Right-Censored Regression for IBD with CatBoost and XGBoost Stacking

Summary

A censor-aware stacking pipeline for an IBD target right-censored at \(c=20\). CatBoost is trained on uncensored cases using \(K=9\)-fold out-of-fold predictions, a robust residual scale is estimated, and Tobit-style tail corrections are applied before XGBoost meta-learning.

Context

The report models the observed label as \(Y_i=\min(Y_i^*,c)\) and keeps NaN-labelled rows as a scoring set rather than using them for fitting.

Main contributions

• Restricts CatBoost fitting to uncensored observations and generates (K=9)-fold out-of-fold predictions.
• Applies a Tobit-style tail-expectation correction to base and meta predictions.
• Fits an XGBoost meta-learner on the corrected out-of-fold signal and preserves a pure scoring set for NaN-labelled rows.