A question about randomly splitting the Planetoid datasets #3040
-
|
How many nodes should we choose for training, validating, and testing when randomly splitting the Planetoid datasets, i.e. Cora, Citeseer and PubMed? I am asking this since if I use 50 as the number of nodes from each class of Cora (all in total 350 nodes for training), I get higher accuracy (around 5% higher) than the case in which I use 20 as the number of nodes from each class (all in total 140 nodes for training). |
Beta Was this translation helpful? Give feedback.
Replies: 1 comment 1 reply
-
|
More training labels usually results in better model performance. As such, there is no rule how many nodes should be used for training (and usually depends on how many labels are given overall). If you want to compare against existing methods, you need to take care of evaluating on the same split ratios to make the comparison fair. |
Beta Was this translation helpful? Give feedback.
-
|
Thank you. |
Beta Was this translation helpful? Give feedback.
More training labels usually results in better model performance. As such, there is no rule how many nodes should be used for training (and usually depends on how many labels are given overall). If you want to compare against existing methods, you need to take care of evaluating on the same split ratios to make the comparison fair.