Classifiers

Classifiers determine whether a comparison vector represents a match, non-match, or possible match. They operate on the similarity scores produced by field comparators.

from reclink.pipeline import ReclinkPipeline

pipeline = (
    ReclinkPipeline.builder()
    .block_exact("last_name")
    .compare_string("first_name", metric="jaro_winkler")
    .classify_threshold(0.85)
    .build()
)

---

Available Classifiers

Classifier	Type	Parameters	Best for
Threshold	Rule-based	threshold	Simple cutoff
Threshold Bands	Rule-based	upper, lower	Triage (match/possible/non-match)
Weighted Sum	Rule-based	weights, threshold	Per-field importance
Weighted Sum Bands	Rule-based	weights, upper, lower	Weighted triage
Fellegi-Sunter	Probabilistic	m_probs, u_probs	Statistical record linkage
EM	Probabilistic	max_iterations	Unsupervised F-S estimation
Logistic Regression	ML	training data	Supervised matching
Decision Tree	ML	training data	Interpretable ML

---

Threshold Classifier

Classifies pairs based on whether the average comparison score exceeds a threshold.

pipeline = ReclinkPipeline.builder().classify_threshold(0.85)

Pairs with avg_score >= 0.85 are classified as Match, otherwise NonMatch.

Threshold Bands Classifier

Three-band classification: Match, Possible, or NonMatch.

pipeline = ReclinkPipeline.builder().classify_threshold_bands(upper=0.85, lower=0.6)

• avg_score >= 0.85 → Match
• 0.6 <= avg_score < 0.85 → Possible (needs manual review)
• avg_score < 0.6 → NonMatch

Weighted Sum Classifier

Assigns different weights to different comparison fields.

pipeline = ReclinkPipeline.builder().classify_weighted(
    weights=[0.4, 0.3, 0.3],
    threshold=0.8
)

Fellegi-Sunter Classifier

Probabilistic classifier based on the Fellegi-Sunter model of record linkage.

pipeline = ReclinkPipeline.builder().classify_fellegi_sunter(
    m_probs=[0.95, 0.9],
    u_probs=[0.1, 0.05]
)

EM Classifier

Estimates Fellegi-Sunter parameters automatically using Expectation-Maximization.

pipeline = ReclinkPipeline.builder().classify_em(max_iterations=100)

---

ML Classifiers

Logistic Regression

Learns per-field weights from labeled training data using gradient descent with L2 regularization.

from reclink import train_logistic_regression

# Training data: comparison vectors + labels
vectors = [[0.9, 0.95], [0.85, 0.9], [0.1, 0.15], [0.2, 0.1]]
labels = [True, True, False, False]

weights, bias, threshold = train_logistic_regression(
    vectors, labels,
    learning_rate=0.1,
    max_iterations=1000,
    regularization=0.01,
)
# Use weights/bias/threshold in pipeline configuration

Decision Tree

CART-style binary decision tree that splits on comparison fields to maximize Gini impurity reduction.

from reclink import train_decision_tree

tree_json = train_decision_tree(
    vectors, labels,
    max_depth=5,
    min_samples_leaf=5,
)
# Returns JSON representation of the trained tree

---

Threshold Optimization

Automatically find the optimal threshold for any classifier:

from reclink import optimize_threshold

scores = [0.9, 0.8, 0.7, 0.3, 0.2, 0.1]
labels = [True, True, True, False, False, False]

result = optimize_threshold(scores, labels, criterion="f1")
print(f"Optimal threshold: {result.threshold}")
print(f"F1: {result.f1}, Precision: {result.precision}, Recall: {result.recall}")

Supported criteria: "f1", "precision", "recall".