Airbnb Kaggle Competition: New User Bookings
This repository contains the code developed for the Airbnb's Kaggle
competition. It's written in Python, some in the form
of Jupyter Notebooks, and other in pure Python 3.
The code produces predictions with scores around 0.88090% in the public
leader-board, enough to be in the best 5% participants(0.001% behind the best)
and 0.88509% in the private leader-board(0.0018% behind the winner)
The entire run should not take more than 4 hours(thanks to the parallel
preprocessing) in a modern/recent computer, though you may run into memory
issues with less than 8GB RAM.
Feel free to contribute to the code or open an issue if you see something wrong.
Description
New users on Airbnb can book a place to stay in 34,000+ cities across 190+
countries. By accurately predicting where a new user will book their first
travel experience, Airbnb can share more personalized content with their
community, decrease the average time to first booking, and better forecast
demand.
In this competition, the goal is to predict in which country a new user
will make his or her first booking. There are 12 possible outcomes of the
destination country and the datasets consist of a list of users with their
demographics, web session records, and some summary statistics.
Data
Due to the Competition Rules, the data sets can not be shared. If
you want to take a look at the data, head over the competition
page and download it.
You need to download train_users_2.csv, test_users.csv and sessions.csv
files and unzip them into the 'data' folder.
Note: Since the train users file is the one re-uploaded by the competition
administrators, rename train_users_2.csv as train_users.csv.
Main Ideas
-
The provided datasets have lot of NaNs and some other random values, so, a
good preprocessing is the primary key to get a good solution:- Replace -unknown- values with NaNs
- Clean age values
- Extract day, weekday, month, year from
date_account_created
andtimestamp_first_active - Add number of missing values per user
- General user session information:
- Number of different values in
action,action_type,
action_detailanddevice_type
- Number of different values in
-
That kind of classification task works nicely with tree-based methods, I
usedxgboostlibrary and the Gradient Boosting Classifier that provides along
scikit-learnto make the probabilities predictions.
Requirements
To replicate the findings and execute the code in this repository you will need
basically the next Python packages:
Resources
- XGBoost Documentation - A library designed
and optimized for boosted (tree) algorithms. - Pattern Classification -
Tutorials, examples, collections, and everything else that falls into the
categories: pattern classification, machine learning, and data mining.
License
Copyright © 2015 David Gasquez
Licensed under the MIT license.