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_namee.g.
python train.py -m Metapath2vec -d acmThe model parameter could be modified in the file ( ./src/config.ini ).
--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...
--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_nameThe 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 ... ...
| ACM dataset | Micro-F1 | Macro-F1 | NMI |
|---|---|---|---|
| DHNE | 0.7201 | 0.7007 | 0.3280 |
| HAN | 0.8401 | 0.8362 | 0.4241 |
| HeGAN | 0.8308 | 0.8276 | 0.4335 |
| HERec | 0.8308 | 0.8304 | 0.3618 |
| HIN2vec | 0.8458 | 0.8449 | 0.4148 |
| Metapath2vec(PAP) | 0.7823 | 0.7725 | 0.2828 |
| MetaGraph2vec | 0.8085 | 0.8019 | 0.5095 |
| PTE | 0.7624 | 0.7543 | 0.3781 |
| RHINE | 0.7699 | 0.7571 | 0.3970 |
| DBLP dataset | Micro-F1 | Macro-F1 | NMI |
|---|---|---|---|
| DHNE | --- | --- | --- |
| HAN | 0.8325 | 0.8141 | 0.3415 |
| HeGAN | 0.9414 | 0.9364 | 0.7898 |
| HERec | 0.9249 | 0.9214 | 0.3412 |
| HIN2vec | 0.9495 | 0.9460 | 0.3924 |
| Metapath2vec(APCPA) | 0.9483 | 0.9448 | 0.7786 |
| MetaGraph2vec | 0.9138 | 0.9093 | 0.6136 |
| PTE | 0.9335 | 0.9301 | 0.3280 |
| RHINE | 0.9360 | 0.9316 | 0.7356 |
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.