CPS Small: Integrating Bayesian Deep Learning, Microeconometrics, and Graph-Based Data for Enhanced Interpretable Choice Models with Heterogeneous Network Interactions

Modern transportation, energy, and service systems increasingly depend on connected infrastructure and interactions among people. In these cyber-physical systems, individual decisions — such as whether to use transit, charge an electric vehicle, adopt a new mobility service, or respond to congestion — are shaped by social influence, infrastructure conditions, digital information, peer behavior, and feedback from the system itself. Predicting these decisions is critical for designing efficient, trustworthy, and resilient systems. Yet many existing modeling tools either rely on simplified behavioral assumptions or use methods that are difficult to interpret and do not provide reliable measures of uncertainty. This project will develop new methods for modeling human decision-making in networked cyber-physical systems in a way that is both flexible and interpretable. The goal is to improve the scientific basis for predicting how people respond to infrastructure technologies and how those responses aggregate into system-level outcomes such as congestion, adoption, crowding, or demand shifts. The work serves the national interest by enabling the design of more reliable, sustainable, and human-centered infrastructure systems for mobility and energy-efficient transportation, while also contributing educational materials and research training at the intersection of engineering, economics, data science, and artificial intelligence. Specifically, the project will develop interpretable Bayesian deep-learning methods for discrete choice modeling in networked cyber-physical systems. The central technical objective is to combine the flexibility of graph neural networks and attention mechanisms with the identification, interpretability, and welfare foundations of microeconometric discrete choice models. The proposed models will represent individual utilities as structured combinations of interpretable private utility components, nonlinear neural components, and network-mediated interaction terms. Skip connections, batch normalization, behavioral regularization, sparsity-inducing priors, and Bayesian posterior approximation methods will be used to preserve economically meaningful quantities such as marginal utilities, elasticities, substitution patterns, and willingness-to-pay measures. The project will advance graph convolutional and graph attention architectures tailored to discrete choice settings with observed, partially observed, or latent social, spatial, and cyber-mediated network structure. Bayesian learning methods, including stochastic-gradient-based posterior approximation and ensemble-based marginalization, will quantify epistemic uncertainty in high-dimensional models and propagate that uncertainty to policy-relevant behavioral functionals. Behavioral regularization will be used to impose or diagnose consistency with economic principles such as monotonicity and interpretable tradeoffs. Empirical validation will use networked mobility and technology-adoption data, including transportation and energy-related case studies, to evaluate predictive performance, behavioral consistency, uncertainty calibration, and usefulness for CPS design and decision support. The expected contribution is a new class of trustworthy, interpretable, and uncertainty-aware AI tools for modeling human-system feedback in cyber-physical infrastructure. This award reflects NSF's statutory mission and has been deemed worthy of support through evaluation using the Foundation's intellectual merit and broader impacts review criteria.