RFFBCGA algorithm beats input noise for time-series prediction

A new paper from researchers at (presumably) Southwest Jiaotong University introduces RFFBCGA, a nonlinear adaptive filter designed to handle both input and output noise in time-series prediction. Current algorithms like bias-compensated kernel least mean square (BCKLMS) only account for output noise or suffer from fixed-size dictionaries that limit their ability to capture input signal characteristics, plus they perform poorly under non-Gaussian output noise. RFFBCGA overcomes these by combining random Fourier features (RFF) with a bias-compensation (BC) term and a general adaptive (GA) function. The RFF framework maintains a fixed network size while still capturing input signal dynamics, the BC term mitigates input noise interference, and the GA function gives the algorithm adaptability to different noise scenarios, including non-Gaussian.

The algorithm is validated through extensive simulations, including real-world time-series prediction tasks, where it outperforms state-of-the-art methods like BCKLMS and other nonlinear filters. The paper, published on arXiv (2607.19902), falls under machine learning (cs.LG) and audio and speech processing (eess.AS), suggesting potential applications in speech enhancement, financial forecasting, or sensor data processing. The authors (Yi Peng, Haiquan Zhao, Jinhui Hu) provide both a PDF and HTML version of the paper. This work offers a practical solution for engineers dealing with noisy data in nonlinear systems, especially when input noise is a significant factor.