Supervisor

Suitable for

Abstract

Supervisor: Michael Benedikt

Co-supervisor: Neo4j (industrial partner)
We have discussed a set of projects concerning graph ML in industry with Brian Shi (an Oxford alumnus) of the Neo4j Graph Data Science team, including (but not limited to):

1. LLM agents with graph tools.
   LLMs equipped with tools can act as agents capable of handling complex tasks. Adding retrieval (Cypher) and graph-algorithm tools enhances the retrieval and reasoning capabilities of LLM-based agents. What types of tasks can be solved by agents that call sequences of such tools? How can we design these tools so that LLMs can use them more effectively out of the box? Can we use techniques such as SFT, RLHF, or RLVR to improve the agent’s ability to use existing algorithmic tools?
2. Bridging graph neural networks and graph databases.
   Basic message-passing equations can be implemented in Cypher, but which GNN architectures can be (partly) expressed as queries, and which queries can be implemented and learned by GNNs? Are there GNNs that cannot be written in Cypher? How can we leverage a graph database for efficient GNN inference?

A student would work with Michael Benedikt and the Neo4j team. The project will be research-focused, and any software produced would be in the public domain. Additional hardware and resources will be provided if needed. The balance between experiment and theory can be tuned to the student’s interests. The main prerequisite is strong knowledge of graph ML (e.g., the GDL course), algorithms, databases, or general ML, depending on the specific topic.