Subject

The lack of data is a current issue in the deep learning community. Sometimes, data scarcity can be due to lost or unmeasured data, a problem that can be solved with data reconstruction techniques. The current approach to this is deep generative models such as GANs or V(G)AEs. Variational (graph) autoencoders are one type of deep (generative) model which can produce synthetic data with similar characteristics to the input data thanks to their inference mechanism. Another kind of variational inference mechanism is the mean-field algorithm (based on Markov random fields (MRF)), which allows for approximate inference for graphical models such as VGAEs. We wish to study the integration of these two mechanisms and observe their effectiveness in the task of data reconstruction.

Kind of work

In this master thesis, the student will be able to investigate different deep generative models, specifically, VGAEs. For such task, first the sudent will need to understand the properties of graphs and the functioning of graph neural networks.
Additionally, she/he will have to understand the functioning of Markov random fields and the unfolding of these into layers of a neural network model.
Finally, the task will be to merge the VGAE model with the MRF layers to produce a state-of-the-art model which can be applied for data reconstruction.
In this thesis, we will work on different datasets for analyzing and benchmarking the designed model. For instance, this can be applied in weather data reconstruction or prediction of people density in the street (busy/not busy and why, e.g. for coronavirus measures), textual or image data reconstruction and many other tasks. It can also be applied in regression tasks as natural phenomena prediction (pollutant concentrations or temperature). The student will be able to choose from our existing datasets or of his/her choice.

Framework of the Thesis

MRF model:
• Fake News Detection using Deep Markov Random Fields. Duc Minh Nguyen, Tien Huu Do, Robert Calderbank, Nikos Deligiannis. NAACL 2019. (https://aclanthology.org/N19-1141/)
VGAEs:
• Variational Graph Auto-Encoders. Thomas N. Kipf, Max Welling NIPS 2016 (https://arxiv.org/abs/1611.07308 )
VGAES for data reconstruction:
• Matrix Completion With Variational Graph Autoencoders: Application in Hyperlocal Air Quality Inference. Tien Huu Do et al. ICASSP 2019 (https://arxiv.org/abs/1811.01662 )
Possible datasets:
• Graph classification: https://paperswithcode.com/task/graph-classification/latest
o MUTAG, collab, IMBD
• Node classification https://paperswithcode.com/task/node-classification
o MovieLens 1M dataset, Netflix dataset (node classification)

Number of Students

1

Expected Student Profile

Good knowledge of Machine Learning, AI and data processing. Good programming skills in Python (particularly PyTorch)