KANs for Time Series Forecasting reintroduce spectral bias with autocorrelation

A new paper reveals that Kolmogorov-Arnold Networks (KANs), previously thought to overcome spectral bias, actually reintroduce it when dealing with time series data due to temporal autocorrelation. Researchers found that this bias intensifies with higher autocorrelation, potentially hindering KANs' performance in time series forecasting. To mitigate this, the study proposes using the Discrete Cosine Transform (DCT) to preprocess inputs, which empirically demonstrated a significant reduction in the low-frequency preference. AI

Summary written by gemini-2.5-flash-lite from 1 source. How we write summaries →

IMPACT Suggests standard KANs may struggle with time series data, requiring preprocessing like DCT for improved performance.

RANK_REASON Academic paper on a specific neural network architecture's limitations and a proposed solution.

Read on arXiv cs.LG →

• Kolmogorov-Arnold Networks
• KANs
• Discrete Cosine Transform
• DCT
• arXiv

• paper
• other