Builds enterprise knowledge graphs from user behavior signals (queries, clicks, edits, co-occurrence in documents) rather than from manually-declared schemas, so the graph captures the relationships people actually care about and adapts as workflows evolve.
Patent Overview
- Filed
- 2020-07-17
- Granted
- 2022-01-20 (published application)
- Application Number
- US 16/932,540
The Challenge
The Challenge
Enterprise knowledge graphs traditionally come from declared schemas: someone defines entity types, relationships, and instances by hand. Schemas miss the live structure of how employees actually work and rapidly become stale. The system needs to learn the graph from behavior.
- Declared Schemas Are Brittle — Hand-built enterprise schemas reflect one team's view at one moment. They miss cross-team relationships, lag changes in workflow, and become abandonware as the people who built them move on.
- User Behavior Is Continuous Signal — Every query, click, edit, and document co-occurrence carries relationship information. Aggregated, behavior reveals the live structure of how people use enterprise content.
- Privacy Constraints Are Real — User behavior signals include sensitive personal and workflow data. The mining pipeline must respect access controls, retention policies, and consent boundaries.
- Sparse Behavior For Long Tail — Most enterprise entities have few interactions. Mining must work despite sparse signal, perhaps by borrowing across similar entities.
- Graph Must Stay Current — Enterprise content turns over fast. A graph built from last quarter's behavior may already be stale. Mining must run continuously and graph contents must update as behavior shifts.
Innovation
How The System Works
The patent ingests behavior signals from enterprise tools (search, document editing, email, chat), extracts entity-relationship inferences from those signals, weighted by signal strength, and continuously updates a graph that reflects current organizational reality rather than declared structure.
- Ingest Behavior Signals — Enterprise tools emit usage signals: search queries, clicked results, document edits, co-membership in folders, email mentions, chat references. Signals stream into the mining pipeline.
- Apply Access-Control Filters — Signals are filtered against access controls and consent settings. Only signals the mining process is authorized to use enter the analysis.
- Extract Entity Mentions — Each signal is parsed for entity mentions (people, projects, documents, products). Entity recognizers map mentions to canonical IDs in the enterprise graph.
- Infer Relationships From Co-Occurrence — When two entities co-occur in signals (same query, same document, same edit), the system infers a relationship. The relationship type is classified by the signal kind.
- Weight By Signal Strength — Strong signals (explicit references, frequent co-occurrence) produce strong relationships. Weak signals contribute proportionally less. Weights aggregate across many signal instances.
- Merge Into Graph — Inferred relationships merge into the enterprise graph, augmenting or refining existing edges. Conflicting inferences resolve by weight.
- Update Continuously — The pipeline runs continuously so the graph reflects current behavior. Relationships that stop being reinforced decay; new relationships appear as behavior evolves.
Behavior Builds The Graph
The patent's load-bearing idea is that user behavior is the richest source of true organizational structure. The graph is not declared; it is observed.
Live Structure, Not Frozen Schema
Declared schemas freeze structure at the moment of declaration. Behavior-mined graphs track live structure as it evolves. The shift removes the maintenance burden of keeping schemas current.
- Signal Aggregation — Many small behavior signals aggregate into strong relationship evidence. No single signal needs to be definitive; the aggregate is.
- Privacy-Aware Mining — Access controls and consent settings shape what signals the pipeline reads. Privacy is baked into the design, not added afterward.
- Continuous Update — The graph never stops evolving. New behaviors create new relationships; abandoned workflows drop relationships. The graph stays current.
Technical Foundation
Technical Foundation
The patent specifies the signal taxonomy, the access-control layer, the entity recognition pipeline, the relationship-inference rules, and the graph-update protocol.
- Signal Taxonomy — Each enterprise tool emits signals categorized by type: search query, document access, edit event, email mention. The taxonomy informs how each signal contributes to graph inference.
- Access Control Layer — Per-user, per-document access controls gate which signals enter the pipeline. The layer ensures mining respects the underlying permission model.
- Entity Recognition — Recognizers map signal content to enterprise entity IDs. Recognizers are tuned to enterprise vocabulary, including project names, product codes, and team names.
