Graph Neural Networks

Outline of Today’s Lecture

• Basics of deep learning
• Deep learning for graphs
• Graph Convoultional Networks and GraphSAGE

Machine Learning as Optimization

• Supervised learning

Loss Function Example

Minibatch SGD

• Concepts:
  • Batch size : the number of data points in a minibatch
  • Iteration: 1 step of SGD on a minibatch
  • Epoch: one full pass over the dataset (iterations is equal to ratio of dataset size and batch size)
• SGD is unbiased estimator of full gradient

Milti-layer perceptron(MLP)

Deep learning of Graph

Local network neighborhoods

Stacking multiple layers

• A Native Approach
• Idea : Convolutional Networks

Real-World Graphs

Neighborhood Aggregation

• The Math : Deep Encoder
  • Basic approach : Average neighbor messages and apply a neural network

Summary

• Recap :Generate node embeddings by aggregating neighborhood information