【作者】 赵军；

【导师】 杨胜；

【作者基本信息】 湖南大学 ， 计算机科学与技术， 2014， 硕士

【摘要】 随着信息技术的不断发展,信息获取能力的不断提高,人们往往需要分析和处理各种高维数据,如:海量web数据、遥感图像、微阵列数据等等。这些高维数据通常会导致机器学习算法的计算量成指数倍增长,引发“维数灾难”问题,因此对于高维数据的特征选择问题已成为数据挖掘领域的一个重要课题。通过特征选择技术将高维数据对象从高维特征空间映射到低维特征空间,这些低维特征更能体现出数据对象的本质含义,同时又能提高数据分析处理的效率。本文以微阵列数据为例作为实验数据,对高维数据的特征选择方法的理论思想和实际应用进行了深入研究和探讨。主要工作包括以下几个方面:提出了一种基于特征相似性的特征选择算法,首先使用局部标准化信噪比方法剔除无关特征,然后对剩余特征进行聚类,并将簇中特征数较少的簇当做噪音特征剔除,噪音特征剔除之后,会得到k个簇,其中簇内特征相互冗余度较高,而簇间特征相互冗余度较低,然后根据本文提出的评价标准,依次对每个簇中的每个特征进行评价,并决定是否剔除,将最终剩下的特征集合到一起,并按照单独分类能力的大小进行排序。实验表明,算法可以很好的剔除无关特征、噪音特征和冗余特征。分析基于特征相似性的特征选择算法和Top-r特征选择算法各自的优缺点,然后将这两种特征选择算法结合起来,取长补短,使得结合后的新算法不仅可以充分考虑不同特征组合带来的分类优势,同时又能保证算法的执行效率较高。新算法首先将特征集合进行删减,得到一个无关特征和冗余特征较少的特征子集;然后将该特征子集中相互冗余的特征划分到不同块中,相互不冗余的特征划分到同一块中;最后对特征块执行Top-r方法,选择最优的特征组合。实验表明,新算法不仅可以选择出较好的特征组合,又能保证算法的执行效率较高。更多还原

【Abstract】 With the continuous development of the information technology and constant improvement of information obtaining ability,people often need to analyze and deal with various high dimensional data,such as the mass of web data,remote sensing images,microarray data etc.These high dimensional data usually lead to the exponential increase of the calculation of machine learning algorithm,causing “the curse of dimensionality”.Therefore,feature selection technology for high dimensional data has become an important subject in the field of data mining.Feature selection technique maps high dimensional data from high-dimensional space to lowdimensional space which can better reflect the essential meaning of the data object and improve the efficiency of data analyzing and processing simultaneously.The paper has a deep research and discussion on the theoretical idea and practical application of feature selection technique for high dimensional data by taking the microarray data as the experimental data.A new feature selection algorithm based on feature similarity is proposed.Firstly,normalized signal to noise ratio algorithm is used for removing irrelevant features.Then surplus features are clustered into several clusters and clusters which only have little features are removed as noise features.After the removal,k clusters will be left,and the intra-cluster features redundancy is high,while features redundancy between the clusters is low.Finally,each feature of the clusters is evaluated successively according to the evaluation criteria which is proposed in this paper to decide whether or not to remove.In this way,the rest would be assembled and sorted according to the individual classification ability.The experiment confirms that the algorithm is valid in removing irrelevant features,noise features and redundancy features.A new algorithm can be achieved by analyzing the advantages and disadvantages of the feature selection algorithm based on feature similarity and the Top-r feature selection algorithm,combing them and learning from each other,which can not only fully consider the classification advantages but also guarantee high execution efficiency.First of all,feature set is cut for getting a feature subset with little irrelevant feature and redundancy features,then features in the same clusters are divided into different blocks and features in different clusters into the same.Finally,feature blocks are processed with Top-r algorithm to choose the optimum feature subset.The experiment confirms that the new algorithm can not only select the superior feature subset,but also guarantee high execution efficiency,fully affirming its superiority.更多还原