Openhgnn releases version 0.3

GAMMA Lab Officially released today OpenHGNN Third edition , You can go from github、 Enlighten the community or through pip Download the latest OpenHGNN. The new version includes the following new features ： Update heterogeneous graph related models 、 Provide API Call mode 、 Simplify the process of customizing datasets and models 、 Provide heterogeneous graph statistical information visualization tools, etc .

One 、 Update heterogeneous graph model

OpenHGNN In the third version, some heterogeneous graph models have been modified and added .

• A heterogeneous graph embedding model with multiple types of edges between new nodes ：GATNE(KDD 2019)
• New knowledge map embedding model ：TransE(NeurIPS 2013) TransH(AAAI 2014) TransR(AAAI 2015) TransD(ACL 2015)

We make use of DGL v0.8 New API——TypedLinear and HeteroLinear, The following attention network models have been added and modified .

modify ：

• Heterograph binding Transformer Your attention network ：HGT(WWW 2020)
• A simple heterogeneous graph attention network ：SimpleHGN(KDD 2021)

newly added ：

• Attention networks for heterostructure learning ：HetSANN(AAAI 2020)
• Explicable , Efficient heterogeneous graph attention network ：ieHGCN(TKDE 2021)

Two 、 New usage

On the basis of running the experiment with one line of command , We provide Experiment The interface provides users with another option to set parameters and run experiments under existing models and datasets . Users can use simple code , Run an experiment by customizing experiment parameters , And through the custom parameter search space to carry on the super parameter optimization .

• Run the experiment

experiment = Experiment(model='RGCN', dataset='acm4GTN', task='node_classification',
                        gpu=0, max_epoch=50, lr=0.01, hidden_dim=64, dropout=0.2, n_layers=2)
experiment.run()

• Superparameter optimization

def search_space(trial):
    return {
        "lr": trial.suggest_categorical("lr", [1e-3, 5e-3, 1e-2]),
        "hidden_dim": trial.suggest_categorical("hidden_dim", [32, 64]),
        "dropout": trial.suggest_uniform("dropout", 0.0, 0.5),
        'n_layers': trial.suggest_int('n_layers', 2, 3)
    }
experiment = Experiment(model='RGCN', dataset='acm4GTN', task='node_classification', gpu=0,
                        hpo_search_space=search_space, hpo_trials=20)
experiment.run()

3、 ... and 、 Simplify the process of customizing datasets and models

We support users to customize data sets and models through simple code , And based on OpenHGNN Construction experiment .

• Custom datasets

suffer DGL v0.8 in Dataset Adapter inspire , We provide heterogeneous graph dataset adapters , By adjusting the DGLDataset To run the OpenHGNN Node classification and link prediction tasks .

For node classification tasks , Users can only provide a heterogeneous graph dataset with node labels , Specify the target node type and partition scale , Apply your own data set to OpenHGNN In the node classification task of .

For link prediction tasks , The user provides the original heterogeneous graph 、 Predict the edge type and partition ratio , The dataset adapter automatically performs dataset partitioning and negative edge sampling on the test set and verification set , Enable user datasets to run OpenHGNN Link prediction task .

• Custom model

Most heterogeneous graph neural network models follow the message passing paradigm , In the process of forward propagation, the graph structure and initial characteristics of nodes are taken as inputs , Output the representation of the central node aggregated by neighbors . Users can develop their own systems that follow the unified forward propagation format HGNN Model , Apply it to OpenHGNN During the training process .

For details and sample code, please refer to ：https://github.com/BUPT-GAMMA/OpenHGNN/tree/main/examples/customization

Four 、 Heterogeneous graph statistical information visualization tool

We provide three heterogeneous graphs for statistical information visualization API, Proportion of nodes respectively 、 Node degree distribution 、 The number of meta paths is counted and visualized , It is convenient for researchers to intuitively display the structure information of heterograph . The following figure shows IMDB Information about the degree distribution of different types of nodes in the data set .

5、 ... and 、 The open source community

With the release of new features , We have gradually collected some feedback from community users , Let's see the user's needs and perspective OpenHGNN Defects and deficiencies in , It also provides us with ideas for future development and evolution . We welcome you to use OpenHGNN A new version of the , Any questions and suggestions are also welcome .

Github Address ：https://github.com/BUPT-GAMMA/OpenHGNN

Qizhi community address ：https://git.openi.org.cn/GAMMALab/OpenHGNN