Master Thesis

Learning event representations on graphs

Work default illustration



  • If you are interested in the proposal, please contact with the supervisors.


Graphs are commonly used to describe the geometry of structured data in a wide range of domains, including social networks, physical and biological systems. Representation learning on graphs, also known as graph embedding, entails to learn graph node representations as low dimensional vectors. Such representations have the advantage of capturing critical information about the data geometry which is encoded in the graph structure. However, in a variety of problems the precise structure of the graph underlying the observed data is also unknown, and only general priors on the graph structure are available.

This project focuses on the problem of characterizing the temporal structure of videos and image sequences (i.e. events) by learning low dimensional node embeddings that capture prior constraints on the underlying graph (which is unknown) jointly with the high-dimensional data geometry.

The student is expected to be fluent in Python and having a good mathematical background. Knowledge of graphical models, graph neural networks and machine learning will be a plus.


[1] Hamilton, W. L., Ying, R., & Leskovec, J. (2017). Representation learning on graphs: Methods and applications. IEEE Data Engineering Bulletin.