视觉注意机制启发的分组排序特征选择方法研究

发布时间：2018-05-21 08:53

本文选题：特征选择 + 视觉注意　；参考：《郑州大学》2017年硕士论文

【摘要】：在大数据的时代背景下,为了从超大规模的高维数据中快速获取有效信息,特征选择至关重要也备受关注。基于特征分组的特征选择方法由于其兼顾了最大化特征与类别之间的相关程度和最小化特征间的冗余度的优势而被广泛接受,而视觉注意机制的显著性计算方式与特异性处理机制又对分组排序特征选择方法产生了奇妙的启发。本文受这种启发影响,分别模拟其对相关信息的特异性处理策略和对显著信息的显著性计算策略,形成分组排序的新思路,提出分组排序特征选择(Grouped Sorting Feature Selection,GSFS)算法。首先从特征分组和排序过程与视觉注意机制两种策略的相似性入手,介绍GSFS算法与视觉注意机制的模拟—启发关系,结合特征分组和排序指标的介绍,确定基于最大信息压缩指数和Fisher分数的GSFS算法并详述其基本原理,构建合适的最佳分组数目确定准则和最优特征子集确定准则并实现了算法流程。随后在8个不同维度,不同类别数的标准数据集上进行了分类实验和去冗余效果测评,通过和6种经典特征选择算法的实验结果对比说明GSFS算法选出的最优特征子集分类能力优良且所含冗余信息最少,从而证明了GSFS算法在解决特征选择问题时的有效性。在GSFS算法的理论基础上,为进一步解决高维数据集特征选择中计算复杂度较高的问题,又提出基于分布式处理的分组排序特征选择(Distributed Processing based Grouped Sorting Feature Selection,DP_GSFS)算法。它采用将原始特征集均分的方式,减少特征对之间相关性度量的计算复杂度以提高计算效率。在6个标准数据集上的分类实验和去冗余效果测评分别验证了DP_GSFS算法分类能力和去冗余效果的基础上,两种算法的对比实验结果表明DP_GSFS算法仅需设置合适的分组数,就可以显著降低计算时间,由此证明了其在提高计算效率方面的作用,这也可以为将来超大规模海量数据集的快速分析处理提供有价值的参考。另外,本文还将算法应用在真实医学图像特征数据集上,特征选择的效果和分类结果也证明了其解决实际问题的能力。
[Abstract]:In the background of large data age, feature selection is very important to obtain efficient information quickly from large scale data. Feature selection based on feature packet is widely accepted because it takes into account the degree of correlation between maximum features and the degree of redundancy between categories and minimization of features. The remarkable calculation method of visual attention mechanism and the specific processing mechanism also have a wonderful inspiration for the selection method of grouping sorting feature. This paper is influenced by this kind of inspiration, and simulates the specific processing strategy of the related information and the significant calculation strategy of the significant information, forming a new idea of grouping sorting and putting forward the grouping. The Grouped Sorting Feature Selection (GSFS) algorithm is first introduced from the similarity between the two strategies of the feature grouping and the sorting process and the visual attention mechanism, and the simulation and heuristic relationship between the GSFS algorithm and the visual attention mechanism is introduced, and the maximum information compression index and Fi are determined based on the introduction of the feature grouping and sorting index. The basic principle of the GSFS algorithm for Sher fraction is described in detail, and the appropriate criteria for determining the optimal number of groups and the optimal subset selection criteria are constructed and the algorithm flow is realized. Then, the classification experiments and redundant fruit evaluation are carried out on the standard data sets of 8 different dimensions and different categories, and the 6 classic feature selection algorithms are adopted. The comparison of the experimental results shows that the optimal subset selection of the GSFS algorithm is excellent and the redundant information is the least, which proves the validity of the GSFS algorithm in solving the problem of feature selection. On the basis of the theory of GSFS algorithm, the basis of the algorithm is to further solve the problem of high computational complexity in the feature selection of high dimensional data sets. The Distributed Processing based Grouped Sorting Feature Selection, DP_GSFS) algorithm in distributed processing. It reduces the computational complexity of the correlation measurement between the feature pairs and improves the computational efficiency by equifying the original feature set to improve the computational efficiency. The classification experiment and the redundant effect on the 6 standard data sets On the basis of verification of the classification ability and the deredundancy effect of the DP_GSFS algorithm, the comparison experiment results of the two algorithms show that the DP_GSFS algorithm only needs to set the appropriate number of packets, which can significantly reduce the calculation time, thus proving that the algorithm is used to improve the computational efficiency. This can also be a large scale of large scale data in the future. The fast analysis processing of set provides valuable reference. In addition, this paper also applies the algorithm to the feature dataset of real medical images. The effect of feature selection and classification results also prove its ability to solve practical problems.
【学位授予单位】：郑州大学
【学位级别】：硕士
【学位授予年份】：2017
【分类号】：TP181

【相似文献】