Implementing and Evaluating Kolmogorov-Arnold Networks in Graph-Based Contexts
Supervised by: Anthony Gillioz
If you are interested in this topic or have further questions, do not hesitate to contact anthony.gillioz@unibe.ch.
Context
This project aims to explore the application of Kolmogorov-Arnold Networks (KANs) in graph-based contexts.
The Kolmogorov-Arnold Network (KAN) architecture is a relatively new and theoretically significant development in neural network research. Based on Kolmogorov’s superposition theorem and extended by Arnold, KANs offer a robust framework for universal function approximation. Despite their theoretical promise, practical implementations of KANs, especially in graph-based contexts, are still underexplored. This project seeks to bridge that gap by demonstrating how KANs can be effectively applied to graph-based data, providing valuable insights into their potential across various domains.
The project will involve researching existing implementations, reproducing a functional graph-based KAN, conducting experiments on graph datasets, and reporting the findings in a scientific paper.
Goal(s)
- Understand the theoretical foundations and practical implementations of KANs.
- Research existing papers and code implementing KANs and their applications with graphs.
Approach
- Reproduce a functional implementation of graph-based KAN.
- Ensure the implementation is robust and can be adapted to various graph datasets and tasks.
- Conduct thorough experiments using commonly used graph datasets.
- Analyze the efficiency and effectiveness of KAN architecture in graph-based contexts.
- Document the research, implementation, and experimental findings.
Required Skills
- Good programming skills
- Basic statistics and a bit of graph theory and algorithms
- Willingness to explore and learn more about these topics
Further Reading(s)
- The original KAN paper
- Presentation from one of the author of the original KAN paper
- Tutorial to reproduce KAN’s basic architecture
- Comprehensive list of resources on KANs
- Scientific paper containing lists of graph datasets and results from prominent GNNs for comparison purposes (useful if the graph classification task is chosen)