Interpretable Machine Learning
Definition
Interpretable machine learning makes relevant relationships from models accessible.
Interpretable machine learning uses models to gain relevant knowledge about relationships in data. Which representation is helpful also depends on the question and the audience.
Simply explained
Like a map that makes certain paths in a landscape visible. The map is not the landscape, and an illustrative representation does not automatically explain every model decision completely.
Why it matters in practice
Before choosing a method, the professional question, target audience, and required accuracy of the representation should be specified.
Example from business
Possible example: A team investigates which input features are related to the predictions of their own test model.
Limits and pitfalls
- A described relationship is not proof of causation. Prediction accuracy alone does not prove that an interpretation correctly describes the model.
Origin & history
Murdoch and coauthors classified model-based and post hoc interpretation methods in 2019 based on prediction accuracy, descriptive accuracy, and relevance.
Use in mid-sized companies
- Possible fields of application are the expert examination of predictive models and communication of their data relationships within a team.
Frequently asked questions
Must a model be simple from the outset?
The work considers both model-based and post hoc methods.
Does an interpretation prove causality?
No, the work distinguishes this task from causal inference.
Sources
- Interpretable machine learning: definitions, methods, and applications · Murdoch et al. (2019): Interpretable machine learning: definitions, methods, and applications
Editorially reviewed on: