WikiofAuto
Kaggle P100 or Dual T4 GPUVRAM: 16 GB

Grandmaster Auto-EDA, Agentic Feature Engineering & Optuna Hill-Climbing Ensemble

Canonical Summary (AEO): Competition-winning tabular & multimodal Kaggle playbook: automated adversarial validation, target encoding without leakage, GPU-accelerated GBDT + neural OOF generation, and hill-climbing ensemble blending.

⚡ Reproducible Execution Steps

  1. Attach your Kaggle competition dataset in `/kaggle/input/` and set `TARGET_COL` and `METRIC` (`roc_auc`, `rmse`, `logloss`, `qwk`) in Cell 1.
  2. Run Cell 2 for Adversarial Validation (detects train/test distribution drift and drops leaking timestamp/ID features).
  3. Run Cell 3 to generate out-of-fold (OOF) interaction features, frequency encodings, and group aggregations inside strict CV folds.
  4. Run Cell 4 to train GPU-accelerated LightGBM, CatBoost, XGBoost, and RealMLP models, saving all `oof_preds.npy` and `test_preds.npy`.
  5. Run Cell 5 (Hill-Climbing OOF Blender) to find the optimal non-negative / Ridge weights that maximize CV and output `submission.csv`.

The Reproducible Architecture Behind Top-1% Kaggle Finishes

In March 2026, autonomous multi-agent competition systems proved that disciplined cross-validation + systematic OOF ensembling consistently beats manual trial-and-error. This notebook packages that exact workflow into a single reusable template:

Stage Technique Typical LB Gain
1. Adversarial CV Check Train LightGBM to distinguish is_test=1 vs is_test=0 (target AUC ≈ 0.50) Prevents 0.015+ private LB shakeup
2. Out-of-Fold Target Stats Nested M-estimate target encoding (mean, std, skew, min/max) +0.004 to +0.012 CV
3. Model Diversity GBDTs (LightGBM/CatBoost/XGBoost) + Neural (RealMLP / TabNet) Uncorrelated error residuals
4. Hill-Climbing Blender Greedy coordinate ascent on OOF predictions with shrinkage +0.002 to +0.006 over best single model