Post by Timo Wilm

📢 The author version of our paper "Industry Insights from Comparing Deep Learning and GBDT Models for E-Commerce Learning-to-Rank" is now available! You can read it here: https://lnkd.in/eQS9QY77 😊 Accepted at ACM RecSys 2025 🎉 this paper presents OTTO's current state-of-the-art ranking systems for search and re-ranking tasks. It explores whether deep neural networks (DNNs) can outperform strong tree-based models like LambdaMART in large-scale e-commerce environments. We benchmarked multiple DNN architectures and loss functions against our production-grade LambdaMART model using a proprietary dataset from OTTO, and validated the results through an 8-week online A/B test. 📈 𝐊𝐞𝐲 𝐟𝐢𝐧𝐝𝐢𝐧𝐠𝐬: A simple DNN architecture outperformed the GBDT baseline in clicks and revenue, while matching it in units sold. Interestingly, we found that transformer-based models underperform compared to these simpler DNN architectures. A promising direction for applying deep learning to real-world ranking systems! 🚀 Huge thanks to everyone who contributed to OTTO's neural ranking operations, which are now set to serve millions of users daily. 🙌 𝐒𝐩𝐞𝐜𝐢𝐚𝐥 𝐭𝐡𝐚𝐧𝐤𝐬 𝐭𝐨, 𝐢𝐧 𝐧𝐨 𝐩𝐚𝐫𝐭𝐢𝐜𝐮𝐥𝐚𝐫 𝐨𝐫𝐝𝐞𝐫: Co-authors: Yunus Lutz and Philipp Duwe Search Team: Mika Wolle, Robert Breetzmann, Sinem Escher and Cosima Heinen, Joscha Harpend, Daniel Basedow, Minh Tuan Nguyen, Leonie Brinkmann, Max Linus Walter and Thorsten Krause Management: Marcel Kollmar, Hendrik Wimmer, Richar Gründer, Andreas Stuht, Dr. Janina Klautke, Benjamin Kurth and Matthias Baumgarten Former team members: Felix Rolf, Andrea Schuett, Anja Zenz and Andreas Wagenmann If someone is missing, please let me know! #RecSys2025 #MachineLearning #DeepLearning #GBDT #LearningToRank #RecommenderSystems #EcommerceAI #ABTesting #IndustryResearch