If Interpretability Is the Answer, What Is the Question?: A Causal Perspective

MCML Authors

Gunnar König

Dr.

* Former Member

→ Group Bernd Bischl
 Statistical Learning and Data Science

Abstract

This thesis addresses fundamental challenges in the field of interpretable machine learning (IML), particularly the lack of a clear definition of 'interpretability', the potential misinterpretation of existing methods, and the computational difficulties of conditional-sampling-based techniques. By disentangling the different goals of interpretability, we provide clearer guidelines for deriving target estimands, with specific examples such as recourse and scientific inference. Additionally, we propose formal interpretation rules for feature importance, highlight common pitfalls in IML, and introduce efficient methods for estimating conditional-sampling techniques by leveraging the data’s dependence structure, with a strong emphasis on causal inference to improve clarity and computational efficiency. (Shortened.)

phdthesis

---

Dissertation

LMU München. Oct. 2023

---

Authors

G. König

---

Links

DOI

---

Research Area

 A1 | Statistical Foundations & Explainability