Fuzzy Matching & Record Linkage: Normalising Names, Similarity Scores & Duplicates

What you'll build

What you will be able to do

Each lab ends with a concrete result that you build yourself and the checker confirms, cell by cell.

• Clean twelve raw company names in three visible steps (spaces and case, punctuation, legal suffixes) and see how many different customers are really left.
• Build a blocking key for each customer, count the candidate pairs inside each block and measure how many comparisons blocking saves, and what it can miss.
• Compare twelve pairs that are known to be the same customer with three methods (EXACT, a simple =, and a reusable Norm LAMBDA) and measure each method's match rate.
• Score ten name pairs with character bigrams: count the bigrams, count the shared ones with one array formula, and turn them into Jaccard and Dice similarities between 0 and 1.
• Split names into words with TEXTSPLIT, count shared words, and compare word-level Jaccard and the overlap coefficient with the bigram score.
• Combine four pieces of evidence into one weighted match score, sort every candidate pair into Match, Review or Non-match, and colour the pairs that need a human look.
• Score your matcher against 20 hand-labelled pairs: predicted matches at a threshold, true and false positives, precision, recall and F1.
• Turn matched records into clusters with MINIFS, pick the most recently updated record of each cluster as the survivor for the golden record, count the real customers and chart the score distribution.