Identifying Indicators of Bias in Data Analysis Using Proportionality and Separability Metrics

Abstract

There have been a number of high-profile cases of artificial intelligence (AI) systems making culturally-inappropriate predictions, mainly due to imbalanced training data and the overall black-box nature of modern deep learning algorithms. In this paper, we consider metrics for analyzing the data to determine the characteristics which may result in a biased model. Specifically, we look at a combination of separability between and proportionality of clusters within a given dataset to identify the presence of implicit biases. We measured the effectiveness of Alpha Diversity, entropy, and euclidean norm as a measure of proportionality. Manually permuting the MNIST and fashion MNIST dataset, we found all three scores strongly correlated with model accuracy. Silhouette scores were used as the separability metric and showed limited but noticeable correlation with predicted accuracy.