Otto
Otto (otto.de) is one of Germany’s largest e-commerce companies, transitioning from a catalogue retailer to a marketplace with millions of products from multiple sellers.
Search & Ranking Work
LTR implementation — motivated by marketplace expansion making manual ranking rules unscalable:
- Model: LambdaMART (pairwise + NDCG-weighted listwise)
- Training framework: RankyMcRankFace (fork of RankLib)
- Infrastructure: Solr with LTR plugin + feature store
- Labels: derived from customer interaction data (clicks, add-to-basket, orders, dwell time)
- Evaluation: custom Offline Metrics Analyzer computing NDCG at multiple configurations
Challenge: crowd-sourced labels don’t work for purchase-intent relevance — labels must come from behavioral signals.