A unified time series classification framework via adaptive Gaussian image representation

Real-world time series classification (TSC) is frequently hindered by significant data heterogeneity, characterized by complex multivariate dependencies and variable sequence lengths. While image-based representations offer a promising direction, existing methods often rely on rigid, deterministic mappings that fail to adapt to diverse input structures or capture intrinsic statistical uncertainty. To address these challenges, we propose TS2Vision, a unified framework driven by Adaptive Time Series Gaussian Mapping (ATSGM). Unlike static encodings, ATSGM transforms temporal signals into 2D images by leveraging Gaussian statistics to robustly model local fluctuations and employing an adaptive “circle packing” optimization to dynamically arrange spatial layouts for multivariate channels. Theoretically, we prove that this mapping satisfies Lipschitz continuity, ensuring representation stability against input perturbations. We conduct a comprehensive evaluation on 158 datasets from the UCR and UEA archives. The results demonstrate that TS2Vision achieves Rank-1 consistency across both univariate and multivariate benchmarks, significantly outperforming state-of-the-art baselines. By bridging sequence modeling with computer vision, our framework offers a robust and generalizable solution for intelligent decision-making systems.