Abstract: Support vector data description (SVDI))is considered as a classical method for novelty detection. As is well known, the parameter setting and the quality of features arc two key points to affect the performance of SVDD. Combi- ning feature extraction and parameter selection of SVDD,this paper proposed a simulated annealing approach for feature extraction and parameter selection of SVDD (SA-SVDD). During the procedure of simulated annealing, the optimal ker- ncl parameter, tradeoff parameters, and number of extracted features arc automatically selected. Experimental results on the UCI benchmark data sets demonstrate that SA-SVDI)has better performance than the traditional parameter selec- tion methods.

Key words: Feature extraction, Simulated annealing, Parameter selection, SVDD, Novelty detection