基于隐高斯模型的多元离散数据异常检测

ANOMALY DETECTION FOR MULTIVARIATE DISCRETE DATA BASED ON LATENT GAUSSIAN MODEL

异常检测在许多实际应用中非常重要,尤其在精密度控制领域。针对异常检测,提出了利用概率分布的思想来解决该问题,即学习出正常行为的概率分布,并将低概率的行为视为异常。传统方法中常用多项式和狄利克雷多项式分布作为学习正常行为概率分布的模型。但是,当面对小规模多元离散数据集的时候,这些模型不得不面对维数灾难的问题,并很难捕获常规行为的分布特性。为解决此难题特提出一种基于贝叶斯学习的技术——隐高斯模型,它能通过使用高斯过程对于这些多变量的分类样本学得一个连续的隐空间从而可以对小规模数据集建模。一系列的试验结果表明,隐高斯模型的方法相比于其他的异常检测技术来说更加有效。

Anomaly detection is very important in many practical applications, especially in the area of precision controlling. Some methods based on probability distribution which try to learn the probabilistic distribution of normal behavior have been proposed to solve this issue. Various traditional methods can be used to learn the probabilistic distribution of normal behavior such as multinomial or Polynomial distribution. However, when faced with discrete small-scale data set, these methods were suffered from the curse of dimensionality and difficult to capture the statistical properties of conventional behavior. Therefore, we proposed a method based on Bayesian which is Latent Gaussian models. LGMs could model for small-scale data set through learning a continuous latent space for multivariate categorical samples using Gaussian process. Experimental results show that our method achieves better detection performance on small-scale data set compared with other anomaly detection methods.