Booking.com Data Challenge 6th Place Solution

This repository contains 6th place solution codes for Booking.com Data Challenge, which is a challenge with a task of predicting travellers' next destination. We trained four types of Long short-term memory (LSTM) model, and archived the final score: 0.5399 by weighted averaging of these predictions. There are some differences in these models in feature engineering, multi-task learning, and data augmentation. Our experiments showed that the diversity of the models boosted the final result.

Our solution is described in the submitted paper, and our code is available at teammate's repository and here. In this repository, our baseline model and the models named LSTM 1-3 are avaliable.

Baseline
LSTM 1: BookingNN
LSTM 2: BookingNNMtl (LSTM 1 + Multi-task learning)
LSTM 3: BookingNNAug (LSTM 2 + Data augmentation)

model	fold0	fold1	fold2	fold3	fold4	average
Baseline	0.4773	0.4388	0.4420	0.4731	0.4837	0.4629
LSTM 1	0.5083	0.4673	0.4728	0.4999	0.5154	0.49274
LSTM 2	0.5086	0.4681	0.4747	0.5012	0.5162	0.49376
LSTM 3	0.5029	0.4575	0.4661	0.4967	0.5082	0.48628

BookingNN

python train_nn.py

LSTM with categorical and numerical features.
Pretrained weights of city_id and hotel_country are calculated by Word2Vec.
Some numerical features are created like days_stay, num_visit_same_city, and so on.
days_stay at the prediction point can be used and useful in this competition.

from src.datasets import load_train_test, BookingDataset, MyCollator
from src.models import BookingNN
from src.utils import seed_everything
from src.runner import CustomRunner

BookingNNMtl

python train_nn_mtl.py

Multi-task learning version BookingNN.
Predict not only city_id but also hotel_country.

from src.datasets import load_train_test
from src.datasets import BookingDatasetMtl as BookingDataset
from src.datasets import MyCollatorMtl as MyCollator
from src.models import BookingNNMtl as BookingNN
from src.utils import seed_everything
from src.runner import CustomRunnerMtl as CustomRunner

BookingNNAug

Each sequence of trips can be flipped. Be sure to remove booker_country in order to keep consistency.

python train_nn_aug.py

from src.datasets import load_train_test
from src.datasets import BookingDatasetAug as BookingDataset
from src.datasets import MyCollatorAug as MyCollator
from src.models import BookingNNAug as BookingNN
from src.utils import seed_everything
from src.runner import CustomRunnerAug as CustomRunner

Graph features

We believe that graph related features are important because it can lead to reconstruct geographical information. Each sequence of trips are just a fragment of it.

The following figure is a scatter plot of city_id vectors calculated by Word2Vec. The number of dimension is compressed by umap. Embedding vectors are used as weights in the model.

python fe_w2v.py

Training tips

Google Colab with GPU (runner.ipynb).
Catalyst for PyTorch model training.
Stratified split by the length of each trip.
Use model with best validation score from epoch 11 to 14.
You need to add PYTHONPATH as follows.

import sys
sys.path.append(YOUR_PROJECT_ROOT)

upura / booking-challenge-2021