Addressing Low-Quality Domain Knowledge in Knowledge-Guided Machine Learning - PhD Dissertation

Aug 2, 2026

Upload by Eden

Many industries collect vast datasets and use machine learning to support decisions, yet scientific and engineering deployments face a core obstacle: training data may not represent the true underlying phenomena, leading to poor generalization and spurious correlations. Knowledge-guided machine learning addresses this by integrating domain expertise to improve accuracy, consistency, and generalizability. The thesis analyzes low-quality domain knowledge, including difficult integration, imperfect representation, and mismatches with task requirements, and proposes evaluation criteria and KGML methods to mitigate these issues.

LanguageEnglish

CategoryResearch & Report

FormatPDF

Pages111

Reading Time4 h 40 min

AccessFree to download

Abstract

Problem: Low-quality domain knowledge in KGML

Causes of low-quality domain knowledge

Thesis contributions

Evaluation criteria for practical collaboration

Handling imperfect domain knowledge

Quantifying mismatches in ontology-guided learning

Addressing Low-Quality Domain Knowledge in Knowledge-Guided Machine Learning Zur Erlangung des akademischen Grades eines Doktors der Ingenieurwissenschaften von der KIT-Fakultät für Informatik des Karlsruher Instituts für Technologie (KIT) genehmigte Dissertation von Pawel Bielski aus Gdansk (Polen) Tag der mündlichen Prüfung: 11 . Februar 2025 Referenten: Prof. Dr. Veit Hagenmeyer Prof. Dr. Jochen Garcke apl. Prof. Dr.-Ing. Ralf Mikut Betreuer: Prof. Dr.-Ing. Klemens Böhm This document is licensed under a Creative Commons Attribution 4.0 International License (CC BY 4 .0): [https://creativecommons.org/licenses/by/4.0/deed.en](https://creativecommons.org/licenses/by/4.0/deed.en) Acknowledgments Completing this PhD has been a challenging journey, and it would not have been possible without the support of other people. First, I would like to thank my supervisor, Prof. Klemens Böhm, for providing me with scientific freedom and for securing the funding of my PhD. I would also like to thank Prof. Veit Hagenmeyer, Prof. Jochen Garcke, and Prof. Ralf Mikut for serving as the dissertation reviewers, with special thanks to Prof. Ralf Mikut for his invaluable guidance during the final stages of my PhD. Next, I would like to thank my co-author and research partner, Jakob, for his strategic advice that has been crucial to the success of my PhD. I also want to acknowledge my student co-authors with whom I spent several months exploring early ideas that ultimately evolved into this thesis: Nico, Florian, Lena, Sönke, and Aleksandr. Additional thanks go to Elaheh, Snigdha, and Angelika for our many conversations throughout the entire PhD journey. To Miki, for organizing the board game evenings. To Haoyu, for his help with the experimental data. To Mrs. Bohlinger, for her support during the final stages of my PhD. To Jan, for his indoor cycling classes accompanied by the upbeat music that helped me stay balanced throughout my PhD. Finally, I want to thank my family and friends for their support and belief in me. The warmest thanks go to my dear Aleksandra, who has been by my side through it all. Thankyou for your patience, understanding, and faith in me, and for all our small and big adventures that make my life so much more memorable. Abstract Many industries collect vast amounts of data, applying machine learning techniques to derive actionable insights and improve decision-making. However, deploying machine learning models in scientific and engineering contexts presents unique challenges. This is because the available data often may not sufficiently represent the true nature of the underlying phenomena. In consequence, models may not generalize well beyond the training data and are may be prone to learning spurious relationships. It is crucial to ensure patterns discovered from data are consistent with relevant domain expertise. Knowledge-guided Machine Learning (KGML, also known as Theory-Guided Data Science, Physics-Guided Machine Learning or Informed Machine Learning) is a subfield of Machine Learning that focuses on systematically integrating domain knowledge into machine learning. Recent studies have shown that even little information integrated in this way can substantially enhance the accuracy, consistency, and generalizability of machine learning models. However, domain knowledge can sometimes degrade model performance, a phenomenon we refer to as low-quality domain knowledge. Low-quality domain knowledge can arise for several reasons: (1) Difficult and ineffective process of integrating domain knowledge, due to the interdisciplinary nature of the task and the need for extensive collaboration with domain experts. (2) Imperfections of domain knowledge, due to difficulties in its collection, definition and representation. (3) A mismatch between domain knowledge and task-specific requirements, where knowledge originally developed for a different purpose proves to be suboptimal. This thesis i

FAQs

What is the main problem addressed in this thesis?

The thesis targets low-quality domain knowledge in knowledge-guided machine learning, which can harm performance through integration difficulty, imperfect representation, or mismatches with task needs.

How does the thesis evaluate knowledge-guided machine learning methods for real collaboration?

It introduces evaluation criteria focused on the amount of domain knowledge required and the intellectual effort needed for implementation, then compares KGML integration methods using a lithium-ion battery time-series voltage prediction case study.

What solutions does the thesis propose for imperfect domain knowledge?

It presents a KGML approach for approximate and incorrect domain knowledge by formalizing a relevant form of domain knowledge in temporal dynamics and incorporating it via a knowledge-guided loss function.

How does the thesis handle mismatches between domain knowledge and task requirements?

It studies mismatches in ontology-guided machine learning, proposes a framework to quantify them, and evaluates the impact on image classification and patient health prediction.

Aug 2, 2026

Upload by Eden