Q-MKL: Matrix-induced Regularization in Multi-Kernel Learning with Applications to Neuroimaging.

Area of Science:

• Machine Learning
• Computational Neuroscience
• Biomedical Informatics

Background:

• Multiple Kernel Learning (MKL) extends Support Vector Machines (SVMs) by learning a linear classifier and an optimal combination of base kernels.
• Current MKL methods primarily use norm regularization for model complexity, neglecting kernel interaction information.
• Higher-order kernel-pair relationships, obtainable through various mechanisms, are underutilized in existing MKL frameworks.

Purpose of the Study:

• To develop a generalized MKL framework that regularizes kernel mixing weights using arbitrary quadratic functions.
• To incorporate kernel interaction information as an inductive bias for improved learning.
• To evaluate the proposed method on a neuroimaging task for Alzheimer's Disease (AD) prediction.

Main Methods:

• Replaced traditional norm penalties in MKL with a quadratic function Q to impose a desired covariance structure on mixing weights.
• This generalized formulation encompasses standard 1- and 2-norm MKL objectives.
• Experimental validation was performed on a neuroimaging dataset for predicting conversion to Alzheimer's Disease (AD).

Main Results:

• The proposed MKL method significantly outperforms state-of-the-art approaches on the Alzheimer's Disease prediction task.
• Experimental results showed statistically significant improvements (p-values ⪡ 10^-3).
• The model effectively leverages aggregate information from multiple neuroimaging modalities.

Conclusions:

• The novel MKL formulation, by incorporating kernel interactions via quadratic regularization, offers a powerful generalization of existing methods.
• This approach provides a flexible way to introduce inductive biases into MKL.
• The method demonstrates high efficacy in complex neuroimaging applications, particularly for predicting neurodegenerative disease progression.