mehakkhan7 / gcn.pytorch

Graph Convolutional Networks for Text Classification.

Introduction

An inofficial PyTorch implementation of Semi-Supervised Classification with Graph Convolutional Networks.

Datasets

• citeseer
• cora
• pubmed

Dataset	Nodes	Edges	Features	Classes	#train	#val	#test
Citeseer	3,327	4,732	3,703	6	120(20 each class)	500(29, 86, 116, 106, 94, 69)	1000(77, 182, 181, 231, 169, 160)
Cora	2,708	5,429	1,433	7	140(20 each class)	500(61, 36, 78, 158, 81, 57, 29)	1000(130, 91, 144, 319, 149, 103, 64)
Pubmed	19,717	44,338	500	3	60(20 each class)	500(98, 194, 208)	1000(180, 413, 407)

Note:

• There are 15 nodes(without features) in Citeseer dataset.
• We use dataset splits provided by https://github.com/tkipf/gcn
• More information see https://linqs.soe.ucsc.edu/data

GCN

A is the adjacency matrix(adding self-loops), D is the degree matrix, X is the features, W is the parameters.

Train

# usage: python train.py --dataset DATASET
python train.py --dataset citeseer

or

python train.py --dataset cora

or 

python train.py --dataset pubmed

Evaluate

# usage: python evaluate.py --dataset DATASET --checkpoint CHECKPOINT
python evaluate.py --dataset citeseer --checkpoint checkpoints/citeseer/gcn_200.pth

Experiment results

• Accuracy

  	Citeseer(%)	Cora(%)	Pubmed(%)
  this repo	70.8	82.5	79.2
  paper	70.3	81.5	79.0

• Accuracy and Loss curve(see experiments/results.png)

  

  • citeseer

    tensorboard --logdir=experiments/citeseer
    

  • cora

    tensorboard --logdir=experiments/cora
    

  • pubmed

    tensorboard --logdir=experiments/pubmed
    

• T-SNE Visualization

  • citeseer(see experiments/citeseer_vis.png)

    

    python visualize.py --dataset citeseer --checkpoint checkpoints/citeseer/gcn_200.pth
    

  • cora(see experiments/cora_vis.png)

    

    python visualize.py --dataset cora --checkpoint checkpoints/cora/gcn_200.pth
    

  • pubmed(see experiments/pubmed_vis.png)

    

    python visualize.py --dataset pubmed --checkpoint checkpoints/pubmed/gcn_200.pth
    

• Dropout and Regularization (results of citeseer for training 200 epoches)

  weight_decay	dropout	train ac(%)	val ac(%)	test ac(%)
  5e-3	0.7	90.8	68.4	70.1
  5e-3	0.5	91.7	70.4	71.5
  5e-3	0.2	93.3	71.0	71.8
  5e-3	0.0	94.2	68.8	69.6
  5e-4	0.7	97.5	70.0	70.8
  5e-4	0.5	98.3	70.2	70.8
  5e-4	0.2	98.3	70.8	70.7
  5e-4	0.0	98.3	71.0	71.7
  5e-5	0.7	100.0	67.8	68.3
  5e-5	0.5	100.0	70.6	69.1
  5e-5	0.2	100.0	68.2	69.5
  5e-5	0.0	100.0	67.4	68.0
  0.0	0.7	100.0	68.0	68.0
  0.0	0.5	100.0	68.0	67.6
  0.0	0.2	100.0	63.8	59.7
  0.0	0.0	100.0	64.8	63.6

  Note: Bold parameters represents the parameters in the paper.

• All GCNLayers with regularization (results of citeseer for training 200 epoches)

  	gcn1	gcn2	accuracy(%)
  experiment	reg	reg	71.9
  paper	reg	no-reg	70.8

• Nums of train samples (results of citeseer for training 200 epoches)

  # train samples	accuracy(%)
  120	70.8
  240	70.6
  360	71.5