csvdedupe

Contents:

Command line tools for using the dedupe python library for deduplicating CSV files.

csvdedupe take a messy input file or STDIN pipe and identify duplicates

csvlink take two CSV files and find matches between them

Read more about csvdedupe on OpenNews Source

Build Status

Installation and dependencies

csvdedupe requires numpy, which can be complicated to install. If you are installing numpy for the first time, follow these instructions. You’ll need to version 1.6 of numpy or higher.

After numpy is set up, then install the following: * fastcluster * hcluster * networkx

git clone git@github.com:datamade/csvdedupe.git
cd csvdedupe
pip install "numpy>=1.6"
pip install -r requirements.txt
python setup.py install

csvdedupe usage

Take a messy input file or STDIN pipe and identify duplicates

Provide an input file and field names bash csvdedupe examples/csv_example_messy_input.csv \           --field_names "Site name" Address Zip Phone \           --output_file output.csv

or

Pipe it, UNIX style bash cat examples/csv_example_messy_input.csv | csvdedupe --skip_training \           --field_names "Site name" Address Zip Phone > output.csv

or

Define everything in a config file bash csvdedupe examples/csv_example_messy_input.csv \             --config_file=config.json

Example config file

{
  "field_names": ["Site name", "Address", "Zip", "Phone"],
  "field_definition" : {"Site name" : {"type" : "String"},
                        "Address"   : {"type" : "String"},
                        "Zip"       : {"type" : "String",
                                       "Has Missing" : true},
                        "Phone"     : {"type" : "String",
                                       "Has Missing" : true}},
  "output_file": "examples/output.csv",
  "skip_training": false,
  "training_file": "training.json",
  "sample_size": 150000,
  "recall_weight": 2
}

Arguments:

Required

  • input a CSV file name or piped CSV file to deduplicate

Either * --config_file Path to configuration file.

Or * --field_names List of column names for dedupe to pay attention to

Optional

  • --output_file OUTPUT_FILE CSV file to store deduplication results (default: None)
  • --destructive Output file will contain unique records only
  • --skip_training Skip labeling examples by user and read training from training_file only (default: False)
  • --training_file TRAINING_FILE Path to a new or existing file consisting of labeled training examples (default: training.json)
  • --sample_size SAMPLE_SIZE Number of random sample pairs to train off of (default: 150000)
  • --recall_weight RECALL_WEIGHT Threshold that will maximize a weighted average of our precision and recall (default: 2)
  • -h, --help show help message and exit

Training

The secret sauce of csvdedupe is human input. In order to figure out the best rules to deduplicate a set of data, you must give it a set of labeled examples to learn from.

The more labeled examples you give it, the better the deduplication results will be. At minimum, you should try to provide 10 positive matches and 10 negative matches.

The results of your training will be saved in a JSON file ( training.json, unless specified otherwise with the --training-file option) for future runs of csvdedupe.

Here’s an example labeling operation:

Phone :  2850617
Address :  3801 s. wabash
Zip :
Site name :  ada s. mckinley st. thomas cdc

Phone :  2850617
Address :  3801 s wabash ave
Zip :
Site name :  ada s. mckinley community services - mckinley - st. thomas

Do these records refer to the same thing?
(y)es / (n)o / (u)nsure / (f)inished

Output

csvdedupe attempts to identify all the rows in the csv that refer to the same thing. Each group of such records are called a cluster. csvdedupe returns your input file with an additional column called Cluster ID, that either is the numeric id (zero-indexed) of a cluster of grouped records or an x if csvdedupe believes the record doesn’t belong to any cluster.

csvlink operates in much the same way as csvdedupe, but will flatten both CSVs in to one output file similar to a SQL OUTER JOIN statement. You can use the --inner_join flag to exclude rows that don’t match across the two input files, much like an INNER JOIN.

Preprocessing

csvdedupe attempts to convert all strings to ASCII, ignores case, new lines, and padding whitespace. This is all probably uncontroversial except the conversion to ASCII. Basically, we had to choose between two ways of handling extended characters.

distance("Tomas", "Tomás')  = distance("Tomas", "Tomas")

or

distance("Tomas, "Tomás") = distance("Tomas", "Tomzs")

We chose the first option. While it is possible to do something more sophisticated, this option seems to work pretty well for Latin alphabet languages.

Testing

Unit tests of core csvdedupe functions bash pip install -r requirements-test.txt nosetests

Community

Recipes

Combining and deduplicating files from different sources.

Lets say we have a few sources of early childhood programs in Chicago and we’d like to get a canonical list. Let’s do it with csvdedupe, csvkit, and some other common command line tools.

Alignment and stacking

Our first task will be to align the files and have the same data in the same columns for stacking.

First let’s look at the headers of the files

File 1 console > head -1 CPS_Early_Childhood_Portal_Scrape.csv "Site name","Address","Phone","Program Name","Length of Day"

File 2 console > head -1 IDHS_child_care_provider_list.csv "Site name","Address","Zip Code","Phone","Fax","IDHS Provider ID"

So, we’ll have to add “Zip Code”, “Fax”, and “IDHS Provider ID” to CPS_Early_Childhood_Portal_Scrape.csv, and we’ll have to add “Program Name”, “Length of Day” to IDHS_child_care_provider_list.csv.

> cd examples
> sed '1 s/$/,"Zip Code","Fax","IDHS Provider ID"/' CPS_Early_Childhood_Portal_Scrape.csv > input_1a.csv
> sed '2,$s/$/,,,/' input_1a.csv > input_1b.csv

> sed '1 s/$/,"Program Name","Length of Day"/' IDHS_child_care_provider_list.csv > input_2a.csv
> sed '2,$s/$/,,/' input_2a.csv > input_2b.csv

Now, we reorder the columns in the second file to align to the first.

> csvcut -c "Site name","Address","Phone","Program Name","Length of Day","Zip Code","Fax","IDHS Provider ID" \
         input_2b.csv > input_2c.csv

And we are finally ready to stack.

> csvstack -g CPS_Early_Childhood_Portal_Scrape.csv,IDHS_child_care_provider_list.csv \
           -n source \
           input_1b.csv input_2c.csv > input.csv

Dedupe it!

And now we can dedupe

> cat input.csv | csvdedupe --field_names "Site name" Address "Zip Code" Phone > output.csv

Let’s sort the output by duplicate IDs, and we are ready to open it in your favorite spreadsheet program.

> csvsort -c "Cluster ID" output.csv > sorted.csv

githalytics.com alpha