How mobile.de brings Data Science to Production...
summary
Personalization at mobile.de uses Hadoop/HDFS, PySpark and TensorFlow to build production systems: a logistic‑regression buyer‑intent model reaches 70% accuracy (sensitivity ~68%, specificity ~76%) using 30‑day windowed features, code shrinks from 2,500 to 1,700 lines and batch jobs reduced runtime to 1–2 hours after migrating to Spark; production serving uses TensorFlow Serving as a microservice, and Spark 2.3 pandas UDFs cut Python‑JVM serialization overhead dramatically.
AI-written summary · about
chapters (8)
- 0:00mobile.de and personalization
- 10:28User intent in car buying
- 17:08Predicting car buying intent
- 22:06Hive to PySpark migration
- 26:00Optimizing Python UDFs with Arrow
- 30:17Python environments in Spark
- 33:50Data science at eBay Classifieds
- 35:07Q&A on recommendations
related talks
02h 00mData Science Driven Business Insights with Python and PySp
39mProductionising Data Science at Scale
17mFrom Jupyter to Spark: Stages of Machine Learning Productionization
01h 48mBig Data made easy with a Spark (3.0)
19mData science for real problems
39mPyData in Production: Lesson Learned from Various PyData Deployment Strategies
37mBig Data Analytics, Machine Learning, Real Time Processing
24mSmart approach in development and deployment process for ML models