AI Case Study
Researchers at Stanford and Columbia predict restaurant affinity among lunch-goers in the Bay area using machine learning
Researchers at Stanford and Columbia universities uscombined SafeGraph’s Movement Panel, a dataset anonymized, high-quality GPS pings from roughly 10% of U.S. smartphones and data from Yelp on Bay Area restaurants to develop the Travel Time Factorization Model. The TTFM is a Bayesian machine learning model used to analyse the complex and dynamic decision making process of choosing a restaurant for lunch. The model considers factors such as willingness to travel, availability of options and anonymous consumers' preferences. In the Bay Area, it showed that people will travel far for Japanese but not for pizza, for example. The system can be used to answer crucial questions for
restaurateurs, retailers and real estate investors.
Industry
Consumer Goods And Services
Food Beverage And Drugs
Project Overview
"As part of a recent research project, Susan Athey, Robert Donnelly and I, Tobias Schmidt (Stanford University), as well as David Blei and Francisco Ruiz (Columbia) built the Travel Time Factorization Model (TTFM), a Bayesian machine learning model that allows us to study the complex, dynamic, often unobserved factors that go into how people choose which restaurant to visit for lunch.
The model uses SafeGraph’s Movement Panel, a dataset of anonymized, high-quality GPS pings from roughly 10% of U.S. smartphones, as well as data on Bay Area restaurants from Yelp.
Like most machine learning models, TTFM can be used to make predictions, e.g. about the number of visitors certain restaurants can expect to see over time.
But TTFM can also imagine. It can tell us how visit patterns would change if certain restaurants close, or what would happen to the local pizza shop if a café opens nearby. It can tell restaurant entrepreneurs which kind of Asian cuisine to serve in a given neighborhood to maximize foot traffic, and it can help urban planners understand how far people are willing to travel to get their favorite sandwich. TTFM can also help retailers, real estate agents, and economists understand what it really means for a neighboring restaurant to be a competitor.
We began by feeding SafeGraph’s anonymous GPS data into TTFM, which then models the choice of where to eat for lunch as a function of two variables:
1) the restaurant options available, and
2) the dynamic preferences that each anonymous consumer has for certain foods, price points, and distances, all of which were inferred from location data. This data allowed us to empirically observe the habitual movement patterns of these consumers at lunchtime.
Immediately, we had the ability to look at market segmentation and competition between restaurants in a way that historically has been difficult for human analysts.
Studying willingness to travel shows stark differences in how far people are willing to travel to get their lunch. We categorized roughly 5,000 restaurants in the Bay Area by cuisine and used TTFM to rank these categories by willingness to travel.
Reported Results
"TTFM lets us understand how far people will travel for different foods. In the Bay Area, people will travel far for Japanese but not for pizza.
Unsurprisingly, people aren’t willing to go far for cheap eats but are more willing to travel farther to upscale restaurants. People are willing to go the proverbial — and literal — extra mile for Japanese, Vietnamese and New American restaurants but not for sandwiches, pizza and Mexican food."
Technology
Travel Time Factorization Model (TTFM) is a Bayesian machine learning model
Function
Strategy
Strategic Planning
Background
"Where should a new restaurant be located? What type of restaurant would be best in a given location? How will competition down the block or across town impact market share? Historically, real estate site selection has been a labor intensive and error prone process for business owners and investors, who must often rely on expensive surveys and focus groups to understand what people want and where they want it."
Benefits
Data
"SafeGraph’s Movement Panel, a dataset of anonymized, high-quality GPS pings from roughly 10% of U.S. smartphones, as well as data on Bay Area restaurants from Yelp."