A generalized Mitchell-Dem'yanov-Malozemov algorithm for one-class support vector machine

The dual problem of one-class support vector machine (OCSVM) can be interpreted as a minimum norm problem associated with the reduced convex hull. Based on this geometric interpretation, a generalized Mitchell-Dem'yanov-Malozemov (GMDM) algorithm is proposed for OCSVM. The GMDM algorithm finds the minimum norm point in the reduced convex hull of training samples and employs such a point to construct the separating hyper-plane. Numerical experiments are conducted to compare the proposed geometric algorithm with some existing algorithms such as two modified sequential minimal optimization algorithms and the generalized Gilbert algorithm. The experimental results show that the GMDM algorithm exhibits better performance in terms of computational efficiency while achieving comparable classification accuracies to other algorithms.