Kernel Mixture Correntropy Conjugate Gradient Algorithm for Time Series Prediction.

Area of Science:

• Machine Learning
• Signal Processing
• Nonlinear System Analysis

Background:

• Kernel adaptive filtering (KAF) is effective for nonlinear time series prediction but traditional methods like stochastic gradient descent (SGD) suffer from slow convergence and low accuracy.
• Kernel conjugate gradient (KCG) offers lower complexity but lacks robust performance, especially in non-Gaussian noise.
• Correntropy, a kernel-based similarity measure, effectively handles outliers in robust signal processing.

Purpose of the Study:

• To propose a novel Kernel Conjugate Gradient (KCG) algorithm, the Kernel Mixture Correntropy Conjugate Gradient (KMCCG), to enhance learning performance and robustness.
• To improve computational efficiency and accuracy in nonlinear time series prediction, particularly in non-Gaussian noise environments.
• To apply the KMCCG algorithm to practical tasks like malware prediction.

Main Methods:

• Developed the Kernel Mixture Correntropy Conjugate Gradient (KMCCG) algorithm utilizing the Mixture Correntropy Criterion (MCC).
• Incorporated a sparsification criterion based on angle between elements in Reproducing Kernel Hilbert Space (RKHS) to manage Radial Basis Function (RBF) network growth.
• Evaluated performance using synthetic chaotic time series and a real benchmark dataset.

Main Results:

• The proposed KMCCG algorithm demonstrates reduced computational complexity compared to existing KAF algorithms.
• KMCCG achieves superior performance in non-Gaussian noise environments and shows better computational performance on benchmark datasets.
• The algorithm achieved high prediction accuracy and short training times in malware prediction tasks.

Conclusions:

• The KMCCG algorithm offers a significant advancement in kernel adaptive filtering for time series prediction.
• The method provides a robust and computationally efficient solution for handling non-Gaussian noise and outliers.
• KMCCG shows practical applicability and high accuracy in real-world problems such as malware prediction.