Refines results using per-user characteristics. Personalization signal layered on top of Navboost — per-user attributes modulate which results best serve the individual.
Patent Overview
- Inventor
- Hyung-Jin Kim, Oleksandr Grushetskyy, Andrei Lopatenko
- Assignee
- Google LLC
- Filed
- 2014
- Granted
- 2017-07-04
The Challenge
The Challenge
Users have characteristics — location, language, expertise level, content preferences. These characteristics inform which results best serve the user. Personalization based on user characteristics improves relevance without overfitting to per-user history.
- User Characteristics Inform Relevance — Location, language, expertise, preferences all shape which results best serve the user.
- Characteristics Are More Stable Than Behaviors — Per-user characteristics evolve slowly; per-user behavior varies day-to-day. Characteristics provide stable personalization basis.
- Characteristics Plus Cohorts Plus Behavior Layer — Characteristics complement cohorts and aggregate behavior. Each layer adds specificity.
- Privacy Preservation Required — Per-user characteristics are sensitive. Privacy preservation in storage and use is essential.
- Characteristic Inference Must Be Reliable — User characteristics often inferred from behavior. Inference accuracy determines personalization quality.
Innovation
How The System Works
The system captures or infers per-user characteristics, scores per-result fit to each characteristic, combines characteristic-specific fit scores, modulates ranking by user-result characteristic alignment, and preserves privacy throughout.
- Capture Or Infer Characteristics — Per user, capture explicit characteristics (locale, language) and infer implicit characteristics (expertise level, preference signals).
- Score Per-Result Characteristic Fit — Per result, score fit to each user characteristic dimension.
- Combine Characteristic Fit Scores — Per (user, result), combine per-characteristic fit scores into composite fit.
- Modulate Ranking — Per user, ranking modulated by characteristic-fit scores. Better-fit results rise.
- Privacy-Preserve Storage — User characteristics stored with privacy preservation. Aggregations use differential privacy.
- Validate Inference Accuracy — Inferred characteristics validated against held-out data. Drift triggers recalibration.
- Recalibrate Continuously — Per-characteristic inference and fit models recalibrate against fresh data.
Characteristics Personalize Beyond Behavior
The patent's load-bearing idea is that user characteristics — distinct from raw behavior — provide a stable personalization basis. Combining characteristic-based fit with behavior-based cohort signal yields multi-layer personalization.
Stable Characteristics Beat Volatile Behavior
Per-user characteristics evolve slowly. Personalization grounded in characteristics produces consistent, stable relevance rather than chasing volatile behavior.
- Characteristic Capture And Inference — Explicit characteristics captured; implicit characteristics inferred from behavior.
- Per-Result Fit Scoring — Per result, fit to each user characteristic scored.
- Composite Fit — Per (user, result), characteristic fits combine into composite fit score modulating ranking.
Technical Foundation
Technical Foundation
The patent specifies the characteristic capturer, characteristic inferrer, fit scorer, composite combiner, ranking modulator, and privacy layer.
- Characteristic Capturer — Captures explicit user characteristics (locale, language).
- Characteristic Inferrer — Infers implicit characteristics from behavior patterns.
- Fit Scorer — Per (result, characteristic), fit scored.
- Composite Combiner — Per (user, result), combines per-characteristic fits into composite.
- Ranking Modulator — Per user, modulates ranking by composite fit.
- Privacy Layer — Differential privacy or comparable safeguards on characteristics and aggregations.
The Process
The Process
Characteristic capture and inference run continuously; fit scoring and ranking modulation run at query time.
- Capture Characteristics — Per user, explicit characteristics captured; implicit inferred.
- Receive Query — User issues query. Personalization layer activates.
- Fetch Candidates — Base ranking returns candidate results.
- Score Per-Result Fit — Per result, fit to user characteristics scored.
- Combine Fit Scores — Composite fit per (user, result) computed.
- Modulate Ranking — Composite fit modulates base ranking. Personalized results emerge.
- Validate And Recalibrate — Inferred characteristics and fit models validated and recalibrated periodically.
Quality Control
Quality Control
Characteristic inference accuracy determines personalization quality. The patent specifies safeguards.
- Inference Accuracy Validation — Inferred characteristics validated against labeled data. Drift triggers recalibration.
- Privacy Preservation — Per-user characteristics stored with privacy preservation. Aggregations use differential privacy.
- Modulation Bounds — Per-characteristic modulation magnitudes bounded. Prevents over-personalization.
- Adversarial Defense — Synthetic characteristics designed to manipulate ranking flagged and filtered.
- Continuous Recalibration — Inference models and fit scorers recalibrate against fresh data.
Real-World Application
Characteristic-based personalization complements behavior-based cohort signal. The combination produces multi-layer personalization that respects stable user characteristics while adapting to evolving behavior.
- Per-user Granularity — Per user, characteristics captured or inferred. Personalization individualized.
- Multi-dimensional Characteristic Scope — Locale, language, expertise, preferences combine. Multi-dimensional fit.
- Composite-scored Fit Method — Per (user, result), per-characteristic fits combine into composite fit.
Why Locale And Language Matter Structurally
Locale and language are stable characteristics that strongly modulate result fit. Content localized appropriately earns characteristic-based boost for matching locale/language users — straightforward but high-leverage.
Why Audience Targeting Compounds With Characteristics
Content written for a specific expertise level or audience fits users with matching inferred characteristics. The match earns characteristic-fit boost; serving the wrong characteristic level earns penalty.
<\/section>What This Means for SEO
What This Means for SEO
This patent personalizes results using stable per-user characteristics (locale, language, inferred expertise, preferences) layered on top of behavioral signals. SEO implication: matching content to clear audience characteristics, especially locale, language, and expertise level, earns a characteristic-fit boost for the users you target.
- Locale And Language Are High-Leverage — Locale and language are stable characteristics that strongly modulate result fit. Properly localizing content earns a straightforward characteristic-based boost for matching users; serving the wrong locale costs you.
- Target A Clear Expertise Level — The system infers expertise and matches content to it. Writing unambiguously for beginners or for experts earns characteristic-fit boost with that audience, while content that straddles levels fits neither well.
- Characteristics Are More Stable Than Behavior — Personalization here is grounded in slowly-evolving characteristics rather than volatile day-to-day behavior. Building for a consistent audience profile produces stable, durable relevance rather than chasing fleeting behavior.
- Composite Fit Combines Many Dimensions — Locale, language, expertise, and preferences combine into a composite fit score. Aligning on several dimensions at once compounds the boost, so localize and pitch the expertise level together.
- Serving The Wrong Characteristic Penalizes — Mismatch on a characteristic dampens ranking for that user. Forcing expert content on a beginner audience, or vice versa, is an active misfit, not a neutral choice.
- Synthetic Characteristics Are Filtered — Attempts to fake characteristics to manipulate ranking are flagged. You earn fit by genuinely matching the audience, not by signaling false attributes.
- It Layers On Behavioral Signals — Characteristic fit complements cohort and aggregate behavior rather than replacing them. The strongest position is content that both fits the user's characteristics and earns their engagement.