Evaluating Fairness and Mitigating Bias in Machine Learning: A Novel Technique using Tensor Data and Bayesian Regression

Skin color requires different fairness approaches than traditional categorical attributes: Unlike gender or race, skin color is represented as continuous tensor data in computer vision systems, meaning existing fairness evaluation methods designed for categorical features may not adequately capture bias in image classification tasks.

Probability distributions better capture fairness nuances than rigid categorization: The paper's technique converts skin color tensor data into probability distributions and uses statistical distance measures, allowing the model to recognize fine-grained variations within and across groups rather than forcing skin tones into discrete categories that may obscure unfair treatment.

Fairness can be evaluated without manual annotation: The proposed method enables fairness assessment in image classification without requiring humans to manually label sensitive attributes, which reduces annotation burden and potential labeler bias in the evaluation process itself.

Bayesian regression with polynomial functions can help mitigate latent bias: The paper proposes using color distance estimates calculated through Bayesian regression to create more equitable training methods that address biases embedded in conventional skin tone categorization systems.

Current fairness research has overlooked continuous visual attributes: Most prior work on fairness in ML has focused on categorical sensitive attributes, leaving a gap in how fairness is evaluated and improved for continuous visual properties like skin color that are inherently represented as tensor data in computer vision applications.