Upgrade to Pro
— share decks privately, control downloads, hide ads and more …
Speaker Deck
Features
Speaker Deck
PRO
Sign in
Sign up for free
Search
Search
model_pipeline_final.pdf
Search
Maxwell
September 18, 2018
Science
1
210
model_pipeline_final.pdf
model pipeline and others in Home Credit Default Risk competition.
Thanks to team mates.
Maxwell
September 18, 2018
Tweet
Share
More Decks by Maxwell
See All by Maxwell
Causal Impact -paper summary-
hoxomaxwell
3
860
Great Barrier Reef Model Pipeline: 15th place
hoxomaxwell
1
220
Lecture materials at the University of Tokyo School of Medicine
hoxomaxwell
1
140
Kaggle Hungry Geese
hoxomaxwell
1
110
HuBMAP 17th place model pipeline
hoxomaxwell
1
99
LT: Shallow Dive into Bayes Factor
hoxomaxwell
6
1.4k
Kaggle APTOS 2019 @ U-Tokyo Med
hoxomaxwell
1
420
Cornell Birdcall 36th place solution
hoxomaxwell
2
240
Kaggle Bengali.AI 6 th place solution
hoxomaxwell
4
8.6k
Other Decks in Science
See All in Science
AIによる科学の加速: 各領域での革新と共創の未来
masayamoriofficial
0
150
データベース14: B+木 & ハッシュ索引
trycycle
PRO
0
470
動的トリートメント・レジームを推定するDynTxRegimeパッケージ
saltcooky12
0
200
地質研究者が苦労しながら運用する情報公開システムの実例
naito2000
0
280
深層学習を用いた根菜類の個数カウントによる収量推定法の開発
kentaitakura
0
180
機械学習 - ニューラルネットワーク入門
trycycle
PRO
0
860
データベース02: データベースの概念
trycycle
PRO
2
910
Transport information Geometry: Current and Future II
lwc2017
0
210
データマイニング - グラフ構造の諸指標
trycycle
PRO
0
190
生成AIと学ぶPythonデータ分析再入門-Pythonによるクラスタリング・可視化をサクサク実施-
datascientistsociety
PRO
4
1.8k
データベース09: 実体関連モデル上の一貫性制約
trycycle
PRO
0
990
データベース06: SQL (3/3) 副問い合わせ
trycycle
PRO
1
640
Featured
See All Featured
Faster Mobile Websites
deanohume
310
31k
It's Worth the Effort
3n
187
28k
How To Stay Up To Date on Web Technology
chriscoyier
791
250k
The Web Performance Landscape in 2024 [PerfNow 2024]
tammyeverts
9
850
JavaScript: Past, Present, and Future - NDC Porto 2020
reverentgeek
52
5.6k
Designing Experiences People Love
moore
142
24k
10 Git Anti Patterns You Should be Aware of
lemiorhan
PRO
657
61k
Responsive Adventures: Dirty Tricks From The Dark Corners of Front-End
smashingmag
252
21k
Measuring & Analyzing Core Web Vitals
bluesmoon
9
610
ピンチをチャンスに:未来をつくるプロダクトロードマップ #pmconf2020
aki_iinuma
127
53k
Large-scale JavaScript Application Architecture
addyosmani
514
110k
Why You Should Never Use an ORM
jnunemaker
PRO
59
9.6k
Transcript
ikiri_DS Model PipeLine 600+1 ( LB804 ) FEATURES 1000+1 (
LB803 ) meta app meta bur Kernel GP Nejumi features Tereka features + LGBM 5 3 tosh 5 + CatBoost 5 2 1 + LGBM * 4 3 1 + CNN 7 Residual 2 + ExtTree 4 3 1 Residual 1 ( corrected with residual regression ) Blending CV 0.8094 Adversarial Stochastic Blending CV 0.8096 Adversarial Stochastic Blending CV 0.81050 * model drawn in next page + NN 1 3 ONODERA Maxwell Nejumi Tereka RK 1 2 3 4 5 6 7 Branden features 8 Branden + NN 1 3 takuoko features 9 Angus features 10 takuoko nejumi feature Angus + Res2 + LGBM 1 6 + Res1 + LGBM 1 6 1 or 2 or 5 + LGBM 1 or 2 or 5 + CatBoost or + LGBM 5 1 or 2 5 + LGBM 8 + LGBM 9 + LGBM 10 Adversarial Stochastic Blending CV : 0.8061 29.Aug.2018 Tam Tam features 11 + LGBM 11 + RGF 1 + LGBM 11 + RNN 7 1 * using