(Machine) Learning Preferences from Complex Choice Sets: An Application to Service Networks

Abstract

Hidden complexity pervades commonplace choices. For example, customers at everyday venues such as supermarkets, shopping centers, and amusement parks routinely choose from the combinatorially many available paths through all or some of the venues' stations (e.g., sections, stores, rides). Such choice complexity, e.g., a factorial rate of growth of paths in venue size, is captured by increasingly rich data and poses a challenge for empirical research. Firstly, it is not clear how customers assess such complex decisions. Secondly, even state-of-the-art choice estimation methods exhibit fundamental issues ranging from intractability to inconsistency when estimating consumer preferences from such data. Under commonplace conditions, we show, a minute fraction of the utility comparisons between choices suffices to capture all customer preference information from data. This provably low-dimensional structure of relevant choice comparisons can be automatically learned and exploited by machine learning (ML) and rationally justifies simple shopping heuristics for customers. We develop a neural network-based estimator of customer preferences and demonstrate it on data tracking shoppers at a large urban hypermarket serving over 1M customers annually. Our theory-backed, ML-based empirical approach contrasts with current estimation methods that model or approximate customers' full consideration of all available choices. Our methods recover consumption preferences from complex choice data. Using such estimates, we uncover the hypermarket's demand structure and counterfactually raise its equilibrium service throughput by 5-25% through modest capacity reallocations. Our approach uses cross-station spillovers in demand from local congestion shocks to identify truly complementary consumption, which is a long-standing challenge in the marketing and operations literatures.

Keywords: Adversarial estimation; Consumer choice; Empirical operations management, Machine learning, Marketplace design, Neural networks, Queueing, Service networks, Structural estimation