Simplicial Methods in Graph Machine Learning
Published in Florida State University, 2024
The Weisfeiler-Leman algorithm is a powerful test of graph isomorphism, and works by iteratively isolating clusters of nodes via a coloring scheme based on binary relationships. This refinement process is akin to message passing on graph neural networks. Graph neural networks are paradigms of computation that yield powerful results for structured data based on binary relationships, and therefore their expressivity shares the limitations of the Weisfeiler-Leman test. One way to circumvent this limitation relies on identifying and working with higher relationships within the data, or by finding a better refinement criterion that can be implemented on a machine. In this thesis, we put forward data structures closely based on the identification of such higher relationships via Kan Extensions of structured data built on binary relations. These data structures are well known in algebraic topology as Simplicial Sets; the theoretical underpinnings based on the combinatorics of simplicial sets and directed graphs are explored in this thesis and, based on of it, algorithms are introduced that can identify higher-order relations. A metric that can serve as a quantification for community detection tools is also introduced, as well as the notion of a pseudoterminal vertex as a proxy for these higher relations. This helps us balance the variance-bias tradeoff necessary for generalizability of the architecture. We introduce functors that show how these higher relationships can be modelled and expanded. Finally, we propose three neural networks that benefit from the topology of these nonbinary relations. These neural networks can be ported into the standard graph neural network library to augment feature classification in a semisupervised manner. All source code is available at https://github.com/abdullahnaeemmalik/simplicial-methods-in-graph-machine-learning
Recommended citation: Malik, A.N. Simplicial Methods in Graph Machine Learning. Diss. Florida State University, 2024. https://www.proquest.com/dissertations-theses/simplicial-methods-graph-machine-learning/docview/3064071895/se-2?accountid=4840