Propagates query-classification labels across the query graph. Lets the system classify a long-tail query by analogy to classified head queries — class labels flow along query-similarity edges.
Patent Overview
- Inventor
- Henele Adams, Hyung-Jin Kim, others
- Assignee
- Google LLC
- Filed
- 2012
- Granted
- 2017-05-23
The Challenge
The Challenge
Query classifications (intent, type, topical category) inform ranking. But classifying every query individually doesn't scale, especially for long-tail queries. Propagating classifications across the query similarity graph lets head-query labels lift long-tail queries.
- Per-Query Classification Doesn't Scale — Manual or model-per-query classification can't cover every query. Propagation is required.
- Similar Queries Share Classifications — Semantically similar queries tend to share intent, type, and topical categories. Class labels transfer along similarity edges.
- Propagation Must Be Bounded — Too-aggressive propagation overgeneralizes; too-conservative misses long-tail coverage. Calibration matters.
- Confidence Decreases With Distance — Propagation confidence decreases along graph distance. Far-from-source classifications less reliable.
- Iterative Convergence Required — Propagation runs iteratively until classifications converge. Single-pass propagation undercoverage.
Innovation
How The System Works
The system builds a query similarity graph, seeds it with manually or model-classified queries, propagates classification labels along similarity edges with confidence decay, iterates until convergence, and exposes the resulting classifications to ranking.
- Build Query Similarity Graph — Per query, identify semantically similar queries. Edges weighted by similarity strength.
- Seed With Classified Queries — Head queries classified manually or by model. Become seed nodes with confidence 1.0.
- Propagate Along Edges — Per iteration, classifications propagate along edges. Confidence decays with edge weight and distance.
- Aggregate Multi-Source Propagation — Per query, classifications from multiple sources aggregate. Convergent classifications strengthen.
- Iterate To Convergence — Propagation runs iteratively. Convergence reached when classification labels stabilize.
- Expose To Ranking — Per query, final classification labels feed ranking. Query-type and intent signals available for all queries.
- Refresh As Queries Evolve — Graph and seed classifications refresh as queries and language evolve.
Labels Flow Along Similarity
The patent's load-bearing idea is that classifications transfer along query similarity edges. Head-query labels flow to long-tail queries via graph propagation, providing classification coverage that per-query classification cannot.
Propagation Scales Classification
Per-query classification doesn't scale. Graph propagation scales naturally — seed at the head, propagate to the tail. The architectural insight is the graph-based scaling.
- Similarity Graph — Per query, semantically similar queries connect via weighted edges.
- Seed Plus Propagate — Head queries classified become seeds. Labels propagate along edges with confidence decay.
- Iterative Convergence — Propagation iterates until labels stabilize. Multi-source convergence strengthens labels.
Technical Foundation
Technical Foundation
The patent specifies the similarity-graph builder, seed classifier, edge propagator, multi-source aggregator, convergence detector, and ranking integrator.
- Similarity-Graph Builder — Per query, identifies similar queries; edges weighted by similarity.
- Seed Classifier — Head queries classified manually or by model. Seeds with confidence 1.0.
- Edge Propagator — Per iteration, classifications propagate along edges. Confidence decays.
- Multi-Source Aggregator — Per query, classifications from multiple sources aggregate.
- Convergence Detector — Detects when labels stabilize. Iteration stops at convergence.
- Ranking Integrator — Per query, final classifications feed ranking.
The Process
The Process
Graph construction and propagation run offline. Final classifications cache for query-time consumption.
- Build Graph — Query similarity graph built.
- Seed With Classified — Head queries seed graph.
- Propagate Per Iteration — Labels propagate along edges with decay.
- Aggregate Multi-Source — Per query, multi-source labels aggregate.
- Check Convergence — Convergence detector monitors stability.
- Cache Classifications — Converged labels cache for query time.
- Refresh Periodically — Graph and seeds refresh as queries evolve.
Quality Control
Quality Control
Propagation correctness depends on similarity graph quality and decay calibration. The patent specifies safeguards.
- Similarity-Edge Validation — Edges validated against held-out labeled data. Wrong edges propagate wrong labels.
- Decay Calibration — Confidence decay calibrated. Too-fast decay leaves tail unclassified; too-slow propagates noise.
- Multi-Source Confirmation — Per query, classifications require multi-source convergence. Single-source flags rejected.
- Seed Quality — Seed classifications validated. Wrong seeds propagate widely.
- Continuous Recalibration — Graph, decay, and seed quality recalibrate against fresh data.
Real-World Application
Query-classification propagation densifies classification coverage across the long tail. The graph-based scaling pattern underpins how modern engines route long-tail queries to appropriate rankers, SERP layouts, and feature surfaces.
- Graph-based Propagation Method — Query similarity graph carries class labels along edges.
- Iterative Convergence Method — Propagation runs iteratively until labels stabilize.
- Multi-source Aggregation Pattern — Per query, multi-source classifications aggregate. Convergence strengthens labels.
Why Topical Coverage Multiplies Classification Reach
Pages covering a topical cluster surface across many queries within that cluster. Propagation means head-query classifications inform long-tail queries. Topical authority earns favorable classification across the cluster's long tail.
Why Query Intent Matching Compounds
Once a query is classified, ranking applies type-aware logic. Pages matching the classified intent earn type-match boost. Propagation extends this benefit to long-tail queries that inherit classifications from head.
<\/section>What This Means for SEO
What This Means for SEO
This patent propagates query-classification labels (intent, type, topical category) along a query-similarity graph so long-tail queries inherit head-query classifications. SEO implication: once a query is classified, ranking applies type-aware logic, so topical coverage earns favorable classification across a cluster's long tail.
- Topical Coverage Multiplies Classification Reach — Head-query classifications propagate to long-tail queries across the similarity graph. Pages covering a topical cluster benefit as those tail queries inherit favorable classifications from the head.
- Match The Classified Intent To Earn Type Boost — Once classified, ranking applies type-aware logic and rewards pages matching that intent. Aligning your page format to the classified intent of the cluster earns a type-match boost that extends to long-tail variants.
- Classification Routes Queries To Surfaces — Labels determine which ranker, SERP layout, and feature surface a query gets. Understanding your target queries' classification tells you which result format to build for to be eligible.
- Long-Tail Inherits Head Classification — Because labels flow from seed head queries outward, your long-tail relevance rides on the classification of the head queries you serve. Winning the head topic shapes how its tail is treated.
- Confidence Decays With Distance — Propagated classifications weaken with graph distance, so far-flung queries are less reliably labeled. Staying tightly within a coherent topic keeps you near high-confidence classifications.
- Multi-Source Convergence Strengthens Labels — Labels reinforced by multiple similar queries are stronger. Comprehensive coverage that connects to many related queries helps anchor your topic's classification firmly.
- Build For Coherent Topic Clusters — Since propagation scales classification along similarity edges, clustered, internally-related content benefits more than scattered pages. Organize content into coherent topical clusters to maximize inherited classification.