If AUC is under 0.50, is there model tuning aside from # trees in RF that can be used?

nliulab / AutoScore

AutoScore: An Interpretable Machine Learning-Based Automatic Clinical Score Generator

31 stars 4 forks source link

If AUC is under 0.50, is there model tuning aside from # trees in RF that can be used? #6

Open shlid007 opened 1 year ago

shlid007 commented 1 year ago

Running AutoScore on test data (fraud dataset), the AUC was under 0.50. Can model tuning address this? I didn't see suggestions for model tuning in the article (other than fine-tuning the cutoff points).

fengx13 commented 1 year ago

Hi, thanks for the question. It depends on how low the AUC is. If it's below 0.4 or 0.3, you can just swap your binary outcome to get AUC > 0.5. But if the AUC is around 0.4-0.5, it may means the predictive value is very low