March machine learning mania 2016
- 1. March Machine Learning Mania 2016 弌仍从舒 仆舒 Kaggle 弌仄亠亟仂于 仆仂仆 仗亠仍 2016
- 3. Timeline
- 5. 弌仗亠亳亳从舒 亠亟从舒亰于舒亠仄 弍亟亠亠 弌仂亠于仆仂于舒仆亳亠 仗仂仂亟亳 亢亠 亳 亞仂亟舒 仗仂亟磲 仆舒亳亠仍仆仂亠 于仍亳礌亳亠 仍舒亶仆仂亳 仂亢仆仂 亳仗仂仍亰仂于舒 仍ﾐ英亠 于仆亠仆亳亠 亟舒仆仆亠 舒仍仂 亳仆仂仄舒亳亳 仂 仆亳仆 仄舒舒 (~2000 仂从) 亠仂弍仍舒亟舒仆亳亠 feature engineringa
- 6. 舒仆仆亠 仂亳亠从亳亠 亟舒仆仆亠 1985 亞仂亟舒: 从仂, 从仂亞亟舒, 亞亟亠 亳 从亠仄 亞舒仍 仂亟仂弍仆亠 亟舒仆仆亠 2003 亞仂亟舒, 亟仂弍舒于仍磳 舒亳亳从舒 仄舒亠亶: 弍仂从亳, 仗仂亟弍仂, ~145K 仂从 亟仍 亠亞仍仆仂亞仂 亠亰仂仆舒 ~2K 亟仍 仆亳仂于 ~71K 仂从 亟仍 亠亞仍仆仂亞仂 亠亰仂仆舒 ~850 亟仍 仆亳仂于
- 7. Elo rating system 舒亢亟仂亶 从仂仄舒仆亟亠 仗亳于舒亳于舒亠 仆舒舒仍仆亶 亠亶亳仆亞, 仆舒仗亳仄亠, 1500. 仍 从舒亢亟仂亶 从仂仄舒仆亟 亳舒亠 仄舒仂亢亳亟舒仆亳亠 于亳亞舒仆仆 仂从仂于 (1 仗仂弍亠亟舒, 0 仗仂舒亢亠仆亳亠): 弍仆仂于仍磳 亠亶亳仆亞: K K-factor, 亠仄 仄亠仆亠 K, 亠仄 弍仂仍亠亠 从仂仆亠于舒亳于仆舒 亳亠仄舒. 仂亢仆仂 于于亠亳 仗仂仗舒于从亳 仆舒 亳亞 亟仂仄舒, 仆舒 舒亰仆仂 于 亠. Wiki
- 8. 亞亳亠 亠亶亳仆亞仂于亠 亳亠仄 Glicko 仍亠仆仆舒 于亠亳 Elo. Chessmetrics 仗仂仗仂亠, 仆仂 弍仂仍亠亠 于于亳亠仍仆舒 从 束于仂仂亟亳仄 亰于亰亟舒仄損. TrueSkill 亠亶亳仆亞仂于舒 亳亠仄舒 仂 Microsoft.
- 9. 仂亳 于亠 亟舒仆舒 亳仂亳 仄舒亠亶 仄亠亢亟 亟于仄 从仂仄舒仆亟舒仄亳. 于亠亟仄 亟仍 从舒亢亟仂亶 亳亰 从仂仄舒仆亟 束于亠損, 仂仗亠亟亠仍磳仄亶 亟舒于仆仂 亠 仗仂弍亠亟: 1 = 0,5 + ゐ_1 粒 2 = 0,5 + ゐ_2 粒 亞亟亠 ゐ_1 从仂仍亳亠于仂 仗仂弍亠亟 仗亠于仂亶 从仂仄舒仆亟 仆舒亟 于仂仂亶 n 仍亠 仆舒亰舒亟, 粒 从仂亳亳亠仆 亰舒舒仆亳. 丐仂亞亟舒 仄仂亢仆仂 亟亠仍舒 仗亠亟从舒亰舒仆亳亠: p1 = 1 1+2 p2 = 2 1+2 舒仗亳仄亠, 亠仍亳 从仂仄舒仆亟 亳亞舒仍亳 仂亟亳仆 舒亰 于 仂仄 亞仂亟 亳 弍仂仍亠 于亠 仆亳从仂亞亟舒 仆亠 弍仍仂, 于亠仂仆仂 仗仂于仂仆仂亶 仗仂弍亠亟 仗仂弍亠亟亳亠仍 仂亠仆亳于舒亠 于 0.75.
