THE IMPACT OF DEMOGRAPHIC COMPOSITION IN UTKFACE AND APPA-REAL DATASETS ON FAIRNESS IN AGE ESTIMATION MODELS

Abstract

This paper analyzes the impact of demographic composition on fairness in facial age estimation models trained on the UTKFace and APPA-REAL datasets. Building on previously published empirical results, the study provides a theoretical and analytical interpretation of how dataset imbalance affects model bias. Through comparative evaluation of group-wise performance metrics including Mean Absolute Error, Standard Deviation, Disparate Impact, and Equality of Opportunity, the paper introduces the concept of a Distributional Fairness Baseline (DFB) as a diagnostic framework for separating dataset-driven bias from model-induced bias. The analysis reveals that fairness is primarily a function of the representativeness and internal structure of training data, rather than model architecture. Contrary to common assumptions, full data equalization through oversampling does not necessarily enhance equity and may even amplify disparities due to overfitting and redundancy. Instead, moderate redistribution, particularly controlled undersampling of dominant groups often achieves an optimal balance between accuracy and fairness. These findings emphasize that equitable model performance depends on both quantitative and qualitative diversity within datasets, establishing data design as the central determinant of fairness in automated age estimation systems.