Quick answer
Identity resolution algorithms fall into four main families: exact matching (deterministic), fuzzy matching (handles typos and formatting variations), graph-based (builds identifier networks), and machine learning (uses trained models to predict matches). Real systems combine algorithms from multiple families. Exact matching handles high-confidence anchors. Fuzzy and graph-based algorithms extend coverage. Machine learning handles the hardest edge cases where explicit rules break down.
Family 1: Exact matching (deterministic)
Compares identifiers character-by-character. If two records have the same email address, they belong to the same person. Simple, fast, accurate, and requires clean data.
Limitations: fails on any variation. “[email protected]” doesn’t match “[email protected]” without pre-normalization. “555-123-4567” doesn’t match “5551234567” without formatting logic.
Family 2: Fuzzy matching
Handles variations that would break exact matching. Case insensitivity, typo tolerance, phonetic similarity (Soundex, Metaphone), transliteration for international names.
Widely used for name matching where spelling variations are common. Also for address matching where formatting varies. More computationally expensive than exact matching but essential in real-world data.
Family 3: Graph-based
Builds a network of identifiers connected by shared data. If email A and phone B both appear on the same purchase, they get linked. If phone B and device ID C both appear together elsewhere, C gets linked to A transitively.
The output is an identity graph where clusters of connected identifiers represent single people. Well-suited to high-volume, multi-channel data where relationships between identifiers are the key signal.
Family 4: Machine learning
Trained models that predict whether two records belong to the same person based on multiple features. Combines behavioral, temporal, geographic, and device signals into a probability score.
Handles cases where explicit rules break down. Also handles the case of new identifier types where rule-based systems would need custom coding.
How systems combine them
A production identity resolution system typically layers algorithms:
- Exact matching first for high-confidence anchors
- Fuzzy matching next for name and address variations
- Graph traversal to extend the network of connections
- Machine learning to handle the remaining edge cases
Each layer processes the output of the previous layer, progressively expanding coverage while maintaining accuracy.
Common follow-up questions
Which algorithm family is most accurate?
Exact matching. But accuracy alone isn’t the goal. Coverage matters too. Real systems combine algorithms to balance both.
Do I need to understand the algorithms to use identity resolution?
Not deeply. But knowing which algorithms your vendor uses helps evaluate their claims about accuracy and coverage.
Are ML-based algorithms always better than rule-based?
Not always. Rule-based algorithms are more auditable and predictable. ML-based algorithms have higher ceilings but harder-to-explain failures.