Choice Prediction With Set-Dependent Aggregation

Abstract

Providing users with alternatives to choose from has become an essential component in many of today’s online platforms, making the accurate prediction of choice crucial to their success. A recently renewed interest in learning choice models has led to significant progress in modeling power, but current methods are either limited in the types of choice behavior they capture, cannot be applied to large-scale data, or both. In this work, we propose a method for learning to predict choice that is expressive, accurate, and scale well. Our key modeling point is that to account for human choice, all items in a choice set must be considered jointly. From this observation, we derive a model that can express any behavioral choice pattern, enjoys favorable sample complexity guarantees, and can be efficiently trained end-to-end. We conclude with a thorough experimental evaluation on three large choice datasets that demonstrates the power of our approach.