Building and Evaluating Interpretable Models using Symbolic Regression and Generalized Additive Models

In this paper we investigate new methods to build and evaluate interpretable predictive models for time series data using symbolic regression and generalized additive models. We propose a novel framework to iteratively build a model while maintaining model interprebility as accuracy and complexity increase. We also propose multiple methods that ease interpretation of the built model, the model building process, and the model output. The proposed methods study the contributions of the model constituents, partial derivatives, and behavior in the frequency domain. Finally, we empirically demonstrate the framework methods by modeling weather phenomena using a weather station observational dataset, and show how the resulting model finds underlying meteorological principles automatically.