snap-stanford/neural-subgraph-learning-GNN

Imbalance Dataset distribution

abhishekrajgaria opened this issue · 1 comments

As mentioned in the following paper of Neural Subgraph Matching paper https://arxiv.org/abs/2007.03092 , the dataset distribution of the imbalanced data is 3:1 (negative_sample: positive_sample),

but when we tried to get the distribution ourselves for the enzyme-imbalanced, we got the following outcome.
Total Sample - 2048
Positive Sample - 48
Negative Sample - 2000

It seems that it is not following distribution as mentioned in the paper.

We recently made an update of the paper (should be available on arxiv in a few days), where we changed the setting to focus on matching neighborhood as well as the entire graph. The label imbalance is close to what you mentioned here (see the label imbalance section in appendix).

But for the table 1 we are using the balanced setting (equal number of positive and negative examples)