The Power of the Bonus Card: Road to Personalised Search
summary
Albert Heijn uses Bonus Card purchase histories to move from popularity to personalized product search, training a listwise Learning-to-Rank model with XGBoost on a million-fold increase in data of single impression events, joined point-in-time via an offline feature store and materialized to an in-memory online store over Kafka for inference under P99 latency 100ms. Relevancy labels combine debiased conversion rates (via inverse propensity scoring), contribution margin and CO2/health metrics; negatives are downsampled (≈3:1), cold-start is handled by temporary boosts and small exploration fractions, and alternatives use product embeddings with cosine similarity.
AI-written summary · about
chapters (10)
- 0:00Introduction and motivation
- 3:51Search importance and bonus card
- 6:19Old search system setup
- 10:56Personalized search changes
- 17:40Beyond customer relevancy
- 22:19AB testing and continuous improvement
- 24:46Relevancy labels and debiasing
- 31:24Semantic search and alternatives
- 35:40Future of search at Albert Heijn
- 37:26Q&A and closing
related talks
27mTransform online shopping experiences with personalized recommendations
31mEverything in its Right Place: Optimising Ranking in Online Grocery
37mFood is personal! Design and impact of personalized recommendations in e-grocery
32mLeveraging recommender systems to personalise search results
28mThe Power of Personalized Customer Experiences Through AI Innovations
17mCalculating recommendations based on product images #bbuzz
23mML-Powered Search & Recommendation 101
35mLearning to Judge