核模糊谱聚类LOF降噪方法研究

为解决核模糊相似性度量谱聚类算法的样本点降噪问题，优化聚类效果和稳定性，本文从分析异常点分布特性出发，引入局部异常因子（LOF）算法，提出聚类中心候选对象的概念，过滤数据集的噪声数据，从而优化初始聚类中心的计算，突出正常样本点在聚类中心调整中的影响力，使聚类算法更易于得出准确的聚类结果. 同时提出一种局部过滤因子以修正相似性度量的方法，该方法通过放大正常数据之间的权值、缩小正常数据与噪声数据间的权值，使优化后的核模糊谱聚类算法大大降低对异常点的敏感度. 算法有效性实验和算法稳定性实验表明：该方法对相似性度量修正的有效性使核模糊谱聚类算法更为稳定和鲁棒.

A Research on Local Outlier Factor De-noising Method for Kernel Fuzzy Spectral Clustering

To deal with noise reduction of kernel fuzzy spectral clustering for obtaining a better cluster capability, a new spectral clustering based on local outlier factor was presented. With the proposed method the distribution features of outliers were analyzed. Based on the analysis result, the cluster center candidate object was proposed by using local filter factor algorithm. And then a process was constructed to filter the noise data and highlighted normal data's influence in the clustering center adjustment. Secondly, the local filter factor was presented by using local outlier factor between two arbitrary objects. Then the local filter factor was used as weighting factor to improve the similarity measure of kernel fuzzy spectral clustering. The improved similarity measure made weight between normal data and normal data became bigger and weight between normal data and noise data smaller. Therefore, the improved kernel fuzzy spectral clustering can reduce greatly the sensitivity of outliers. The validity experiment and stability experiment results show the proposed method has better clustering accuracy and robustness.