Learning the determinants of human behavior

This paper introduces moderational learning, a data-driven framework for identifying generative mechanisms underlying human behavior and cognition. Built on deep latent variable models, the framework jointly learns both a generative model that yields responses and a set of latent variables that encode individual differences. The approach can also integrate information across multiple tasks or modalities to estimate individual-level trait representations that predict behavior across tasks. Two simulation studies demonstrate that moderational learning accurately predicts behavior, recovers true latent factors, and identifies population heterogeneity that arises when participants adopt distinct strategies. Applied to real data from value-based decision tasks, the framework outperforms both subject- and group-level neural models in predicting human choices and reveals surprising theoretical insights. In multiple decision tasks, a single latent factor accounted for most individual variability, providing simpler and more accurate models of behavior than existing theories. Across tasks, we identified three common latent factors---risk discounting, delay discounting, and bidding proclivity. Sensitivity analysis shows that responses in pricing tasks do not always decrease monotonically with aversive attributes. Together, these results demonstrate that the proposed moderational learning framework can pave the way for a new data-driven paradigm that integrates and automates theory construction, model discovery, and representation learning of individual differences.