Fairness in human activity recognition from wearable inertial sensor data

Abstract

Human Activity Recognition (HAR) has gained significant attention in recent years due to its potential applications in various domains such as healthcare, smart environments, and human-computer interaction. As HAR technologies become more pervasive, there is a growing concern about the fairness and equity aspects associated with their deployment. In this study, we address the issue of fairness in human activity recognition models. We begin by analyzing the data and identifying an unfairness gap in our initial model, which was based on a neural network architecture known as Multi-Layer Perceptron (MLP). To quantify the unfairness, we employ two well-established fairness metrics: equalized odds (EO) gap and demographic parity (DP) gap. To bridge this gap, two approaches were proposed, incorporating fairness losses: demographic parity loss and equalized odds loss. These losses aim to eliminate gender bias in the model’s predictions. To optimize the model with fairness losses, Lagrangian multiplier is employed, achieving a balance between accuracy and fairness in each training epoch. The results demonstrate the successful transformation of the initial unfair model into a fairer one, particularly addressing gender bias. The proposed approach attains a strong accuracy rate, approximately 95% for certain activities and over 75% for others. Furthermore, we compare our approach with state of art machine learning methods such as logistic regression and decision trees, further validating its effectiveness. This work contributes to advancing fairness-aware machine learning techniques for human activity recognition, promoting ethical and unbiased AI systems.