Identifies queries by association patterns across users and sessions, supporting query-canonicalization and intent-cluster identification for the downstream ranking, personalization, and answer-extraction layers.
Patent Overview
- Inventor
- Ramanathan V. Guha
- Assignee
- Google LLC
- Filed
- 2009-07-31
- Granted
- 2014-01-14
- Application Number
- US 12/512,907
The Challenge
The Challenge
The same intent appears in many surface forms across users: 'cheap flights NYC', 'inexpensive plane tickets to New York', 'low-cost airfare NYC' all want the same outcome. The system needed to identify these as one intent cluster so downstream layers (ranking, personalization, answer extraction) operate on canonical intent rather than surface variation.
- Surface Variation Hides Shared Intent — Many queries express the same intent in different words. Treating each surface form independently fragments signal that downstream layers could use coherently.
- Association Patterns Reveal Equivalence — When users click similar results, follow similar journeys, or reformulate to canonical forms, the patterns reveal which surface forms share intent.
- Canonicalization Improves Downstream Quality — Once intent clusters are identified, ranking, personalization, and answer extraction can operate on the canonical intent. Signal aggregates across surface forms.
- Identification Must Resist Drift — Intent clusters evolve as topics shift and new phrasings emerge. The system must update clusters as patterns change.
- Privacy Boundaries Constrain Pattern Mining — Cross-user association mining must respect privacy. Patterns aggregate without exposing individual users.
Innovation
How The System Works
The system mines query logs for association patterns across users and sessions, clusters surface-form queries by shared intent, identifies a canonical form per cluster, and exposes intent clusters to downstream layers so ranking, personalization, and answer-extraction operate on coherent intent rather than fragmented surface form.
- Collect Query Logs — Aggregate query logs across users and sessions. Logs are pseudonymized and aggregated; raw individual data is not used directly.
- Mine Association Patterns — Patterns include shared click-throughs (different queries clicking the same results), shared journeys (one query followed by another), and shared topical reformulations.
- Cluster By Shared Intent — Queries with strong association patterns cluster together. Each cluster represents one canonical intent expressed across many surface forms.
- Identify Canonical Form — Per cluster, identify the canonical query form: usually the most-common, most-explicit, or most-engagement-bearing variant. Canonical form represents the cluster.
- Expose Clusters Downstream — Cluster membership and canonical form expose to downstream layers via the query-feature store. Ranking, personalization, answer extraction can consume.
- Update Continuously — As new queries arrive and patterns evolve, clusters update. The system stays current with shifting phrasings and emerging intents.
- Respect Privacy Boundaries — Aggregation respects privacy: minimum-cluster-size thresholds, no individual-user trace, sensitive-category exclusions.
Surface Form To Canonical Intent
The patent's load-bearing idea is to bridge the gap between surface query variation and canonical intent through association-pattern mining. Downstream layers benefit from coherent intent rather than fragmented surface forms.
Patterns Across Users Reveal Intent Equivalence
No single user issues all surface variants of an intent. But across users, association patterns make the equivalences visible. Cross-user aggregation reveals what individual analysis cannot.
- Association Pattern Mining — Shared click-throughs, journeys, and reformulations across users reveal which queries share intent. Patterns are the substrate for clustering.
- Intent Clusters With Canonical Forms — Per cluster, one canonical form represents the intent. Downstream layers reference the canonical rather than surface variants.
- Privacy-Bounded Aggregation — Patterns mine in aggregate. Minimum cluster sizes plus pseudonymization protect individual users.
Technical Foundation
Technical Foundation
The patent specifies the query log aggregator, the association pattern miner, the clustering algorithm, the canonical-form selector, the cluster store, and the privacy enforcement.
- Query Log Aggregator — Pseudonymized query logs aggregate across users and sessions. Aggregation handles billions of queries per day at scale.
- Association Pattern Miner — Identifies association patterns: queries sharing click-throughs, queries appearing in shared journeys, queries reformulated from common origins.
- Clustering Algorithm — Queries with strong association patterns cluster together. Hierarchical or graph-based clustering produces intent clusters at appropriate granularity.
- Canonical Form Selector — Per cluster, selects the canonical form based on frequency, explicitness, and engagement. The canonical represents the cluster downstream.
- Cluster Store — Per cluster, stores membership (all surface forms), canonical form, and metadata (size, age, engagement profile). Fast lookup at query time.
- Privacy Enforcement — Minimum cluster size thresholds, sensitive-category exclusions, and pseudonymization protect users. No individual queries are exposed.
The Process
The Process
The pipeline runs as periodic batch over aggregated query logs. Output is updated intent clusters that downstream layers consume at query time.
