有噪条件下的新类检测算法

A Novelty Detection Algorithm in the Presence of Noise

针对有噪条件下新类检测性能较差的问题,提出一种基于核零空间判别局部保持投影算法(KNDLPP).首先通过核函数将样本隐式映射到高维特征空间,在核空间中利用距离加权机制对样本进行权重赋值,在保持局部结构的同时降低离群噪声样本的影响;然后利用样本类内零空间使同类样本坍塌为一点,实现对已知类分布的有效约简;最后基于零空间再求得使类间距最大化的变换矩阵,以上步骤得到一个判别性变换矩阵以刻画样本的分布信息、描述样本之间的相似性.该算法能刻画样本潜在结构,提升已知类与新未知类之间的判别性.在11个公开数据集上的实验结果表明,该算法是有效和鲁棒的,具有较好的新类检测性能.在局部保持性实验中,KNDLPP在4个UCI数据集上的整体平均AUC值为90.656%;在复杂结构保持性实验中,KNDLPP在Banana,Moon及3个UCI数据集上的整体平均AUC值为91.949%;在2个无噪高维数据集的新类检测实验中,KNDLPP平均AUC值为86.214%,高于次优算法4个百分点;在4个UCI数据集的4种有噪条件下,KNDLPP性能排名第1.

To address the poor performance of novelty detection in the presence of noisy samples,a method named kernel null space discriminant locality preserving projections(KNDLPP)is proposed.Firstly,the training samples are transformed into a high dimensional space through a kernel function implicitly,and different weights are assigned to these samples according to the distance weighted scheme in the kernel space,to preserve the locality while reducing the impacts of noisy samples.Then,through the kernel null space of intra-class,each class collapses to a point,which makes each known class concise efficiently.Finally,a projection matrix maximizing the distance among inter-classes can be computed based on the null space,thus after these steps a discriminative transformation matrix is got to characterize the distribution and similarity of samples.This method can grasp the underlying structure of samples,and improve the discrimination between the known classes and the unknown novelty.The comparison experiments are based on eleven public datasets,the results validate the effectiveness and robustness of proposal during the testing,and this method performs well for novelty detection.During the experiments about locality preserving on 4 UCI UCI datasets,the whole mean AUC of KNDLPP is 90.656%.During the experiments about complex structure on Banana,Moon and 3 UCI datasets,the whole mean AUC of KNDLPP is 91.949%.During the experiments on 2 clean high dimensional datasets for novelty detection,the whole mean AUC of KNDLPP is 86.214%,which is 4 percent higher than the second best algorithm.On 4 UCI datasets with 4 different kinds of noise,the performance of KNDLPP ranks first.