样本自适应的多核学习算法研究

Multiple kernel learning (MKL) provides an elegant theoretical framework for automatically learning an optimal kernel from data. It optimally combines multiple channels of samples to improve the performance of a classifier. However, due to the lack of sample-adaptive mechanism, existing MKL algorithms cannot effectively handle the common situations where some channels of samples are absent or noisy, leading to unsatisfactory learning performance and a relatively narrow application scope. To address the above issues, this project intends to focus on the research on sample-adaptive MKL algorithms, aiming to make MKL algorithms with sample-adaptive mechanism. Its research contents include: 1) Designing absent MKL algorithms to handle the issue of channel absence among samples; 2) Designing MKL algorithms with latent variables to handle the issue of noisy channels among samples; and 3) Theoretically analyzing the generalization performance of the proposed algorithms. The theoretical analysis and implementations on the mechanism of sample-adaptive MKL in this project can promote the development of theoretical research on MKL, help to remove the application barriers in the presence of absent or noisy channels, and further broaden its application prospects. Some preliminary research results have been accepted by some famous international journals, including IEEE TNNLS, IEEE TCyb and PR, which has laid a solid foundation for the research work of this project.

多核学习为从数据中学习最优核提供了优美的理论框架，它通过最优地集成多个通道信息来改进算法的性能。然而，现有的多核学习算法忽略了样本间潜在的自适应性，不能有效地处理现实应用中常见的缺失或噪声通道等问题，导致令人不满意的学习性能并极大地制约了其应用范围。为解决这一问题，本项目拟开展样本自适应多核学习算法研究，以增强多核学习算法的样本自适应机制。其主要研究内容包括: 1）设计缺失多核学习算法以解决样本中含有缺失通道问题；2）设计隐变量多核学习算法以解决样本中含有噪声通道问题；及3）从理论上分析拟提出算法的泛化性能。本项目关于多核学习中样本自适应机制的理论分析与算法实现，可推动多核学习研究的理论发展，有助于清除它在有缺失或噪声情形下的应用障碍，并进一步扩大其应用前景。该项目的部分前期研究成果已被国际著名期刊IEEE TNNLS，IEEE TCyb及PR所接收，为本项目的研究工作奠定了坚实的基础。

多核学习为从数据中学习最优核提供了优美的理论框架，它通过最优地集成多个通道信息来改进算法的性能。然而，现有的多核学习算法忽略了样本间潜在的自适应性，不能有效地处理现实应用中常见的缺失或噪声通道等问题，导致令人不满意的学习性能并极大地制约了其应用范围。为解决这一问题，本项目拟开展样本自适应多核学习算法研究，以增强多核学习算法的样本自适应机制。其主要研究内容包括: 1）设计缺失多核学习算法以解决样本中含有缺失通道问题；2）设计隐变量多核学习算法以解决样本中含有噪声通道问题；及3）从理论上分析拟提出算法的泛化性能。本项目关于多核学习中样本自适应机制的理论分析与算法实现，可推动多核学习研究的理论发展，有助于清除它在有缺失或噪声情形下的应用障碍，并进一步扩大其应用前景。该项目的部分研究成果已被国际著名期刊IEEE TNNLS，IEEE TCyb及PR和顶级会议AAAI及IJCAI所接收，为本项目的研究工作奠定了坚实的基础。