(bill.com) MiniGCN for Link Prediction!

What it does

Using Graph Convolutional Network to predict whether two web pages on the graph are connected.

It is a mini-implementation of GraphSAGE, a popular learning algorithm for graph data.

Accomplished ~91.5% classification accuracy on the test set!

How we built it

Data

• nodes (number id): webpage
  • (22470 linked, 1655 isolated)
• edge: exists if two pages link to each other (132039)
• Page’s text description
• Page type (label)

Pre-processing

• Node features
  • labels: provided, 4 types
  • Embedding text one-hot vectors
  • Use Doc2Vec, decide the output feature dimension based on the raw sentence length

Problem Abstraction: Link Prediction in Graph

• Small model — increase complexity
  • Deeper GraphSAGE
  • higher number of channels
  • longer text embedding

Graph

• Nodes: pages
• Edges: connectivity of pages
• Node feature: label + (embedded) text

Built With

• dgl
• doc2vec
• gcn
• networkx
• pandas
• pytorch