TensorTonicTensorTonic
Problems
Study PlansProjectsInterviewPricingFeedback
Back to Projects

Build from first principles

Classify Research Papers by Their Citations

Build a graph neural network that predicts a paper's topic from its words and the papers it is linked to, then find out why stacking many layers makes every paper look alike.

What is a graph neural network?Papers as nodes, citations as edges, and why neighbors are evidence
1Score Predictions on One SplitScore a model on one split of the papers
2Build the Normalized Adjacency MatrixTurn citation links into an averaging matrix
3Pass Messages Through One GCN LayerPass messages along citation links
4Average the Neighbors, Keep Yourself ApartAggregate neighbors by their mean
5Measure Over-smoothingWatch papers become alike
From Layers to a ClassifierSee how your pieces become a trained GCN before running it on Cora
Classify Cora and Break ItTrain a GCN on real citations, watch deep networks over-smooth, fix it, and compare across seeds

One paper, many messages

2,708 Cora papers
0123456

Paper 0

links4gathersitself + 4 papersstepcollect messages
Wordswhat a paper says
Linkswho it cites
Averageitself and neighbors
Topicone of 7
Start project