How to win the SIGIR eCommerce Challenge with Transformers (Gabriel Moreira / NVIDIA)
The NVIDIA team performed well on the Session-based recommendation task of the SIGIR eCommerce Coveo Data Challenge 2021. NVIDIA achieved 1st place on the ‘Subsequent Items Prediction Leaderboard’ and 2nd on the ‘Next Item Prediction’ Leaderboard. These successes follow the previous successes of NVidia at the ACM RecSys Challenge and other recommender-system challenges.
Gabriel Moreira from NVIDIA explains in a blog post, how the team (Gabriel de Souza P. Moreira, Sara Rabhi, Ronay Ak, Md Yasin Kabir, Even Oldridge) won the challenge. It’s quite an in-depth explanation and a good read for anyone being interested in recommender systems.
Our solution included data augmentation and feature engineering techniques and models consisting of an ensemble of two different Transformer architectures — Transformer-XL and XLNet — trained with autoregressive and autoencoding approaches inspired by Natural Language Processing (NLP). We leveraged the rich information provided by the dataset, and explored different ways to combine tabular data of user interactions events (e.g., clicks, add-to-card, remove-from-cart, purchases, search queries) with unstructured data (product description and images) in a multi-modal approach.
Our detailed solution is described in our paper and solution code is available on GitHub. In this post, we will briefly describe the competition and our solution and provide some takeaways that might be instrumental to help you build your own session-based recommender system.
https://medium.com/nvidia-merlin/winning-the-sigir-ecommerce-challenge-on-session-based-recommendation-with-transformers-v2-793f6fac2994
About The Author
Joeran Beel
I am the founder of Recommender-Systems.com and head of the Intelligent Systems Group (ISG) at the University of Siegen, Germany https://isg.beel.org. We conduct research in recommender-systems (RecSys), personalization and information retrieval (IR) as well as on automated machine learning (AutoML), meta-learning and algorithm selection. Domains we are particularly interested in include smart places, eHealth, manufacturing (industry 4.0), mobility, visual computing, and digital libraries. We founded or maintain, among others, LensKit-Auto, Darwin & Goliath, Mr. DLib, and Docear, each with thousand of users; we contributed to TensorFlow, JabRef and others; and we developed the first prototypes of automated recommender systems (AutoSurprise and Auto-CaseRec) and Federated Meta Learning (FMLearn Server and Client).