- Aggregate Query Logs — Pseudonymized logs aggregate across users. Aggregation prepares logs for pattern mining.
- Mine Patterns — Association pattern miner identifies shared-click-through, shared-journey, and reformulation patterns across query pairs.
- Cluster Queries — Clustering algorithm produces intent clusters. Each cluster contains surface-form queries sharing intent.
- Select Canonicals — Per cluster, canonical form is selected based on frequency, explicitness, engagement.
- Publish To Cluster Store — Per cluster, membership and canonical publish to the store. Downstream layers consume via lookup.
- Apply In Downstream Layers — Ranking, personalization, answer extraction read cluster membership at query time. Operations use canonical intent.
- Refresh — Periodic refresh updates clusters as patterns evolve. New queries get cluster assignments; stale clusters merge or split.
Quality Control
Quality Control
Wrong cluster assignments confuse downstream layers. The patent specifies safeguards.
- Pattern Mining Calibration — Association patterns must clear significance thresholds. Spurious co-occurrence is filtered out before contributing to clusters.
- Cluster Coherence Validation — Clusters are validated for internal coherence: members really share intent. Bad clusters split or refined.
- Canonical Form Validation — Canonical selection respects frequency, explicitness, engagement. Wrong canonicals propagate errors downstream.
- Privacy Threshold Enforcement — Minimum cluster sizes prevent inference about individuals. Sensitive categories excluded from mining.
- Downstream Outcome Monitoring — Quality of cluster-aware ranking, personalization, and answer extraction is monitored. Drops trigger cluster reanalysis.
Real-World Application
Query identification primitives underpin Google's query understanding stack: intent clusters inform ranking weights per query type, personalization profiles per intent, and answer-extraction targeting. The patent's mining-to-clustering pipeline is the substrate for much of modern query interpretation.
- Cross-user Mining Scope — Patterns mine across users in aggregate. Cross-user analysis reveals what single-user analysis cannot.
- Cluster-based Output Form — Intent clusters with canonical forms are the output. Downstream layers reference canonicals.
- Privacy-bounded Aggregation Model — Mining respects privacy: minimum thresholds, sensitive-category exclusions, pseudonymization.
Why Targeting Canonical Intent Beats Targeting Surface Keywords
SEO that targets surface keyword variations splits effort across many micro-queries. Targeting canonical intent (the cluster's identity, not individual surface forms) compounds effort because all surface variants flow to the same cluster's results.
Why Long-Tail Surfaces Roll Up To Clusters
Long-tail queries that look distinct often share intent with high-volume queries through cluster membership. Content that serves the canonical intent earns visibility on long-tail surfaces through cluster-aware retrieval.
<\/section>What This Means for SEO
What This Means for SEO
The patent mines query logs across users and sessions to cluster surface-form queries into canonical intents that downstream ranking, personalization, and answer layers operate on. SEO implication: targeting the canonical intent of a cluster compounds, because every surface variant flows to the same cluster's results.
- Target Canonical Intent, Not Surface Keywords — Many surface variants roll up to one intent cluster. Optimizing for the cluster's underlying need beats splitting effort across dozens of micro-keyword pages, because all variants resolve to the same canonical intent the downstream layers serve.
- One Strong Page Per Intent Cluster — Since variants converge, fragmenting coverage across near-duplicate pages dilutes signal and risks cannibalization. Consolidate into one authoritative page per canonical intent so the cluster's traffic concentrates rather than splits.
- Long-Tail Variants Inherit Cluster Visibility — Distinct-looking long-tail queries often share intent with high-volume queries via cluster membership. Content serving the canonical intent earns visibility on long-tail surfaces through cluster-aware retrieval without separate optimization for each tail phrase.
- Cross-User Patterns Reveal Equivalence — No single user issues every variant, but cross-user association makes equivalences visible. Reflecting how a broad audience phrases the same need (synonyms, reformulations) on your page aligns you with the full cluster rather than one phrasing.
- Canonicalization Feeds Personalization — Clusters are exposed to the personalization layer. Content matched to a canonical intent benefits when personalization promotes that intent for a user. Clean intent alignment is what lets personalization route the right users to you.
- Answer Extraction Reads The Cluster — The answer-extraction layer operates on canonical intent. Structuring your page to directly answer the cluster's core question positions you for direct-answer surfaces across all of the cluster's surface forms.
- Reformulation History Defines The Cluster — Clusters form partly from session reformulation patterns. Understanding the sequence of how users refine toward an intent helps you cover the resolved intent, which is where the cluster ultimately points.