How Intel and Burger King built an order recommendation system that preserves customer privacy

The pandemic has placed enormous strains on the restaurant and fast food industry. Within a month of the health crises, 3% of restaurants had closed for good and another 11% anticipated doing so within the next month, according to a National Restaurant Association study. While fine, idlevel, and casual dining establishments suffered the bulk of the impact, the fast food industry wasn’t immune. A Datassential survey found that sales among fast food operators declined 42% during the first few weeks of the pandemic.

As more and more customers began relying on take-out and drive-thru options versus indoor dining, fast food retailers like Burger King turned to AI and machine learning for solutions. In collaboration with Intel, Burger King developed an AI system that recommends items on touchscreen menu boards to customers as they’re about to order. It can predict if a customer will order a hot or cold drink or a light or large meal, potentially saving time and leading to a better customer experience.

Burger King and Intel say the solution has already been piloted in over 1,000 Burger King locations. 

Burger King isn’t the first fast food chain to experiment with AI in customer service. McDonald’s has been using AI in the drive-thru since acquiring tech company Dynamic Yield in 2019. Dunkin’ Donuts is testing drive-thrus that can recognize a loyalty member as soon as they pull up. Some Sonic drive-ins recently got AI-powered menu kiosks. And Chick-fil-A is using AI to spot signs of foodborne illness from social media posts.

As Luyang Wang, director of advanced analytics and machine learning at Burger King, explained to VentureBeat via email, fast food recommendation has its own set of unique challenges compared with, say, ecommerce. There’s no easy way to identify customers and retrieve their profiles, because all of the recommendations happen offline. Moreover, context features like location, time, and weather conditions have to be preprocessed before they can be loaded into a model.

To solve these challenges, TxT was built with what’s called a “double” Transformer architecture that learns both real-time order sequence data as well as features like location, weather, and order behavior. TxT leverages all data points available in a restaurant without having to identify customers prior to the order-taking process. For example, if a customer puts a milkshake as the first item in their basket, it’ll influence what TxT suggests based on what’s been sold in the past, what’s selling today, and what’s sold at that store.

TxT was developed within Analytics Zoo, Intel’s open source platform for big data analytics workloads running in datacenters. Intel and Burger King collaborated to create an end-to-end recommendation pipeline, which includes distributed Apache Spark data processing and Apache MXNet training on an Intel Xeon cluster. The TxT model was deployed using Intel’s RayOnSpark library, which allows enterprises to directly run programs on existing clusters.

According to Wang, TxT has already led to surprising sales insights. For one, Burger King customers will order milkshakes in any weather — even when it’s cold out. And people are much more willing to add a dessert when they have a high-calorie basket versus low-calorie basket.

“At Burger King, we are always looking to improve our guests’ experience,” Wang said. “This AI recommender system — Transformer Cross Transformer (TxT) — allows Burger King to better learn customer habits and, essentially, better communicate with guests.”