Supervisors

Suitable for

Abstract

Background:

Feature propagation is a technique to impute missing features in graphs. However, many real world graphs are temporal, changing over time. Accordingly, a node that has features at one point may lack them in another. This can be dealt with by treating the problem as a relaxation over the graph+time instead of only the graph. Doing so is relevant also to non-classical graph problems that contain missing data, eg. noisy video streams.

Focus:

Research questions – how can one effectively impute features in temporally evolving graphs.

Expected contribution – developing, implementing, and testing different methods of temporal feature propagation.

Method:

The student would build upon https://arxiv.org/abs/2111.12128, with inspiration from other works that treat graphs continuously or analyse them using differential equations, eg. https://arxiv.org/pdf/2202.02296.pdf and https://arxiv.org/abs/2106.10934.

Goals:

• Fully formulate temporal feature propagation approaches.
• Empirically verify on standard temporal graph learning benchmarks.
• Test on graph signal processing problems, eg. video streams, evolving social networks, etc.

Pre-requisites: a Graph Representation Learning/Geometric Deep Learning course, ideally also a differential equations/computational numerics course