Kaggle Competition: Santander Product Recommendation

This repository reflects the source code of my submission to the Santander Product Recommendation Kaggle Competition, for which I ended up on 26th place, out of 1809 total participants.

I used R (in particular data.table) for data munging, and LightGBM for training a gradient boosting machine model. LightGBM turned out to be very fast and allowed for rapid iterations.

key ideas of this approach:

• use a single GBM model for binary classification (the product becomes yet another variable)
• use 2015-06 data for the two "seasonal" products ind_reca_fin_ult1 and ind_cco_fin_ult1
• use last three months for all other products
• add 5-month lags of product usage (p1,..,p5)
• add flag whether each product was used at all in p2 to p5
• add count how many times each product was activated in p1 to p5
• add some customer covariate lags
• sanitize the customer covariates a bit
• leverage the information from the leaderboard probes

many thanks to the Kaggle forum users and the Vienna Kaggle meetup for inspiring ideas!!