輪講_Kaggleで勝つデータ分析の技術_第2章

輪講用のまとめ

Transcript

1. ୈ2ষɹλεΫͱධՁࢦඪ 2020/04/15 Yu Sato @Yust ୈ4ճྠߨ
2. ࣗݾ঺հ ▪ 2019೥2݄͜Ζ͔ΒkaggleʹࢀՃɻ ▪ ޷͖ɿ΋΋Ϋϩɺ kaggleɺα΢φɺ͏ͳ͗ ▪ ؾʹͳΔɿ͋ͭ৿
3. KaggleͰউͭσʔλ෼ੳͷٕज़_໨࣍ 3 ୈ1ষɹ෼ੳίϯϖͱ͸? ୈ2ষɹλεΫͱධՁࢦඪɹˡࠓ೔͸ίί ୈ3ষɹಛ௃ྔͷ࡞੒ ୈ4ষɹϞσϧͷ࡞੒ ୈ5ষɹϞσϧͷධՁ ୈ6ষɹϞσϧͷνϡʔχϯά ୈ7ষɹΞϯαϯϒϧ
4. ϞσϧͷධՁͱ͸ʁ 4 test model ϞσϧͷධՁ͸2ͭ͋ΓɺͦΕͧΕͷࢦඪΛʮ໨తؔ਺ʯʮධՁࢦඪʯͱ͍͏ɻ fit predict evaluate train Leaderboard f(x)Λ ܭࢉ͢Δ f(x):໨తؔ਺ ύϥϝʔλ update f(x)͕࠷খͱ ͳΔ·Ͱloop f(x)Λ ܭࢉ͢Δ f(x):ධՁࢦඪ … ໨తؔ਺͸Ϟσϧʹґଘ ධՁࢦඪ͸ίϯϖʹґଘ p.87
5. ۩ମతͳؔ਺ 5 λεΫͷछྨʹΑͬͯ࢖༻͞ΕΔؔ਺͸ҟͳΔɻ ճؼ ೋ஋ ෼ྨ ଞΫ ϥε ෼ྨ ɾRMSE ɾlogloss ɾmulti-class logloss ճؼ ೋ஋ ෼ྨ ଞΫ ϥε ෼ྨ ɾRMSE ɾRMSLE ɾMAE ɾR2(ܾఆ܎਺) ɾaccuracy ɾerror rate ɾF1-score, Fβ-score ɾlogloss ɾAUC ɾmulti-class accuracy ɾmean-F1, macro-F1, micro-F1 ɾquadratic weighted kappa(QWK) ɾmulti-class logloss ໨తؔ਺ ධՁࢦඪ p.62~83
6. ໨తؔ਺ͱධՁࢦඪͷؔ܎ 6 ɾRMSE ɾlogless ɾmulti-class logloss 1.ධՁࢦඪͱಉ ؔ͡਺Λ໨తؔ ਺ʹઃఆՄೳ 2.ධՁࢦඪΛ໨ తؔ਺ʹ਺ֶత ʹม׵Մೳ 3.ධՁࢦඪ͕0/ 1Λ༧ଌ͢Δ෼ ྨλεΫ ɾRMSE ɾlogless ɾmulti-class logloss ɾ(logม׵)RMSE ɾRMSE ɾlogless ɾmulti-class logloss ɾaccuracy ɾerror rate ɾF1-score, Fβ-score ɾFair ɾqwkΛ࿈ଓؔ਺ۙࣅ ɾRMSLE 4.ධՁࢦඪͷྨ ࣅؔ਺Λ໨తؔ ਺ʹઃఆՄೳ ɾMAE ɾQWK ໨తؔ਺ ٯlogม׵ ᮢ஋Ͱ෼ྨ 0.23 0.88 0.67 0.12 0 1 1 0 (0.60) ͦͷ·· ΄΅ͦͷ·· ධՁࢦඪ େ͖͘෼͚ͯҎԼͷ4ύλʔϯ͕͋Δɻ ৄ͘͠ղઆ ৄ͘͠ղઆ p.90
7. ᮢ஋͸Ͳ͏΍ͬͯಋग़͢Δʁ 7 trainͰ࠷దԽ͞Εͨᮢ஋ΛٻΊɺͦΕΛ༻͍ͯtestΛ෼ྨɻ threshold 0.00 ~ 0.80 → 0 0.81 ~ 1.80 → 1 1.81 ~ 2.50 → 2 2.51 ~ 3.00 → 3 train LGBM OptimizedRounder test LGBM OptimizedRounder prob_target target 0.45 1 1.12 2 2.90 2 2.04 0 target 0.45 1.12 2.90 2.04 ɾ࠷దͳᮢ஋ͷಋग़ ɾճؼ஋͔Β෼ྨ Λ࣮ࢪ͢Δɻ ίʔυ͸ΞϥΠ͞Μ͕·ͱΊͯ ͘Ε͍ͯΔ(*)ɻ (*)https://qiita.com/kaggle_master-arai-san/items/d59b2fb7142ec7e270a5 target 0 1 3 2 p.91,100,101
8. ྨࣅؔ਺ͱ͸ʁ 8 ͦΕͧΕͷؔ਺ͷಛ௃Λ௫Έɺۙࣅ͢Δؔ਺Λબ୒͢Δɻ ɾQWKͷۙࣅ ɾMAEͷۙࣅ p.101~103 p.103 ਤ2.23
9. ͋Γ͕ͱ͏͍͟͝·ͨ͠ɻ ▪ Kaggle: @Yust ▪ Twitter: @yust_kaggle ▪ e-mail: Yu.Sato@nttdata.com