基于自然邻居邻域图的无参数离群检测算法

数据挖掘领域，基于最近邻居思想的离群检测算法在面对复杂数据时，很难在没有足够先验知识条件下进行适当的参数选择。为了解决这个问题，本文在自然邻居方法的基础上，提出一种利用加权自然邻居邻域图进行离群检测的算法。该算法在整个过程不需要人为设置参数，并且能在不同分布特征的数据中准确找到数据集中的全局离群点和局部离群点。人工数据集和真实数据的离群检测结果均证明，本算法能够取得和有参数的算法中最优参数相近的效果，算法检测结果远好于对参数敏感算法的大部分情况，且更优于对参数不敏感的算法，具有更强的普适性和实用性。

A parameter-free outlier detection algorithm based on natural neighborhood graph

This study aims to deal with the practical shortages of nearest-neighbor-based data mining techniques, particularly outlier detection. In particular, when data sets have arbitrarily shaped clusters and varying density, determining the appropriate parameters without a priori knowledge becomes difficult. To address this issue, on the basis of the natural neighbor method, which can better reflect the relationship between elements in a data set than the k-nearest neighbor method, we present a graph called the weighted natural neighborhood graph for outlier detection. The weighted natural neighborhood graph does not need to set parameters artificially in the entire process and can identify global and local outliers in the data set with different distribution characteristics. The outlier detection results of artificial dataset and real data prove that the algorithm can obtain an effect similar to that of the optimal parameter in the algorithm with parameters. The algorithm detection result is far better than that of most parameter-sensitive algorithms and is much better than that of the parameter-insensitive algorithm, which has stronger universality and more practicality.