# 24th Place Solution

First, congratulations to the competition winners, and sincere thanks to our team members
[@gentlezdh](https://www.kaggle.com/gentlezdh)
,
[@yoshifumimiya](https://www.kaggle.com/yoshifumimiya)
,
[@lizhecheng](https://www.kaggle.com/lizhecheng)
,
[@nandodmelo](https://www.kaggle.com/nandodmelo)
,
[@kekshibata](https://www.kaggle.com/kekshibata)
, as well as the competition organizers.
We present a short summary of our solution (
24th place, private LB 0.694
). More detailed version is
[here](https://www.kaggle.com/competitions/equity-post-HCT-survival-predictions/discussion/566809)
.
Overview
We trained regressors and classifiers separately and combined their outputs using ensembling & post-processing. And the “Trust CV” strategy led us to a final LB score of 0.694 to win silver.
![](assets/ad337f0c7419dd246a3bc300.jpg)
Model Training
Regressors:
GBDTs (CAT, LGB, XGB) and AutoGluon on 8 targets with basic feature engineering.
For specific 4 targets, applied extensive feature engineering with CatBoost and AutoGluon
PRL NNs, including customized losses
Classifiers:
GBDT (CAT, LGB, XGB) classifiers
Stacked logistic regression and combined Autogluon classifier.
Ensemble & Post-Processing
Combined regressors’ outputs using a
multi-task Elastic Net
on 8 targets. (excluding some custom PRL NNs as inputs).
Post-processing: Adjusted predictions based on classifier probabilities with race-specific parameters optimized by Optuna.
Blended in custom PRL NNs for high efs probability data and weaker races.
Submission Strategy
Best CV:  CV 0.6908 / public LB 0.690 / private LB 0.694
Best LB:  CV 0.6853 / public LB 0.694 / private LB 0.692
Selected the best CV and LB for final submissions. “Trust CV” approach ultimately secured us the silver medal.