Learning Interpretable Feature Context Effects in Discrete Choice

Abstract

Individuals are constantly making choices - -purchasing products, consuming Web content, making social connections - -so understanding what contributes to these decisions is crucial in many settings. A major interest is understanding context effects, which occur when the set of available options itself affects an individual's relative preferences. These violate traditional rationality assumptions but are commonly observed in human behavior. At the same time, identifying context effects from choice data remains a challenge; existing models posit a specific context effect a priori and then measure its effect from (often effect-targeting) data. Here, we develop discrete choice models that capture a broad range of context effects, which are learned from choice data rather than baked into the model. Our models yield intuitive, interpretable, and statistically testable context effects, all while being simple to train. We evaluate our model on several empirical choice datasets, discovering, e.g., that people are more willing to book higher-priced hotels when presented with options that are on sale. We also provide the first analysis of context effects in online social network growth, finding that users forming connections place relatively more emphasis on shared neighbors when popular users are an option.