Abstract: Outlicrs arc objects that do not comply with the general behavior of the data. SVM(support vector machine)finds the maximal margin hyperplane in feature space for the purpose of distinguishing the outliers from normal samp1es. Based on the high performance of SVMs in tackling small sample size, high dimension and its good generalization,we proposed a new method for outlicr detection, which combines a novel unsupervised algorithm GPLVM(Gaussian process latent variable model) with standard SVM. GPLVM provides a smooth probabilistic mapping from latent to data space, embeds the dataset in a low-dimensional space which is used for cross validation of SVM I'he proposed approach was applied to KDD99 benchmark problems, and the simulation results show its validity.

Key words: Outlier detection, Support vector machine, Dimensionality reduction, GPLVM