一种基于核空间局部离群因子的离群点挖掘方法

Outlier Mining Based on Kernel Local Outlier Factor

提出了一种基于核空间局部离群因子(KLOF)的离群点挖掘方法。该方法通过核函数将数据集映射到特征空间,然后在特征空间计算每个模式的局部离群因子。该方法继承了基于密度的局部离群因子(LOF)的优点,可以定量地描述每个模式的离群程度;同时又克服了LOF的不足,对线性不可分的数据,可以取得比较好的分析结果。通过两个仿真的和两个真实的数据集对KLOF及LOF方法进行了比较,结果表明,KLOF具有适应的数据集范围宽,识别率高等优点。

A novel method for outlier mining based on kernel local outlier factor（ KLOF） is proposed. KLOF of each pattern is computed after the dataset is mapped to the feature space by a certain kernel function. KLOF inherits the advantages of traditional LOF. For example,it can describe the outlying degree of the pattern quantitatively. Meanwhile,KLOF overcomes LOF for the dataset with non-linear separable clusters. The KLOF method is compared with LOF using two simulated datasets and two real datasets. The results show that KLOF outperforms other methods in terms of dataset applicability and recognition rate.