Contact: dkelesis@iit.demokritos.gr
Network flow optimization focuses on finding the most efficient way to route flow through a network to meet demand at minimum cost. This problem is crucial in areas such as transportation, telecommunications, and supply chain logistics. Traditional algorithms like the Ford-Fulkerson method, particularly the Edmonds-Karp implementation for maximum flow problems, and the Successive Shortest Path algorithm for minimum cost flow problems, have set the foundation. However, these methods often face scalability issues and are less adaptive to dynamic network conditions. This thesis proposes to investigate the integration of Graph Neural Networks (GNNs) with traditional network flow algorithms to enhance scalability and adaptability.
Key research questions include: How can GNNs be effectively combined with traditional network flow optimization techniques to improve performance in large-scale and dynamic environments? What are the potential benefits in terms of computational efficiency and solution quality? Can GNN-enhanced methods provide more robust and adaptive solutions to real-time network flow problems compared to traditional algorithms alone? This research aims to develop and evaluate novel GNN-based hybrid models that leverage the strengths of both GNNs and classical optimization methods to address these challenges. Explore more GNN architectures, create new attention schemes, explore the transformation to heterogenious graph (edge to node) and apply node classification methods.
Relevant literature:
[1] https://arxiv.org/pdf/2209.05208
[2] https://people.cs.uchicago.edu/ tkannan/graph flow neural networks.pdf