Research On Support Vector Data Description Based On Relative Density Degree

Novelty detection is a common problem which exists in modern medical research, newagricultural research and engineering security. Support vector data description (SVDD) is amethod for novelty detection. SVDD tries to establish a hypersphere and include all theknown normal samples. About detection problems, there are two kinds of data or the datawith only normal samples and the data with both normal and abnormal samples. Note thatthe number of abnormal samples is very small even if the data contains abnormal samples.Lee et al. introduced the concept of the relative density degree of samples, which canimprove the traditional SVDD algorithm. Based on the relative density degree and SVDD,this thesis has the following contributions.(1) Based on the Density-induced Support Vector Data Description (D-SVDD)algorithm, this thesis analyzes the parameter T and its influence on the performance ofalgorithm. We present a method for selecting the value of parameter T.(2) This thesis puts the abnormal sample information into the training process ofDensity-punished Support Vector Data Description (DP-SVDD) algorithm, and obtains amore stable and better result.(3) In this thesis, we introduce fuzzy membership function into ODP-SVDD to reducethe effect of noise normal points on the model. The new method effectively improves theability of ODP-SVDD.