Fairness Metric Calculation for Spark

Calculate fairness metrics using Spark. This implementation is based on Aequitas and includes the following metrics:

• Equal Parity
• Proportional Parity
• False Positive Rate Parity
• False Discovery Rate Parity
• False Negative Rate Parity
• False Omission Rate Parity

How to build

This project requires sbt (tested against version 0.13.17) and scala (version 2.11.x).

git clone https://github.com/eubr-atmosphere/spark-fairness.git
cd spark-fairness
sbt build

Spark Packages

WIP

You can use the Fairness Metric Calculation in Spark by adding its JAR file (using option --jars) or by using Spark Packages:

spark-shell --conf spark.jars.packages=spark-fairness:spark-fairness:0.1.0-spark2.1-s_2.11

Usage

FairnessEvaluatorTransformer is a Spark transformer, with the following parameters:

• sensitiveColumn: Name of the column with the sensitive information.
• labelColumn: Name of the column with the label information (expected result).
• scoreColumn: Name of the column with the score information (predicted result).
• baselineValue: Value used as baseline value.
• tau: Define the range of assessment for the metrics.

Python (WIP)

import ...
df = spark.read.csv('samples/compas.csv')
transformer = FairnessEvaluatorTransformer(
    sensitiveColumn='race', labelColumn='label_value',
    scoreColumn='score', baselineValue='Caucasian',
    tau=0.8)
result = transformer.transform(df)
result.show()