Graph neural networks (GNNs) have become a popular machine learning model for graph prediction tasks. Considerable progress has been made to characterize the connection between the representation power of GNNs and the Weisfeiler-Leman (WL) Algorithm (and accordingly the k-WL hierarchy).
Through this connection, a beautiful relation between GNNs and finite-variable logics has also been established. Nonetheless, some interesting questions still remain underexplored, such as the locality of GNNs and first-order logic/finite model theory, the generalization ability, and the interpretability of GNNs. In this project, one specific issue relating to graph neural networks will be explored.
- This project requires students to have a solid background in machine learning, algorithms (and logic if working on logic-related topics). ANU students are expected to have finished COMP4670 or have an equivalent experience.
Graph Research Lab#
Graph algorithms provide fundamental and powerful ways of exploiting the structure of graphs. Recent advances in machine learning, particularly deep learning, are also achieving remarkable results in a wide variety of application domains. Graph Research Lab @ ANU aims to investigate graph-related problems by marrying the best of two worlds: traditional graph algorithms and new machine learning techniques to bridging the gap between combinatorial generalisation and deep learning.
For further information about our team, visit our Github.