cleanzr / dblink

Distributed Bayesian Entity Resolution in Apache Spark
Other
57 stars 9 forks source link
apache-spark bayesian-inference distributed-machine-learning entity-resolution mcmc record-linkage

dblink: Distributed End-to-End Bayesian Entity Resolution

dblink is a Spark package for performing unsupervised entity resolution (ER) on structured data. It's based on a Bayesian model called blink (Steorts, 2015), with extensions proposed in (Marchant et al., 2021). Unlike many ER algorithms, dblink approximates the full posterior distribution over clusterings of records (into entities). This facilitates propagation of uncertainty to post-ER analysis, and provides a framework for answering probabilistic queries about entity membership.

dblink approximates the posterior using Markov chain Monte Carlo. It writes samples (of clustering configurations) to disk in Parquet format. Diagnostic summary statistics are also written to disk in CSV format—these are useful for assessing convergence of the Markov chain.

Documentation

The step-by-step guide includes information about building dblink from source and running it locally on a test data set. Further details about configuration options for dblink is provided here.

Example: RLdata

Two synthetic data sets RLdata500 and RLdata10000 are included in the examples directory as CSV files. These data sets were extracted from the RecordLinkage R package and have been used as benchmark data sets in the entity resolution literature. Both contain 10 percent duplicates and are non-trivial to link due to added distortion. Standard entity resolution metrics can be computed as unique ids are provided in the files. Config files for these data sets are included in the examples directory: see RLdata500.conf and RLdata10000.conf. To run these examples locally (in Spark pseudocluster mode), ensure you've built or obtained the JAR according to the instructions above, then change into the source code directory and run the following command:

$SPARK_HOME/bin/spark-submit \
  --master "local[*]" \
  --conf "spark.driver.extraJavaOptions=-Dlog4j.configuration=log4j.properties" \
  --conf "spark.driver.extraClassPath=./target/scala-2.11/dblink-assembly-0.2.0.jar" \
  ./target/scala-2.11/dblink-assembly-0.2.0.jar \
  ./examples/RLdata500.conf

(To run with RLdata10000 instead, replace RLdata500.conf with RLdata10000.conf.) Note that the config file specifies that output will be saved in the ./examples/RLdata500_results/ (or ./examples/RLdata10000_results) directory.

How to: Add dblink as a project dependency

Note: This won't work yet. Waiting for project to be accepted.

Maven:

<dependency>
  <groupId>com.github.cleanzr</groupId>
  <artifactId>dblink</artifactId>
  <version>0.2.0</version>
</dependency>

sbt:

libraryDependencies += "com.github.cleanzr" % "dblink" % "0.2.0"

How to: Build a fat JAR

You can build a fat JAR using sbt by running the following command from within the project directory:

$ sbt assembly

This should output a JAR file at ./target/scala-2.11/dblink-assembly-0.2.0.jar relative to the project directory. Note that the JAR file does not bundle Spark or Hadoop, but it does include all other dependencies.

Contact

If you encounter problems, please open an issue on GitHub. You can also contact the main developer by email <GitHub username> <at> gmail.com

License

GPL-3

Citing the package

Marchant, N. G., Kaplan, A., Elazar, D. N., Rubinstein, B. I. P. and Steorts, R. C. (2021). d-blink: Distributed End-to-End Bayesian Entity Resolution. Journal of Computational and Graphical Statistics, 30(2), 406–421. DOI: 10.1080/10618600.2020.1825451 arXiv: 1909.06039.