I have a complete graph with N nodes and (N*N-1)/2 links. Total links with positive label are N-1. I use GINEConv for link prediction. I need to predict which links are part of the graph (positive label). Since the number of positive links is much lower than the negative links, the model cannot learn to predict + links. It just predicts - links with high accuracy. I also tried to give higher weight to positive links. With higher weight, the model just predicts + links! I am beginner in GNN field. Any idea is appreciated. I think I can perform undersampling or oversampling techniques. Any other solutions?
I tried this loss function with different weights: torch.nn.BCEWithLogitsLoss(pos_weight=torch.tensor([100])) When the weight is high, the model predicts only + and when the weight is low, the model predicts only -.