Customizable probabilistic record linkage with Name Match
summary
Probabilistic record linkage tool Name Match trains a supervised model using records that include first name last name and DOB plus some records with a ground truth unique ID (for example SSN) to create predicted match probabilities, cluster connected records into person identifiers, and emit precision/recall metrics. It speeds work with blocking to avoid trillions of pairs, supports custom constraints in Python to veto implausible links, runs incrementally, and reports how many pairs and false positives the configuration produces.
AI-written summary · about
chapters (14)
- 0:00Customizable probabilistic linkage
- 1:49Name Match tool background
- 8:55Traditional linkage challenges
- 16:40Name Match solutions
- 21:16Name Match requirements
- 27:01Configuring Name Match
- 34:02Interpreting results
- 36:29Why linkage quality matters
- 37:58Q&A: Stability and ordering
- 39:47Q&A: Business names adaptation
- 40:49Q&A: Ground truth without SSN
- 42:13Q&A: Name changes and fairness
- 42:58Q&A: Common names and bias
- 44:44Closing remarks
related talks
39mRecord linkage: Join for real life
41mSimilar entity detection in large data
34mPropensity Score Matching: A Non-experimental Approach to Causal...
29m1 + 1 = 1 or Record Deduplication with Python
38mKeynote: Using the Python Data Science Stack to Determine
27mRapid deduplication and fuzzy matching of large datasets using Splink
22mPredict the future (or don't) with machine learning in applications
21mUsing Semarchy to Master Data Management