preclink

High-precision record linkage library implementing a 7-step pipeline with multi-pass support.

Installation

pip install preclink

Quick Start

import pandas as pd
from preclink import Pipeline, StringComparison, ExactComparison

df_left = pd.DataFrame({
    "id": [1, 2, 3],
    "first_name": ["John", "Jane", "Bob"],
    "last_name": ["Smith", "Doe", "Johnson"],
    "state": ["CA", "NY", "CA"],
})

df_right = pd.DataFrame({
    "id": [101, 102, 103],
    "first_name": ["Jon", "Jane", "Robert"],
    "last_name": ["Smith", "Doe", "Johnson"],
    "state": ["CA", "NY", "CA"],
})

result = (
    Pipeline()
    .preprocess(normalize_unicode=True, lowercase=True)
    .block(on="state")
    .score(comparisons=[
        StringComparison("first_name", algorithm="jaro_winkler"),
        StringComparison("last_name", algorithm="jaro_winkler"),
    ])
    .filter(min_score=0.7)
    .decide(method="hungarian")
    .build()
    .link(df_left, df_right)
)

print(result.matches)

The 7-Step Pipeline

1. Preprocess: Normalize text (unicode, case, whitespace)
2. Deduplicate: Remove within-table duplicates
3. Block: Reduce comparison space using blocking keys
4. Score: Compute pairwise similarity scores
5. Filter: Remove low-confidence pairs
6. Decide: Apply matching algorithm (Hungarian, Greedy, Row-Sequential)
7. Inspect: Generate diagnostics and reports

Multi-Pass Matching

For complex datasets, use multi-pass matching with progressively relaxed thresholds:

from preclink import MultiPassOrchestrator, StringComparison

orchestrator = MultiPassOrchestrator()
result = orchestrator.run(
    df_left,
    df_right,
    passes=[
        {"min_score": 0.95, "method": "hungarian"},
        {"min_score": 0.85, "method": "hungarian"},
        {"min_score": 0.70, "method": "greedy"},
    ],
    comparisons=[
        StringComparison("first_name"),
        StringComparison("last_name"),
    ],
)

Decision Rules

• Hungarian: Optimal assignment maximizing total score (recommended for precision)
• Greedy: Best-global-pair first, fast and precision-optimized
• Row-Sequential: Process left records in order (baseline)

Features

• Type-safe with full mypy strict mode support
• Extensible via protocols (custom comparisons, blockers, decision rules)
• Native pandas DataFrames throughout
• Crosswalk support for blocking key normalization
• Margin-based filtering for ambiguity removal

Benchmarks

Comparison against recordlinkage on standard Febrl datasets:

Dataset	Library	Precision	Recall	F1
febrl1	preclink	100.0%	76.4%	86.6%
febrl1	recordlinkage	99.5%	79.0%	88.1%
febrl2	preclink	97.3%	39.5%	56.2%
febrl2	recordlinkage	95.0%	80.0%	86.9%
febrl3	preclink	99.2%	35.4%	52.2%
febrl3	recordlinkage	98.1%	79.6%	87.9%
febrl4	preclink	99.9%	79.0%	88.2%
febrl4	recordlinkage	94.3%	80.8%	87.0%

febrl4 is the true record linkage scenario (linking two separate tables). On this dataset, preclink achieves 99.9% precision with higher F1 than recordlinkage. The other datasets (febrl1-3) are deduplication tasks where records are split artificially.

When to use preclink: When false positives are costly (merging administrative records, survey linking, research applications) and you need provably optimal 1:1 matching.

Reproduce:

pip install recordlinkage
python examples/benchmark_febrl.py

Documentation

Full documentation at finite-sample.github.io/preclink

License

MIT License