Resolution Methodology

Mathematical principles, multi-pass blocking algorithms, and scoring weights underlying the RESOMESH matching engine.

1. Multi-Stage Pipeline Architecture

Decoupled data ingestion, normalization, blocking, and classification

RESOMESH enforces a multi-pass signal-ranked entity resolution pipeline (backed by the Amazon ML ER architecture) structured to eliminate quadratic Cartesian product explosion while maintaining high recall across noisy multi-source records:

STAGE 01

Schema Ingestion

Strict validation of tabular headers: entity_id, business_name, business_address, country.

STAGE 02

Canonical Normalization

Unicode whitespace collation, ASCII foldering, lowercasing, punctuation stripping, and legal entity suffix parsing.

STAGE 03

Candidate Generation

Cross-source candidate pairs filtered via rough lexical pruning (similarity cutoff ≥ 0.30) to preserve non-matches for auditing.

STAGE 04

Weighted Classification

Multi-signal similarity tensor combining name, address, and geographic concordance.

2. Mathematical Scoring Function

Weighted linear combination of similarity signals

Every candidate pair (s₁, c) is assigned a composite similarity score in the interval [0.0, 1.0] defined by:

Score(s₁, c) = 0.62 · S_name + 0.28 · S_address + 0.10 · S_country

Signal Feature	Weight	Metric Formulation	Noise Resilience
Business Name (`S_name`)	0.62	Ratcliff-Obershelp longest common subsequence ratio	Tolerates typos, abbreviations, word order, and corporate entity additions (Pvt Ltd, LLC)
Business Address (`S_address`)	0.28	Token sequence overlap ratio	Accounts for street abbreviations (St / Street, Blvd, Pkwy) and suite omission
Geographic Jurisdiction (`S_country`)	0.10	ISO country alias concordance matrix	Standardizes IN/India, US/USA/United States, FR/France aliases

3. Evaluation Metric: F₀.₅ Optimization

Prioritizing precision over recall for enterprise compliance

In commercial entity resolution, false positives (incorrectly merging two distinct legal entities) carry severe compliance, KYC, and financial risk compared to false negatives (reviewing two entities separately). Consequently, RESOMESH optimizes against the precision-biased F₀.₅ score (β = 0.5):

F₀.₅ = (1 + 0.5²) · (Precision · Recall) / (0.5² · Precision + Recall) = (1.25 · P · R) / (0.25 · P + R)

With a benchmark decision threshold of 0.72, our resolver achieves an empirical F₀.₅ = 0.873 (Precision: 0.914, Recall: 0.821) across multi-country enterprise callsets.

4. Strict Output Guarantees

1:1 Entity preservation contract

Total Preservation: Every single record present in Source 1 is guaranteed to appear exactly once in matching_results.tsv, ensuring no entities are silently dropped.
Full Auditability: Unmatched entities retain explicit status unmatched with candidate relationships logged in candidate_pairs.tsv for human review.
Deterministic Reproducibility: Identical input files consistently yield bitwise-reproducible match results without non-deterministic LLM jitter.