Background
Research has shown how data sets convey social bias in Artificial Intelligence systems, especially those based on machine learning. A biased data set is not representative of reality and might contribute to perpetuate societal biases within the model. To tackle this problem, it is important to understand how to avoid biases, errors, and unethical practices while creating the data sets. In order to provide guidance for the use of data sets in contexts of critical decision-making, such as health decisions, we identified six fundamental data set features (balance, numerosity, unevenness, compliance, quality, incompleteness) that could affect model fairness. These features were the foundation for the FanFAIR framework.
Results
We extended the FanFAIR framework for the semi-automated evaluation of fairness in data sets, by combining statistical information on data with qualitative features. In particular, we present an improved version of FanFAIR which introduces novel outlier detection capabilities working in multivariate fashion, using two state-of-the-art methods: the Empirical Cumulative-distribution Outlier Detection (ECOD) and Isolation Forest. We also introduce a novel metric for data set balance, based on an entropy measure.
Conclusion
We addressed the issue of how much (un)fairness can be included in a data set used for machine learning research, focusing on classification issues. We developed a rule-based approach based on fuzzy logic that combines these characteristics into a single score and enables a semi-automatic evaluation of a data set in algorithmic fairness research. Our tool produces a detailed visual report about the fairness of the data set. We show the effectiveness of FanFAIR by applying the method on two open data sets.
