润色论文网提供毕业论文写作、职称论文写作、职称论文发表及期刊论文发表服务!QQ:526140740

电话:18796337551
当前位置:润色论文网系统论文 → 文章正文

基于超图学习的多标记降维算法研究

作者:润色论文网  来源:www.runselw.com  发布时间:2019/10/8 10:19:51  

摘要:多标记降维是机器学习领域的重要研究内容。现有的多标记降维方法主要通过挖掘标记间的相关性,提高学习模型的性能。然而,样本的标记信息中往往存在噪声标记问题以及缺失标记问题,导致所学得的标记相关性不准确,进而降低学习算法的性能。为了解决该问题,本文主要以样本特征空间作为切入点,在挖掘标记相关性的同时,通过嵌入样本流形结构信息,辅助提高多标记学习算法的性能。具体地,本文定义特征空间的概率超图模型,并将其与基于标记空间的传统超图相结合,提出联合降维算法。基于此,本文的主要工作如下:

提出基于样本流形结构与标记相关性的多标记降维算法。该算法基于样本的原始特征,定义特征空间的概率超图模型,嵌入样本的流形结构信息,从而削弱噪声标记的影响,弥补标记缺失的问题,并通过传统超图挖掘标记间的高阶相关性。在多个基准数据集上的实验结果表明了所提算法的有效性。

提出基于鲁棒特征学习策略的多标记降维算法。该算法提出采用矩阵分解技术学习样本的鲁棒特征,削弱原始特征空间的噪声以及离群点影响,从而构建具有高层语义特征的概率超图,并联合标记空间的概率超图一同嵌入多标记降维模型。通过理论分析和实验证明,该算法能够解决概率超图模型的时空复杂度问题。

Multi-label dimensionality reduction is animportant research in the field of machine learning. Existing multi-labeldimensionality reduction methods mainly improve the performance of learningmodel by mining the correlation between labels. However, the problem of noisylabel and missing label often exists in the label information of samples, whichleads to the inaccurate label correlation obtained, and then reduces theperformance of the learning algorithm. In order to solve this problem, thispaper mainly takes the sample feature space as the breakthrough point. Whilemining the label correlation, it also improves the performance of multi-labellearning algorithm by embedding the manifold structure information of samples.Specifically, this paper defines the probabilistic hypergraph model of featurespace and combines it with the traditional hypergraph based on label space topropose a joint dimensionality reduction algorithm. Based on this, the mainwork of this paper is as follows:

A multi-label dimensionality reductionalgorithm based on sample manifold structure and label correlation is proposed.Based on the original features of samples, the algorithm defines aprobabilistic hypergraph model of feature space and embeds the manifold structureinformation of samples, thus weakening the influence of noisy label, making upfor the problem of missing label, and mining high-order correlation betweenlabel through traditional hypergraph. The experimental results on severalbenchmark datasets show the effectiveness of the proposed algorithm.

A multi-label dimensionality reductionalgorithm based on robust feature learning strategy is proposed. This algorithmuses matrix decomposition technology to learn robust features of samples,weakens the noise of original feature space and the influence of outliers, soas to construct probabilistic hypergraph with high-level semantic features, andembeds multi-label dimensionality reduction model together with probabilistichypergraph of label space. The theoretical analysis and experiments show thatthe algorithm can solve the space-time complexity problem of the probabilistichypergraph model.

关键词:多标记降维;标记相关性;流形结构;超图结构;鲁棒特征

multi-label dimensionality reduction; labelcorrelation; manifold structure; hypergraph structure; robust feature

联系方式

客服QQ 526140740
客服热线18796337551
网站地址 www.runselw.com
郑重承诺 原创,包修改,包通过!
润色论文网真诚欢迎新老客户的光临与惠顾!