dblink is a Spark package for performing unsupervised entity resolution
(ER) on structured data.
It's based on a Bayesian model called
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
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.
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
Both contain 10 percent duplicates and are non-trivial to link due to added
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:
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
$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
Note that the config file specifies that output will be saved in
Note: This won't work yet. Waiting for project to be accepted.
<dependency> <groupId>com.github.cleanzr</groupId> <artifactId>dblink</artifactId> <version>0.2.0</version> </dependency>
libraryDependencies += "com.github.cleanzr" % "dblink" % "0.2.0"
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
relative to the project directory.
Note that the JAR file does not bundle Spark or Hadoop, but it does include
all other dependencies.
If you encounter problems, please open an issue
You can also contact the main developer by email
<GitHub username> <at> gmail.com
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.