Follow Datanami:
June 17, 2016

Behavior I/O: Using Machine Learning to Empower Human Learning

Satyen Sangani


At lunch last week, I learned that a couple colleagues were engaged in a little duel—trying to out-walk each other, as tracked by their new Fitbits. Self-improvement was definitely the goal, but seeing peer performance and benchmarks provided the required motivation to achieve that goal.

If you’re a data science or analytics leader, your job is to manage analysts who produce insights.  Insights drive innovation which in turn drive profits.  So, if you want to drive your business through these insights, you have two options.  You can either hire more analysts or you can increase the productivity of your existing analysts.  The first option is expensive and difficult. So, you’re left with the question of how to change the behavior of your analysts (or, really, anyone seeking self-service data).

Historically, we’ve solved this problem in two ways–first by training our analysts and second, by commissioning knowledge management initiatives requiring experts to document knowledge.  Anyone who has participated in a knowledge management project knows how fundamentally unscalable the approach is – no one likes to write documentation.

Data scientists are familiar with the concept of a supervised learning system.  Algorithms get better as they receive more feedback.  A well known example of a supervised algorithm is LinkedIn’s “People you may know.”  As people confirm good matches, they’re also teaching the machine what a good match is.  The machine subsequently gets better and better at suggesting matches.  The humans train the machines.

You can think of this virtuous cycle as “Behavior I/O:”  In the consumer world, many companies like Fitbit (NYSE: FIT) and LinkedIn (NYSE: LNKD) learn from people’s behavior to help train other people to behave better.  My co-founders and I asked the question: could you do the same thing in the realm of analysis?  Could you use machine learning to observe behavior of analysts, ultimately using those observations to improve how their colleagues use data?

But how could we possibly observe the behavior of analysts? By observing the queries they write and the reports they generate.

Even with such a rich corpus,  there are still a lot of problems to solve in implementing a behavioral learning system that actually helps drive behavior.

  • Incorrect behavior – One of the most insidious problems is that people often behave incorrectly.  BI/DW professionals lament that people often use the wrong old report, even when the new correct report has been published for weeks.  Alternatively, they use an incorrect customer list, even when a new master has been published.  People can behave incorrectly, and the last thing you want to do is perpetuate bad behavior.
  • Lack of specificity – The tough thing about data is that, if you’ve ever worked with it, you realize that you quickly get in to the weeds.  If you’re going to a Tableau workbook that everyone else uses, you’d better tell me what they used down to the sheet and perhaps even the metric.  Otherwise, I might come to the wrong conclusion.
  • Inadequate context – Even if I know that 323 people ran a specific query in the last 20 days, can you tell me who used it and why?  What if 90% of those runs were from an outdated workload, initiated by an employee who left the company 5 months ago?  What if I’m in marketing and I’m looking at a customer report where 60% of the users were in finance and the other 40% are in operations. If I could know that no one on my team uses this report, maybe there’s some context that I’m missing.
  • Data Gaps– Sometimes there’s just no data at all.  Maybe you can’t find the data set, the data set just does not exist inside of the company, or it’s available only for purchase?  Perhaps my use case does not, by itself, justify the expense of purchasing or producing this data.  However, if I knew that my colleagues had similar data needs or were asking similar questions, maybe we could collectively justify the investment.  Whose keeping track of all those unanswered questions?

Notwithstanding the challenges above, I’m a big believer in the value of Behavior I/O and it’s power to aid both data consumption and the governance of self-service data.

We’ve historically applied embedded analytics to front line workers like customer  service and telesales where we can prompt hundreds, if not thousands of workers, that have the same effective job.  Behavior I/O allows us to broaden the application of just-in-time analytics to a variety of high order job functions.  The opportunity exists to change the way people think, by highlighting the best thoughts, ideas and questions of their peers.  Behavior I/O is a much faster form of collaboration, allowing us to collectively learn more easily and quickly that we’ve done before.


About the author:  Satyen Sangani is the CEO and co-founder of Alation.  Before Alation, Satyen spent nearly a decade at Oracle, ultimately running the Financial Services Warehousing and Performance Management business where he helped customers get insights out of their systems.  Prior to Oracle, Satyen was an Associate with the Texas Pacific Group and an Analyst with Morgan Stanley & Co.  Satyen holds a Masters from the University of Oxford and a Bachelors from Columbia College, both in Economics.