Mitigating Class Imbalance in Time Series with Enhanced Diffusion Models

This study introduces a novel approach to mitigate class imbalance in time series data using enhanced diffusion models by integrating oversampling techniques and classifier-free guidance to generate high-quality synthetic time series data. Our results indicate significant improvements not only concerning data quality but also in handling class imbalances, showcasing the potential of the proposed approach in improving the performance of machine learning models in scenarios where data annotation distribution is skewed. The efficacy of our approach was demonstrated using the UniMiB SHAR dataset with a focus on enhancing the automatic fall detection for patients. This research opens new avenues for data augmentation addressing critical challenges in training algorithms with balanced data representation. Such advancements hold significant implications for a variety of real-world contexts, especially within the healthcare sector.