多视角学习 (Multi-View Learning)

最新推荐文章于 2024-12-08 23:03:37 发布

转载最新推荐文章于 2024-12-08 23:03:37 发布 · 2.1w 阅读

文章标签：

#Multi-View

机器学习专栏收录该内容

39 篇文章

订阅专栏

本文介绍了多视角学习的基本概念及应用，重点讲述了协同训练、多核学习和子空间学习等几种核心算法，并讨论了共识与互补原则。文章还详细解释了协同训练的不同变种及其依赖的假设。

多视角学习 (Multi-View Learning) 简介

Views from (1) multiple sources (2) different feature subsets;

视角来自于 (1) 多个源 (2) 多个特征子集

Multi-view learning algorithms: (1) co-training (2) multi-kernel learning (3) subspace learning;

多视角学习算法: (1) 协同训练 (2) 多核学习 (3) 子空间学习

Co-training(协同训练) :

trains alternately to maximize the mutual agreement on two distinct views of the unlabeled data.

Multiple kernel learning(多核学习 ) :

exploits kernels that naturally correspond to different views and combine kernels either linearly or non-linearly to improve learning performance.

Subspace learning(子空间学习 ) :

obtains a latent subspace shared by multiple views by assuming that the input views are generated from this latent subspace.

Principles: (1) consensus principle (共识准则) (2) complementary principle (互补准则);

准则: (1) 共识准则 (2) 互补准则

Multi-view learning: introduces one function to model a particular view and jointly optimizes all the functions to exploit the redundant views of the same input data and improve the learning performance.

多视角学习: 引入了一个函数去模型化一个特定的视角, 并且利用相同输入的冗余视角去联合优化所有函数, 最终提高学习效果.

Co-training: trains alternately(轮流地) to maximize the mutual agreement on two distinct views of the unlabeled data.

协同训练: 在未标记数据的两个不同视角下, 轮流的训练, 使相互一致性最大化.

Co-training Variants(变种):

(1) Expectation-Maximization (EM): assigning changeable probabilistic labels to unlabeled data; (聚类算法-OpenCV有相关函数)

(2) Semi-Supervised Learning Algorithm: Muslea et al;

(3) Bayesian undirected graphical model & Gaussian process classifiers: Yu et al;

(4) Combinative label propagation (结合的标签传播): Wang & Zhou;

(5) A data-dependent “co-regularization(协同正则化)”norm: Sindhwani;

(6) Data clustering and designed effective algorithms: Bickel & Scheffer and Kumar et al;

Relying on three assumptions (依赖的三个假设):

(1) sufficiency (充足性) - each view is sufficient for classification on its own;

(2) compatibility (兼容性) – the target function of both views predict the same labels for co-occurring features with a high probability;

(3) conditional independent (条件独立性) – views are conditionally independent given the label.