How to handle cofound variables? [D]

edit: confound

Hello all,

I am working on a object classification with a automotive radar point clouds. I compared many models and feature vectors.

Once i used range as feature, all models scored higher f1 in all K validation sets and on the final test set.

One particular artifact of a radar, is that as the farther the object is the less number of points it returns to the radar. Although the performance improved and there is no overfit in the classical sense, i am afraid my model is learning the environment not the class distribuiton and even worse, its learning that big range means big object.

How can i stress test this claim? Should i try to split the data sets so range distribution differs? Or not even using the feature at all and accept lower performance?

Would appreciate your insights.

Thank you.