Perceptron-as-Opinion-Dynamics (POD): A Unified and Interpretable Machine Learning Framework for Opinion Dynamics

TMLR Paper5978 Authors

Abstract: We introduce Perceptron-as-Opinion-Dynamics (POD), a dual-activation perceptron framework that learns opinion dynamics directly from data with full interpretability. Each parameter in POD corresponds to a concrete social mechanism, including agent-specific inertia, dynamic influence networks, persistent bias, and nonlinear modes of perception and expression. With appropriate parameter choices, POD exactly recovers canonical linear models such as DeGroot, Friedkin–Johnsen, and Altafini, closely approximates nonlinear models like Hegselmann–Krause, and naturally extends to kinetic and extremist cases. When trained end-to-end, POD generalizes beyond these settings, achieving up to 83% faster convergence and reducing prediction error by 40–82% compared to canonical baselines on real-world data. By unifying fragmented opinion dynamics models into a single trainable and interpretable neural framework, POD lays a robust foundation for modeling complex belief evolution in social systems, addressing challenges that have persisted for decades.

Submission Type: Regular submission (no more than 12 pages of main content)

Assigned Action Editor: ~Manuel_Gomez_Rodriguez1

Submission Number: 5978