Why You’re Not Getting Value from Your Data Science

Businesses generate enormous amounts of data that doesn’t always translate to actionable information. Over the past several years, my MIT research team and I have sought to understand how businesses can realize the potential of their data repositories through machine learning.

One problem is that machine learning experts often don’t build their work around the final objective — which is to derive business value. In most cases, predictive models are meant to improve efficiency, increase revenue or reduce costs. But the people actually working on the models rarely ask, “What value does this model provide, and how can we measure it?”

Asking this question about value proposition often leads to a change in formulating the original problem, which can be more useful than tweaking the process later.

If companies want to get value from their data, they need to focus on accelerating human understanding of data, scaling the number of modeling questions they can ask of that data in a timely fashion, and assessing the results. In our work with companies, we decided that machine learning’s impact will come from four principles:

1. Stick with simple models: Focus on reducing the time between the data acquisition and the development of the first simple predictive model.

2. Explore more problems: Data scientists need to able to define and explore multiple prediction problems quickly and easily. Instead of exploring one business problem with a super-sophisticated machine learning model, companies should explore dozens of problems, building a simple predictive model for each one and assessing their value.

3. Learn from a sample of data, not all the data: Instead of tackling big data all at once, invest in techniques to draw conclusions from subsamples. Thus you can explore more hypotheses.

4. Emphasize automation: Frequently, across different data problems, we find ourselves applying similar data processing techniques, whether to aggregate data or to prepare it for predictive modeling. It’s time to streamline these processes by developing algorithms that perform them automatically.