Airbnb New User Bookings Prediction

This project deals with predicting the country in which a new Airbnb user will make his/her first booking. The problem statement and the dataset are part of a recently conducted Kaggle competition. The dataset consists of information about more than two hundred thousand Airbnb users. The information about users include the demographics of the user,web session activity logs, age, gender, signup information etc. Each user instance in the training dataset is also providedwith the country where they made their first booking makingthis a supervised learning problem. The problem is very interesting as it involves identifying the right features to fit an appropriate learning classifier to make predictions on the probable destination country. We have conducted extensive preprocessing of the data and extracted many features that are useful for the problem. Various classifiers with different learning algorithms are trained on top of the extracted features. LightGBM classifier trained on the extracted features gives an NDCG score of 0.88524 which is very close compared to the best ever score achieved in the competition 0.88697. Through the experiments conducted in this work, we observed that Boosting techniques outperform other classification methods used in this work. To evaluate and understand the importance of the developed features, we use gradient boosting techniques to calculate the feature importance. The experiments indicate that the age of an user, date of account creation, date of first activity along with certain features from sessions file are the most important ones for this problem. We have contributed an end-to-end machine learning framework to solve a real-world problem efficiently and accurately.