Better search relevance using Learning to Rank at mobile.de
summary
mobile.de replaces the Elasticsearch LTR plugin with a Python ranking microservice (FastAPI) that reads Kafka events, writes to BigQuery, uses Airflow to sync and caches predictions in MongoDB, serving a LightGBM LambdaMART pairwise ranker with monotonic constraints and multi‑objective weighting. Offline NDCG validation matches live A/B gains: 5% more user contacts, 2.07% more clicks and 1.3% more saves, while the service hits ~6 ms median latency, >99% success rate and peaks at ~1,500 requests/min.
AI-written summary · about
chapters (11)
- 0:00Introduction to search
- 1:57Identifying architecture gaps
- 3:52Defining the new solution
- 5:26Building the ranking service
- 7:11Integration and data flow
- 10:29Advanced ranking techniques
- 12:47Multi-objective approach
- 13:16AB test and performance
- 14:24Operational best practices
- 15:35Project reflections
- 20:04Q&A session
related talks
20mArchitecture of relevancy search at mobile.de
47mLearning to rank search results
43m[Machine Learning] Creating more relevant search results with Learn to Rank
29mBuilding Learning to Rank models for search using LLMs
27mA smooth ride: Online car buying and selling at mobile.de
38mLearning to Rank journey at GetYourGuide: Our Logbook #bbuzz
40mLevel Up: Everything you need to know about search, ...
41mBuilding MLOps Infrastructure at Japan's Largest C2C E-Commerce Site