record linkage using ECM in python

Question

I am working on a record linkage problem and applying unsupervised algorithm since I do not have external labels.

I have applied ECM alogorithm. Code used is:

import recordlinkage

indexer = recordlinkage.BlockIndex(on=['FirstName_CD','LastName_CD'])
pairs = indexer.index(data1, data2)

compare_cl = recordlinkage.Compare()

compare_cl.string('FirstName_CD', 'FirstName_CD', method='jarowinkler', threshold=0.50,label='given_name')
compare_cl.string('LastName_CD', 'LastName_CD', method='jarowinkler', threshold=0.50, label='surname')
compare_cl.exact('Date.Of.Birth_CD', 'Date.Of.Birth_CD', label='date_of_birth')
compare_cl.exact('Gender_CD', 'Gender_CD', label='gender')
compare_cl.exact('Profession_CD', 'Profession_CD', label='profession')
compare_cl.string('Address_CD', 'Address_CD', threshold=0.85, label='address_1')

features = compare_cl.compute(pairs,data1)

ecm = recordlinkage.ECMClassifier()
result_ecm=ecm.learn(features)

Now it returns a multiindex. My question is what inference I can draw from it? How to get the matches/mismatches information?

Answer 1

In order to get ECM classifier work on 'compare vectors' OR 'features' is to fit the model only on the columns which has unique>1.

Here is the python code:

cl = recordlinkage.ECMClassifier()

compare_vectors=compare_vectors[[col for col in compare_vectors.columns if compare_vectors[col].nunique() >= 2]]

cl.fit(compare_vectors)

Answer 2

I couldn't find information for this method either, but as far as I could understand, ecm_learn returns a MultiIndex, which is the same datatype of pairs returned by the indexer.index() method.

So, this is how I used it (and I can be completely wrong, btw!)

import recordlinkage

indexer = recordlinkage.BlockIndex(on=['FirstName_CD','LastName_CD'])
pairs = indexer.index(data1, data2)

compare_cl = recordlinkage.Compare()

compare_cl.string('FirstName_CD', 'FirstName_CD', method='jarowinkler', threshold=0.50,label='given_name')
compare_cl.string('LastName_CD', 'LastName_CD', method='jarowinkler', threshold=0.50, label='surname')
compare_cl.exact('Date.Of.Birth_CD', 'Date.Of.Birth_CD', label='date_of_birth')
compare_cl.exact('Gender_CD', 'Gender_CD', label='gender')
compare_cl.exact('Profession_CD', 'Profession_CD', label='profession')
compare_cl.string('Address_CD', 'Address_CD', threshold=0.85, label='address_1')

features = compare_cl.compute(pairs,data1)

ecm = recordlinkage.ECMClassifier()
result_ecm=ecm.learn(features)

#reprocess the compute() call with the newly adjusted match information
features = compare_cl.compute(result_ecm,data1)

#now, do your stuff..
#...

If someone else has further information, I'd really appreciate some feedback.

Thanks!

Answer 3

The MultiIndex returned by ecm.learn() is basically a bunch of index pairs that are the classifier's guesses for which items are matches. (Note that ecm.learn() is deprecated. The new function name is ecm.fit_predict().)

I don't know what your data looks like, but here's an example with one of the sample data sets:

from recordlinkage.datasets import load_febrl4

dfA, dfB = load_febrl4()

# Indexation step
indexer = recordlinkage.Index()
indexer.block('given_name')
candidate_links = indexer.index(dfA, dfB)

# Comparison step
compare_cl = recordlinkage.Compare()
compare_cl.string('surname', 'surname', method='jarowinkler', threshold=0.85, label='surname')
compare_cl.exact('date_of_birth', 'date_of_birth', label='date_of_birth')
compare_cl.exact('suburb', 'suburb', label='suburb')
compare_cl.exact('state', 'state', label='state')
compare_cl.string('address_1', 'address_1', threshold=0.85, label='address_1')

features = compare_cl.compute(candidate_links, dfA, dfB)

ecm = recordlinkage.ECMClassifier()
matches = ecm.fit_predict(features)

(Note that this example is based off the one in the documentation at here.)

The matches object is indeed a Pandas MultiIndex. We can convert it to a list of tuples to get a better idea of what information it contains.

# Look at the first 5 matches
list(matches)[:5]

[('rec-2371-org', 'rec-2371-dup-0'),
 ('rec-3024-org', 'rec-3024-dup-0'),
 ('rec-4652-org', 'rec-4652-dup-0'),
 ('rec-4795-org', 'rec-4795-dup-0'),
 ('rec-1016-org', 'rec-1016-dup-0')]

These are the names of indices in the dfA and dfB data frames. We can look them up to see that they are in fact matches:

dfA.loc['rec-2371-org'], dfB.loc['rec-2371-dup-0']

 given_name           michaela
 surname              dunstone
 street_number              37
 address_1        deane street
 address_2            rosedown
 suburb              woodcroft
 postcode                 2065
 state                     vic
 date_of_birth        19121018
 soc_sec_id            3166178
 Name: rec-2371-org, dtype: object

 given_name           michaela
 surname              dunstone
 street_number              37
 address_1        deane street
 address_2            rosedlwn
 suburb              woodcroft
 postcode                 2065
 state                     vic
 date_of_birth        19121018
 soc_sec_id            3166178
 Name: rec-2371-dup-0, dtype: object