基于深度信念网络和线性单分类SVM的高维异常检测

High-dimensional outlier detection based on deep belief network and linear one-class SVM

针对目前高维数据异常检测存在的困难,提出一种基于深度信念网络和线性单分类支持向量机的高维异常检测算法。该算法首先利用深度信念网络具有良好的特征提取功能,实现高维数据的降维,然后基于线性核函数的单分类支持向量机实现异常检测。选取UCI机器学习库中的高维数据集进行实验,结果表明,该算法在检测正确率和计算复杂度上均有明显优势。与PCA-SVDD算法相比,检测正确率有4.65%的提升。与自动编码器算法相比,其训练和测试时间均有显著下降。

Aiming at the difficulties in high-dimensional outlier detection at present,an algorithm of high-dimensional outlier detection based on deep belief network and linear one-class SVM was proposed.The algorithm firstly used the deep belief network which had a good performance in the feature extraction to realize the dimensionality reduction of high-dimensional data,and then the outlier detection was achieved based on a one-class SVM with the linear kernel function.High-dimensional data sets in UCI machine learning repository were selected to experiment,result shows that the algorithm has obvious advantages in detection accuracy and computational complexity.Compared with the PCA-SVDD algorithm,the detection accuracy is improved by 4.65%.Compared with the automatic encoder algorithm,its training time and testing time decrease significantly.