Neural Graph Scoring
Loading...
Date
Authors
Journal Title
Journal ISSN
Volume Title
Publisher
University of the Witwatersrand, Johannesburg
Abstract
Learning graph structures has long been a focal point of graphical modelling, serving as a way of analyzing the relationship between concepts and observations and using that information to make predictions or to reason about some situation. Not only will more efficient solutions benefit graphical modelling but can broaden the applicability of graphical models as a whole. But this problem is N P-hard and many solutions put forward are either too inefficient (spatially or temporally) or do not produce optimal approximations. This necessitates the need for novel and innovative solutions that can circumvent these issues faced by traditional structure learning methods. With the continuing development of deep learning technologies, an increase in the popularity of structure discovery using deep learning has surface. This led to the devel- opment of continuous optimization based approaches, using techniques such as Graph Neural Networks and Reinforcement Learning. This dissertation, motivated by the use of deep learning for causal discovery in directed acyclic graphs, proposes the integra- tion of deep learning techniques for the association of graphical structures with their numerical scores. Conducted on synthetically generated data, this framework shows to accurately score graphs of varying sizes with data belonging to a variety of data distributions. This mitigates the issue experienced by many continuous optimization methods of causal discovery that are held back by inefficient scoring methods.
Description
A research report submitted in fulfillment of the requirements for the Master of Science, in the Faculty of Science, School of Computer Science and Applied Mathematics, University of the Witwatersrand, Johannesburg, 2025
Citation
Gathoo, Mohammed. (2026. Neural Graph Scoring [Master’s dissertation, University of the Witwatersrand, Johannesburg]. WIReDSpace. https://hdl.handle.net/10539/50280