This repository contains the code of the winning solution of the WSDM2021 Booking.com Challenge - achieving 59.39 Precision@4. Our solution is a blend of 3 different neural network architectures, each trained multiple times with different seeds.
The code for XLNet-SMF will be added soon.
- Download the data from https://www.bookingchallenge.com/
- Pandas==1.2.0
- TensorFlow>=2.x
- PyTorch
- cuDF
This nvtabular:0.3 docker container has most libaries pre-installed. Pandas requires an update to 1.2.0 version.
The code is written in multiple jupyter notebooks.
Files:
- 00_Data is the folder, containing the original train and test csv files. Preprocessed data will be stored there, as well.
- 01_Preprocess/Preprocess.ipynb splits data in 5-folds for cross-validation and assigns the fold to each example
- 02_Models/GRU_SM_SMF/GRU-SM-SMF.ipynb trains the GRU-SM-SMF model (see some details below)
- 02_Models/GRU_SM_SMF/GRU-SM-SMF-Ensemble.ipynb blends (average) the trained model with GRU-SM-SMF.ipynb to intermediate prediction
- 02_Models/MLP_SMF/MLP-SMF.ipynb trains the MLP-SMF model (see some details below)
- 02_Models/MLP_SMF/MLP-SMF-Ensemble.ipynb blends (average) the trained model with MLP-SMF.ipynb to intermediate prediction
- 03_Ensemble/Ensemble.ipynb ensembles MLP-SMF, GRU-SM-SMF and XLNet-SMF
MLP-MF is trained 8 times with different seed.
GRU-MF is trained 7 times with different seeds and a small variation on input features.
| Ver | Remove Features |
|---|---|
| 31 | 'gap_','isweekend_', 'season_' |
| 32 | 'gap_' |
| 33 | 'isweekend_', 'season_' |
| 36 | None |
| 40 | 'isweekend_', 'season_' |
| 41 | 'isweekend_', 'season_' |
| 44 | None |
