the detail is about that when i create a model which can predicts the stability of landslide dams with missing values.I know that XGB or lgb model is good at dealing with missing data ,but i dont know if the missing values are more in the training data or in the test data may bring uncertainty to my result. I've tried lots of times , the result deffers from the random state I set when I do the train_test split. Some time the AUC of test can be 90%, some time it is 69% even.
i tried make a progress to the lgb model .I replace the loss funtion of lgb by Focal loss ,and it doesn't make sense ,I have no idea what i can do next