Anmerkungen:
Diese Datenquelle enthält auch Bestandsnachweise, die nicht zu einem Volltext führen.
Beschreibung:
In this work, we investigate how to improve machine learning algorithms intended to work on graph structure. First, we leverage ideas from distributed computing algorithms to improve graph neural networks on graph classification tasks by enhancing their expressive power and efficiency. In the second part of the work, we focus on generative tasks. Generative tasks, such as generating novel examples similar to a set of provided examples, can be viewed as an ultimate test of the understanding of the given domain. For this, we propose multiple ways to improve generic generative methods by tailoring them more to the graph domain than in the previous methods. For example, by first generating the large-scale graph structure by generating the corresponding graph spectra. Or by ensuring that the diffusion process used to generate adjacency matrices remains discrete. We further showcase the real-world applicability of such work of generating discrete structures in two different domain adaptations of graph generation. Namely, the generation of crease patterns for rigid origami and the generation of novel antibodies that have been shown to perform their function well in vitro.