Patterns for Personalization in Recommendations and Search
Source: https://eugeneyan.com/writing/patterns-for-personalization/ Author: Eugene Yan
Summary
Survey of five core ML patterns for personalization, with industry examples from Airbnb, Netflix, YouTube, DoorDash, Alibaba, and others.
The Five Patterns
1. Bandits
Continuously balance exploration vs. exploitation. Netflix uses contextual bandits for image personalization; DoorDash applies them to cuisine recommendations with multi-level geolocation context.
2. Embeddings + MLP
Map sparse features (user ID, item ID, behavioral signals) into dense vectors, pool variable-length sequences, feed through MLP. Used by TripAdvisor (experiences), YouTube (videos).
3. Sequential Models
Capture item order using RNNs or Transformers. Alibaba’s Behavioral Sequence Transformer achieved +4.5% CTR over mean pooling baseline by modeling behavioral sequences.
4. Graph-Based
Represent users and items as nodes in weighted graphs. Uber uses graph convolutional networks for food recommendations; Alibaba uses graph intention networks.
5. User Embeddings
Learn direct user representations. Airbnb generates user-type embeddings for search (see Listing Embeddings in Search Ranking); Tencent builds user lookalikes for long-tail content.
Cross-Cutting Concerns
- Cold-start problem: handling new users/items with limited data
- Exploration vs. exploitation: discovery vs. known preferences
- Pooling strategies: mean/sum/max compress variable behavioral sequences
- Attention mechanisms: weight history by relevance to current context