Adaptive Diffusion Model-Based Data Augmentation for Unbalanced Time Series Classification

Classification models that provide good generalization are trained with sufficiently large datasets, but these are often unavailable due to restrictions and limited resources. It is challenging to learn the distribution that approximates the actual sample distribution when using the generation model for data generated for such unbalanced or small data sets. In this paper, an adaptive diffusion model-based data augmentation (ASE-DDPM) is present for unbalanced time series classification (TSC). One of the features of ASE-DDPM is that it clusters the time series sets with a specific label into multiple clusters and continuously injects noise until the time series between clusters are indistinguishable, then stops injecting noise. The superiority of the proposed data enhancement method is substantiated through comparative experiments on the UEA datasets, where it is shown to improve classification accuracy compared to using only the original train set.