University of British Columbia
Explainable artificial intelligence: causal, conceptual, and counterfactual approaches
Abstract
dc:descriptionDeep neural networks achieve remarkable predictive performance but remain largely opaque, limiting their trustworthy deployment in high-stakes domains. This thesis develops three complementary, post-hoc approaches to explainability - causal, conceptual, and counterfactual - that together provide a comprehensive toolkit, named the 3C explanation framework, for understanding deep learning systems without altering their architecture or degrading their predictive capacity. For causal explainability, we propose a data-based explanation method that answers “what underlying causal factors generate the data?” by discovering cause-and-effect relations among variables in data. Our method addresses the over-reconstruction problem in existing causal discovery approaches, where intrinsic noise is inadvertently absorbed into the learned causal structure, by proposing a noising-and-denoising process inspired by diffusion models. Experiments on synthetic and real-world datasets demonstrate the effectiveness of the proposed approach. For conceptual explainability, we introduce a framework that answers “why does this user receive this prediction?” through discovering human-understandable semantic concepts for predictions, e.g., “red cross” for ambulance classification. Unlike existing explanation methods, the proposed explainer not only discovers semantic concepts for a prediction but also decomposes each concept's predictive contribution into uniqueness, redundancy, and synergy, offering practitioners a richer understanding of the model's reasoning. Real-world experiments confirm the advantages of this decomposition. For counterfactual explainability, we target the question “why is a different prediction not made for this user?” To answer this question, counterfactual explainers aim to find minimal changes to an input sufficient to flip the model's prediction. We design a counterfactual explainer based on overcomplete generative models (GMs), applying a geometry-aware optimization scheme in the GM's latent space and proving that the resulting counterfactuals are valid and realistic. By further equipping the GM with a semantic-stochastic disentanglement mechanism, sparsity is achieved simultaneously, satisfying the three main desiderata of high-quality counterfactuals. Experiments on three datasets with eleven metrics demonstrate an improved trade-off among the three properties.
Degree
thesis:*- Name thesis:degree_name
- Doctor of Philosophy - PhD
- Level thesis:degree_level
- doctoral
- Discipline thesis:degree_discipline
- Mathematics
- Grantor dc:publisher
- University of British Columbia
- Year dc:date
- 2026
Author and committee
dc:creator, dc:contributor.*- Author dc:creator
-
- Ji, Chaojie
- Contributors dc:contributor
-
- Cao, Yankai
Rights
dc:rights- Statement dc:rights
-
- Attribution-NonCommercial-NoDerivatives 4.0 International
- Language dc:language
- eng
Identifiers
dc:identifier.*- Handle dc:identifier
- http://hdl.handle.net/2429/95829