Source: Production methodology at e-commerce and consumer search companies; Tunkelang on personalized search; Coveo and Algolia personalization documentation
Classification — Pattern for integrating user-specific, session-specific, and contextual signals as features in LTR ranking models.
Adjust ranking based on context features that capture who the user is, what they've done recently, and their current environment, producing per-user-per-query ranking that outperforms uniform ranking on personal-relevance metrics.
Two users entering the same query string often want different results. A returning customer searching "running shoes" probably wants results filtered by their prior preferences (size, brand affinity, price tier). A user in a specific region wants results adjusted for local availability. A user with prior session context (just viewed Nike products) probably wants related items boosted. Without personalization, the ranking treats all users identically, missing per-user relevance signals that improve outcomes meaningfully.
User features. Long-term signals derived from accumulated user behavior: purchase history (what brands, categories, price ranges they've bought), viewing history (what they've looked at without buying), preference signals (saved searches, wish lists, ratings), demographic signals where available and legitimate (age band, region). These features are computed offline and stored in a user profile that the ranking pipeline reads at query time.
Session features. Short-term signals from the current session: queries earlier in the session, items viewed, filters applied, time spent on results, items added to cart. Session features capture the user's current intent more directly than long-term profile data; a user who just searched "red running shoes" and is now searching "women's" is probably looking for women's red running shoes specifically. Session features are typically computed on-the-fly from a session store (Redis, custom in-memory infrastructure).
Contextual features. Time, device, geographic, and environmental signals: hour of day (some queries shift meaning by time — "restaurant" at noon means lunch, at 10pm means dinner), day of week, season, device type (mobile vs desktop affects result presentation needs), geographic location and locale, traffic source. Contextual features are derived from request metadata; they're typically the simplest to compute and add.
Feature integration as LTR features. The personalization signals join other features in the LTR model. A model trained on (query, document, user/session context, relevance) tuples learns to use the context features alongside query, document, and query-document features. The LTR model decides per-query how much weight to give the personalization features; for navigational queries (where the user wants a specific item regardless of personalization) the model learns to weight personalization less; for discovery queries (where personal preference matters) the model weights personalization more.
Cold-start handling. New users without profiles, anonymous sessions, no available context: personalization needs to degrade gracefully. Common patterns: fall back to non-personalized model (when no context available, model uses query and document features only); use weak available signals cautiously (locale-based defaults); let session signals accumulate within the session to enable per-session personalization even without long-term profile.
Privacy considerations. User behavior data is sensitive. Personalization implementations must comply with privacy regulations (GDPR, CCPA, sector-specific rules), retention policies, and user consent frameworks. Best practice: keep personalization signals visible to users ("Recommended based on your recent activity"), allow opt-out, log the signals used for any specific result. The discipline overlaps with the agentic AI series' compliance volume; for search specifically, personalization opacity erodes user trust over time.
Filter bubble considerations. Aggressive personalization can collapse results to a narrow set the user has previously engaged with, missing valuable discovery. Production patterns: diversification rules (Section F) that limit how much personalization narrows results; exploration injection that surfaces non-personalized candidates; periodic re-evaluation that catches when personalization is hurting outcomes. Multi-objective ranking (Section G) provides the framework for balancing personalization against discovery.
E-commerce search where user history and preferences meaningfully affect what users want. Enterprise search where user role, team, and access patterns affect relevance. Consumer search where session context shapes intent. Multi-region deployments where locale-based personalization is mandatory.
Alternatives — non-personalized ranking for narrow use cases or cold-start situations. Lightweight personalization (locale, device) only for privacy-sensitive deployments. The right level of personalization depends on the use case; defaulting to maximum personalization is typically wrong.
- Tunkelang's writing on personalized search
- Grainger, AI-Powered Search, chapters on personalization
- Coveo personalization documentation
- Algolia personalization documentation