Generalized Convolution Spectral Mixture for Multitask Gaussian Processes.

Area of Science:

• Machine Learning
• Statistical Modeling
• Time Series Analysis

Background:

• Multitask Gaussian processes (MTGPs) are valuable for joint regression modeling of related functions.
• Existing kernels for MTGPs struggle to capture complex nonlinear correlations and dependencies between tasks.
• Current methods often impose restrictive alignment constraints on components across different tasks.

Purpose of the Study:

• To address limitations in current MTGP kernels for modeling nonlinear task correlations.
• To propose an enhanced spectral mixture (SM) kernel, the multitask generalized convolution SM (MT-GCSM) kernel.
• To enable MT-GCSM to model nonlinear task correlations, inter-component dependencies, and remove component alignment constraints.

Main Methods:

• Focused on spectral mixture (SM) kernels as a foundation for enhancement.
• Developed the multitask generalized convolution SM (MT-GCSM) kernel by integrating generalized convolution structures.
• Employed inner and outer full cross-convolution between base components to decouple task alignment.

Main Results:

• The MT-GCSM kernel demonstrates the ability to model nonlinear task correlations and dependencies, including time and phase delays.
• It successfully removes the alignment constraint present in previous MTGP kernels.
• Experiments on synthetic and real-life datasets show MT-GCSM's superiority over existing SM kernels.

Conclusions:

• The proposed MT-GCSM kernel significantly advances the capabilities of multitask Gaussian processes.
• It offers a more flexible and powerful approach for modeling complex dependencies in multi-output regression.
• MT-GCSM proves effective in practical applications, outperforming prior spectral mixture-based methods.