基于自适应距离的离群点检测算法

Adaptive Distance Based Outlier Detection Algorithm

基于近邻的离群点检测方法根据数据对象周围的邻居来挖掘离群点,但该类方法受阈值参数的影响较大,且大多只在数据分布单一的情况下表现良好。针对数据分布多样的情况下离群点检测困难以及阈值参数的敏感性问题,提出了一种基于自适应距离的离群点检测算法。首先,通过动态地调整数据属性的贡献因子,使得关键属性在离群点检测中具有更大的影响力,能够准确反映关键属性与离群点之间的关联性;其次,综合考虑属性贡献因子和密度来计算数据对象之间的距离,以便更好地识别数据对象之间的位置关系和密度分布特征;最后,为了降低阈值参数的影响,逐步增大邻居的大小来计算数据对象的自适应距离的变化之和,将其累加作为离群得分。通过在人工合成数据集和公共数据集上进行实验,验证了提出的算法检测精度更高。

Near-neighbour based outlier detection methods mine outlier points based on the neighbours around the data object,but this type of method is greatly affected by the threshold parameter and mostly performs well only in the case of a single data distribution.Aiming at the difficulty of outlier detection in the case of diverse data distribution and the sensitivity of threshold parameters,an adaptive distance-based outlier detection algorithm is proposed.Firstly,by dynamically adjusting the contribution factor of data attributes,the key attributes have more influence in outlier detection,which can accurately reflect the correlation between the key attributes and outliers.Secondly,the distance between data objects is calculated by comprehensively considering the contribution factor of attributes and the density,so as to better identify the positional relationship between data objects and the density distribution characteristics.Lastly,in order to reduce the threshold parameter's influence,the size of neighbours is gradually increased to calculate the sum of changes in adaptive distances of data objects,which is accumulated as the outlier score.The proposed algorithm is verified to have higher detection accuracy through experiments on synthetic datasets and public datasets.