problem with metattack about greatx

[EdisonLeeeee]

Thank you for reporting. There are some reasons for the poor results:

• Randomness. Current implementation is unstable and you need to run it multiple times with different random seeds to get a decent performance.
• Validation with model picking. As mentioned here, you can disable the validation step to improve the poisoning attack performance:

# Disable the validation step and the model picking
trainer.fit(attacker.data(), mask=splits.train_nodes)