UK address data rarely arrives in a single, consistent format. This practical record-linkage workflow shows how to clean address fields, compare likely candidates, set confidence bands, and route uncertain pairs to review without treating every similar-looking address as a safe match.
Overview
Fuzzy matching for addresses is the process of deciding whether two records refer to the same property when their text differs. One system may store “Flat 3, 18 High St”, another “3/18 High Street”, and a third may contain only “18 High St” with a postcode. The task is not simply to find the closest strings. It is an entity-resolution problem: combine several fields, account for their reliability, and preserve enough evidence to explain the decision.
A useful workflow has five stages:
- Profile the source data. Measure missing fields, common abbreviations, unusual characters, postcode errors, and how addresses are split across columns.
- Normalise without destroying meaning. Standardise case, whitespace, punctuation, and common terms while retaining flat, unit, building, and house-number information.
- Generate candidate pairs. Use blocking so each record is compared with plausible matches rather than every address in the database.
- Score fields separately. Compare postcodes, numbers, units, street names, localities, and towns using rules suited to each field.
- Apply confidence bands. Automatically accept clear matches, reject clear non-matches, and send ambiguous pairs to a human review queue.
This separation matters. A single overall text-similarity score can hide a serious conflict, such as an identical street name but a different flat number. For wider entity-resolution principles, see the repeatable customer-record workflow.
Checklist by scenario
1. CRM import or customer migration
- Map every source column to a standard schema: address lines, unit or flat, house number, street, locality, town, region, and postcode.
- Keep the original values alongside cleaned values. Originals are essential when a reviewer needs to understand a match.
- Normalise case and whitespace, remove harmless punctuation, and standardise selected abbreviations such as “rd” and “road” only where your data supports that rule.
- Compare postcode outward and inward parts separately if the source often contains spacing or partial-postcode errors. Treat a postcode as strong evidence, not an unconditional identifier.
- Give unit and house-number agreement high importance. “Flat 2” and “Flat 3” should not be treated as a minor spelling difference.
2. Billing or payment records
- Decide whether the objective is matching the bill-to address, service address, correspondence address, or account holder. These may legitimately differ.
- Require stronger evidence before merging records that could affect invoices, refunds, credit decisions, or service eligibility.
- Use exact or near-exact number and postcode checks as gates where appropriate, then use fuzzy comparison for street and locality text.
- Do not infer that two households are the same customer merely because they share a building or postcode.
3. Property, maintenance, or field-service data
- Preserve property identifiers, building names, plot numbers, and development names; they may be more useful than a town field.
- Handle new-build and rural addresses separately. Missing locality fields, named properties, and non-standard delivery points can make generic rules unreliable.
- Include operational aliases only when they are verified. A site nickname should not silently replace the official address.
- Block on a combination such as postcode plus house number, or street plus town, then score the remaining fields.
4. Duplicate-address detection
- Define what “duplicate” means before matching. Identical properties may have multiple legitimate occupants, accounts, or service relationships.
- Use address matching to identify possible duplicate properties, not automatically to merge business records.
- Return the candidate pair, field-level scores, and reasons for the decision so an operator can validate the result.
What to double-check
Build a deliberate normalisation checklist
Typical transformations include trimming whitespace, converting text to a consistent case, standardising Unicode characters, and removing punctuation that carries no matching value. Tokenisation can make “18 High Street” and “High Street, 18” easier to compare. However, do not remove all numbers, unit labels, or directional terms. In address matching, those tokens often distinguish neighbouring properties.
Create explicit rules for common variants rather than applying unlimited fuzzy correction. Examples include “st” versus “street”, “apt” versus “flat”, and inconsistent postcode spacing. Keep a log of each transformation and test it against known examples. If your data contains Welsh, Scottish Gaelic, Irish, or other multilingual forms, review transliteration and language-specific normalisation separately; the guidance in multilingual fuzzy matching is relevant to that design decision.
