Identity Stitching

Links records using hard identifiers that uniquely and absolutely verify a user.

• Login credentials: User ID, email address, phone number.
• Account linking: OAuth-based cross-platform connections.
• Precision: Near-zero false positives, but limited to authenticated sessions.
• Example: A user logs into a mobile app and later on a desktop browser; the system merges both sessions instantly via the common User ID.

Uses statistical models to link records based on soft identifiers and behavioral patterns.

• Signals: IP address, device fingerprint, browser version, geolocation, Wi-Fi SSID.
• Fuzzy logic: Weights attributes by uniqueness and stability.
• Confidence scoring: Matches above a threshold (e.g., 95%) are automatically merged; lower scores queue for review.
• Example: A user browses anonymously on a work laptop, then later on a home tablet; the system infers a match via shared IP subnet and device proximity patterns.

Manages the half-life of identity signals to prevent stale data from corrupting profiles.

• Time-based rules: A cookie from 2 years ago carries less weight than a login from 2 minutes ago.
• Session boundaries: Inactivity timeouts (e.g., 30 minutes) define distinct sessions.
• Churn detection: Identifiers that haven't been seen in a defined window are deprecated.
• Example: A device ID unused for 18 months is automatically unlinked from the golden record to prevent merging a new device owner into the old profile.

Replaces direct personally identifiable information (PII) with non-identifying tokens while preserving linkability.

• Hashing: One-way cryptographic hashes (SHA-256) of email/phone before matching.
• Tokenization: Vendor-specific persistent IDs replace raw PII in downstream systems.
• Consent management: Stitching rules respect opt-out signals; if a user revokes consent, their graph is partitioned.
• Example: An email [email protected] is hashed to b4c9a... before any cross-channel matching occurs, ensuring raw PII never enters the analytics pipeline.

Models identities as a connected graph where nodes are identifiers and edges are match events.

• Transitive closure: If ID-A matches ID-B, and ID-B matches ID-C, then ID-A and ID-C belong to the same entity.
• Graph traversal: Algorithms like connected components find all identifiers belonging to one user.
• Edge pruning: Weak or old edges are removed to prevent graph collapse (everyone connected to everyone).
• Example: A household with shared devices is correctly split into individual profiles by analyzing disjoint subgraphs of personal email logins.

Balances latency requirements with computational complexity.

• Real-time (streaming): Matches incoming events against a cache of known identifiers within milliseconds for personalization use cases.
• Batch (offline): Runs complex graph algorithms and transitive closure over massive datasets (billions of events) for analytics and golden record generation.
• Lambda architecture: Combines a speed layer for immediate linking and a batch layer for nightly reconciliation.
• Example: A streaming pipeline stitches a new session for ad targeting instantly, while a nightly batch job corrects any missed merges using full historical data.

The computational process of identifying, linking, and merging disparate records that refer to the same real-world entity. This is the umbrella discipline that encompasses identity stitching.

• Deterministic Matching: Exact field comparisons (e.g., SSN, account number)
• Probabilistic Matching: Statistical likelihood scores based on weighted attribute similarity
• Transitive Closure: Inferring that if A matches B and B matches C, then A matches C

The single, best-curated version of a data entity that serves as the authoritative, canonical source of truth. Identity stitching's end goal is to produce a golden record for each individual.

• Survivorship Rules: Logic that selects the best value from conflicting sources (e.g., most recent email, most frequently used phone)
• Merge Purge: The deduplication operation that collapses matched records into the golden record
• Source Priority: Weighting trusted systems (CRM) over less reliable ones (web forms)

A technique for identifying non-identical but probabilistically similar strings, critical for linking records with typographical errors, abbreviations, or transliterations.

• Levenshtein Distance: Measures minimum single-character edits (insert, delete, substitute) between two strings
• Jaro-Winkler: Optimized for name matching, giving higher scores to strings with matching prefixes
• Phonetic Algorithms: Soundex and Metaphone encode words by pronunciation to catch spelling variants

A rule-based approach that links records only when specific identifiers match exactly. This method offers zero false positives but may miss valid matches due to data entry errors.

• Match Keys: Composite fields like 'First Name + Last Name + Date of Birth' used as a unique join key
• Exact Matching: Binary true/false comparison with no tolerance for variance
• Use Case: Financial compliance where a false match could trigger a Suspicious Activity Report

A statistical approach that calculates the likelihood that two records refer to the same entity based on weighted attribute agreement and disagreement.

• Fellegi-Sunter Model: The foundational mathematical framework for probabilistic record linkage
• M-Probability: The probability that a field matches given the records belong to the same entity
• U-Probability: The probability of a random match coincidence
• Match Score: A composite weight threshold above which records are automatically linked

A specialized data structure that stores all known identifiers and their relationships to a unified profile. Unlike a simple golden record, the graph preserves the full linkage history.

• Nodes: Individual identifiers (email, cookie, device ID, loyalty card)
• Edges: Weighted connections representing the confidence of each link
• Temporal Decay: Edges weaken over time if identifiers become inactive
• Real-Time Resolution: Graph traversal at query time to resolve identity on the fly