Behavior Learning

Core task: learning interpretable and identifiable optimization structures from data. This field addresses the challenge of extracting transparent, verifiable models and decision rules from complex datasets, spanning diverse application domains and methodological traditions. The taxonomy reveals seven major branches that organize research by the type of structure being learned. Identifiable Latent Variable and Generative Models focus on recovering hidden factors with provable uniqueness guarantees, often drawing on statistical theory. Interpretable Predictive Modeling and Feature Extraction emphasizes transparent mappings from inputs to outputs, including symbolic regression and rule-based classifiers. Causal Structure Learning and Heterogeneity Analysis targets the discovery of cause-effect relationships and subgroup-specific patterns. Physics-Informed and Hybrid Data-Driven Modeling integrates domain knowledge with learning algorithms to ensure physically meaningful solutions. Optimization-Driven Interpretable Learning—where Behavior Learning[0] resides—centers on inferring objective functions, constraints, or utility models that explain observed decisions or behaviors. Interpretable Machine Learning for Engineering and Applied Optimization applies these ideas to real-world systems such as energy management, manufacturing, and materials design, as seen in works like SHAP Bayesian Metamaterial[3] and Battery Life Interpretable[11]. Finally, Interpretability in Complex Data Structures handles high-dimensional or structured inputs like graphs and sequences. Within Optimization-Driven Interpretable Learning, a central theme is reverse-engineering decision-making processes: given observed choices or trajectories, can we recover the underlying utility function or optimization criterion? Behavior Learning[0] exemplifies this utility-based framework, aiming to infer interpretable objective structures from behavioral data. This contrasts with neighboring branches that prioritize forward modeling—such as Physics-Informed approaches that embed known equations—or purely predictive methods that may sacrifice interpretability for accuracy. Closely related efforts like Explainable Data-Driven Optimization[19] and Intelligent Data-Driven Optimization[6] also blend optimization with transparency, but they often emphasize solution explanation or hybrid solver design rather than behavioral inference. The trade-off between model fidelity and interpretability remains a key open question: richer utility models can capture nuanced preferences but may become less transparent, while simpler forms risk oversimplifying real decision processes. Behavior Learning[0] navigates this tension by focusing on identifiable structures that remain both expressive and interpretable.