基于变分高斯过程模型的快速核偏标记学习算法

发布时间：2018-03-16 02:26

本文选题：偏标记学习　切入点：核方法　出处：《计算机研究与发展》2017年01期 　论文类型：期刊论文

【摘要】：偏标记学习(partial label learning)是人们最近提出的一种弱监督机器学习框架,由于放松了训练数据集的构造条件,只需知道训练样本的真实标记的一个候选集合就可进行学习,可以更方便地处理很多领域的实际问题.在该框架下,训练数据的标记信息不再具有单一性和明确性,这就使得学习算法的构建变得比传统分类问题更加困难,目前只建立了几种面向小规模训练数据的学习算法.先利用ECOC技术将原始偏标记训练集转换为若干标准二分类数据集,然后基于变分高斯过程模型在每个二分类数据集上构建一个具有较低计算复杂度的二分类算法,最终实现了一种面向大规模数据的快速核偏标记学习算法.仿真实验结果表明,所提算法在预测精度几乎相当的情况下,训练时间要远远少于已有的核偏标记学习算法,利用普通的PC机处理样本规模达到百万级的问题只需要40min.
[Abstract]:Partial label learning (partial label learning) is a recently proposed a weakly supervised machine learning framework, due to structural conditions to relax the training data set, only need to know a real candidate set of labeled training samples can be carried out on learning, practical problems can be more convenient to handle in the framework of many fields. Under the training data label information is no longer single and clear, which makes more difficult than the traditional classification problem of constructing learning algorithm, currently only established several small training data for learning algorithm. By using the ECOC technology to the original partial labeled training set is converted to the number of standard two classification data set, and then based on the variational Gauss model in each of the two classification data set to construct a two classification algorithm has low computational complexity, finally realizes an oriented large-scale data quickly Simulation results show that the training time of the proposed algorithm is much less than that of the existing kernel biased markup learning algorithm when the prediction accuracy is almost the same. Using the ordinary PC machine to deal with the problem of sample size reaches to a million level, it only needs 40min..

【作者单位】：大连理工大学电子信息与电气工程学部;大连民族大学信息与通信工程学院;
【基金】：国家自然科学基金项目(61503058,61502074,U1560102) 辽宁省自然科学基金项目(201602190) 中央高校基本科研业务费专项资金项目(DC201501055,DC201501060201)~~
【分类号】：TP181

【相似文献】