“Which car fits my life?” - mobile.de’s approach to recommendations
summary
mobile.de runs a hybrid recommender that combines collaborative matrix factorization with content vectors to handle a rapidly changing 1.6 million-item inventory and cold starts, serving ranked results via ElasticSearch + Cassandra + Mahout and a Python/Java pipeline. Models (they use LightFM matrix factorization) compress sparse user×item data, retrain weekly while user preference vectors update near real time, and drive higher engagement across a site with 13.5 million monthly users.
AI-written summary · about
chapters (9)
- 0:00Introduction
- 3:38Why recommendations matter
- 8:24Traditional recommendation algorithms
- 12:17Collaborative filtering challenges
- 14:01User preferences and content-based filtering
- 18:01Hybrid recommender approach
- 23:03LightFM real-world example
- 31:11Deep learning for recommendations
- 34:26Q&A
related talks
42mWhich Car Fits My Life? - mobile.de’s Approach to Recommendations
27mA smooth ride: Online car buying and selling at mobile.de
35mHybrid Recommender Systems in Python
58mRecommender Systems – A Walkthrough
41mRecommender Systems: Crafting Effective Personalised Experiences
20mArchitecture of relevancy search at mobile.de
37mA primer on recommendation systems.
29mThe Challenges of Building a Personalized Recommendation System in E-Commerce