There are two main projects running at the moment.

Generating Trees

The project on generating trees aims to create a generative model for dendritic trees of neurons and botanical trees. In contrast to many works to generate such shapes, we aim to use graph-learning methods that focus on the structure (i.e. branching pattern) of the tree first and the positions second. A core challenge for these kinds of datasets is that branches may be positioned close to each other without being connected – in contrast for molecules, which are currently the main application of graph generation, atoms that are close by do share a bond (or rather: the distance between two atoms in a molecule defines whether they are connected by a chemical bond, the direction does not matter here). In contrast to generating e.g. an office scene, trees and neurons contain much finer detail and are way more sparse. The main methods used in the project are based on graph diffusion.

Machine Learning on Relational Databases

One (surprisingly common) way to apply machine learning to databases is to essentially join all tables into a very big one which can then be send to an LLM. While this approach works to a degree and for some kind of databases, it completely ignores the structure of the database and creates a lot of redundant data (e.g. the address and capacity of the Leipzig Stadium will be repeated for all home matches of RB Leipzig). By working directly on the hypergraph defined by the database, one can train machine learning models that can extract information directly from the database. We are currently investigating how to best create the hypergraph in a way that it is useful for machine learning while also not becoming too large so it will still work on large databases.