1. Keshav Balasubramanian
Inductive Representation Learning on Large Graphs William L. Hamilton, Rex Ying, Jure Leskovec
Keshav Balasubramanian

2. Outline Main goal: generating node embeddings Survey of past methods
GCNs
GraphSAGE
Algorithm
Optimization and learning
Aggregators
Experiments and results
Conclusion

3. Node Embeddings
Fixed length vector representations of nodes in a graph (similar to word embeddings)
Can be used in downstream machine learning and graph mining tasks
Need to learn smaller dense embeddings from higher dimensional information

4. Methods Non Deep Learning based models: DeepWalk, node2vec
Biased random walk based approaches
Attempt to linearize a graph by viewing the results of the walks as sentences
Deep Learning models: Vanilla graph neural networks, GCNs
Neural Networks learn the representations

5. Graph Convolution Networks
Type of deep learning model that learns node representations
Two approaches:
Spectral – Involves operating in the spectral domain of the graph, specifically on the Laplacian and Adjacency matrix.
Spatial – Convolution is defined directly based on the spatial neighborhood of a node
Drawbacks of the spectral approach
Involves operating on entire Laplacian
Transductive, generalizes poorly to unseen nodes

6. GraphSAGE Spatial and inductive graph convolution network
Nodes learn from neighborhood in graph
A fixed size neighborhood is randomly sampled for each node

7. Algorithm
GraphSAGE forward pass

8. Learning the parameters of the network
Supervised
Final embeddings are passed through a dense layer to get class probabilities
Eg: Labelling the nodes of a graph
End goal is labelling the node, not generating the embedding
Unsupervised
Use custom loss function that ensure “nearby” nodes have similar representations, “far away” nodes have dissimilar representations
Negative sample to ensure latter
Nearness defined by cooccurrence score on a fixed length random walk

9. Aggregators
Mean:
Pooling:
LSTM based aggregator

10. Experiments and Results

11. Conclusion
Proposed an inductive framework to learn node representations based on spatial graph convolutions