Skip to content

Record matching and entity resolution at scale in Spark

License

Notifications You must be signed in to change notification settings

ing-bank/spark-matcher

Repository files navigation

Version Downloads Docs - GitHub.io

spark_matcher_logo

Spark-Matcher

Spark-Matcher is a scalable entity matching algorithm implemented in PySpark. With Spark-Matcher the user can easily train an algorithm to solve a custom matching problem. Spark Matcher uses active learning (modAL) to train a classifier (Scikit-learn) to match entities. In order to deal with the N^2 complexity of matching large tables, blocking is implemented to reduce the number of pairs. Since the implementation is done in PySpark, Spark-Matcher can deal with extremely large tables.

Documentation with examples can be found here.

Developed by data scientists at ING Analytics, www.ing.com.

Installation

Normal installation

As Spark-Matcher is intended to be used with large datasets on a Spark cluster, it is assumed that Spark is already installed. If that is not the case, first install Spark and PyArrow (pip install pyspark pyarrow).

Install Spark-Matcher using PyPi:

pip install spark-matcher

Install with possibility to create documentation

Pandoc, the general markup converter needs to be available. You may follow the official Pandoc installations instructions or use conda:

conda install -c conda-forge pandoc

Then clone the Spark-Matcher repository and add [doc] like this:

pip install ".[doc]"

Install to contribute

Clone this repo and install in editable mode. This also installs PySpark and Jupyterlab:

python -m pip install -e ".[dev]"
python setup.py develop

Documentation

Documentation can be created using the following command:

make create_documentation

Dependencies

The usage examples in the examples directory contain notebooks that run in local mode. Using the SparkMatcher in cluster mode, requires sending the SparkMatcher package and several other python packages (see spark_requirements.txt) to the executors. How to send these dependencies, depends on the cluster. Please read the instructions and examples of Apache Spark on how to do this: https://spark.apache.org/docs/latest/api/python/user_guide/python_packaging.html.

SparkMatcher uses graphframes under to hood. Therefore, depending on the spark version, the correct version of graphframes needs to be added to the external_dependencies directory and to the configuration of the spark session.
As a default, graphframes for spark 3.0 is used in the spark sessions in the notebooks in the examples directory. For a different version, see: https://spark-packages.org/package/graphframes/graphframes.

Usage

Example notebooks are provided in the examples directory. Using the SparkMatcher to find matches between Spark dataframes a and b goes as follows:

from spark_matcher.matcher import Matching

myMatcher = Matcher(spark_session, col_names=['name', 'suburb', 'postcode'])

Now we are ready for fitting the Matcher object using 'active learning'; this means that the user has to enter whether a pair is a match or not. You enter 'y' if a pair is a match or 'n' when a pair is not a match. You will be notified when the model has converged and you can stop training by pressing 'f'.

myMatcher.fit(a, b)

The Matcher is now trained and can be used to predict on all data. This can be the data used for training or new data that was not seen by the model yet.

result = myMatcher.predict(a, b)