Skip to content

Latest commit

 

History

History
 
 

NARS

Folders and files

NameName
Last commit message
Last commit date

parent directory

..
README.md
README.md
 
 

README.md

NARS

Paper:Scalable Graph Neural Networks for Heterogeneous Graphs

Code from author:https://github.com/facebookresearch/NARS

How to run

Clone the Openhgnn-DGL

python main.py -m NARS -t node_classification -d acm4NARS -g 0 --use_best_config

Candidate dataset: acm4NARS

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

candidate dataset

acm4NARS

NOTE: NARS can handle mag oag dataset, we will add these two datasets in our further work.

description

  • Number of nodes

    paper 4025
    author 17431
    field 73

  • Number of edges

    paper-author 13407
    paper-field 4025

  • Subsets: paper-author, paper-field

performance

Node classification

accuracy
acm4NARS 0.93

TrainerFlow: node_classification

model

  • NARS

    NARS is composed of WeightedAggregator and SIGN.

  • WeightedAggregator

    Get new features by multiplying the old features by the weight matrix.

  • SIGN

    The MLP classifier. It is composed of a several linear layers. Then project the node embeddings to the vector space to predict the type of the nodes.

Hyper-parameter specific to the model

R = 2
input_dropout = True
cpu_preprocess = True
ff_layer = 2

Best config can be found in best_config

More

Contirbutor

Tianyu Zhao, Yibo Li[GAMMA LAB]

If you have any questions,

Submit an issue or email to liushiliushi0@gmail.com.