Skip to content

Latest commit

 

History

History
 
 

HetGNN

Folders and files

NameName
Last commit message
Last commit date

parent directory

..
README.md
README.md
 
 

README.md

HetGNN[KDD2019]

Paper: Heterogeneous Graph Neural Network

Code from author: https://github.com/chuxuzhang/KDD2019_HetGNN

How to run

Clone the Openhgnn-DGL

python main.py -m HetGNN -t node_classification -d academic4HetGNN -g 0

If you do not have gpu, set -g -1.

Performance

Node classification for academic4HetGNN

Node classification Macro-F1 Micro-F1
paper 0.978 0.979
OpenHGNN 0.9701 0.9705

Author link prediction for academic4HetGNN

A-II(type-1) authors link prediction AUC F1
paper 0.717 0.669
OpenHGNN 0.7235 0.7205

Dataset

We process the Academic II dataset given by HetGNN. It saved as dgl.heterograph and can be loaded by dgl.load_graphs

You can download the dataset by

wget https://s3.cn-north-1.amazonaws.com.cn/dgl-data/dataset/academic4HetGNN.zip

Or run the code mentioned above and it will download automaticlly.

Description: academic4HetGNN

TrainerFlow: HetGNNTrainer

  • Sampling Heterogeneous Neighbors (C1)
    • It designs a heterogeneous neighbors sampling strategy based on random walk with restart (RWR). Build a hetgnn_graph, which is used in aggregation neighbours in form of a full graph. So it limits the scale of graph though we give the mini-batch trainer. [openhgnn/sampler/HetGNN_sampler.py]
    • [TODO] Combine the sampler with NeighborSampler without generating the hetgnn_graph.
  • Encoding Heterogeneous Contents (C2)
    • encoder_het_content(nn.Module)
    • Note: A node can carry unstructured heterogeneous con- tents,e.g.,attributes,textorimage. So if you build a new dataset, overwrite the extract_feature func in HetGNN(BaseModel) [openhgnn/models/HetGNN.py]
  • Aggregating Heterogeneous Neighbors (C3)
    • aggregate_het_neigh(nn.Module)

Hyper-Parameter

You can modify the parameters in openhgnn/config.ini

Description

batch_size # the batch_size of nodes sampled in SkipGramBatchSampler
window_size # window size for relation extration
num_workers	# how many subprocesses to use for data loading. ''0'' means that the data 										will be loaded in the main process.
batches_per_epoch # the batches trained in a epoch

# the three parameters decided in what nodes aggregated.[used in building hetgnn_graph]
rw_length # length of each walk
rw_walks # window size for relation extration
rwr_prob # the probability of restart

#train epoch and early stop patience
max_epoch
patience
mini_batch_flag = True # we only support the mini-batch training.

Best config can be found in best_config

More

Contirbutor

Tianyu Zhao[GAMMA LAB]

If you have any questions,

Submit an issue or email to tyzhao@bupt.edu.cn.