Interpreting Time-Series Machine Learning Models through Domain-Informed Basis Functions
Yasser Mehmood Qureshi, Vitaly Voloshin, Philip J. McCall, James Anthony Covington, Cathy Towers, David Peter Towers · 2025
Interpreting machine learning models for time-series data is a critical challenge, particularly in fields where decisions have real-world implications, such as biology. In this work, we introduce a novel approach for interpreting time-series machine learning models using domain-defined basis functions. We apply this method to the classification of mosquito flight trajectories as either insecticide-susceptible (IS) or insecticide-resistant (IR), a study into behavioural resistance, which may inform vector control strategies against malaria and other mosquito-borne diseases. By generating synthetic trajectories based on relevant parameters, we systematically probe a trained classifier to reveal the patterns and features it has learned. This approach enhances model interpretation, providing a new perspective on mosquito movement analysis. Furthermore, our findings offer valuable insights into distinguishing between IS and IR mosquito populations, contributing to more targeted and effective mosquito control efforts. Our methodology can be extended and adapted beyond mosquito trajectory analysis, demonstrating how synthetic data can be used to probe and understand complex time-series classifiers thus contributing to the growing field of explainable AI (XAI).