February 27, 2017
Slides
名詞
- Out of fold prediction
- Lv1, Lv2, Lv3, …
Stacked Generalization
資料的切割
- 事先決定K-Fold 與 Subsets
- 各自進行Lv1 Model Training
- Step 1, 2, and 3
- 收集大家的Predictions
- Prediction of Model 1.1 on Fold 2
- Prediction of Model 1.2 on Fold 1
- Prediction on Test Dataset
- 進行Stacked Generalization
- New Submission
How to do Local CV on Stacked Generalization?
- Fold 1, Fold 2, Fold 3
- Local CV
- Lv 1 Model:
- Fold 1 + Fold 2 ==> Fold 3 and evaluate
- Fold 1 + Fold 3 ==> Fold 2 and evaluate
- Fold 2 + Fold 3 ==> Fold 1 and evaluate
- Lv 2 Model:
- Fold 1 + Fold 2 ==> Fold 3 with Stacked Generalization and evaluate
- Fold 1 + Fold 3 ==> Fold 2 with Stacked Generalization and evaluate
- Fold 2 + Fold 3 ==> Fold 1 with Stacked Generalization and evaluate
- Lv 1 Model:
Summary
- 事先決定Folds, 至少 3 以上
- 如果要玩高Lv, 考慮Local CV的執行，Fold 數量 = Lv + 1 (?)
- 所有人用相同的Fold 切割法
- 各自帶開、各自建模、各自上傳、各自觀察Local CV v.s. Public Leader Board
- 可能需要重新分配CV
- 組隊
- Stacked Generalization…
Criteo Competition 的心得
- 想清楚自己的優勢
- 所有上面的知識所有人都會
- 計算能量
- 演算法實做
- 演算法實做
- Dropout + FTPRL
- 那時候我們完全沒有用上面的Kaggle技巧
How To Win?
- 學會上述的技巧
- 擁有比別人多的Domain Knowledge
Homework
- Read Paper、Report Paper
- Awesome Deep Learning Papers
- 我們以實做的觀點來讀paper
- Pick Competition
Next Homework
- Read Paper、Report Paper
- Survey the Winner of Similar Competition
- Lv 1 Modeling