Sanitation Inspector Allocation in San Francisco Using Yelp Reviews

• Activity Number: 180
• Type: Contributed
• Date/Time: Monday, August 4, 2014 : 10:30 AM to 12:20 PM
• Sponsor: Section on Statistical Learning and Data Mining
• Author(s): Guillaume Pouliot*+ and Michael Luca
• Companies: and Harvard Business School
• Keywords: latent Dirichlet allocation ; Bayesian Support Vector Machines ; Variational Inference ; Topic Models ; Supervised Learning

Abstract

The city of San Francisco has a limited number of sanitation inspectors. It would be more efficient to send inspectors to restaurants that are more likely to get caught for sanitation norm violations. Our data set consists of a comprehensive list of San Francisco restaurants, their general characteristics and Yelp reviews (obtained from Yelp), as well as their health inspection records. The first-order task is to predict who will commit what infraction in the future. Many methods are compared. In particular, we develop a latent Dirichlet allocation (LDA) topic model supervised with multicategory Bayesian support vector machines. We compare performance of inference via Gibbs sampling and variational inference. We find that this method improves on LDA supervised with a generalized linear model.