Leveraging Cognitive-Inspired Machine Learning to Understand Multi-Attribute Preference Construction

Abstract How people form preferences in naturalistic environments—where options are defined by latent attributes—is not well understood. We introduce a modeling framework that combines image embeddings with cognitively structured neural network (NN) models to study multi-attribute choice using naturalistic visual stimuli. In a large-scale behavioral experiment, participants (N = 938) make ternary choices among images of places representing different weekend activities. In our modeling approach, images are first transformed into embeddings, which are then used as inputs to NN models that predict choice behavior. Six NN models were developed, each of which structurally encodes distinct cognitive assumptions, differing along three key dimensions: contextual sensitivity (whether an option’s value depends on the other items in the choice set), attribute independence (whether attributes are evaluated independently or jointly), and linearity of attribute evaluation (whether attribute evaluation is linear or nonlinear). We further vary the level of dimensionality reduction in the embedding step to assess whether individuals rely on a small subset of attributes or integrate across a larger set. Results show that (1) non-contextual models perform as well as contextual ones, (2) there is modest evidence for joint attribute evaluation, (3) nonlinear attribute transformations improve predictive accuracy, and (4) relatively few attributes are needed to model behavior. These findings suggest that, in the present naturalistic multi-attribute choice task, preferences are constructed from low-dimensional, nonlinear attribute representations and most decisions appear to be context independent.