Graph Neural Networks – harvesting relations
Modern machine learning advances by drawing more information from the context of the task. Multimodal Learning combines text, image, and author into a shared space to reason about all these different entities together. Still, this and a few other applications may have more entangled context than one to one mapping between entities, such as recommendation systems, social networks, knowledge bases, and molecular structures. The graph between entities provides inductive bias imposed by the structure, which is necessary to improve many relation-dependent machine learning tasks.
In the course, we will discuss how many problems map to interaction and dependency graphs. Recent developments of Deep Learning resulted in a family of Graph Neural Networks (GNN) models, which harvest the value of these relations for classification, link prediction, and representation learning. The GNN explicitly models entities and relationships between them to create an efficient learning space for relation-dependent tasks (such as clients, goods, reviews – all linked by purchases, views, reactions). We will dive into the state-of-the-art GNN algorithms, their differences, and applications.
- Jupyter notebook
- Basic calculus
- Linear algebra
Level of complexity of course
Dr. Oleksandr Pryymak
Dr. Oleksandr Pryymak works on the problems of Fake Accounts, previously on content personalization. His academic interests started in distributed learning, trust, and focused on conflicting opinion sharing in large societies. He is dived into dynamic processes on graphs, multi-agent learning, now converting into relationship learning.
Fields of interests: Applied Machine Learning, Deep Learning, Graphs, Privacy