Very large graphs are ubiquitous on the Internet, and graph data is often essential to solving applied computational social science problems. One exceedingly common such problem is that of supervised classification (or regression) on the nodes of edges of a graph. Picking a scalable modeling strategy is a key practical challenge to solving such graph-based supervised machine learning problems. Broadly speaking, modeling approaches can be divided into attempts to model the graph structure explicitly (e.g. through Loopy Belief Propagation) or those approaches that use dimensionality reduction (e.g. Non-Negative Matrix Factorization or Latent Dirichlet Allocation, etc.) to extract node-level features.

A more recent development comes from the field of neural network research, where several new techniques have been used to derive graph embeddings. In particular, the use of automated differentiation has opened up new scalable ways of thinking about graphs, which also promise to revolutionize how we do research on social networks.

In this workshop I will focus on learning graph embeddings using PyTorch, a Python-based framework for stochastic computation. Because of the elegance of PyTorch's semantics (which include straightforward integration with GPUs), graph embeddings generalize readily to weighted and signed graphs, as well as to hypergraphs. While packages like PyTorch present us with a step change in our ability to process graph data, they are still limited by computational resources: I will end my discussion with an overview of the challenges involved in processing graph data at scale.

As there is limited capacity, please register your interest here: http://vuw.qualtrics.com/jfe/form/SV_aVRjNwPaub3UKLb