FAIM: Frequency-Aware Interactive Mamba for Time Series Classification

TL;DR

FAIM tackles time series classification under noise and limited data by fusing frequency-domain adaptive filtering with a light-weight Interactive Mamba backbone, enhanced by self-supervised pre-training. The Adaptive Filtering Block uses learnable global/local spectral filters and thresholds to denoise in the Fourier domain, while the Interactive Mamba Block enables multi-granularity temporal feature interaction through dual-causal convolutions. Extensive experiments on 85 UCR and 26 UEA datasets demonstrate state-of-the-art accuracy and robustness, with favorable efficiency and scalability, supported by thorough ablations and visualizations. The approach offers a practical, robust, and efficient framework for diverse TSC applications, especially in noisy environments and resource-constrained settings.

Abstract

Time series classification (TSC) is crucial in numerous real-world applications, such as environmental monitoring, medical diagnosis, and posture recognition. TSC tasks require models to effectively capture discriminative information for accurate class identification. Although deep learning architectures excel at capturing temporal dependencies, they often suffer from high computational cost, sensitivity to noise perturbations, and susceptibility to overfitting on small-scale datasets. To address these challenges, we propose FAIM, a lightweight Frequency-Aware Interactive Mamba model. Specifically, we introduce an Adaptive Filtering Block (AFB) that leverages Fourier Transform to extract frequency-domain features from time series data. The AFB incorporates learnable adaptive thresholds to dynamically suppress noise and employs element-wise coupling of global and local semantic adaptive filtering, enabling in-depth modeling of the synergy among different frequency components. Furthermore, we design an Interactive Mamba Block (IMB) to facilitate efficient multi-granularity information interaction, balancing the extraction of fine-grained discriminative features and comprehensive global contextual information, thereby endowing FAIM with powerful and expressive representations for TSC tasks. Additionally, we incorporate a self-supervised pre-training mechanism to enhance FAIM's understanding of complex temporal patterns and improve its robustness across various domains and high-noise scenarios. Extensive experiments on multiple benchmarks demonstrate that FAIM consistently outperforms existing state-of-the-art (SOTA) methods, achieving a superior trade-off between accuracy and efficiency and exhibits outstanding performance.

Figures (9)

• Figure 1: Pipeline of how frequency transform acts in TSC framework.
• Figure 2: The structure of our proposed FAIM. FAIM processes time series data with a hierarchical architecture consisting of multiple layers, each including an Adaptive Filtering Block (AFB), an Interactive Mamba Block (IMB), and Layer Normalization (LN). The model starts by dividing the input into patches and applying positional embeddings to retain temporal information. The AFB transforms the data to the frequency domain using a Fourier Transform, where high- and low-frequency noise is suppressed using adaptive and learnable filters. After iFFT reconstruction, the IMB further enhances the features via interactive convolution-activation, improving sensitivity to temporal dynamics. The final features are processed through the classification layer to obtain the final result.
• Figure 3: CD diagram of SOTA methods on UCR (left) and UEA (right) datasets with a confidence level of 95%.
• Figure 4: Robustness to noise levels on two UEA datasets: ArticularyWordRecognition and Libras.
• Figure 5: Results on AWR and SRSCP under different data sizes and layer numbers.