JSAI Cup2017報告会
?
4 likes?1,832 views
This document discusses using XGBoost for a machine learning competition called JSAI Cup 2017. It provides details on: - Using XGBoost to predict transportation demand using historical data from 500km of roads in Japan from 2012-2013. - Preprocessing the data, which included one-hot encoding of categorical features and splitting the data into 5 periods. - Training XGBoost models and evaluating their performance on the test data, achieving a log loss score of 0.1985 using default parameters and 0.1922 after additional hyperparameter tuning. - The top 20 predictions made by the best XGBoost model, with performance increasing compared to benchmark methods.
JSAI Cup2017報告会
- 2. 2015 4 n : n STAIR Captions: ACL2017 short paper n STAIR Actions ( ): ? NEDO n ? NLP CV
- 5. n : n 10 n : n 2012/01/01 – 2013/6/30 500km n n 5
- 6. 301 : 7 : 5 n (5 ): 2 -4 , 5 -10 , 11 -17 , 18 -22 , 23 -1 : 55 5 = 275 n 5 (55 ) one-hot : 2 n ID . : 2 n 6 : 2 n : 3 n : 4 n (CG) 3 : 1 n 1
- 7. ( n XGBoost (Gradient Tree Boosting) v0.60 https://github.com/dmlc/xgboost : 3 n XGBoost 1: 3 XGBoost n XGBoost 2: XGBoost 1 4 (5 4=20 ) n : XGBoost 1 XGBoost 2 + 0.3 0.3 0.4 t+1 t
- 8. J S A n XGBoost n XGBoost (log loss) = 0.1985 n XGBoost + (log loss) = 0.1922 n 1-fold 1 5-fold XGBoost : 10000 n
- 10. ? Log loss ? Log loss ( ) benchmark ( ) 1.38629 0.21121 + 0.20698 (+2.00%) + 0.20113 (+2.82%) + 0.20089 (+0.12%) + , 0.19972 (+0.58%) + 0.19949 (+0.11%) + 0.19958 (-0.04%) + 0.19841 (+0.59%) 1: XGBoost Log loss ( )
- 11. ? 6/20 ? 20 2: XGBoost 1 Top-20 0.00 0.01 0.02 0.03 0.04 0.05 0.06 0.07 0.08 0.09 0.10 (uchibou, ) ( ) ( ) (keiyou, ) (uchibou, ) (keihintohoku, ) (keiyou, ) (keihintohoku, ) ( ) ( ) (keihintohoku, ) (keiyou, ) ( ) ( ) (uchibou, ) ( ) ( ) (tyuou, ) (keihintohoku, ) ( )
- 12. 1. n 2. n ( ) 3. n 4. n 1 2 Top- 20 0
- 13. B C n XGBoost 3 n 9 n n XGBoost