【作者】 张婷；

【导师】 袁方；

【作者基本信息】 河北大学 ， 软件工程， 2018， 硕士

【摘要】 随着信息技术的快速发展和广泛应用,各行各业积累了大量的数据信息,如何对这些海量数据进行有效处理,并从中挖掘潜在的有用信息,是目前的一个重要研究领域。聚类分析作为数据挖掘领域重要的工具之一显得愈发重要,迄今为止,研究人员提出了多种聚类算法,并被广泛应用在图像处理、模式识别、自然语言处理等领域。然而,传统的聚类算法存在一些不足,如对离群点敏感、鲁棒性不佳等,聚类性能有待进一步提高。本文把截断思想融入聚类算法研究,用以提高聚类算法的鲁棒性和聚类性能。论文主要工作如下:1.给出了截断式鲁棒K-means算法(Capped Robust K-means Algorithm,CRK-means)。针对传统K-means聚类算法对离群点敏感这一不足,结合截断式思想,将去噪比例因子引入目标函数中,同时,通过在目标函数求解过程中添加辅助变量使得在每一次迭代中实现动态更新样本权重值,从而降低离群点对算法的影响并提高了算法的鲁棒性和准确度。2.给出了截断式鲁棒子空间聚类方法(Capped Robust Subspace Clustering,CRSC)。为了解决子空间聚类对噪声敏感的问题,结合截断式函数惩罚噪声项,减少了数据样本中存在的离群点对构造系数矩阵的影响,提高了算法的鲁棒性。同时,该方法从样本间相关性出发对所有样本进行自表示,并且加入局部相似性约束,在保证全局约束的同时加强了样本间局部结构表达。该方法能得到更优的邻接矩阵,从而得到更加鲁棒的聚类结果。3.对CRK-means和CRSC算法进行了实验验证。选取人工数据集、UCI标准数据集以及图像数据集,对所给出的算法进行了实验研究以及参数分析。实验结果表明,本文给出的CRK-means算法和CRSC算法在提高算法鲁棒性的同时提高了聚类的准确度。更多还原

【Abstract】 With the rapid development and widespread application of information technology,various industries have accumulated a large amount of data.How to effectively deal with these data and obtain potential and useful information from it is an important research area at present.Cluster analysis is becoming more and more important as one of the important tools in the data mining field.So far,researchers have proposed many clustering algorithms,and are widely used in image processing,pattern recognition,natural language processing and other fields.However,the traditional clustering algorithm has some shortcomings,such as being sensitive to outliers,poor robustness,etc.And the clustering performance needs to be further improved.This thesis combines the capped function with the clustering algorithm to improve the robustness and clustering performance of the clustering algorithm.The main research work of this thesis is as follows:1.This thesis proposed a capped robust K-means algorithm(CRK-means).For the disadvantage of the traditional K-means clustering algorithm which is sensitive to outliers,the proposed method adds the idea of capped function,and introduces the denoising value factor into the objective function.At the same time,by adding auxiliary variables during the objective function solving process.In each iteration,the dynamic update of sample weight values is implemented,thereby reducing the impact of outliers on the algorithm and improving the robustness and accuracy of the algorithm.2.The capped robust subspace clustering Algorithm(CRSC)is proposed.In order to solve the problem of subspace clustering is sensitive to noise,this thesis combines the capped function to punish the noise term,which reduces the influence of outliers in the data sample on the construction coefficient matrix and improves the robustness of the algorithm.At the same time,this method self-represents all samples from the correlation between samples and adds local similarity constraints.While ensuring global constraints,it strengthens the expression of local structures among samples.This method can obtain a better affinity matrix,so as to obtain more robust clustering results.3.Experiments on the capped robust K-means algorithm and capped robust subspace clustering Algorithm are studied in this thesis.By choosing UCI standard data sets,synthetic data sets and image data sets,experimental analysis and parameter analysis of the proposed algorithms are performed.The results show that the CRK-means algorithm and CRSC algorithm presented in this thesis can improve the robustness of the algorithm and improve the clustering accuracy.更多还原