Model instability

[wpy678]

Hello, thanks for your work! Now I have some questions. I have 1000 training samples, each sample consisting of 100 different characteristic vector, my goal is to classify graph, so I use two layers of GAT, two full connection layer, but I find that each run to end the results of the test set is different, and quite big difference (that is, the model can't get stable results), can you provide your idea?

I hope you can give me some advice. Thank you very much.