Explainable Machine Learning via Argumentation (EMLA)

This repo will host a generalized version of ArgEML, a general Explainable Machine Learning Framework and Methodology based on Argumentation. For more information please refer to paper Explainable Machine Learning via Argumentation.

At the moment, the repo includes the oner library that learns single condition rules from a given dataset using the concepts of "One Rule" classification algorithm OneR. The library currently supports datasets with categorical and numerical features.

OneR

OneR algorithm generates one rule for each predictor in a dataset, then selects the rule with the smallest total error as its "one rule". The oner implementation in this repo generates "one rule" for each predictor, by assessing the product of its coverage and accuracy (coverage * (1-error)) and provides various options for the consuming application to select one rule.

To create a rule for a predictor, a frequency table is constructed for each predictor against the target. For categorical features the test conditions are predictor=value_i (for each discrete predictor value). For example:

For each condition the most frequent (majority) class is assigned to the rule:

>> IF outlook=sunny THEN Yes [coverage=5/14, accuracy=3/5]
>> IF outlook=overcast THEN Yes [coverage=4/14, accuracy=1]
>> IF outlook=rainy THEN No [coverage=5/15, accuracy=2/5]

The different rules that can be generated from a single predictor are assessed based on their coverage*accuracy score:

>> IF outlook=sunny THEN Yes [coverage=5/14, accuracy=3/5] [score=0.21]
>> IF outlook=overcast THEN YES [coverage=4/14, accuracy=1] [score=0.28]
>> IF outlook=rainy THEN No [coverage=5/15, accuracy=2/5] [score=0.13]

The rule with the highest score is selected as the one rule for the predictor.

For numerical features splitting points are calculated and used to test conditions like predictor >= split_point_i AND predictor < split_point_j.

• Add emla module as a dependency to your project

    <dependency>
        <groupId>org.emla</groupId>
        <artifactId>emla</artifactId>
        <version>0.0.2-SNAPSHOT</version>
    </dependency>

• Load data into a Dataset object, create a new LearningSession and calculate frequencies.

Dataset ds = new Dataset("./src/test/resources/agentRequests.csv", "resourceAccess", "access", 1, 0);
LearningSession emlaSession = new LearningSession(ds,"resourceAccess");
List<FrequencyTable> frequencyTables = emlaSession.calculateFrequencyTables(ds, "train",null);
frequencyTables.forEach(f -> System.out.println(f.toString()));

• Frequency examples:

** Frequency Table for predictor 'role':

>> role == admin (target=allow), (coverage=0.33, accuracy=1, assessment=0.33):
	[allow, 3]
>> role == contributor (target=allow), (coverage=0.44, accuracy=0.5, assessment=0.22):
	[allow, 2]	[deny, 2]
>> role == guest (target=deny), (coverage=0.22, accuracy=1, assessment=0.22):
	[deny, 2]

** Frequency Table for predictor 'age':

>> age < 35.0 (target=deny), (coverage=0.33, accuracy=1, assessment=0.33):
	[allow, 0]	[deny, 3]
>> age >= 35.0, age < 45.0 (target=allow), (coverage=0.33, accuracy=1, assessment=0.33):
	[allow, 3]	[deny, 0]
>> age >= 45.0 (target=allow), (coverage=0.33, accuracy=0.67, assessment=0.22):
	[allow, 2]	[deny, 1]

• Select a frequency to create a rule:

Frequency newFrequency = emlaSession.calculateFrequencyHighCoverageLowError(frequencyTables, "allow");