Solution for Recruit Restaurant Visitor Forecasting Competition
This is a complete solution for the Recruit Restaurant Visitor Forecasting Competition that I'll present at Kaggle Days in Warsaw on May 19th, 2018. The ideas behind it are quite general and can be used for any time-series competition. Actually, the solution here is almost a one-to-one copy of my solution of the Masters Caesars competition which my team won. The submission scores 0.505 on private LB. This would have been good for the 3rd place.
To run the solution, download the git repository and run bash setup_directory.sh first. This generates all the folders that you will need. Download the train/test data from the competition homepage and put the data in input/. From Hunter McGushion's weather dataset download the file air_store_info_with_nearest_active_station.csv and put it in weather_data/. Also download 1-1-16_5-31-17_Weather.zip from there and put it in weather_data/stations.
You have all the data you need now. You also have to have python installed with the following packages: numpy, pandas, scipy, scikit-learn, lightgbm and keras.
Run preprocessing.py first. This is just basic preprocessing of the data. Then, generate all features by running feat_gen_standard.py, feat_gen_visitors.py, feat_gen_res_visitors.py, feat_gen_res_visitors_type2.py and feat_gen_weather.py (the order of the scripts does not matter).
Run the following scripts in the given order to obtain Lightgbm predictions for the validation set, for the test set and the submission file: LGB_01_CV.py, LGB_01_LB.py and LGB_01_SUB.py.
Run the following scripts in the given order to obtain Keras predictions for the validation set, for the test set and the submission file: KERAS_01_CV.py, KERAS_01_LB.py and KERAS_01_SUB.py.
Run weighted_average_SUB.py for a weighted average of the Lightgbm and Keras model.
Performance of the models:
| Model | Public LB | Private LB |
|---|---|---|
| Lightgbm | 0.468 | 0.509. |
| Keras | 0.474 | 0.513 |
| weighted average | 0.468 | 0.505 |
Have fun with it!