Abstract
dc:descriptionThe question of causality "What if?'' has been a cornerstone of scientific thought since antiquity, from Aristotle’s early explorations to modern empirical research. This question encompasses the core challenge of causal inference, which is the understanding of how an intervention or a treatment affects the outcomes within complex systems. The present thesis approaches this challenge through the development of machine learning methodologies for causal effect estimation, with particular emphasis on applications where reliable causal insights can inform critical decision-making processes, particularly in healthcare. Although Randomized Controlled Trials are considered the gold standard for establishing causal relationships through controlled experimentation, they are often impractical due to real-world constraints. Financial limitations, ethical concerns and logistical challenges often necessitate to rely on observational data for causal analysis. This shift from experimental to observational settings introduces significant methodological challenges, including confounding bias, selection bias as well as the inherent absence of counterfactual outcomes, which is well-known as the fundamental problem of causal inference. The emergence of machine learning has provided opportunities to address these challenges of causal inference through advanced data-driven approaches. In particular, neural network-based architectures and tree-based models have demonstrated remarkable capabilities in capturing the complex relationships within observational data, and demonstrate potential for mitigating confounding and capturing treatment effect heterogeneity. However, existing methodologies exhibit significant limitations constraining their practical applicability and theoretical robustness. Contemporary machine learning approaches for causal inference primarily process individual instances independently, thereby failing to exploit valuable information contained in similar instances within the dataset. This limitation becomes especially critical in regions where few observations exist, such as areas with extreme covariate values, uncommon combinations of patient characteristics, or boundary regions of the covariate distribution. In these sparse data regions, neighboring instances with similar characteristics could provide crucial insights about expected outcomes under different treatment conditions. Furthermore, traditional methods often underperform in settings with limited data availability, which is a common challenge in medical research when rare conditions are involved, in studies focused on specific population subgroups as well as in investigations of early-stage phenomena where sample sizes are inherently restricted. This dissertation addresses these limitations through four principal scientific contributions: 1. We investigate how neighboring information can be systematically integrated into neural network-based causal inference models to enhance estimation accuracy and robustness. The Nearest Neighboring Information for Causal Inference methodology represents an approach that augments existing state-of-theart models with aggregated outcome information from similar instances, thereby enriching the models’ capacity to capture local patterns and improve treatment effect estimation. Additionally, we present a data analysis workflow for handling real-world and observational data to facilitate research on breast cancer patients’ quality of life. This workflow is developed in the context of the REBECCA project and specifically focuses on demonstrating its effectiveness in analyzing quality of life aspects of breast cancer patients suffering from osteopenia and osteoporosis, which is a chronic comorbidity of breast cancer treatment. 2. We explore the critical challenge of causal inference under data scarcity through the development of the Causal-Siamese framework. This approach fundamentally reconceptualizes the causal inference problem by learning pairwise similarity relationships rather than direct mappings from covariates to outcomes. By leveraging metric learning principles, the Causal-Siamese model can effectively estimate treatment effects even when traditional sample size requirements are not satisfied, making it particularly valuable for applications involving rare events or specialized populations. 3. We introduce C-XGBoost, a novel tree-based boosting model that combines the superior performance of XGBoost on tabular data with specialized architectural features designed for causal effect estimation. This model addresses the limitations of neural network approaches on structured data while incorporating the ability to learn representations suitable for predicting outcomes across both treatment and control conditions. The development includes a specialized loss function tailored for causal inference applications. 4. We extend this work through C-XGBoost with targeted regularization, which incorporates theoretical principles from non-parametric estimation theory to achieve both practical performance improvements and statistical optimality guarantees. This advanced framework integrates efficient influence functions and targeted regularization procedures into the tree-based boosting paradigm, bridging the gap between practical machine learning applications and rigorous statistical theory.
Degree
thesis:*- Grantor dc:publisher
- Harokopio University
- Year dc:date
- 2025
Author and committee
dc:creator, dc:contributor.*- Authors dc:creator
-
- Kyriakidou, Niki
- Κυριακίδου, Νίκη
Subjects
dc:subject × 16- Αιτιώδεις επιδράσεις
- Εκτίμηση αιτιωδών επιδράσεων
- Αιτιώδης συμπερασμός
- Δυνητικά αποτελέσματα
- Αιτιότητα
- Causal effects
- Treatment effect estimation
- Causal inference
- Potential outcomes
- Causality
- Επιστήμη Ηλεκτρονικών Υπολογιστών και Πληροφορική
- Φυσικές Επιστήμες
- Τεχνητή νοημοσύνη
- Computer and Information Sciences
- Natural Sciences
- Artificial Intelligence
Rights
- Language dc:language
- eng
Identifiers
dc:identifier.*- Identifier
- 10.12681/eadd/60508