason for failure relates to the previous one. A client who learned it on his own mistakes said:

“Prioritization of business projects is a bit more difficult because we are so siloed in business units. We do not do a good job justifying the platform as a whole.  Whoever screams loudest gets it.”

Solution avoidance. The most typical example is pharmaceutical industry required to report any known adverse drug effects. This whole industry does not conduct sentiment analysis, because they have to report to FDA any event when, for example, a patient complains about a headache in the same paragraph where a particular drug is mentioned.

My list of big data failures can go on, and on, and on. I especially want to stress the need to understand the data, no matter if it’s big or not. There are tons of cases of not knowing data, and, as a result, inability to deliver anything new, or having so much data and no experience of  how to manage, analyze or query it. I will talk about data, big data and greater data in two weeks from now, at our Catalyst conference in San Diego. Come over!

Follow Svetlana on Twitter @Sve_Sic

Comments are closed

1 Comment

  • CA says:

    Excellent article. I certainly appreciate this site. Thanks!