Identity Graph

The graph must simultaneously support two distinct matching methodologies to achieve both scale and precision.

• Deterministic Matching: Creates edges based on hard, verified matches such as a hashed email key or login credential. This provides absolute certainty but limited reach.
• Probabilistic Matching: Infers connections using statistical models that analyze non-personal signals like IP address, browser type, and behavioral patterns, assigning a confidence score between 0 and 1.
• Hybrid Fusion: A mature graph merges these approaches, using deterministic anchors to calibrate and validate the probabilistic clusters.

At the core of the graph lies the Canonical ID, a single, persistent master key that resolves all disparate identifiers to one unified profile.

• Entity Resolution: The process of deduplicating and merging records from various sources (CRM, web, mobile) into a Golden Record.
• Survivorship Rules: Logic that dictates which attribute value (e.g., email vs. loyalty number) takes precedence when source systems conflict.
• Graph Traversal: The Canonical ID acts as the root node, allowing real-time lookups to instantly retrieve all linked devices, cookies, and offline identifiers.

An identity graph is a living structure that must model the passage of time to prevent profile pollution.

• Identity Decay: A temporal model that progressively reduces the linkage confidence of an identifier as it ages without fresh validation. An inactive cookie from 30 days ago has a lower weight than a login event from 5 minutes ago.
• Session Stitching: The algorithmic process of connecting discrete web sessions interrupted by timeouts or device switches into a single, continuous behavioral journey.
• Recency Windows: Configurable time frames that define how long a probabilistic link remains valid before requiring re-verification.

Modern identity graphs must enforce privacy constraints structurally, not just through policy overlays.

• Differential Privacy: Injecting calibrated statistical noise into aggregate queries to guarantee that the presence or absence of any single individual in the dataset remains indistinguishable.
• k-Anonymity: Ensuring that any released personal information is indistinguishable from at least k-1 other individuals, preventing re-identification by grouping users into sufficiently large cohorts.
• Consent Integration: The graph must natively consume signals from a Consent Management Platform (CMP) to suppress or sever edges where the legal basis for processing has been revoked.

Advanced identity graphs leverage deep learning architectures designed to operate directly on graph-structured data for superior linkage prediction.

• Node Embeddings: GNNs learn dense vector representations of devices and users that capture complex, multi-hop relationships.
• Link Prediction: The model predicts missing edges in the graph, identifying that a new mobile device likely belongs to an existing user profile based on structural similarity rather than just attribute matching.
• Scalability: Unlike traditional Fellegi-Sunter probabilistic models, GNNs can scale to graphs with billions of nodes and trillions of edges without a linear increase in computation time.

A walled-garden graph is useless. The structure must support Cookie Syncing and mapping to external frameworks to activate audiences.

• Unified ID 2.0 (UID2): Mapping internal Canonical IDs to this open-source, hashed-email framework enables targeted advertising without third-party cookies.
• Match Rate Optimization: The graph must constantly monitor the percentage of user records successfully linked between disparate platforms, as this is the critical KPI for activation value.
• Data Clean Room Integration: The graph should be queryable within a secure, neutral environment where multiple parties can combine first-party data without exposing raw, user-level records.

The absolute anchor of an identity graph. This process links records using exact, verified matches of personally identifiable information (PII), such as a hashed email key or a login credential. Unlike probabilistic methods, deterministic matching provides 100% certainty, making it the highest-fidelity input for building a canonical ID. It is the preferred method for stitching authenticated sessions but requires a direct user login event to function.

The statistical engine that fills the gaps when a user is anonymous. This method infers device ownership by analyzing non-personal signals like IP address, browser type, operating system, and behavioral patterns. It assigns a confidence score to each linkage, allowing the graph to include anonymous sessions. Key techniques include household IP matching and device fingerprinting, which are essential for top-of-funnel personalization but require careful identity decay models to prevent stale links.

The secure, neutral environment where identity graphs are enriched without violating privacy. Multiple parties—like a retailer and a CPG brand—can join their first-party data sets for cross-device attribution and resolution. The clean room allows queries to match hashed email keys and analyze overlap, but strictly prevents the export of raw, user-level data. This architecture is the primary solution for post-third-party cookie collaboration.

The temporal logic that prevents an identity graph from becoming a data swamp. Identity decay is a model that progressively reduces the linkage confidence of an identifier—such as a cookie or a device ID—as it ages without fresh validation. A deterministic login might reset the decay clock, while an inactive probabilistic match based on an old IP address will eventually be pruned. This ensures the graph reflects current reality, not a historical snapshot.