StellarGraph demos

StellarGraph provides numerous algorithms for graph machine learning. This folder contains demos of all of them to explain how they work and how to use them as part of a TensorFlow Keras data science workflow.

The demo notebooks can be run without any installation of Python by using Binder or Google Colab - these both provide a cloud-based notebook environment. The whole set of demos can be opened in Binder or you can click the Binder and Colab badges within each notebook.

Find algorithms for a task

• Introduction to StellarGraph and its graph machine learning workflow (with TensorFlow and Keras): GCN on Cora

• Predicting attributes, such as classifying as a class or label, or regressing to calculate a continuous number:

  • For nodes/vertices/entities: node classification

  • For edges/links/connections: link prediction (includes knowledge graph completion)

  • For graphs/networks: graph classification

  • Adjusting predictions scores to be probabilities (for any model): calibration

  • Interpreting/introspecting models, for node classification: interpretability

• Representation learning or computing embedding vectors (including unsupervised tasks):

  • For nodes/vertices/entities, edges/links/connections and graphs: embeddings

• Time series or sequence prediction for nodes within a graph (including spatio-temporal data): time series

• Ensembling models to reduce prediction variance: ensembles

• Loading data into a StellarGraph object, with Pandas, NumPy, Neo4j or NetworkX: basics

• Experimental: running GraphSAGE or Cluster-GCN on data stored in Neo4j: neo4j connector

Find a demo for an algorithm

Algorithm	Heterogeneous	Directed	Edge weights	Time-varying, temporal	Node features	Node classification	Link prediction	Unsupervised	Inductive	Graph classification
Graph Convolutional Network (GCN)	see RGCN		yes	see T-GCN	yes	demo	demo	UnsupervisedSampler, DeepGraphInfomax	via Cluster-GCN	demo
Cluster-GCN			yes		yes	demo	yes	DeepGraphInfomax	yes
Relational GCN (RGCN)	multiple edges types		yes		yes	demo	yes	DeepGraphInfomax
Temporal GCN (T-GCN), implemented as GCN-LSTM				node features	time series, sequence	demo
Graph ATtention Network (GAT)			yes		yes	demo	yes	UnsupervisedSampler, DeepGraphInfomax	via Cluster-GCN
Simplified Graph Convolution (SGC)			yes		yes	demo	yes
Personalized Propagation of Neural Predictions (PPNP)			yes		yes	demo	yes	UnsupervisedSampler, DeepGraphInfomax
Approximate PPNP (APPNP)			yes		yes	demo	yes	UnsupervisedSampler, DeepGraphInfomax	via Cluster-GCN
GraphWave						via embedding vectors	via embedding vectors	demo
Attri2Vec					yes	demo	demo	demo	yes
GraphSAGE	see HinSAGE	demo	yes		yes	demo	demo	UnsupervisedSampler, DeepGraphInfomax	demo
HinSAGE	yes				yes	yes	demo	DeepGraphInfomax	yes
Node2Vec			demo			via embedding vectors, keras layer, gensim	via embedding vectors	keras layer, gensim
Metapath2Vec	yes					via embedding vectors	via embedding vectors	demo
Continuous-Time Dynamic Network Embeddings				yes		via embedding vectors	via embedding vectors	yes
Watch Your Step			yes			via embedding vectors	via embedding vectors	demo
ComplEx	multiple edges types	yes				via embedding vectors	demo	yes
DistMult	multiple edges types	yes				via embedding vectors	demo	yes
Deep Graph CNN			yes		yes					demo

See the root README or each algorithm’s documentation for the relevant citation(s).

Download the demos

You can run download a local copy of the demos using the curl command below:

curl -L https://github.com/stellargraph/stellargraph/archive/master.zip | tar -xz --strip=1 stellargraph-master/demos

The dependencies required to run most of our demo notebooks locally can be installed using one of the following:

• Using pip: pip install stellargraph[demos]

• Using conda: conda install -c stellargraph stellargraph

Table of contents

• StellarGraph basics
  • Loading data into StellarGraph from NetworkX
  • Loading data into StellarGraph from NumPy
  • Loading data into StellarGraph from Pandas
  • Loading and saving data between StellarGraph and Neo4j
• Model calibration
  • Calibrating a GraphSAGE link prediction model
  • Calibrating a GraphSAGE node classification model
• Connector
  • Neo4j
    • Cluster GCN on Neo4j
    • Directed GraphSAGE with Neo4j
    • Loading Cora dataset into Neo4j database
    • GraphSAGE with Neo4j
• Unsupervised representation learning
  • Node representation learning with attri2vec
  • Node representation learning with Deep Graph Infomax
  • Unsupervised graph classification/representation learning via distances
  • Node representation learning with GraphSAGE and UnsupervisedSampler
  • Node representations with GraphWave
  • Node2Vec representation learning with Stellargraph components
  • Node representation learning with Metapath2Vec
  • Node representation learning with Node2Vec
  • Node representation learning with Watch Your Step
• Ensemble learning
  • Ensemble models for link prediction
  • Ensemble models for node classification
• Graph classification
  • Supervised graph classification with Deep Graph CNN
  • Supervised graph classification with GCN
• Interpretability of node classification results
  • Interpreting nodes and edges with saliency maps in GAT
  • Interpreting nodes and edges with saliency maps in GCN
  • Interpreting nodes and edges with saliency maps in GCN (sparse)
• Link prediction
  • Link prediction via inductive node representations with attri2vec
  • Knowledge graph link prediction with ComplEx
  • Link prediction with Continuous-Time Dynamic Network Embeddings (CTDNE)
  • Knowledge graph link prediction with DistMult
  • Link prediction with GCN
  • Link prediction with GraphSAGE
  • Link prediction with Heterogeneous GraphSAGE (HinSAGE)
  • Comparison of link prediction with random walks based node embedding
  • Link prediction with Metapath2Vec
  • Link prediction with Node2Vec
• Node classification
  • Node classification via node representations with attri2vec
  • Node classification with Cluster-GCN
  • Node classification with directed GraphSAGE
  • Node classification with Graph ATtention Network (GAT)
  • Semi-supervised node classification via GCN, Deep Graph Infomax and fine-tuning
  • Node classification with Graph Convolutional Network (GCN)
  • Inductive node classification and representation learning using GraphSAGE
  • Node classification with GraphSAGE
  • Node classification with Node2Vec using Stellargraph components
  • Node classification with Node2Vec
  • Node classification with weighted Node2Vec
  • Node classification with Personalised Propagation of Neural Predictions (PPNP) and Approximate PPNP (APPNP)
  • Node classification with Relational Graph Convolutional Network (RGCN)
  • Node classification with Simplified Graph Convolutions (SGC)
• Graphs with time series and sequence data
  • Forecasting using spatio-temporal data with combined Graph Convolution + LSTM model
• StellarGraph internal development
  • Resource usage of the StellarGraph class