Personalization

Definition

Personalization adapts search results to individual user preferences, history, and context — returning different results for different users submitting the same query, based on signals like their browsing history, purchase behavior, location, and demographics.

Why Personalize?

The same query “shoes” means different things to different users:

• A runner → technical running shoes
• A parent → children’s shoes
• A fashion-conscious shopper → trendy sneakers
• A formal dresser → dress shoes

Generic ranking serves the “average user” — personalization serves the actual user.

Personalization Signals

Signal Type	Examples	Freshness
Explicit preferences	Saved searches, wishlists	Static
Purchase history	Bought brands, price range	Medium
Browse history	Categories, products viewed	Fresh
Search history	Previous queries, clicked results	Fresh
Demographics	Age, gender, location	Static
Session context	Current session queries	Real-time
Social proof	Friends’ purchases	Medium

Personalization Approaches

Re-ranking

Run generic retrieval → re-rank based on user preferences.

• Simple to implement (add user features to Learning to Rank)
• Doesn’t change candidate set — limited recall personalization

Query Rewriting

Modify the query before retrieval based on user profile:

• “shoes” → “Nike women’s running shoes” (based on user history)
• Improves both recall and precision

Embedding-Based Personalization

Airbnb’s approach: learn user embeddings alongside listing embeddings in the same space.

• Query embedding → user-contextualized embedding
• Personalized ANN search over listing embeddings
• Published as “Listing Embeddings in Search Ranking”

Sequence Pooling over Behavioral History

A user’s interaction history is variable-length, but most ranking models want a fixed-size input. The simplest bridge is pooling — mean, sum or max over the sequence of interacted items — which compresses the history into one preference vector at the cost of discarding order.

Sequence models (RNNs, Transformers) are the order-preserving alternative. Alibaba’s Behavioral Sequence Transformer reported +4.5% CTR over a mean-pooling baseline by modeling the sequence instead of collapsing it — the same fidelity-vs-cost tradeoff that pooling forces in embedding construction. See Patterns for Personalization.

Feature Engineering for Personalized LTR

Add user features to the Learning to Rank model:

• User’s average purchase price range
• User’s brand affinities
• User’s category preferences
• Session recency signals

See LTR Feature Engineering for a full taxonomy of user affinity features and selection strategy.

In-Engine Click Profile Matching (Kleinanzeigen / Vespa)

Kleinanzeigen moved user behavioral profiles entirely inside Vespa, eliminating the external orchestration layer. Click events fire as Vespa document updates, processed in-process by a document processor. Profiles decay over time, are L2-normalized, and are stored as Vespa documents keyed {userId}:category:{categoryId}. At query time: read profile → expand tokens via a relations similarity graph → fire a WAND query. No external profile store. No service hop.

This exemplifies retrieval-personalization colocation: profile maintenance and retrieval in the same platform removes service-boundary latency entirely.

See: From Elasticsearch to Vespa - Rebuilding the Kleinanzeigen Homepage Feed Part 1, Kleinanzeigen - Vespa Migration for Homepage Feed

Privacy Considerations

Personalization requires collecting and using personal data:

• Compliance: GDPR, CCPA require consent and right to erasure
• User trust: explicit personalization signals (wishlists) are trusted; opaque behavioral targeting is not
• Security: personalization profiles are sensitive data — must be protected

Facebook/Meta’s 98 personal data points for ad targeting is the extreme example of what’s possible (and controversial).

Airbnb Case Study

Airbnb’s personalization system (Eugene Yan et al.):

• Trains user embeddings from booking sequences
• Listing embeddings from viewing sequences
• Online: user embedding → approximate nearest neighbor search over listings
• Result: +21% booking lift in A/B test

The key insight: user preferences are encoded in their past interactions, not in their profile form. Embeddings capture implicit preferences.

Related Concepts

• Learning to Rank — personalization features in LTR models
• Dense Vector Retrieval — user and item embeddings for personalized ANN
• Semantic Search — personal context modifies query meaning
• Click Signals — behavioral personalization signals
• Session-Based Evaluation — personalization evaluated at session level
• Search Intent — personalization refines intent interpretation
• Metarank — open-source real-time click-based personalization / re-ranking service
• Implicit Judgments — behavioral labels that drive personalized LTR
• Pooling — collapsing a variable-length interaction history into one preference vector

People

• Daniel Tunkelang — personalization in search; QueryUnderstanding.com series

Elastic Governed Personalization (Ecommerce)

Two mechanisms extending a governance control plane without replacing it:

1. Purchase history boosting — query user_purchases index in parallel with governance; logarithmic frequency + exponential decay recency weights; governance still controls what appears
2. Cohort-aware policy activation — cohort policies stored in same policy engine; terms filter {"cohorts": ["_all", "vegan"]} narrows candidate policy set

Ordering (innermost → outermost)

1. Base query (keyword/semantic)
2. Governance policy layer (hard filters + soft boosts)
3. Business-signal boosts (margin, popularity)
4. Purchase history boosts (outermost)

Articles

• Elasticsearch Personalized Search in Ecommerce - Improve Relevance — Alexander Marquardt, Honza Král, Taylor Roy
• Using LLMs to Amplify Human Labeling and Improve Dash Search Relevance — Dmitriy Meyerzon (Dropbox Dash enterprise search)
• Metarank - Personalized Ranking That Actually Reads Your Clicks — Florian Narr; real-time click personalization service

People (additional)

• Alexander Marquardt — Elastic governed personalization
• Honza Král — Elastic governed personalization
• Taylor Roy — Elastic governed personalization
• Dmitriy Meyerzon — Dropbox Dash relevance labeling