A TensorFlow implementation of GraphSTONE, as described in our paper:
Graph Structural-topic Neural Network (KDD 2020, Research Track)
python main.py
Tensorflow 1.10.0, Networkx 1.11, Python 3
We provide cora and ppi datasets as examples under data/cora and data/ppi.
Note that the data/dataset_name/features.npy has undergone a dimensionality reduction via PCA, and is not identical to the original cora features.
For parameter settings, please see conf.json.
Some parameter definitions:
| Name | Default | Note |
|---|---|---|
| dataset | cora | dataset name |
| input_node_feature | True | input original node features ("True") or not ("False") |
| PreProcess/number_paths | 50 | number of paths from a center node, for generating "word" and "document" concepts on graphs |
| PreProcess/path_length | 15 | max length of random walks from a center node, for generating "word" and "document" concepts on graphs |
| TopicModel/number_topic | 5 | number of structural-topics |
| TopicModel/max_features_dim | 2500 | max topic_features (for the input of structural-topic GNN) dimension |
| TopicGCN/max_training_steps | 5000 | max steps for training |
@inproceedings{long2020graph,
title={Graph Structural-topic Neural Network},
author={Long, Qingqing and Jin, Yilun and Song, Guojie and Li, Yi and Lin, Wei},
booktitle={Proceedings of the 26th ACM SIGKDD International Conference on Knowledge Discovery \& Data Mining},
pages={1065--1073},
year={2020}
}
Certain parts of this project are partially derived from GraLSP and AnchorRecovery.