When: Thursday, January 18, 2018 from 9:30 AM - 11:00 AMWhere: University of Toronto, MC 102
Graph-structured data is prevalent across many domains, including social networks, biological networks, and e-commerce. In these settings, one might wish to infer information missing from the graph, for example, predict user interests in a social network or predict if two proteins interact in a biological network. However, modern Machine Learning techniques, such as Deep Learning, usually operate on continuous-valued inputs (e.g. floating-point numbers) while Graphs are naturally described in discrete form (e.g. nodes and edge).
In this talk, Sami will describe modern embedding techniques that project Graphs onto a continuous vector space. In addition, he will describe Graph Convolution and how it can be used for Semi-supervised node classification, where the labels are observed for only a fraction of the nodes and one wishes to recover all unobserved labels. In this talk, he will brief some of Google's work in these domains, which include learning an edge function, and combining random walks with graph convolution for semi-supervised learning.
Students who might not be familiar with Machine Learning or Representation Learning are encouraged to join, as Sami will cover a review of the field.
Background: - What is a Graph? - Machine Learning Applications on Graphs
Graph Embeddings:- Review: Laplacian methods, Random Walks- Structure-preserving Probabilistic Objective- Edge Representation- Context Distributions
Graph Convolution:- Review: Derivation, Approximations, Graph ConvNets- Graph Convolution on Random Walks
----------------------[IMPORTANT] By completing this form, you are not applying for a role at Google. If you would like to apply, please see opportunities at google.com/students.