Choose field weights and gates
There is no universal weighting template, but a practical starting point is:
- Postcode: high weight when present and plausibly formatted, with penalties for conflicting values.
- House number or property number: high weight and often a required agreement for automatic acceptance.
- Flat, apartment, or unit: high weight when present; missing is not the same as matching.
- Street: medium to high weight, using token-aware similarity rather than raw character distance alone.
- Town, locality, and region: supporting evidence, especially where street names repeat.
- Address lines: useful as a combined view, but risky as the only comparison field.
Use a scoring model that can express conflicts. For example, a close street match plus a different flat number should fall into review or rejection even if the combined similarity is high. Levenshtein distance, Jaro-Winkler, and token-based measures each behave differently; compare their strengths before selecting one, as explained in this algorithm comparison.
Block before you compare
Pairwise comparison across two large address tables is expensive and creates many false candidates. Blocking narrows the search. Options include exact postcode, postcode district, house-number prefix, street token, or a combination such as town plus street initial. Use more than one blocking path when data is incomplete, then deduplicate candidate pairs before scoring. Blocking should improve efficiency without becoming an invisible rejection rule: if a badly entered postcode prevents a genuine candidate from entering the block, the match can never be recovered.
Set confidence bands
Use labelled examples where possible. Mark pairs as confirmed match, confirmed non-match, or uncertain, then inspect the score distribution. A sensible operating model has three bands:
- Auto-match: strong agreement on critical fields and no material conflicts.
- Manual review: partial or conflicting evidence, missing unit data, or a score near the decision boundary.
- Do not match: clear disagreement on critical identifiers or insufficient supporting evidence.
Do not choose thresholds solely to maximise the number of matches. Measure false merges as well as missed matches, and keep a sample of decisions for later audit.
Common mistakes
- Fuzzing the whole address as one string. This can let a long matching street name outweigh a conflicting flat or house number.
- Deleting meaningful tokens. Removing all punctuation is often harmless; removing numbers, unit markers, or directional words is not.
- Treating missing as equal. Two blank flat fields do not prove that two records describe the same unit.
- Using the postcode as an absolute key. Postcodes can be missing, mistyped, shared across multiple properties, or represented inconsistently.
- Skipping candidate review. A threshold created on one import may behave differently after a new source system, cleansing rule, or address format is introduced.
- Overwriting source data. Store raw values, normalised values, scores, rules, and reviewer outcomes separately.
- Measuring only match rate. A high match rate is not useful if it creates costly false merges. Review precision, recall, coverage, and the volume of ambiguous cases.
When a result looks wrong, inspect the full decision path rather than only the final score. Search-log investigation techniques can help identify whether the problem comes from normalisation, blocking, ranking, or threshold selection; see how to diagnose fuzzy-search failures.
When to revisit
Revisit the workflow before each major CRM import, billing migration, property-data refresh, or duplicate-address audit. Also review it when a source system changes its column structure, begins exporting different abbreviations, adds or removes postcode fields, or introduces new property types.
Keep a small, protected regression set containing difficult examples: missing postcodes, reordered address lines, flats in the same building, named properties, rural addresses, spelling variations, and deliberately similar non-matches. Run it whenever you change normalisation, blocking, library versions, field weights, or thresholds. Record acceptance, rejection, and review outcomes so the rules can improve from evidence rather than intuition.
Before the next production run, use this final checklist:
- Confirm the matching objective and what a “same address” decision permits.
- Profile the latest input fields and compare them with the previous batch.
- Test normalisation on real edge cases, retaining raw values.
- Verify blocking coverage and inspect candidates lost through incomplete fields.
- Check critical-field conflicts before applying the combined score.
- Review confidence bands against labelled examples.
- Prepare a human-review queue with side-by-side fields and decision reasons.
- Monitor false merges, missed matches, review volume, and unusual changes after release.
Address record linkage works best as a controlled data-quality process, not a one-off fuzzy matching query. Start conservatively, make decisions explainable, and update the workflow whenever the underlying address data changes.