hidden layer as additional features to correct residuals. + CNN 7 + hidden + Res3 + LGBM 1 6 + RGF 1 + Res2 + LGBM 1 6 + LGBM 5 RK features 12 + LGBM 12 1 or 2 12 + LGBM 8 1 or 2 8 + LGBM 3 1 5 or 3 2 5 + LGBM 8 1 12 or 8 2 12 Public 0.8085 17 th Private 0.8017 18 th + LGBM 8 + LGBM 9 + LGBM 10 Ireko DAE 13 Ireko8 + NN 1 13 + NN 1 + NN 1 13 Nejumi prediction Public 0.8093 10 th Private 0.8016 18 th Public 0.8080 23 th Private 0.8028 14 th + RNN 7 1 Public 0.8110 3 rd Private 0.8042 5 th Giba Post Processing Public 2nd 0.81241 Private 2nd 0.80561 Home Credit Default Risk partial partial partial + LGBM 8 1 or 2 8 or 12 + LGBM 3 1 or 2 3 or 12 3 + LGBM 6 1 Residual 3 + hidden + LGBM 1 6' or 6' 1 + LGBM 6' 2 Blending
ikiri_DS Model PipeLine 600+1 ( LB804 ) FEATURES 1000+1 (
LB803 ) meta app meta bur Kernel GP Nejumi features Tereka features tosh + LGBM * 4 3 1 + CNN 7 Residual 2 Residual 1 ( corrected with residual regression ) Blending CV 0.8085 Adversarial Stochastic Blending CV 0.8085 Adversarial Stochastic Blending CV 0.8097 * model drawn in next page ONODERA Maxwell Nejumi Tereka RK 1 2 3 4 5 6 7 Branden features 8 Branden + NN 1 3 takuoko features 9 Angus features 10 takuoko nejumi feature Angus + Res2 + LGBM 1 6 + Res1 + LGBM 1 6 + LGBM 8 + LGBM 9 + LGBM 10 Adversarial Stochastic Blending CV : 0.8061 29.Aug.2018 Tam Tam features 11 + LGBM 11 + LGBM 11 + RNN 7 1 * using hidden layer as additional features to correct residuals. + CNN 7 + hidden + Res3 + LGBM 1 6 + RGF 1 + Res2 + LGBM 1 6 + LGBM 5 RK features 12 + LGBM 12 1 or 2 12 + LGBM 8 1 or 2 8 Public 0.8071 26 th Private 0.8009 37 th + LGBM 8 + LGBM 9 + LGBM 10 Ireko DAE 13 Ireko8 + NN 1 13 + NN 1 + NN 1 13 Nejumi prediction Public 0.8082 23 th Private 0.8022 18 th Public 0.8080 23 th Private 0.8028 14 th Public 0.8099 7 th Private 0.8040 6 th Giba Post Processing Home Credit Default Risk partial + LGBM 8 1 12 or 8 2 12 partial 1 or 2 + LGBM + LGBM 6 1 Residual 3 + hidden + LGBM 1 6' or 6' 1 + LGBM 6' 2 Blending + ExtTree 4 3 1 + NN 1 3 + RGF 1 + LGBM 4 3 2 + XGB 4 3 1 + NN 1 + RNN 7 1 + hidden + Res3 + LGBM 1 6 + Res1 + LGBM 1 6 + hidden + Res4 + LGBM 1 6 stacking with LGBM CV 0.8080 Public 0.8070 / Private 0.8015 Stacking prediction Stacking + LGBM 3 1 or 2 3
application bureau bureau balance AUC : 0.683 (SEED71) 0.683 (SEEDs
avg) AUC 0.772 (SEED71) 0.773 (SEEDs avg) XGBoost app meta feature XGBoost prev meta feature 229 features 300 features all data stacking-like Light GBM 5 stratified fold ( shuffle = True ) 5 / 8 SEEDs rank averaged SEED : 71 for model fit SEED : 710, 711, 712, 713, 714 ( 715, 716, 717 ) for OOF prediction hyper parameter tuned for 603 features (reflected on meta features) XGBoost bureau meta feature ONODERA BASIC FEATURES 600 features NEJUMI FEATURES ( interest rate ) 1 feature 603 ( 604 ) features Local CV 0.80641 Public LB / Private LB 0.80569 / 0.79853 100 th / 105 th AUC 0.710 (SEED71) 0.712 (SEEDs avg) previous inst POS_CASH credit 952 features Local CV 0.80646 LB 0.804 ( ~ 0.805 ) Maxwell 603 ( 604 ) selected features based on ONODERA criteria w/o feature selection Stacking-like Light GBM