Understanding Customer Journeys with Deep Learning
summary
Zalando aggregates terabytes of clickstream into per-cookie timelines, stores Parquet on S3 and pivots with Spark on EMR to a compact feature matrix (terabytes to gigabytes), then trains models on a Berlin GPU HPC. Models predict order probability (session or 7-day) using time-deltas and simple event encodings; recurrent nets trained in Torch produce a ~10‑dimensional latent state that rises on cart/wishlist events and falls after purchase, and RNNs (Torch) on GPU HPC match or beat a sophisticated production feature/logistic baseline by AUC. Pipeline: Kafka -> S3 -> Parquet -> EMR Spark. Privacy: advertising uses cookie IDs only, sessions use 30‑minute inactivity.
AI-written summary · about
chapters (17)
- 0:00Introduction
- 1:10Who are we
- 2:08Tech hubs
- 2:47Adtech
- 3:51Online Advertising
- 10:20Customer Journeys
- 18:49Cookie History
- 19:05Traditional Approach
- 20:40What is Deep Learning
- 23:14How to use Deep Learning
- 23:45Recurrent Neural Network
- 25:38Recap
- 26:45Experiments
- 27:14Results
- 29:20Wrapup
- 38:21Conclusion
- 40:35Interview
related talks
35mLife, Death, and Shopping
01h 35mDeep Dive into the Big Data Landscape
42mHow mobile.de brings Data Science to Production...
53mMachine Learning in E‑commerce
30mMobile Order Click-Through Rate (CTR) Recommendation with Ray on Apache Spark at Burger King
44mBest practices for analytics and generative AI on AWS(ANT302)
17mLaying the foundation for deep learning
29mMachine Learning in Production: What to do after you’ve trained the models?