A new weighted SVDD algorithm for outlier detection

Traditional Support Vector Domain Description (SVDD) and some improved SVDD algorithms are suboptimal in detecting some complex outliers. To solve the difficulty, we proposed a novel adaptive weighted SVDD (AW-SVDD) algorithm. Firstly, a weighting is computed for each data point based on their spatial position distribution and density distribution in a training dataset. The weighting can be used to measure the degree of the data point to be an outlier. Then the weighted data are trained by traditional SVDD. Lastly, a sphere shaped data description can be obtained for the training data. Experimental results demonstrate that AW-SVDD can overcome the interference from some complex outliers objectively, therefore the algorithm has a better performance than traditional SVDD and other improved SVDD algorithms.