Version | Private LB | Public LB | CV | AUC | CV Strategy | Fold | Seed | Execution Time | Blend Weight |
---|---|---|---|---|---|---|---|---|---|
v128 | 0.01622 | 0.01840 | 0.01661682437083058 | 0.6732857787671046 | new | 5 | 4 | 5898s | [0.48228004 0.29394546 0.2237745 ] |
v131 | 0.01618 | 0.01836 | 0.01512241439772212 | 0.8415474216651967 | old | 7 | 2 | 5964s | average |
- Preprocessing
- Remove
ctrl_vehicle
- RankGauss
- PCA + Existing Features
- KMeans
- Basic stats
- Remove
- Model
- Multi head ResNet (tensorflow)
- TabNet (pytorch)
- NODE - Neural Oblivious Decision Ensembles (tensorflow)
- Training
- Pre-train with scored and non-scored target
- Train with public test pseudo label
- Add swap noise for train data
- Optimizer: Adam/AdamW with
weight_decay
- Loss: BCE with Label smoothing + Logits
- Prediction
- Ensemble above with weight optimization