Diffusion-Based Data Augmentation for Imbalanced Multivariate Time Series Classification

Abstract

The proliferation of sensor networks and Internet of Things (IoT) devices has led to an exponential increase in the availability of multivariate time series (MTS) data. However, in critical domains such as healthcare monitoring, industrial fault detection, and financial anomaly prediction, datasets are intrinsically imbalanced; the events of interest are rare compared to normal operational states. This class imbalance severely degrades the performance of deep learning classifiers, which tend to bias towards the majority class. Traditional oversampling techniques, such as SMOTE, often fail to capture the complex temporal dependencies and inter-variable correlations inherent in MTS data. While Generative Adversarial Networks (GANs) have been proposed as a solution, they suffer from training instability and mode collapse. This paper presents a novel framework for diffusion-based data augmentation specifically tailored for imbalanced multivariate time series classification. We leverage a Conditional Denoising Diffusion Probabilistic Model (CDDPM) effectively conditioned on class labels to generate high-fidelity synthetic samples of the minority class. By modeling the data distribution through a gradual denoising process, our approach preserves the intricate temporal dynamics and cross-channel correlations better than adversarial counterparts. Extensive experiments on multiple benchmark datasets demonstrate that our method significantly improves classification performance, particularly in terms of F1-score and Geometric Mean, compared to state-of-the-art augmentation techniques.