Discovering Novel Social Preferences Using Simple Artificial Neural Networks

Abstract

How do we make decisions about allocating scarce resources to others when we face multiple alternatives? While people usually evaluate multiple allocations jointly or non-independently, distributional preferences have been analyzed through behavioral models that assume independent evaluation. Here, we explore distributional preferences that can be identified when we analyze allocation decisions with a flexible model class that can factor in joint evaluation. In Study 1, we recruited 3,006 Japanese crowd workers and provided them with 13,041 theoretically designed resource allocation problems. Using artificial neural networks, we discovered two types of allocation problems among 13,041 problems where the choices made by the participants could only be predicted with a model that factored in joint evaluation of two options. For example, one of these problems showed that the participants became sensitive to differences in their self-reward between two allocations which pitted the participants against disadvantageous equity, which could be naturally supported by our intuition but could not have been discovered unless joint evaluation was considered. In preregistered Study 2, we recruited 185 Japanese participants to conduct a conceptual replication of the machine-discovered distributional preferences. We had participants evaluate two allocation options on the same (i.e., jointly) or different (i.e., separately) screens, confirming that the distributional preferences discovered in Study 1 (i.e., joint-evaluation situation) were more often observed in the former joint-screen situation. Our findings showcase the usefulness of a prediction-oriented machine learning approach to the exploration of novel behavioral theories in social decision-making.