摘要
自动聚类算法(AutoClass)是一种非监督的能对复杂数据进行精确的自动聚类的有效分类方法,可以事先设定好类别数目让AutoClass自动寻找,在寻找结束后,能够得到每一条数据分别属于每一类别的几率,这样可以根据专业知识,选出比较好的分类效果.描述了使用AutoClass对SDSS DR6的恒星/星系测光数据进行分类,将868974条测光数据进行处理,通过去离群数据和自动聚类的方法,将最终的812613条数据分成两类,其中星系和恒星的数据分别是680361和126988条.对于去掉离群后的数据,星系和恒星的分类正确率分别达到99.51%和98.52%,表明AutoClass算法对去掉离群数据后的恒星/星系数据分类有很好的效率.因此,可以将该算法应用于天文中的其他分类问题,另外基于该算法的非监督性,可以帮助天文学家去掉离群数据或发现一些特殊天体.
AutoClass is an unsupervised valid classification algorithm which can carry on accurately automated clustering on complex data, set the number of classification in advance and perform AutoClass to search after, then get a probability of every data belonging to some type, and finally decide a better classification result by means of professional knowledge. Here the AutoClass algorithm is used to classify stars/galaxies with the photometric data of SDSS DR6. 868974 photometric data records are selected for classification. Firstly Autoclass is applied on these data to delete outliers, then utilized on the rest of 812613 data records to classify stars and galaxies. The number of galaxies and stars is 680361 and 126988, respectively. Their accuracy for galaxies and stars adds up to 99.51% and 98.52%, respectively. Obviously, the AutoClass algorithm obtains a better efficiency and effect on this classification problem. Therefore this algorithm can be applied for other classification problems in astronomy. In addition, given the unsupervised characteristic of this algorithm, it may help astronomers to remove the outliers or find some unusual objects.
基金
国家自然科学基金(批准号:10778724
10778616)
国家高科技研究发展计划(编号:2006AA01A120)资助项目