AlvinWen428 / FeatureCP

The official codebase for Predictive Inference with Feature Conformal Prediction
33 stars 9 forks source link

Error in Finding the Best Step Number #4

Open Soothysay opened 5 months ago

Soothysay commented 5 months ago

In your code from lines 334-348 (https://github.com/AlvinWen428/FeatureCP/blob/71c1b1ffba742ba90729bcc97997973a6a3c0110/conformal/icp.py#L334) you seem to iterate over the different step size validation coverages and simply choose the first index>tolerance which is greater than (1 - val_significance) * 100. This can lead to trivial bands. Consider the case where your smallest step gives you 99% validation coverage. By the current logic, you will choose the index as your step size and will thus get a trivial band.

Am I missing something?