Winning Amazon KDD Cup'23

Chris Deotte, Kazuki onodera, Jean-Francois Puget, Benedikt Schifferer, Gilberto Titericz

Keywords: recsys, gradient boosting, CNN, Transformers

TL;DR: The paper presnts the winning solutions to all tasks of the Amazon KDDCup'23 competition form NVIDIA-Merlin team.

Abstract:

This paper describes the winning solutions of all tasks in Amazon KDD Cup ’23 from the NVIDIA MERLIN team. The challenge was to build a multilingual recommendation system. From each user, we are given a history of item interactions and we need to predict the next item
interaction. Our solution for tasks 1 and 2 is a pipeline of candidate generation, reranking, and ensemble. For candidate generation we leveraged statistical models, representation learning with embedding loss, pre-trained language models, multi-task learning with transformers, and more. Candidate sets were merged and ranked using gradient boosting (XGBoost and CatBoost) to maximize MRR score. Task 3 solution is based on multiple classifiers to maximize BLEU score.

Submission Number: 16