Scientific Inference With Interpretable Machine Learning: Analyzing Models to Learn About Real-World Phenomena
This work addresses the gap for scientists who need to use machine learning models to learn about real-world phenomena, but it is incremental as it builds on existing IML methods.
The authors tackled the problem of using interpretable machine learning (IML) for scientific inference, rather than just model auditing, by developing a framework called 'property descriptors' that reveals properties of the underlying data distribution. They demonstrated that these descriptors, based on statistical learning theory, can effectively uncover relevant aspects of the joint probability distribution, enabling scientists to leverage ML models for understanding real-world phenomena.
To learn about real world phenomena, scientists have traditionally used models with clearly interpretable elements. However, modern machine learning (ML) models, while powerful predictors, lack this direct elementwise interpretability (e.g. neural network weights). Interpretable machine learning (IML) offers a solution by analyzing models holistically to derive interpretations. Yet, current IML research is focused on auditing ML models rather than leveraging them for scientific inference. Our work bridges this gap, presenting a framework for designing IML methods-termed 'property descriptors' -- that illuminate not just the model, but also the phenomenon it represents. We demonstrate that property descriptors, grounded in statistical learning theory, can effectively reveal relevant properties of the joint probability distribution of the observational data. We identify existing IML methods suited for scientific inference and provide a guide for developing new descriptors with quantified epistemic uncertainty. Our framework empowers scientists to harness ML models for inference, and provides directions for future IML research to support scientific understanding.