Question about PubMed performance

[ltz0120]

Hi,

In your paper, GRACE achieves 86.7% in Pubmed, and DGI achieves 86% in Pubmed. However, in the DGI paper, the performance of DGI only achieves 76.8% in Pubmed. I also notice you follow the DGI setting in your experiments. How did you make an improvement of almost 10% on DGI?

[downeykking]

The dataset spilt setting is different. DGI uses public spilt while GRACE uses 10% train, 10% valid, and 80% test random spilt.

[gaoshihao333]

The dataset spilt setting is different. DGI uses public spilt while GRACE uses 10% train, 10% valid, and 80% test random spilt.

GRACE train/test split follows a 90% / 10% . Isn't that it?

[downeykking]

The dataset spilt setting is different. DGI uses public spilt while GRACE uses 10% train, 10% valid, and 80% test random spilt.

GRACE train/test split follows a 90% / 10% . Isn't that it?

In paper,

On these citation networks, we randomly select 10% of the nodes as the training set, 10% nodes as the validation set, and leave the rest nodes as the test set.

train/valid/test split follows a 10%/10%/80%.

And in code, train/test split follows a 10%/90%.
https://github.com/CRIPAC-DIG/GRACE/blob/master/train.py#L41C4
https://github.com/CRIPAC-DIG/GRACE/blob/master/eval.py#L58

[gaoshihao333]

The dataset spilt setting is different. DGI uses public spilt while GRACE uses 10% train, 10% valid, and 80% test random spilt.

GRACE train/test split follows a 90% / 10% . Isn't that it?

In paper,

On these citation networks, we randomly select 10% of the nodes as the training set, 10% nodes as the validation set, and leave the rest nodes as the test set.

train/valid/test split follows a 10%/10%/80%.

And in code, train/test split follows a 10%/90%. https://github.com/CRIPAC-DIG/GRACE/blob/master/train.py#L41C4 https://github.com/CRIPAC-DIG/GRACE/blob/master/eval.py#L58

Thanks for answering! So, no validation sets are used in the code?

[gaoshihao333]

The dataset spilt setting is different. DGI uses public spilt while GRACE uses 10% train, 10% valid, and 80% test random spilt.

GRACE train/test split follows a 90% / 10% . Isn't that it?

In paper,

On these citation networks, we randomly select 10% of the nodes as the training set, 10% nodes as the validation set, and leave the rest nodes as the test set.

train/valid/test split follows a 10%/10%/80%.
And in code, train/test split follows a 10%/90%. https://github.com/CRIPAC-DIG/GRACE/blob/master/train.py#L41C4 https://github.com/CRIPAC-DIG/GRACE/blob/master/eval.py#L58

Thanks for answering! So, no validation sets are used in the code?

GCA？Are validation sets used in GCA?

[downeykking]

The dataset spilt setting is different. DGI uses public spilt while GRACE uses 10% train, 10% valid, and 80% test random spilt.

GRACE train/test split follows a 90% / 10% . Isn't that it?

In paper,

On these citation networks, we randomly select 10% of the nodes as the training set, 10% nodes as the validation set, and leave the rest nodes as the test set.

train/valid/test split follows a 10%/10%/80%.
And in code, train/test split follows a 10%/90%. https://github.com/CRIPAC-DIG/GRACE/blob/master/train.py#L41C4 https://github.com/CRIPAC-DIG/GRACE/blob/master/eval.py#L58

Thanks for answering! So, no validation sets are used in the code?

GCA？Are validation sets used in GCA?

GCA utilizes validation sets, whereas GRACE does not incorporate validation sets. You can refer to their eval.py.