Find (fuzzy) matches between two CSV files in the terminal.
There's a tutorial on how CSV Match can be used for investigative journalism here.
pip install csvmatch
Get an error saying 'command not found'? Sometimes
pip has a different name -- try typing
pip3 instead. If you get an error saying 'permission denied', try prepending
Say you have one CSV file such as:
name,location,codename George Smiley,London,Beggerman Percy Alleline,London,Tinker Roy Bland,London,Soldier Toby Esterhase,Vienna,Poorman Peter Guillam,Brixton,none Bill Haydon,London,Tailor Oliver Lacon,London,none Jim Prideaux,Slovakia,none Connie Sachs,Oxford,none
And another such as:
Person Name,Location Maria Andreyevna Ostrakova,Russia Otto Leipzig,Estonia George SMILEY,London Peter Guillam,Brixton Konny Saks,Oxford Saul Enderby,London Sam Collins,Vietnam Tony Esterhase,Vienna Claus Kretzschmar,Hamburg
You can then find which names are in both files:
$ csvmatch data1.csv data2.csv \ --fields1 name \ --fields2 'Person Name'
You can also compare multiple columns, so if we wanted to find which name and location combinations are in both files we could:
$ csvmatch data1.csv data2.csv \ --fields1 name location \ --fields2 'Person Name' Location
By default, all columns are used to compare rows. Specific columns can be also be given to be compared -- these should be in the same order for both files. Column headers with a space should be enclosed in quotes. Matches are case-sensitive by default, but case can be ignored with
Other things can also be ignored. We can ignore non-alphanumeric characters (
-a), characters from non-latin alphabets (
-n), the order words are given in (
-s), and the order letters are given in (
-e), and common English name prefixes such as Mr and Ms (
-t). Terms specific to your data can be ignored by passing a text file containing a regular expression on each line (
By default the columns used in the output are the same ones used for matching. Other sets of columns can be specified using the
--output parameter. This takes a space-separated list of column names, each prefixed with a number and a dot indicating which file that field is from:
$ csvmatch data1.csv data2.csv \ --fields1 name location \ --fields2 'Person Name' Location \ --output 1.name '2.Person Name' 2.Location \ > results.csv
There are also some special column definitions.
2* expand into all columns from that file. Where a fuzzy matching algorithm has been used
degree will add a column with a number between 0 - 1 indicating the strength of each match.
By default the two files are linked using an inner join -- only successful matches are returned. However using
-j you can specify a
left-outer join which will return everything from the first file, whether there was a match or not. You can also specify
right-outer to do the same but for the second file, and
full-outer to return everything from both files.
We can combine some of the above options to perform operations alike Excel's
VLOOKUP. So if we wanted to add a column to
data2.csv giving the codename of each person that is specified in
$ csvmatch data1.csv data2.csv \ --fields1 name \ --fields2 'Person Name' \ --join right-outer \ --output 2* 1.codename \ > results.csv
CSV Match also supports fuzzy matching. This can be combined with any of the above options.
The default fuzzy mode makes use of the Dedupe library built by Forest Gregg and Derek Eder based on the work of Mikhail Bilenko. This algorithm asks you to give a number of examples of records from each dataset that are the same -- this information is extrapolated to link the rest of the dataset.
$ csvmatch data1.csv data2.csv --fuzzy
The more examples you give it, the better the results will be. At minimum, you should try to provide 10 positive matches and 10 negative matches.
Damerau-Levenshtein is a string distance metric which counts the number of changes that would have to be made to transform one string into another.
For two strings to be considered a match, we require 60% of the longer string to be the same as the shorter one. This threshold can be modified by passing a number between 0.0 and 1.0 with
$ csvmatch data1.csv data2.csv --fuzzy levenshtein name,Person Name George Smiley,George SMILEY Toby Esterhase,Tony Esterhase Peter Guillam,Peter Guillam
Here this matches Toby Esterhase and Tony Esterhase -- Levenshtein is good at picking up typos and other small differences in spelling.
Jaro-Winkler is a string distance metric which counts the number of transpositions that would be required to transform one string into another. It tends to work better than Levenshtein for shorter strings of text.
$ csvmatch data1.csv data2.csv --fuzzy jaro name,Person Name George Smiley,George SMILEY Percy Alléline,Peter Guillam Percy Alléline,Sam Collins Toby Esterhase,Tony Esterhase Peter Guíllam,Peter Guillam Connie Sachs,Konny Saks
Here we can see a couple of incorrect matches for Percy Alléline, but Connie Sachs has matched.
Double Metaphone is a phonetic matching algorithm, which compares strings based on how they are pronounced:
$ csvmatch data1.csv data2.csv --fuzzy metaphone name,Person Name George Smiley,George SMILEY Peter Guillam,Peter Guillam Connie Sachs,Konny Saks
This shows a match for Connie Sachs and Konny Saks, despite their very different spellings.
No module named 'numpy'
If you're on a Mac, this could mean you need to install the Xcode command line developer tools. These can be installed by:
$ xcode-select --install
Then click install on the prompt that appears. After it's finished, try installing CSV Match again.
Broken toolchain: cannot link a simple C program
If you're on a Mac, this could mean you need to accept the Xcode licence. To do this:
$ sudo xcodebuild -license accept
You'll be asked for your password. After it's finished, try installing CSV Match again.
Both with exact matches and fuzzy matching a name being the same is no guarantee it refers to the same person. But the inverse is also true -- even with CSV Match, a combination of first inital and last name is likely to be sufficiently different from forename, middle names, and surname together that a match is unlikely. Moreso if one name includes a typo, either accidential or deliberate.