CORELS: Learning Certifiably Optimal RulE ListS

Run CORELS entirely from your browser!

Here, you can upload a file containing any training dataset with all-binary features and a single binary classification, and then press 'Submit' to have CORELS generate a rule list model from it. To do this, CORELS will generate rule antecedents from the features of your dataset - first one antecedent will be generated for each feature, and then further antecedents will be generated by combining permutations of features. For example, if the dataset contains the features is age > 25 and is female, then this rule list mining would produce three rule antecedents: is age > 25; is female; and is age > 25 AND is female.

In the above example, the third rule antecedent has a 'cardinality' of 2. For a dataset of m features, CORELS can generate, or mine, antecedents with cardinalities 1 through m, combining any number from 1 to all the features. However, limiting the maximum cardinality of these generated antecedents is often prudent, since for even a small numbers of features the number of total rules possible with high cardinalities is enormous. We provide an option to do this - CORELS will generate antecedents with cardinalities less than or equal to the value provided in the 'maximum cardinality' field.

We also provide an option to throw out antecedents that capture too many or too few samples - any antecedent that captures a fraction of the training samples that is less than or equal to the minimum support or that captures a fraction of the training samples that is greater than or equal to 1 minus the minimum support is discarded. Therefore the minimum support can have a value between 0 and 0.5.

For specific information about the data file formatting:

CSV data format

Here are three example csv files containing datasets. Simply download one, upload it to the 'CSV data' panel below, and press 'Submit' to see CORELS in action. Or, you can run CORELS with a default dataset (COMPAS) without having to download anything by clicking on the link below the file upload field.

COMPAS Monks1 Haberman

The above datasets were derived from the following locations: compas, monks1, haberman.

General data format:

The input must contain comma-delimited text.
All data, including features and classification, must be binary (besides the first row, the only value fields may have are the characters '0' and '1').
The first row of data must contain the list of features, each one followed by a comma, and be terminated by the classification label name.
Each subsequenct row, with one row per sample, must contain a list of bits (of length the number of features plus 1 [the classification]), with each bit separated by a comma. The nth bit in this list must be the value of the nth feature (as described in the first row) for that particular sample (row).

For example, a dataset with 4 samples and 4 features could look like the following:

age>25,likes_pie=true,lives_west=true,drives=true,uses_windows=true
1,0,0,1,0
0,1,1,1,0
1,1,1,0,1
0,0,1,1,1

In this example, the classification is uses_windows=true, and the first two samples (individuals) don't use windows, while the last two do use windows.

Data and Parameters

Sample data (extension ".csv")

Using default dataset (COMPAS)

Use default dataset (COMPAS)

Minimum support

Maximum cardinality

Secondary Parameters

Regularization Coefficient

Max number of nodes

Search Policy

Symmetry-aware map type

Debug level

Algorithm progress

Log to files

Print rules

Print labels

Print samples (Warning: LOTS OF OUTPUT)

Output

Corels data

Training data (extension ".out")

Training labels (extension ".label")

Equivalent Points file (optional, extension ".minor")

Generate Equivalent Points file if not provided