andyjzhao / hgsl Goto Github PK

View Code? Open in Web Editor NEW

112.0 3.0 18.0 56.67 MB

Source code of AAAI21-Heterogeneous Graph Structure Learning for Graph Neural Networks

License: Apache License 2.0

Python 100.00%

hgsl's Introduction

HGSL

Source code of AAAI submission "Heterogeneous Graph Structure Learning for Graph Neural Networks"

Requirements

Python Packages

Python >= 3.6.8
Pytorch >= 1.3.0

GPU Memmory Requirements

ACM >= 8G
DBLP >=5G
Yelp >=3G

Usage

Take DBLP dataset as an example: python train.py --dataset='dblp'

FAQ

Code of preprocessing data?

Please kindly note that the data is originally preprocessed by the GTN project (https://github.com/seongjunyun/Graph_Transformer_Networks).

I received quite a lot emails asking me about the dataset. I will not respond to them anymore as I cannot provide the code.

How to generate semantic embeddings?

The semantic embeddings, i.e. $\mathcal{Z}$ in the paper, are generated by metapath2vec algorithm. Users may refer to https://github.com/dmlc/dgl/tree/master/examples/pytorch/metapath2vec for an implementation.

hgsl's People

Contributors

Stargazers

Watchers

Forkers

fendouwangshan xrosliang shengguanwsu willamjie shigen97 ag027592 baiyazi zhh0998 mcdragon liujing1023 hookk shuowang-ai knighticon eiivy juke-morant weiwei1206 peiliangwang

hgsl's Issues

semantic embeddings

I wondering how did you get the semantic embeddings Z for each metapath Pm,such as the apcpa_emb.pkl in dblp directory

Threshold Setting and the labeled data

Dear Authors:
I have two questions about the experiment.
(1) Given a new dataset, how to set the threshold for different graph construction?
(2) In the experimental results, why the number of train + val + test set is not equal to the number of nodes? Are the test data randomly selected?
Thanks :)

EarlyStopping counter?

Hello,
excuse me, i have a question to ask for your help
when i ran the 'train.py' in my pycharm, i found each Epoch follewed by 'EarlyStopping counter:...' ,what is that mean?

A question about DBLP dataset

Dear HGSL authors,

I find that the DBLP dataset you used in your paper does not include term nodes.

Is there any reason to do so? I think there are about 8000 term nodes in DBLP dataset used by HAN, MAGNN and other paper as well. I am a little bit confused.

Look forward to your reply!

How to generate semantic embedding matrices?

Dear Jianan Zhao, I have run your code but I do not know how to generate Semantic Embedding Matrices, could you please tell me how to constructs the semantic embedding matrices Z by metapath-based node embeddings from M metapaths？

通道注意力权重

你好，请问论文中的通道注意力权重是自己设置的吗，比如最后您HGSL.py中代码中的
【1，1，10】
self.overall_g_agg[r] = GraphChannelAttLayer(3, [1, 1, 10])

通道注意力权重

您好，请问论文中的通道注意力权重是自己设置的吗，比如最后您HGSL.py中代码中的
【1，1，10】self.overall_g_agg[r] = GraphChannelAttLayer(3, [1, 1, 10])
如果最后整合三个图的权重需要自己设置，那么
self.sg_agg[r] = GraphChannelAttLayer(len(cf.mp_list))
和self.fg_agg[r] = GraphChannelAttLayer(3)的权重也需要自己手动设置吗？
拜托看到了请回复下，谢谢谢谢！！！

A question about yelp data set

Hello, I have read your paper and am very interested. The node features of the yelp data set are constructed by the word bag vector of keywords, but I don’t know where the keywords in the data set are. I hope to find the original (with key Word) data set, can you help me？
There are where I can find the current data source:
https://github.com/librahu/HeGAN
https://github.com/Andy-Border/HGSL
https://github.com/dingdanhao110/Conch
https://github.com/librahu/HIN-Datasets-for-Recommendation-and-Network-Embedding

but now I can not find the keyword in business (or feature).....\thanks a lot!

Adaption

Your job is great. I would like to ask if this method is suitable for multimodal sentiment analysis?(There exist three modality, eg Audio, text and video.And the data appear heterogeneous nature)