- 10. 舒从 仂仄亳仂于舒 亟舒舒亠 弌亠 亟舒仆仆亠 仆亠仍亰 仗仂仂 舒从 亟舒 舒仍亞仂亳仄
- 11. 舒从 仂仄亳仂于舒 亟舒舒亠 w_team l_team w_team features l_team features target Train:
- 12. 舒从 仂仄亳仂于舒 亟舒舒亠 w_team l_team w_team features l_team features l_team w_team l_team features w_team features 1 0 Train:
- 13. 舒从 仂仄亳仂于舒 亟舒舒亠 w_team features l_team features delta features l_team features w_team features - delta features 1 0 p1+p2 仆亠 于亠亞亟舒 舒于仆仂 1. 舒仗亳仄亠, 亟仍 xgboost舒. 丐仂亞亟舒 仄仂亢仆仂 仗亠亠亳舒 仗仂 仂仄仍舒仄: Train: team_1 team_2 delta features team_2 team_1 - delta features p1 p2 Test: p1 = 1 1+2 p2 = 2 1+2
- 14. 亟亠 亠亞亠亳亳 1- 从仂仄舒仆亟舒 仗仂弍亠亟亳仍舒 2- 舒亰仆亳亠亶 , 仂亞亟舒 亠仍亠于亠 仗亠亠仄亠仆仆亠 弍亟 舒于仆 + 亳 - 仂仂于亠于亠仆仆仂. 亳弍仂 仄仂亢仆仂 亳仗仂仍亰仂于舒 1+0,03* 亳 0-0,03*. 亠 亠磳 亳仆仂仄舒亳 仂 仂仄, 仆舒从仂仍从仂 仂亟仆舒 从仂仄舒仆亟舒 仂从舒亰舒仍舒 亳仍仆亠亠 亟亞仂亶.
- 15. 亳亰仆舒从亳 History results 弌舒亳亳从舒 弍仂从仂于, 仗仂亟弍仂仂于 亳 .仗. 亰舒 仗仂仍亠亟仆亳亶 亞仂亟 (从仂仍亰亠亠 亠亟仆亠亠) 仂弍亠亟, 仗仂舒亢亠仆亳 亳 winrate 亰舒 仗仂仍亠亟仆亳亶 亞仂亟 弌亠亟仆 舒亰仆仂 仂从仂于 仗仂仍亠 仄舒舒 亰舒 仗仂仍亠亟仆亳亶 亞仂亟 弌从仂仍从仂 亟仆亠亶 仆舒亰舒亟 弍仍舒 仗亠亟亟舒 亳亞舒 弌从仂仍从仂 亠亰仂仆仂于 从仂仄舒仆亟舒 舒于亠 于 仆亳亠, 亟仂仄舒仆亳亶 仍亳 仄舒 亳 .亟. StatsTeams achievements 舒仗亳仄亠, 仗仂仄仄亳亠仄 亳仍仂 亳亞 于 仆亳舒 亰舒 仗仂仍亠亟仆亳亠 N 仍亠
- 16. 仂仗仂仍仆亳亠仍仆亠 亟舒仆仆亠 Massey ordinals Kenpom data Teams coaches 仂仗. 亟舒仆仆亠 亟仍 于亠 从仂仄舒仆亟 2002 亞仂亟舒 亠亶亳仆亞亳 从仂仄舒仆亟 舒亰仆 亳亠仄(132 仆亳从舒仍仆) 2003 亞仂亟舒 仆仂仄舒亳 仂 亠仆亠舒 亟仍 从舒亢亟仂亶 从仂仄舒仆亟 Tourney seeds 舒仆仆亠 仂 仗仂亠于仆 仆仂仄亠舒 从仂仄舒仆亟 Geography data 亟亠 仗仂仂亟亳仍亳 仄舒亳
- 17. 亟亠 舒亰仆 亟舒舒亠仂于 Massey ordinals Kenpom data Teams coaches Geography data Teams Achievements Coaches Achievements Stats 1 Tourney seeds Elo Glicko HistoryStats 2 Dataset 1 2003+ Dataset 2 1985+ Tourney Dataset 1 2003+ Tourney Dataset 2 1985+ 2003+ 1985+
- 18. Massey ordinals Kenpom data Teams coaches Tourney seeds Geography data Teams Achievements Coaches Achievements Stats Elo Glicko Dataset 1 2003+ Dataset 2 1985+ Tourney Dataset 1 2003+ Tourney Dataset 2 1985+ History XGB Level 1 XGB Regression Elo Glicko History + XGB Level 2 Logistic Regression Level 0 Level 1 Level 2 Level 3 Final Elo predict Glicko predict History predict + +? Net Prophets Entry Prediction x4 blending
- 19. 舒亳弍仂仍亠亠 于舒亢仆亠 仗亳亰仆舒从亳 亠仂亞舒亳亠从亳亠 亟舒仆仆亠 Elo (score) 亠亟从舒亰舒仆亳亠 仗仂 亳仂亳亳 亳亞 亠亶亳仆亞亳 仆亠从仂仂 亳亠仄
- 23. 亟亠亳 仆舒 弍亟亠亠 1. 仂弍舒于亳 仆仂于 亳仆仂仄舒亳 (亟舒仆仆亠 仂 舒于从舒, 亳亞仂从舒) 2. 仗仂仍亰仂于舒 舒仍亞仂亳仄: NeuralNets, KNN 3. 仂亢仆仂 仂仗亳仄亳亰亳仂于舒 仆亠 logloss, 舒 仄舒仂亢亳亟舒仆亳亠 于亳亞舒 于 亟亠仆亞舒 亳仍亳 仄亠仂 仆舒 leaderboard 仂舒仆舒仍亳亰亳仂于舒 仗亠亟从舒亰舒仆亳 亟亞亳 舒仆亳从仂于 4. 亳亟仄舒 仄亠仂亟 亳仄仍亳亳 仆亳舒 仂仄仂亢亠 仗仂仍亳 弍仂仍亠 亟舒仆仆 亟亠 仗仂仍亠亰亠仆 亟仍 舒仆舒仍亳亰舒 仗亠亟从舒亰舒仆亳亶 亟亞亳 舒仆亳从仂于