Bridging the Gap Between Data-Driven And Theory-Driven Modelling - Leveraging Causal Machine Learning for Integrative Modelling of Dynamical Systems

TL;DR

This paper explores how integrating causal discovery algorithms with domain knowledge enhances the robustness of machine learning models in dynamical systems, especially for complex time series, by improving feature selection and predictive reliability.

Contribution

It demonstrates that combining causal discovery with human expertise leads to more robust predictions than traditional feature selection methods in dynamical systems modeling.

Findings

Causal feature selection improves prediction robustness.

Simulated data validates causal methods over traditional techniques.

Combining causal discovery with domain knowledge enhances model reliability.

Abstract

Classical machine learning techniques often struggle with overfitting and unreliable predictions when exposed to novel conditions. Introducing causality into the modelling process offers a promising way to mitigate these challenges by enhancing predictive robustness. However, constructing an initial causal graph manually using domain knowledge is time-consuming, particularly in complex time series with numerous variables. To address this, causal discovery algorithms can provide a preliminary causal structure that domain experts can refine. This study investigates causal feature selection with domain knowledge using a data center system as an example. We use simulated time-series data to compare different causal feature selection with traditional machine-learning feature selection methods. Our results show that predictions based on causal features are more robust compared to those derived from traditional methods. These findings underscore the potential of combining causal discovery algorithms with human expertise to improve machine learning applications.