0
$\begingroup$

I am running a binary LASSO logistic regression using glmnet. The initial data I work with is raster spatial data. When I create an ROC (AUC ~ 0.72) curve based on the test data, the resulting curve appears to curve early and has a very strange shape (shown below).

Does anyone know how I can interpret this curve and apply changes to my model to improve it?

A ROC curve flattening around 0.8 FPR

When I generate the ROC using the training data (AUC ~ 0.93), it does not appear like this.

Also, when I run the same script on coarser resolution data (30 m as opposed to the 5 m currently being used), AUC curves on training and testing data look as expected (AUC's of ~0.94 and 0.90).

$\endgroup$
1
  • $\begingroup$ It sounds like your LASSO model might be overfitting your training data. Your train/test split might also have introduced a strange bias just by chance. Try re-splitting the data. $\endgroup$
    – jdobres
    Jun 2, 2023 at 22:06

1 Answer 1

0
$\begingroup$

I was able to generate an ROC curve where the TPR curved at a higher value by adjusting the weights on my rarer outcome response from 5 to 10.

It seems my finer resolution data requires higher weights for optimum results

$\endgroup$
1
  • $\begingroup$ Your answer could be improved with additional supporting information. Please edit to add further details, such as citations or documentation, so that others can confirm that your answer is correct. You can find more information on how to write good answers in the help center. $\endgroup$
    – Community
    Jun 3, 2023 at 3:47

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service and acknowledge you have read our privacy policy.