Castor: A Relational Learning System

v0.1

The Castor relational learning system is described in the paper Schema Independent Relational Learning (SIGMOD 2017).

Installation

Install VoltDB

Currently, Castor only works on top of the in-memory RDBMS VoltDB.

1. Download and install VoltDB Community Edition. Instructions available here.

Set environment variables

1. Set VOLTDB_HOME environment variable to installation directory of VoltDB.
2. Add $VOLTDB_HOME/bin to PATH environment variable.

Compile Castor

1. Compile Castor by running:

ant

It will generate a dist folder, containing Castor.jar file and the dependencies in the lib folder.

Start database

1. Create a VoltDB database and insert data. For an example, see examples folder.

Run Castor

1. Run Castor's JAR file. There are two options to set training examples.

• Option 1: Training examples are stored in CSV files. These files are specified by arguments: posTrainExamplesFile, negTrainExamplesFile, posTestExamplesFile, negTestExamplesFile.

java -jar Castor.jar -dataModel <data_model_file> -parameters <parameters_file> -posTrainExamplesFile <pos_train_examples_file> -negTrainExamplesFile <neg_train_examples_file> -posTestExamplesFile <pos_test_examples_file> -negTestExamplesFile <neg_test_examples_file>

• Option 2: Training examples are stored in tables in the database. Castor assumes that the names of these tables are the name of the target relation followed by a suffix. The name of the target relation is extracted from the headMode in the dataModel file. The suffixes are specified by arguments: trainPosSuffix, trainNegSuffix, testPosSuffix, testNegSuffix.

java -jar Castor.jar -dataModel <data_model_file> -parameters <parameters_file> -trainPosSuffix <train_pos_suffix> -trainNegSuffix <train_neg_suffix> -testPosSuffix <test_pos_suffix> -testNegSuffix <test_neg_suffix>

Castor command line arguments

• dataModel <data_model_file> (required): JSON file containing mode declarations (language bias). See an example here. A short explanation on mode declarations can be found in Section 3 of this paper.
• parameters <parameters_file> (required): JSON file containing parameters (explained below).
• inds <inds_file>: JSON file containing inclusion dependencies. See an example here.
• trainPosSuffix <train_pos_suffix>: Suffix of table containing positive training examples.
• trainNegSuffix <train_neg_suffix>: Suffix of table containing negative training examples.
• testPosSuffix <test_pos_suffix>: Suffix of table containing positive testing examples.
• testNegSuffix <test_neg_suffix>: Suffix of table containing negative testing examples.
• posTrainExamplesFile <pos_train_examples_file>: CSV file containing positive training examples.
• negTrainExamplesFile <neg_train_examples_file>: CSV file containing negative training examples.
• posTestExamplesFile <pos_test_examples_file>: CSV file containing positive testing examples.
• negTestExamplesFile <neg_test_examples_file>: CSV file containing negative testing examples.
• test: Test learned definition using testing examples.
• outputSQL: Output learned definition in SQL format.
• sat: Only build bottom-clause for example specified by argument e.
• groundSat: Only build ground bottom-clause for example specified by argument e.
• e: Index of example to build (ground) bottom-clause when sat or groundSat arguments are specified (default: 0).

Castor parameters

These parameters are specified inside the file pointed by the parameters argument.

• dbURL: VoltDB server URL. (default: "localhost")
• port: VoltDB client port. (default: 21212)
• iterations: Number of iterations in bottom-clause construction algorithm. Equivalent to maximum depth of variables in a bottom-clause.
• minprec: Minimum precision that a clause must satisfy to be included in the learned definition (computed based on uncovered positive examples). In other words, how precise each clause should be. (default: 0.5)
• minrec: Minimum recall that a clause must satisfy to be included in the learned definition (computed based on all positive examples). In other words, the minimum percentage of positive examples that a clause should cover. (default: 0).
• minPos: Minimum number of positive examples that a clause must cover to be included in the learned definition. (default: 2)
• sample: Number of examples to use when generalizing a clause using ARMG. (default: 1)
• beam: Number of candidate clauses to keep. (default: 1)
• recall: Maximum number of literals added to a bottom-clause for each application of a mode declaration. (default: 10)
• groundRecall: Maximum number of literals added to a ground bottom-clause for each application of a mode declaration. Ground bottom-clauses are used to evaluate coverage using theta-subsumption. If this parameter is restricted, result of coverage is approximate. (default: Integer.Max_VALUE)
• threads: Number of threads; used to parallelize coverage operations. (default: 1)
• randomSeed: Random seed. (default: 1)
• createStoredProcedure: Create stored procedures that run bottom-clause construction algorithm. (default: true)
• useStoredProcedure: Use stored procedures to run bottom-clause construction algorithm. (default: true)

Castor assumptions

Castor makes the following assumptions. We may remove some of these assumptions in the future.

• The schema contains unique relation names.
• All attributes in schema are strings.
• Only one attribute is input (+) in mode declarations (language bias).

Notes

• Castor is under development.
• Castor has only been tested in macOS and Linux (Red Hat).
• Castor is memory intensive. If you get OutOfMemoryError, increase Java heap size.

Citation

If you use Castor, please cite the paper Schema Independent Relational Learning (SIGMOD 2017) (ACM Digital Library):

@inproceedings{Picado2017SchemaIR,
  title={Schema Independent Relational Learning},
  author={Jose Picado and Arash Termehchy and Alan Fern and Parisa Ataei},
  booktitle={SIGMOD Conference},
  year={2017}
}