M.Sc. seminar

Representation Learning on Dynamic Graphs

Abstract: Graphs are a common language in modeling several problems, from social and economic networks to interactions in cells and brain neurons. According to the availability of an enormous amount of data from graphs, Machine Learning algorithms gained lots of attention in this area. But the main challenge is how to represent and encode nodes so that such models could interpret the data and the representation preserve the main features of nodes.

There are different types of methods for this problem, some are based on a distance function defined on nodes, some are based on random walks like DeepWalk and node2vec, there is also a range of methods based on Deep Learning models like Graph Convolutional Networks and GraphSAGE. These methods are originally designed for static graphs and learn representations for one snapshot of a graph, while in most problems graphs evolve in time and new nodes and edges arrive and leave and the model has to encounter these events and update representations meanwhile.

In this research, we are going to first examine the latest methods proposed for dynamic graph representation learning and then try to enhance them in tasks like link prediction and node classification.

Presentation file