Date of Award

2026-08-01

Degree Name

Doctor of Philosophy

Department

Computer Science

Advisor(s)

Meagan R. Kendall

Second Advisor

Martine Ceberio

Abstract

Artificial Intelligence (AI) and machine learning (ML) models are increasingly being deployed to support decision-making in high-stakes domains such as healthcare, criminal justice, and education, where trust, accountability, and transparency are critical. However, increasing model complexity has made many modern systems insufficiently transparent. Existing approaches to explainable AI (XAI) typically emphasize either intrinsic model simplicity or post-hoc attribution methods that estimate feature importance for predictions. While these approaches provide valuable insights into model behavior, they do not necessarily establish whether the identified importance is grounded in the underlying data patterns or in the structural relationships that generate model behavior. Many post-hoc methods can also be computationally expensive, particularly when explanations must be generated across large datasets or complex models.

To address these limitations, this work proposes a holistic approach to explainable AI that integrates three complementary modeling frameworks: statistical, causal, and argumentative. Each framework addresses a distinct explanatory gap. Statistical methods provide a data-level reference for evaluating whether model-learned feature importance aligns with the discriminative structure of the training data. Causal methods uncover directional and mediating relationships among features that associative techniques cannot reveal. Computational argumentation reasons about conflicting and uncertain relationships among variables, enabling competing evidence to be evaluated within a coherent explanatory framework. Together, these frameworks provide complementary answers to three fundamental explanatory questions: which features matter, how those features influence one another, and why a particular prediction could have been made.

Three research questions structure this work. RQ1 asks how a computationally efficient, data-driven approach can assess whether features are meaningfully important in binary classification independently of any particular model. RQ2 asks how causal discovery methods can produce explanation graphs that reveal directional and mediating relationships between features and outcomes. RQ3 asks how causal and argumentative modeling can be integrated to explain not only which features influence a prediction, but why a particular decision is reached, including how supporting and opposing factors interact.

The dissertation aims to make three contributions, each addressing a corresponding research question. First, a Standardized Mean Difference (SMD) baseline is introduced to assess model-data alignment in binary classification. Second, a dual-encoding approach for constraint-based causal discovery is proposed to address numerical instability caused by categorical variables. The resulting causal graphs capture directional and mediating pathways that are not represented by post-hoc attribution methods. Third, a procedure is introduced for translating the causal graphs recovered by the second contribution into Bipolar Argumentation Frameworks and computing semi-stable extensions. This enables dialectically structured, instance-level explanations that represent both supporting and opposing evidence without requiring predefined domain rules.

These contributions have been disseminated through peer-reviewed venues: the model-data alignment baseline at the International Workshop on Causality, Agents and Large Models (CALM 2025) in Luxembourg, the dual-encoding causal discovery approach at the same workshop the following year (CALM 2026) in Istanbul, and the causal-argumentation framework at the 4th World Conference on eXplainable Artificial Intelligence (XAI 2026) in Brazil.

The proposed methods are validated on diverse benchmark datasets. Building on these contributions, future work aims to make the proposed methods accessible through an interactive interface in which users can upload their own datasets, tune model hyperparameters, and visualize the resulting causal and argumentation graphs. This would lower the barrier to producing statistically grounded, causally informed, and dialectically structured explanations.

Language

en

Provenance

Received from ProQuest

File Size

181 p.

File Format

application/pdf

Rights Holder

Henry Salgado

Download

Share

COinS