Towards externally valid machine learning: A spurious correlations perspective

As machine learning models permeate critical real-world systems, their reliability becomes paramount. However, their performance often critically depends on how representative their training data is of the deployment domain. This can lead to significant challenges when faced with distribution shifts—situations where the real-world data diverges from the training data. A significant issue under distribution shifts is the presence of spurious correlations—statistical associations in the training data that do not hold in all real-world domains. This dissertation develops methodologies that leverage causality to create machine learning models without spurious correlations, yielding more robust, fair, and ethical decision-making. Causality is crucial because it helps disentangle true causal relationships from mere statistical associations, ensuring that models learn robust, externally valid reasoning. This dissertation also establishes construct validity conditions for evaluating machine learning models under distribution shifts, enabling reliable generalization of robustness from distribution shift benchmarks to the real world. Domain-general models and robust evaluation methods enable the development of externally valid machine learning systems.

Name thesis:degree_name

Ph.D.

Level thesis:degree_level

Dissertation

Discipline thesis:degree_discipline

Computer Science

Grantor

University of Illinois at Urbana-Champaign

Year dc:date

2024

Author dc:creator

• Salaudeen, Olawale Elijah

Contributors dc:contributor

• Koyejo, Oluwasanmi
• Forsyth, David
• Zhao, Han
• D'Amour, Alexander

• Domain Generalization
• External Validity
• Spurious Correlations
• Machine Learning Benchmarks

Statement dc:rights

• Copyright 2024 Olawale Salaudeen

Language dc:language

en, eng

Handle dc:identifier

https://hdl.handle.net/2142/125683

OAI record oai:identifier

https://www.ideals.illinois.edu/oai-pmh?verb=GetRecord&metadataPrefix=etdms&identifier=oai%3Awww.ideals.illinois.edu%3A2142%2F125683