Overview of Graph Representation Learning

• How to apply power of deep learning to more complex domains?

• Graphs are the new frontier of deep learning (connect things)

  • GNN: hottest subfield in ML

    • Flexibility to model complex data

• Graphs

  • Molecules, knowledge graph, software

  • Applications: recommender system, neutrino detection, LHC, fake news detection, drug repurposing, chemistry, computer graphics, VR, robotics, autonomous driving, medicine

• Stanford: deployed technology

• Workshop: representation learning for Graph

Discuss

• Tools and frameworks

• Short talks about a wide range of application domains

  • Graphics and Vision

  • Fraud and intrusion detection

  • Financial networks

  • Knowledge Graph and Reasoning

  • ...

• Software tools

Goal: Representation Learning

• Map nodes to d-dim embedding such that similar nodes in the network are embedded close together

  • Feature representation, embedding

• DL in graphs

  • Input: network

  • Predictions: node labels, new links, generated graphs and subgraphs

• Why hard

  • Networks are complex

    • Arbitrary size and complex topological structure (no spatial locality like grids)

    • No fixed node ordering or reference point

    • Often dynamic

• Problem setup

  • Graph G, vertex set V, adjacency matrix A, node features X

• Key idea: network is a computation graph

  • Learn how to represent through the network

  • Each node defines a computation graph

  • Train on a set of nodes, i.e. a batch of computation graphs

    • Back prop through the computation graphs

    • Scale (small # of parameters)

• Benefits

  • No manual feature engineering is needed

  • End-to-end learning results in optimal features

  • Any graph ML task

    • Node-level, link-level

  • Scalable

• Key idea

  • GNN adapts to the shape of the data

    • Other DL assume fixed input

    • GNN makes no such assumptions