Pessimistic Rescaling and Distribution Shift of Boosting Models for Impression-Aware Online Advertising Recommendation

In this paper, we provide an overview of the approach we used as team Gabibboost for the ACM RecSys Challenge 2023, organized by ShareChat and Moj. The challenge focused on predicting user activity in the online advertising setting based on impression data, in particular, predicting whether a user would install an advertised application using a high-dimensional anonymized feature vector. Our proposed solution is based on an ensemble model that combines the strengths of several machine learning sub-models, including CatBoost, LightGBM, HistGradientBoosting, and two hybrid models. Our proposal is able to harness the strengths of our models through a distribution shift postprocessing and fine-Tune the final prediction via a custom build pessimistic rescaling function. The final ensemble model allowed us to rank 1st on the academic leaderboard and 9th overall.