Industrial Solution in Fashion-domain Recommendation by an Efficient Pipeline using GNN and Lightgbm

In this paper, we present our 1st place solution for the ACM RecSys 2022 Challenge. The annual challenge focus on the nuances of fashion domain recommendation this year. Dressipi, the organizer, is the fashion-AI expert, providing products and outfit recommendations to leading global retailers. Putting forward the challenge of accurately predicting which fashion item will be bought at the end of the session, given a dataset of user sessions, purchase data, and content data about items. As the public dataset contains millions of pretty short sessions, both accuracy and scalability should be taken into account. Our solution is a fast prediction pipeline consisting of feature extraction, retrieval and rank. Specifically, the critical factor of our high score submission is the combination of collaborative filtering varieties and pre-trained embedding from GNN, which are carefully designed for the challenge dataset. Finally, we illustrate our experiments on different hyper-parameters and loss functions.