Kaggle Paris Meetup – Otto challenge feedback

A quick presentation of the winner's solution of the most popular Kaggle challenge (yet): the Otto Group Product Classification Challenge.

Presented at Kaggle Paris Meetup @OCTO Technology.

Transcript

1. Otto challenge feedback june 23rd 2015 @OCTO Technology
2. 93 features 9 classes Challenge: product classification evaluation: multi-class log loss
3. Neural networks Gradient boosting Best models Lasagne Keras ... XGBoost
4. Calibration Bagging Best tricks Hyperparameter optimization Grid search Random search Bayesian optimization ... Stacking
5. Level 1 Winning solutions #1 Level 2 Level 3 33 models 8 engineered features XGBoost Neural network Adaboost 0.85 * [XGBoost^0.65 * NN^0.35] + 0.15 * [Adaboost] private leaderboard 0.38243
6. Level 1 Winning solutions #2 Level 2 Level 3 Raw data Term frequency inverse document frequency k-nearest neighbors Neural network XGBoost Neural network Random Forest XGBoost private leaderboard 0.38656
7. • 1st place solution: Gilberto Titericz & Stanislav Semenov • 2nd place solution: Alexander Guschin • Kaggle scripts: Don't Miss These Scripts • Christophe’s post: 6 Tricks I Learned From The OTTO Kaggle Challenge • MLWave’s post: Kaggle ensembling guide Resources