Connecting Black-Box Models to Benchmark Dynamical Systems Models

Google map

The last decade has evidenced increased integration of machine learning (ML) methods into traditional psychometric and quantitative modeling approaches to address longstanding data-analytic challenges. Using targeted illustrative simulations using Extreme Gradient Boosting (XGBoost) and variational autoencoder for stochastic differential equation model (VAE-SDE), I address several issues that arise in using ML methods to analyze the dynamics of intensive longitudinal data (ILD), including use of dynamic summary features to capture multi-temporal dynamics, hyperparameter tuning, balance between theoretical priors and data-driven evidence, and the importance of considering multiple performance criteria beyond just forecast or classification errors. The overarching perspective is that hybrid ML approaches combining the strengths of theory- and data-informed analytic strategies are both promising and increasingly central to the future of ILD modeling, but their success depends on the thoughtful integration of theory, data, and computational resources.