DiffKANformer: Diffusion KAN Transformer for General Time Series Analysis

GVS Mothish, J Rishi, Omkar Kabde, Deepak N. Subramani

Keywords: Diffusion models, Time Series analysis, Kolmogorov-Arnold Networks

Abstract: Time series analysis tasks such as forecasting, imputation, anomaly detection, and classification are crucial for applications spanning climate science, financial domain, retail, and cloud infrastructure. We present DiffKANformer, a conditional diffusion model that integrates Kolmogorov-Arnold Networks (KAN) for feature projection and a Diffusion KAN Transformer architecture for denoising, specifically engineered for time series analysis. DiffKANformer introduces two key innovations: (i) a KAN-based projection mechanism in the forward diffusion process that captures complex correlation between features, and (ii) a Diffusion KAN Transformer architecture that effectively models complex long-term dependencies through adaptive univariate functions. Our model achieves superior performance across four fundamental time series analysis tasks, significantly outperforming existing prominent models in forecasting (eight datasets), imputation (six datasets), classification (ten datasets) and anomaly detection (five datasets). Comprehensive ablation studies across all tasks validate the utility of each DiffKANformer component, demonstrating the model's robustness in diverse time series challenges.

Primary Area: learning on time series and dynamical systems

Submission Number: 24029