Dynamic community detection using class preserving time series generation with Fourier Markov diffusion.

Area of Science:

• Artificial Intelligence
• Machine Learning
• Time Series Analysis

Background:

• Generating class-consistent time series data is challenging for current Generative Adversarial Network (GAN) and diffusion models.
• Maintaining both overarching structure and detailed temporal dynamics is crucial for realistic time series generation.

Purpose of the Study:

• To introduce FMD-GAN, a Fourier-Markov diffusion framework for generating realistic and class-consistent time series.
• To address the limitations of existing models in capturing temporal dynamics and class consistency.

Main Methods:

• FMD-GAN integrates spectral clustering, state-conditioned frequency-domain noise modulation, and a dual-branch temporal-spectral discriminator.
• The framework leverages spectral priors and probabilistic diffusion for enhanced sequence generation.

Main Results:

• FMD-GAN achieved state-of-the-art or competitive results on four UCR datasets (ECG200, GunPoint, FordA, ChlorineConc).
• Demonstrated up to a 50% reduction in Fréchet Inception Distance (FID) and consistent enhancements in Dynamic Time Warping (DTW), Class Consistency Accuracy (CCA), and Spectral Distance (SD).
• Ablation studies confirmed the effectiveness of spectrum masking, Markov-guided diffusion, and adversarial learning.

Conclusions:

• The integration of spectral priors with probabilistic diffusion enables the production of time series that preserve structure and class distinctions.
• FMD-GAN shows promise for applications in biomedical monitoring, sensor analytics, and Tiny AI systems.
• The framework exhibits resilience to hyperparameters, indicating practical applicability.