Fuzzy 0.4.0 deduplicator findclusters

Lodestar.Fuzzy 0.4.0. This page is frozen at that release. Read the current documentation for what main says now. A link to a decision or a migration page follows main, and leaves the archive.

Deduplicator.FindClusters

Group near-duplicate records, comparing only within blocks.

public static IReadOnlyList<IReadOnlyList<int>> FindClusters<T>(IReadOnlyList<T> records, Func<T, string> blockingKey, Func<T, T, double> similarity, double threshold)

Parameters — records is the dataset. blockingKey maps a record to the key that decides which records it is compared against. similarity scores two records. threshold is the score at which two are the same thing.

Returns — IReadOnlyList<IReadOnlyList<int>>: clusters of indices into records.

Example — two typo pairs, blocked on the first letter.

using Lodestar.Fuzzy;

string[] records = ["apple pie", "appel pie", "banana bread", "banana bred"];

IReadOnlyList<IReadOnlyList<int>> clusters = Deduplicator.FindClusters(
    records, record => record[..1], (a, b) => Fuzz.Ratio(a, b), threshold: 80);

int clusterCount = clusters.Count;  // => 2
int firstSize = clusters[0].Count;  // => 2

Remarks — indices rather than records, so the caller can reach whatever else the row carries — an id, a timestamp, the field that decides which duplicate to keep.

Two records in different blocks are never compared, whatever their similarity. That is the trade the blocking key buys and the way this is misused: a key of record => record[..1] misses every duplicate whose first character differs, typos included. Keys worth considering are a sorted initials string, a phonetic encoding — Soundex or Metaphone — or a truncated postcode.

threshold and similarity travel together: 80 means something different under Fuzz.Ratio than under Fuzz.TokenSetRatio, which scores far more generously. Changing the scorer without revisiting the threshold silently changes what counts as a duplicate.

Applies to — net10.0, netstandard2.0.

See also — Deduplicator, Fuzz.Ratio, the matching index.