From Trees to Transformers - Adopting Deep Learning for Ranking

Learn how GetYourGuide successfully migrated their ranking pipeline from XGBoost to a PyTorch transformer-based deep learning architecture in just nine months while maintaining strict latency and throughput requirements. Discover the phased approach used to transition from traditional tree-based models to neural networks in production, starting with a minimal viable model and gradually increasing complexity through over 50 offline iterations and more than 10 live A/B tests. Explore the operational and modeling challenges encountered during the migration, including maintaining real-time performance constraints while implementing deep learning solutions for a global online marketplace. Understand the strategic decision-making process behind adopting deep learning when traditional model improvements plateaued, and examine the flexibility benefits of neural networks for capturing complex patterns and interactions in ranking data. Gain practical insights into testing methodologies, production deployment strategies, and the business impact achieved through transformer-based ranking systems. Walk away with actionable strategies for advancing machine learning capabilities and implementing neural network solutions that unlock new dimensions of relevance and personalization in real-time ranking applications.