基于模糊大间隔最小球分类模型的恒星光谱离群数据挖掘方法

Study on Stellar Spectral Outliers Mining Based on Fuzzy Large Margin and Minimum Ball Classification Model

由于人类对宇宙的认识有限,因此,如何通过对光谱数据分析发现一些新的、特殊的天体成为天文学家面临的重要课题。目前,常见特殊天体发现方法的基本思想是利用智能分类算法对离群数据进行分析。然而,当前主流分类算法大多对离群数据不敏感,分类性能甚至受离群点影响较大,因而无法完成特殊天体发现任务。鉴于此,提出基于模糊大间隔最小球分类模型的离群数据挖掘方法,该方法利用部分一般样本和离群样本建立最小球模型,并在此基础上引入模糊技术,通过降低噪声的权重,尽量减少噪声的影响。与C-SVM,SVDD,KNN等传统分类方法在SDSS恒星光谱数据集上的比较实验表明所提方法的有效性。

It＇s one of the main goals in universe exploration to find unknown and special celestial bodies.The spectra outlier data is analyzed based on the traditional classification approaches,which is a general method of special celestial body exploration.But it＇s depressed that many traditional classification approaches are insensitive to the outlier data,which even influence the classification efficiencies,therefore,these methods can＇t accomplish the task of special celestial body exploration.In view of this,Fuzzy Large Margin and Minimum Ball Classification Model（FLM-MBC）is proposed in this paper.In FLM-MBC,part of general data and outlier data are trained to construct the minimum ball model and the fuzzy technique is introduced to reduce the noise influence to classification.Comparative experiments with C-SVM,KNN,and SVDD on the SDSS spectral datasets verify the effectiveness of the proposed FLM-MBC.