- Relationship Inference Rules — Rules map signal patterns to relationship types. Co-edit suggests collaboration; co-occurrence in a query suggests semantic relatedness. The rule set is extensible.
- Weighted Aggregation Store — Inferred relationships accumulate weights as signals reinforce them. The store supports rapid weight updates and decay over time.
- Graph Update Protocol — Periodic batch jobs merge accumulated inferences into the enterprise graph. Updates are atomic and versioned for rollback.
The Process
The Process
The pipeline runs as a continuous stream from enterprise tool signals to graph updates. Latency from signal to graph reflection is minutes to hours.
- Tool Emits Signal — An enterprise tool produces a usage event. Events stream to the mining pipeline through a standardized event bus.
- Access Control Check — The pipeline filters signals against access policies. Authorized signals proceed; others are dropped.
- Entity Extraction — The signal content is parsed for entity mentions. Mentions resolve to canonical IDs.
- Relationship Inference — Rules generate candidate relationships from the entity set and signal type. Each candidate carries a weight reflecting signal strength.
- Aggregate Weights — Weights for the same (entity, entity, type) triple aggregate over time. Persistent reinforcement strengthens the relationship; isolated single signals contribute minimally.
- Merge To Graph — Periodic merge jobs incorporate accumulated weights into the live graph. Strong relationships become graph edges; weak ones remain in the candidate pool.
- Decay Over Time — Relationship weights decay slowly when no longer reinforced. Stale relationships eventually drop from the active graph.
Quality Control
Quality Control
Behavior-mined graphs risk encoding incorrect patterns or violating privacy. The patent specifies safeguards.
- Privacy Boundary Enforcement — Access controls are enforced at signal ingestion. Sensitive signals cannot leak across user or team boundaries even if they would produce stronger inferences.
- Confidence Thresholds For Merge — Only relationships with weight above a threshold merge into the live graph. Weak inferences stay in the candidate pool until reinforced.
- Anomaly Detection — Unusual signal patterns (bot traffic, scripted access) are detected and excluded so they cannot poison the graph.
- Decay Calibration — Decay rates balance freshness against stability. Too fast and the graph thrashes; too slow and stale relationships linger.
- User Override Channel — Users can mark inferred relationships as wrong. Overrides feed back into the mining pipeline to refine inference rules over time.
Real-World Application
User-mined enterprise graphs power features in Google Workspace (smart suggestions, related documents, expertise discovery) and parallel surfaces in Drive, Docs, and Gmail. The same primitives generalize to other enterprise knowledge platforms.
- Continuous Mining Cadence — Signals stream continuously and graph updates happen in near-real-time. The graph reflects current organizational behavior, not last quarter's.
- Access-aware Privacy Model — Access controls are enforced throughout the pipeline. Mining respects the underlying permission model and cannot leak data across boundaries.
- Self-correcting Decay Mechanism — Relationships decay when not reinforced. Outdated patterns drop out automatically as workflows evolve.
Why Search Behavior Mirrors Public Web
The same user-mined-graph primitives apply to public web search at much larger scale. Aggregate user behavior across the web reveals entity relationships even when no source explicitly declares them.
Why Co-Occurrence Patterns Matter
On the public web, when two entities co-occur frequently in queries or documents, the engine infers a relationship between them. Sites that explicitly bridge entity pairs the user-mined graph already considers related earn the most visibility for those entity-pair queries.
<\/section>What This Means for SEO
What This Means for SEO
The mechanic of mining a knowledge graph from user behavior, rather than declaring it manually, applies to public search too.
- Search Users Build Your Knowledge Graph — When users connect two entities by searching for them together, the system learns that connection. Build content that explicitly bridges the entity pairs your users actually combine in queries.
- Behavioral Graphs Drift Faster Than Static Ones — A graph built from current user behavior captures emerging entities before declared schemas do. Watch your search queries for new entity names appearing alongside known ones.
- Co-Occurrence In Queries Matters — Two entities co-mentioned in the same query create a relationship in the user-mined graph. Optimize for the co-occurrence pattern, not just the individual entity.