CogDL v0.1.1

New Features

• Support link prediction task on knowledge graphs
• Support hyper-parameter search using optuna

New Models

• GCC for graph classification: GCC is a contrastive learning framework that implements unsupervised structural graph representation pre-training.
• GRAND for node classification (thanks to @wzfhaha): GRAND randomly drops node features in training process to implement data augmentatoin and achieves sota in benchmarks.
• DGI for unsupervised node classification: DGI applies local-global contrastive learning methods to train GNN and first achieves results comparable to semi-supervised methods in benchmarks.
• MVGRL for unsupervised node classification: MVGRL is a self-supervised approach based on contrastive multi-view learning to learn representations.
• ProNE++ for unsupervised node classification: ProNE++ employs graph filter and AutoML to help enhance node embeddings.
• GraphSAGE for unsupervised node classification: unsupervised version of GraphSAGE.
• DisenGCN for node classification: DisenGCN disentangles node representations by separating different factors.
• CompGCN/RGCN for KG link prediction: RGCN and CompGCN are GNNs for knowledge graph embedding considering the type of edges.

New Results

• GCC results for heterogeneous node classification task

New Datasets

• Add some knowledge graph datasets (FB/WN datasets)

New Examples

• Add an example using hyper-parameter search
• Add an example using a custom dataset/model

Bug Fixes

• Fixed "division by zero" bug in Sparse GAT model

Requirement Update

• CogDL now requires optuna
• CogDL does not require dgl.model_zoo anymore.

Miscellaneous

• Add a check whether tuples of (task, model, dataset) are matching in the training script
• Add a GCC pre-trained model in saved/