Efficient binary classification through energy minimisation of slack variables

Slack variables are utilized in optimisation problems in order to build soft margin classifiers that allow for more flexibility during training. A robust binary classification algorithm that is based on the minimisation of the energy of slack variables, called the Mean Squared Slack (MSS), is proposed in this paper. Initially, the algorithm is analysed for the linear case, where the minimum mean squared slack is attained as a separating vector. Next, the kernel trick is exploited to facilitate computation of non-linear separating hyperplanes. For this paper, two kernels are tested, namely the radial basis function (RBF) and the polynomial kernel. In order to ensure a time and memory efficient system that converges in a few iterations four strategies are applied so as to withhold just a subset of feature vectors that are misclassified during training. Aiming to the automatic optimisation of the kernel parameters a modern combination of particle swarm optimisation (PSO) with artificial immune system (AIS) is tested. The aforementioned evolutionary methods are combined in a parallel architecture. Four datasets of diverse nature are exploited for performance evaluation, namely the iris, the SPECTheart, the vertebral column, and the wine quality datasets. Simulation experiments demonstrate high classification accuracy in a number of benchmark datasets.