Enhancing Minority Class Prediction in Wearable Sensor-Based Activity Recognition Using SMOTE Oversampling

Abstract

Wearable sensor-based activity recognition has become increasingly important in various domains, particularly healthcare and sports. However, a significant challenge in this field is the issue of class imbalance, where minority activity classes are underrepresented compared to majority classes in datasets. This imbalance leads to biased classifiers that struggle to accurately identify rare but critical activities, which is especially problematic in health monitoring scenarios. This study evaluates the effectiveness of the Synthetic Minority Oversampling Technique (SMOTE) to address class imbalance in the mHealth dataset, which contains multi-sensor data from wearable devices placed on the chest, left ankle, and right lower arm. We employ the XGBoost classifier combined with SMOTE oversampling to improve recognition performance for minority classes. Model evaluation is conducted using precision, recall, F1-score, Area Under the Precision-Recall Curve (AUC-PR), ROC curve, and calibration analysis. The results demonstrate that applying SMOTE improves minority class recall from 0.75 to 0.85 and F1-score from 0.796 to 0.865, despite a slight decrease in overall accuracy from 97% to 96.5%. The AUC-PR also increases from 0.81 to 0.88, indicating a better balance in detecting minority and majority classes. Calibration curves reveal that probability estimates still require refinement to be more reliable for decision-making. This study confirms the efficacy of SMOTE in mitigating class imbalance in wearable sensor-based activity recognition and provides valuable insights for developing more accurate and fair health monitoring systems.