Jaro Winkler Distance

Finds a non-euclidean distance or similarity between two strings.

Jaro and Jaro-Winkler equations provide a score between two short strings where errors are more prone at the end of the string. Jaro's equation measure is the weighted sum of the percentage of matching and transposed characters from each string. Winkler's factor adds weight in Jaro's formula to increase the calculated measure when there is a sequence of characters (a prefix) in both strings.

This version is based on the original C implementation of strcmp95 implementation but does not attempt to normalize homoglyph (e.g.: O vs. 0).

• Impact of the prefix is limited to 4 characters, as originally defined by Winkler.
• Input strings are not modified beyond whitespace trimming.
• In-word whitespace and characters case will optionally impact score.
• Supports optional UTF-8 normalization and homoglyph sanitization.
• Returns a floating point number rounded to the desired decimals (defaults to 2) using Python's round.
• Consider usual floating point arithmetic characteristics when working with this module.

Implementation

The complexity of this algorithm resides in finding the matching and transposed characters. That is because of the interpretation of what are the matching conditions and the definition of transposed. Definitions of those two will make the score vary between implementations of this algorithm.

Here is how matching and transposed are defined in this module:

• A character of the first string at position N is matching if found at position N or within distance on either side in the second string.
• The distance is calculated using the rounded down length of the longest string divided by two minus one.
• Characters in the first string are matched only once against characters of the second string.
• Two characters are transposed if they previously matched and aren't at the same position in the matching character subset.
• Decimals are rounded according to the scientific method.

Example

Calculate the Jaro Winkler similarity ($sim_{w}$) between PENNSYLVANIA and PENNCISYLVNIA:

$$ s_{1}=\text{PENNSYLVANIA} \qquad\text{and}\qquad s_{2}=\text{PENNCISYLVNIA} $$

    P E N N C I S Y L V N I A
  ┌-─────────────────────────
P │ 1          ╎
E │   1          ╎
N │     1          ╎
N │       1          ╎           Symbols '╎' represent the sliding windows
S │             1      ╎        boundary in the second string where we look
Y │ ╎             1      ╎           for the first string's character.
L │   ╎             1      ╎
V │     ╎             1                   d = 5 in this example.
A │       ╎                 1
N │         ╎           1
I │           ╎           1
A │             ╎

$$ \begin{split} d &= \left\lfloor {\max(12, 13) \over 2} \right\rfloor - 1 \newline &= 5 \newline \end{split} \qquad \text{ and } \qquad \begin{split} |s_{1}| &= 12 \newline |s_{2}| &= 13 \newline \end{split} \qquad \text{ and } \qquad \begin{split} \ell &= 4 \newline m &= 11 \newline t &= 3 \newline p &= 0.1 \newline \end{split} $$

Considering the input parameters calculated above:

$$ \begin{split} sim_{j} &=\begin{cases} 0 & \text{if } m = 0 \newline {1 \over 3} \times \left({m \over |s_{1}|} + {m \over |s_{2}|} + {{m - t} \over m} \right) & \text{otherwise} \end{cases} \newline &={1 \over 3} \times \left({11 \over 12} + {11 \over 13} + {{11 - 3} \over 11}\right) \newline &= 0.83003108003 \newline \end{split} \qquad \text{then} \qquad \begin{split} sim_{w} &= sim_{j} + \ell \times p \times (1 - sim_{j}) \newline &= 0.83003108003 + 4 \times 0.1 \times (1 - 0.83003108003) \newline &= 0.89801864801 \newline \end{split} $$

We found that the $\lceil sim_{w} \rceil$ is $0.9$.

Benchmark

Benchmarking ran on a 2024 Macbook Pro with an M4 Pro chip running macOS 26.2.

Usage

from pyjarowinkler import distance

distance.get_jaro_similarity("PENNSYLVANIA", "PENNCISYLVNIA", decimals=12)
# 0.830031080031
distance.get_jaro_winkler_similarity("PENNSYLVANIA", "PENNCISYLVNIA", decimals=12)
# 0.898018648019
distance.get_jaro_distance("hello", "haloa", decimals=4)
# 0.2667
distance.get_jaro_similarity("hello", "haloa", decimals=2)
# 0.73
distance.get_jaro_winkler_distance("hello", "Haloa", scaling=0.1, norm_case=False)
# 0.4
distance.get_jaro_winkler_distance("hello", "HaLoA", scaling=0.1, norm_case=True)
# 0.24
distance.get_jaro_winkler_similarity("café", "cafe\u0301", norm_utf8=True)
# 1.0
distance.get_jaro_winkler_similarity("pаypal", "paypal", norm_ambiguous=True)
# 1.0
distance.get_jaro_winkler_similarity("hello", "haloa", decimals=2)
# 0.76

Contribute

You need to have installed mise on your system. Then, running the commands below will install python, uv, and github-cli.

Typical order of execution is as follow:

$ cd ./jaro-winkler-distance
$ mise install
$ uv venv
$ source .venv/bin/activate
$ uv pip install '.[dev]'

Other helpful commands:

• uvx --python=3.12 python -m unittest discover -s tests/
• uvx ruff check --diff
• uvx ruff format --diff
• uvx mypy
• uvx coverage run -m unittest discover -s tests/
• uvx coverage report

Release

$ ./release.sh help
Usage: release.sh [help|major|minor|patch]
$ PYPI_REPO=main ./release.sh minor