This is the code for the 1st place ACM Recsys competition. Paper is uploaded to the repository (https://github.com/logicai-io/recsys2019/blob/master/Boosting%20algorithms%20for%20session-based%2C%20context-aware%20recommender%20system%20in%20online%20travel%20domain.pdf).

If you have any questions please write to pawel@[address of our company website]

• Paweł Jankiewicz (https://www.linkedin.com/in/pjankiewicz/)
• Liudmyla Kyrashchuk (https://www.linkedin.com/in/liudmylakyrashchuk/)
• Paweł Sienkowski (https://www.linkedin.com/in/pawel-sienkowski/)
• Magdalena Wójcik (https://www.linkedin.com/in/claygirl/)

Below there is a process to generate best single model.

1. cd data/
2. ./download_data.sh
3. cd ../src/recsys/data_prep
4. ./run_data_prep.sh
5. C++ feature generation

cd ../../../cpp
make
./build/price # extracts price features
./build/scores # extracts incremental features for each impression (|-separated format)
./build/extractor # extracts comparison features for |-separated features (extracted in scores)
python select_features.py

6. cd ../src/recsys/data_generator/
7. cd data_generator; python generate_data_parallel_all.py; cd .. (pypy is good)
8. python split_events_sorted_trans.py
9. python vectorize_datasets.py
10. python model_val.py (validate model)
11. python model_submit.py (make test predictions)
12. python make_blend.py (prepare submission file)

1. cd data/
2. ./download_data.sh
3. cd ../../src/recsys/data_prep
4. ./run_data_prep.sh
5. cd ..
6. cd data_generator; python generate_data_parallel_quick.py; cd - (pypy is good)
7. python quick_validate.py

To create the blend we merged 37 LGBMLight models with different datasets and different hyper parameters.

To combine the models we used make_blend.py script which optimizes directly MRR using validation predictions.