Businesses today are constantly generating enormous amounts of data, but that doesn’t always translate to actionable information. Over the past several years, my research group at MIT and I have sought answers to a fundamental question: What would it take for businesses to realize the full potential of their data repositories with machine learning?
As we worked to design machine learning–based solutions with a variety of industry partners, we were surprised to find that the existing answers to this question often didn’t apply. Why?
First, whenever we spoke with machine learning experts (data scientists focused on training and testing predictive models) about the most difficult part of their job, they said again and again, “the data is a mess.” Initially, taking that statement literally, we imagined it referred to well-known issues with data — missing values or a lack of coherence across databases. But as we dug deeper, we realized the problem was slightly different. In its rawest form, even clean data is too overwhelming and complex to be understood at first glance, even by experts. It has too many tables and fields and is often collected at a very high granularity (for example, online clickstreams generate new data with every click, and sensor data is collected at 125 observations per second). Machine learning experts are used to working with data that’s already been aggregated into useful variables, such as the number of website visits by a user, rather than a table of every action the user has ever taken on the site.
At the same time, we often heard business experts complain that “we have a lot of data and we are not doing anything with it.” Further investigation revealed that this was not strictly correct either. Instead, this frustration stems from two problems. For one thing, due to the time it takes to understand, formulate, and process data for a machine learning problem, machine learning experts often instead focus on the later parts of the pipeline—trying different models, or tuning the hyperparameters of the model once a problem is formulated, rather than formulating newer predictive questions for different business problems. Therefore, while business experts are coming up with problems, machine learning experts cannot always keep up.
For another, machine learning experts often didn’t build their work around the final objective—deriving business value. In most cases, predictive models are meant to improve efficiency, increase revenue, or reduce costs. But the folks actually working on the models rarely ask “what value does this predictive model provide, and how can we measure it?” Asking this question about value proposition often leads to a change in the original problem formulation, and asking such questions is often more useful than tweaking later stages of the process. At a recent panel filled with machine learning enthusiasts, I polled the audience of about 150 people, asking “How many of you have built a machine learning model?” Roughly one-third raised their hands.