Why machine learning is still getting the human touch in retail

Handing power back to the people who know their business is becoming a more common practice, even in environments where big data has taken a driver’s seat on the technology side. That’s especially true in retail, where marketing campaigns, overstocked inventory or other concerns might trump straight-up algorithmic recommendations.

On Tuesday, for example, software-as-a-service startup BloomReach announced a new product that tries to give marketing and merchandising experts more control over their websites. Called SNAP (short for search, navigation and personalization) it lets the people in charge of product placement on retail websites exercise their judgment over BloomReach’s machine learning algorithms via a visual interface for determining what content is shown and whether personalization is turned on at all.

We have covered BloomReach’s service before, but the short explanation is that it’s a cloud service for determining what content retail websites should show visitors. That might mean displaying a sponsored search result that more closely matches what someone is seeking, or recommending new product terminology that’s in line with how competitors are describing their wares. Underpinning the service is an expansive data pipeline that analyzes billions of behavioral data points per day to learn what consumers want and how websites are delivering it to them.

Previous incarnations of the BloomReach service acted like a “black box,” BloomReach head of marketing Joelle Kaufman said, whereas SNAP is designed to give users much more input. Tuned to maximum personalization and automation, it will recommend personalized content across devices (BloomReach analyzes web behavior to determine if someone is likely the same person) and use its huge set of synonym pairs to intelligently match what users are looking for with other products. However, merchandising specialists have the option of tuning down personalization all the way down to deciding what products are displayed and where in search results.

Deb uses BloomReach. How results are displayed depends on what rules are in place.
Deb uses BloomReach. How results are displayed depends on what rules are in place.

However, BloomReach hasn’t stumbled upon some novel idea for improving the application of machine learning in e-commerce. Christophe Bisciglia, co-founder and CEO of machine learning startup WibiData recently explained to me how his company launched a similar rules engine for its retail customers. The recommendation system and data scientists might want to do one thing, he explained, but might not realize the company is trying to push Nike products, for example.

Or, sometimes, the recommendations from products such as BloomReach are just way off base and need to be corrected.

We examined this theme at a broader level during our Structure Data conference last year in a session (embedded below) that includeded another e-commerce machine learning specialist, Scott Brave of Baynote. The topic will likely come up again at this year’s event, which takes place next month in New York. Bruce Yen of clothing company Guess will discuss his company’s use of analytics to make merchandising decisions, and we’ll focus on advanced artificial intelligence techniques that will likely make their way into retail and other industries soon.

As smart as the models might be, though, many companies still (and probably wisely) view them as a guide rather than the word of god, and still want to put the ultimate control in the hands of the people they hired to make such decisions.