Evaluation and Validation Approaches for Simulation of Social Behavior: Challenges and Opportunities
While models of social systems have grown in scope to represent increasingly complex phenomena, efforts to validate these models against observations suffer from a number of challenges that include the lack of shared approaches, the use of single-dimensional evaluation criteria, and the difficulty in evaluating causal mechanisms. Moreover, current evaluation approaches suffer from sensitivity of the existing measurements to initial conditions and model assumptions, and failure to account for uncertainty. These effects lead to the lack of generalizability to new contexts, inability to replicate the validation approach for new models, and difficulties in interpretation of validation results. These challenges can be addressed by building a set of diverse and robust evaluation metrics and by developing best practices for comparative, reproducible validation of varied social simulation approaches against real-world data. Herein, we outline several best practices for data-drive model validation, such as defining evaluation strategies consistent with the intended use case and defining broad comparison criteria to highlight both model strengths and weaknesses. Leveraging common validation practices will enable robust analysis of simulation performance across diverse social scenarios and modeling approaches.