String Distance Matrix in Python using pdist

问题

How to calculate Jaro Winkler distance matrix of strings in Python?

I have a large array of hand-entered strings (names and record numbers) and I'm trying to find duplicates in the list, including duplicates that may have slight variations in spelling. A response to a similar question suggested using Scipy's pdist function with a custom distance function. I've tried to implement this solution with the jaro_winkler function in the Levenshtein package. The problem with this is that the jaro_winkler function requires a string input, whereas the pdict function seems to require a 2D array input.

Example:

import numpy as np
from scipy.spatial.distance import pdist
from Levenshtein import jaro_winkler

fname = np.array(['Bob','Carl','Kristen','Calr', 'Doug']).reshape(-1,1)
dm = pdist(fname, jaro_winkler)
dm = squareform(dm)

Expected Output - Something like this:

          Bob  Carl   Kristen  Calr  Doug
Bob       1.0   -        -       -     -
Carl      0.0   1.0      -       -     -
Kristen   0.0   0.46    1.0      -     -
Calr      0.0   0.93    0.46    1.0    -
Doug      0.53  0.0     0.0     0.0   1.0

Actual Error:

jaro_winkler expected two Strings or two Unicodes

I'm assuming this is because the jaro_winkler function is seeing an ndarray instead of a string, and I'm not sure how to convert the function input to a string in the context of the pdist function.

Does anyone have a suggestion to allow this to work? Thanks in advance!

回答1:

You need to wrap the distance function, like I demonstrated in the following example with the Levensthein distance

import numpy as np    
from Levenshtein import distance
from scipy.spatial.distance import pdist, squareform

# my list of strings
strings = ["hello","hallo","choco"]

# prepare 2 dimensional array M x N (M entries (3) with N dimensions (1)) 
transformed_strings = np.array(strings).reshape(-1,1)

# calculate condensed distance matrix by wrapping the Levenshtein distance function
distance_matrix = pdist(transformed_strings,lambda x,y: distance(x[0],y[0]))

# get square matrix
print(squareform(distance_matrix))

Output:
array([[ 0.,  1.,  4.],
       [ 1.,  0.,  4.],
       [ 4.,  4.,  0.]])

回答2:

For anyone with a similar problem - One solution I just found is to extract the relevant code from the pdist function and add a [0] to the jaro_winkler function input to call the string out of the numpy array.

Example:

X = np.asarray(fname, order='c')
s = X.shape
m, n = s
dm = np.zeros((m * (m - 1)) // 2, dtype=np.double)

k = 0
for i in xrange(0, m - 1):
    for j in xrange(i + 1, m):
        dm[k] = jaro_winkler(X[i][0], X[j][0])
        k = k + 1

dms = squareform(dm)

Even though this algorithm works I'd still like to learn if there's a "right" computer-sciency-way to do this with the pdist function. Thanks, and hope this helps someone!

回答3:

Here's a concise solution that requires neither numpy nor scipy:

from Levenshtein import jaro_winkler
data = ['Bob','Carl','Kristen','Calr', 'Doug']
dm = [[ jaro_winkler(a, b) for b in data] for a in data]
print('\n'.join([''.join([f'{item:6.2f}' for item in row]) for row in dm]))

  1.00  0.00  0.00  0.00  0.53
  0.00  1.00  0.46  0.93  0.00
  0.00  0.46  1.00  0.46  0.00
  0.00  0.93  0.46  1.00  0.00
  0.53  0.00  0.00  0.00  1.00

来源：https://stackoverflow.com/questions/46452724/string-distance-matrix-in-python-using-pdist