Causality Concepts in Machine Learning - Heterogeneous Treatment Effect Estimation with Machine Learning & Model Interpretation with Counterfactual and Semi-factual Explanations - Dissertation

Aug 4, 2026

Upload by Hazel

This dissertation investigates how causality concepts can be integrated into machine learning for two connected goals: estimating heterogeneous treatment effects and interpreting predictive models. It studies random-forest-based methods, identifying which components are effective for different data settings such as randomized controlled trials and observational data with confounding. It further develops and analyzes counterfactual and semi-factual explanation methods, addressing pitfalls, the multiplicity of valid explanations, and practical ways to generate and evaluate multiple explanations. Contributions include formal optimization approaches and an R package to facilitate comparison and reuse.

LanguageEnglish

CategoryResearch & Report

FormatPDF

Pages300

Reading Time12 h 36 min

AccessFree to download

Summary

Heterogeneous Treatment Effect Estimation

Causal Interpretation with Counterfactual and Semi-factual Explanations

Multiple Explanations, Pitfalls, and Practical Tools

Causality Concepts in Machine Learning: Heterogeneous Treatment Effect Estimation with Machine Learning & Model Interpretation with Counterfactual and Semi-factual Explanations Susanne Dandl M¨unchen 2023 Causality Concepts in Machine Learning: Heterogeneous Treatment Effect Estimation with Machine Learning & Model Interpretation with Counterfactual and Semi-factual Explanations Susanne Dandl Dissertation an der Fakult¨at f¨ur Mathematik, Informatik und Statistik der Ludwig–Maximilians–Universit¨at M¨unchen eingereicht von Susanne Dandl am 18.09.2023 Erster Berichterstatter: Prof. Dr. Bernd Bischl Zweiter Berichterstatter: Prof. Dr. Torsten Hothorn Dritter Berichterstatter: Prof. Dr. Marvin N. Wright Tag der Disputation: 06.12.2023 Acknowledgments I would like to express my sincere thanks to all those who have supported and advised me throughout this incredible journey of pursuing my Ph.D. In particular, I am deeply grateful to . . . . . . Prof. Dr. Bernd Bischl for arousing my interest in machine learning during my Master’s program and for providing invaluable supervision, advice, and support throughout my Ph.D. . . . Prof. Dr. Torsten Hothorn for his insights and guidance, which opened up new perspectivesand greatly shaped my research path. . . . Prof. Dr. Marvin N. Wright for his willingness to be the third reviewer of my Ph.D. thesis. . . . Prof. Dr. Frauke Kreuter and Prof. Dr. Christian Heumann for their availability to be part of the examination committee . . . . Prof. Dr. Achim Zeileis, Prof. Dr. Stefan Wager, and Dr. Erik Sverdrup for the fruitful discussions and valuable contributions to our joint projects despite the physical distance between our universities. . . . Giuseppe Casalicchio, Ludwig Bothmann, and Andreas Bender for their guidance throughout the research projects and their supervision of the research subgroups I was part of, which heavily inspired me . . . . Heidi Seibold, Cornelia F¨utterer, Christoph Molnar, and Martin Binder for their encour agement to pursue this academic path. . . . all my current and former colleagues at the chair of Statistical Learning and Data Science for being a source of inspiration and for their contributions to joint projects. . . . all research fellows at the Department of Statistics for creating a supportive environment for exchanging ideas in the form of mensa visits, tea talks, summer retreats, . .. . . . my friends and my family for their unwavering support throughout this journey. Summary Over decades, machine learning and causality were two separate research fields that developed independently of each other. It was not until recently that the exchange between the two intensified. This thesis comprises seven articles that contribute novel insights into the utilization of causality concepts in machine learning and highlights how both fields can benefit from one another. One part of this thesis focuses on adapting machine learning algorithms for estimating heterogeneous treatment effects. Specifically, random forest-based methods have demonstrated to be a powerful approach to heterogeneous treatment effect estimation; however, understanding the key elements responsible for that remains an open question. To provide answers, one contribution analyzed which elements of two popular forest-based heterogeneous treatment effect estimators – causal forests and model-based forests – are beneficial in case of real-valued outcomes. A simulation study reveals that model-based forests’ simultaneous split selection based on prognostic and predictive effects is effective for randomized controlled trials, while causal forests’ orthogonalization strategy is advantageous for observational data under confounding. Another contribution shows that combining these elements yields a versatile model framework applicable to a wide range of application cases: observational data with diverse outcome types, potentially under different forms

FAQs

What is the main focus of this dissertation?

It connects causality concepts with machine learning to estimate heterogeneous treatment effects and to interpret model predictions using counterfactual and semi-factual explanations.

How are heterogeneous treatment effects estimated in this thesis?

It adapts random-forest-based approaches and studies which elements work well for real-valued outcomes, using insights from simulation for both randomized trials and observational data under confounding.

What are counterfactual and semi-factual explanations?

Counterfactual explanations describe minimal feature changes needed to alter a prediction, while semi-factual explanations describe maximal feature changes needed to keep the prediction unchanged.

Why is generating multiple explanations important, and what tools are provided?

The thesis emphasizes that multiple equally good explanations can exist and may be overlooked; it proposes methods to generate multiple counterfactual/semi-factual explanations and includes an R package to apply and compare them.

Aug 4, 2026

Upload by Hazel