This is an open-source toolkit for Heterogeneous Information Network Embedding(OpenHINE) with version 0.1. We can train and test the model more easily. It provides implementations of many popular models, including: DHNE, HAN, HeGAN, HERec, HIN2vec, Metapath2vec, MetaGraph2vec, RHINE. More materials can be found in www.shichuan.org.

We build a new toolkit OpenHGNN, which is a high-level package built on top of DGL. It will have Better Extensibility, Better Encapsulation and More Effiencient. And it includes two embedding models, Metapath2vec and HeRec.

convenience provided:

• ​ easy to train and evaluate
• ​ able to extend new/your datasets and models
• ​ the latest model available: HAN、HeGAN and so on

DMGroup from BUPT: Tianyu Zhao, Meiqi Zhu, Nian Liu, Jiawei Liu, Hongrui Liu, Guanyi Chu, Jiayue Liu, Jianan Zhao, Xiao Wang, Cheng Yang, Chuan Shi.

• Python version >= 3.6

• PyTorch version >= 1.4.0

• TensorFlow version >= 1.14

• Keras version >= 2.3.1

python train.py -m model_name -d dataset_name

e.g.

python train.py -m Metapath2vec -d acm

The model parameter could be modified in the file ( ./src/config.ini ).

• common parameter

​ --alpha: learning rate

​ --dim: dimension of output

​ --epoch: the number of iterations

​ --num_workers:number of workers for dataset loading (It should be set to 0, if you are in trouble with Windows OS.)

​ etc...

• specific parameter

​ --metapath: the metapath selected

​ --neg_num: the number of negative samples

​ etc...

If you want to train your own dataset, create the file (./dataset/your_dataset_name/edge.txt) and the format is as follows：

​ src_node_id dst_node_id edge_type weight

​ e.g.

	19	7	p-c	2
	19	7	p-a	1
	11	0	p-c	1
	0	11	c-p	1

PS：The input graph is directed and the undirected needs to be transformed into directed graph.

​ number_of_nodes embedding_dim

​ node_name dim1 dim2

e.g.

11246	2
a1814 0.06386946886777878 -0.04781734198331833
a0 ... ...

​ Structural Deep Embedding for Hyper-Networks

​ src code:https://github.com/tadpole/DHNE

​ Heterogeneous Graph Attention Network

​ Add feature.txt into the input folder or set the parameter "featype": "adj", which means that you will use adjacency matrix as your feature.

​ src code:https://github.com/Jhy1993/HAN

​ Adversarial Learning on Heterogeneous Information Network

​ src code:https://github.com/librahu/HeGAN

​ Heterogeneous Information Network Embedding for Recommendation

​ src code:https://github.com/librahu/HERec

​ metapath_list: pap|psp (split by "|")

​ HIN2Vec: Explore Meta-paths in Heterogeneous Information Networks for Representation Learning

​ src code:https://github.com/csiesheep/hin2vec

​ metapath2vec: Scalable Representation Learning for Heterogeneous Networks

​ src code:https://ericdongyx.github.io/metapath2vec/m2v.html

​ the python version implemented by DGL:https://github.com/dmlc/dgl/tree/master/examples/pytorch/metapath2vec

​ MetaGraph2Vec: Complex Semantic Path Augmented Heterogeneous Network Embedding

​ src code:https://github.com/daokunzhang/MetaGraph2Vec

​ PTE: Predictive Text Embedding through Large-scale Heterogeneous Text Networks

​ src code:https://github.com/mnqu/PTE

​ Relation Structure-Aware Heterogeneous Information Network Embedding

​ only supported in the Linux

​ src code:https://github.com/rootlu/RHINE 

python test.py -d dataset_name -m model_name -n file_name

The output embedding file name can be found in (./output/embedding/model_name/) .

e.g.

python test.py -d dblp -m HAN -n node.txt

​ number_of_nodes embedding_dim

​ node_name dim1 dim2

e.g.

11246	2
a1814 0.06386946886777878 -0.04781734198331833
a0 ... ...

HAN uses the dataset without features.

We will use the dgl as our backend. And the OpenHINE will not be updated. We will be dedicated in building the new toolkit OpenHGNN, which is a high-level package built on top of DGL. It will have Better Extensibility, Better Encapsulation and More Effiencient.