Know Your Neighbours: Machine Learning on Graphs
Machine learning has become ubiquitous in many applications. There are many accessible tools available to apply standard machine learning models to make predictions on data.
Typically, machine learning problems aim to predict something about an entity using some data about each entity. However, in the real world, entities - people, places or things - are connected to each other and can have complex interactions with their neighbours. We can use this information to improve our predictions, and to gain more insight into the network structure and how entities can affect each other.
This talk will introduce machine learning on graphs, give some examples where graph approaches give large improvements over standard machine learning techniques, and will demonstrate some tools that make graph machine learning more approachable.
Outline/Structure of the Talk
Predicting attributes with machine learning:
- How can we predict something from data?
- Assumptions of this approach.
Graphs and networks:
- Data in the real-world is connected.
- Machine learning without features.
- Homogeneous and heterogeneous graphs.
Examples of graph-based machine learning.
- Feature-based models in a graph framework.
- Collaborative filtering models in a graph framework.
You'll learn how how to go from standard machine learning to graph-based machine learning, and when and why you would want to use graph-based algorithms.
Data scientists, developers, and others interested in new techniques in machine learning.
Prerequisites for Attendees
A knowledge of basic machine learning, including logistic regression and linear regression, is assumed. Code samples will be presented in Python, so some familiarity with Python would be useful.