Some of the largest companies in banking, defense, energy, pharmaceuticals and other industries sent tech leaders to Cambridge earlier this month to help answer questions many of them have in common, despite their disparate industries. Among those questions were several we heard over and over again: How can we empower our data scientists to be more productive, to conserve resources through automation and move machine learning products into production faster?
The Massachusetts Institute of Technology (MIT) sponsored the gathering – a meetup of its 25 Startup Exchange (STEX 25) accelerator companies and multinational corporations eager to leverage new technologies. STEX25 supports MIT-affiliated, industry-ready companies poised for significant growth. As a STEX 25 company, Feature Labs attended to showcase our feature engineering capabilities and hear the feedback of professionals in the field.
What we learned is that companies from Northrop Grumman to Bayer to HSBC and others are trying to figure out how to leverage automation to move their data scientists from the backroom to the forefront of their business strategy efforts.
The use cases varied among the companies we spoke to. Some wanted the means to make their credit card fraud systems more secure. Others wanted to analyze border control data to better predict threats. We had another conversation about leveraging data to assess pilot condition. The thread between every meeting we had at the event was about leveraging machine learning for everyday work in a way that is easier for domain experts and data scientists to use in their workflow.
Why most machine learning projects never see the light of day
The hype around machine learning grows every minute on a global scale, yet many projects in most enterprises never see the light of day. In his presentation at the conference, Feature Labs co-founder and CEO Max Kanter explained the reason for that.
To get value from data science, companies have to:
- Identify historical examples through prediction engineering;
- Calculate explanatory variables through feature engineering; and
- Learn rules to map features to outcomes through machine learning.
Essentially, data science is a human-driven, iterative process, which can lead to imprecise results – issues that often aren’t discovered until after the fact, requiring teams to pursue tedious post-mortem evaluations. By automating the feature engineering process, data scientists can accelerate the most error prone, time intensive and costly part of their work to intelligently transform raw data for machine learning algorithms.
Feature Labs’ founders Kanter, Ben Schreck and Kalyan Veeramachaneni wrote about this in a recent Harvard Business Review article that focused on our work with Accenture as an example. Along with two other co-authors, they write, “Currently, the AI project manager (tested and integrated in Accenture's myWizard Automation Platform across delivery projects) serves predictions on a weekly basis and correctly predicts red flags 80% of the time, helping to improve KPIs related to project delivery.”
The value of removing the feature engineering bottleneck from data science resonates with companies. Whether the data they want to leverage relates to customer behavior, sales and marketing, enterprise operations or other areas, enterprises across industries are compelled by the idea of building more accurate machine learning models, more quickly.