Extending MDFT-RNN to Multi-Attribute Decisions: Learning Latent Preferences in the Insider Attack Game

Multi-alternative Decision Field Theory (MDFT) models preference construction as stochastic evidence accumulation with inter-option inhibition driven by similarity. Historically, MDFT has been applied to simple choice tasks where options are described by two attributes. Our goal is to extend MDFT to complex, real-world decisions where options are defined by higher-dimensional, context-rich cues. We present an MDFT extension that replaces the classical distance term with a multi-attribute distance D used to populate the similarity/inhibition matrix S while preserving its k×k form and the standard inhibition mechanism. The distance construction better reflects how higher-dimensional attributes shape option similarity under comparison, shifting predicted choice distributions and learning gradients. Building on neural learning approaches for MDFT (Venable; Rahgooy; Busemeyer), we implement the model as MDFT‑NN, which learns participant-specific attention weights and attribute preferences. A second contribution is selective attribute learning via column-wise gradient masking on the preference matrix M, enabling targeted learning of latent or emergent preference dimensions while holding other dimensions fixed to values derived from modeled experimental signals. We evaluate MDFT‑NN on a trial-level insider attack game dataset where participants repeatedly choose between Attack and Withdraw across trials with experimentally manipulated cues. We train participant-specific models across variants (standard learning, restricted attribute sets, selective preference learning) and validate learned parameters via classical MDFT forward simulation to predict trialwise Attack/Withdraw probabilities, comparing predictions to observed distributions across behavioral archetypes. Together, these advances enable preference modeling in rich contexts, separating overt context from latent subjective signals to personalize learned predictions.