Hey there! Great question—setting Levenshtein distance thresholds for name matching is super context-dependent, but there are some tried-and-true rules of thumb based on real-world use cases. Let’s break this down specifically for your scenario of matching concatenated full names (t1.first_name||last_name vs t2.first_name||last_name):

General Threshold Guidelines by Name Length #

• Short full names (total length ≤ 10 characters, e.g., "AnnSmith" or "BobJones"):
  Stick to a threshold of 0 or 1. A distance of 1 covers common tiny typos like a missing letter ("AnnSmit" vs "AnnSmith") or a single character transposition ("JonhDoe" vs "JohnDoe"). Going higher here risks false positives with unrelated names.
• Medium-length names (11–20 characters, e.g., "ElizabethTaylor" or "MichaelJackson"):
  A range of 0–2 works well. This accounts for misspellings of less common name parts ("Elisabeth" vs "Elizabeth"), extra letters, or swapped character pairs that don’t change the core identity of the name.
• Longer names (>20 characters, e.g., "AlexandrosPapadopoulos"):
  You can safely stretch to 0–3. Longer names have more room for small errors—like a misspelled suffix or an extra vowel—without being a completely different name.

Hard Maximum Threshold #

As a general hard stop, anything above 4 is almost never a valid match for concatenated full names. Even for long names, a distance of 4 usually means the names are fundamentally distinct (e.g., "RobertJohnson" vs "RichardThompson").

Pro Tips for Better Accuracy #

• Normalize by name length: Instead of fixed numbers, use a normalized threshold (Levenshtein distance divided by the length of the longer name). A value of ≤ 0.15 tends to strike a great balance—this auto-adjusts for name length (1 for 10-character names, 3 for 20-character names).
• Combine with other logic: If you’re dealing with nicknames or cultural name variations (e.g., "Mike" vs "Michael", "Sofia" vs "Sophia"), don’t rely solely on Levenshtein distance. Pair it with a nickname lookup table to catch those edge cases.
• Test with your data: Grab a sample of known matches and non-matches from your database, tweak the threshold, and measure precision (fewer false positives) and recall (fewer missed real matches) to find the sweet spot for your specific dataset.

内容的提问来源于stack exchange，提问作者user8834780