Granular Support Vector Machines Using Linear Decision Hyperplanes for Fast Medical Binary Classification
Yuchun Tang, Bo Jin, Yanqing Zhang, Hao Fang, Binghe Wang · 2005
In many life science applications, due to the requirement of real time data processing and/or the very large size of the available dataset, a classifier with high efficiency is usually more preferable, or even necessary, with the prerequisite of not deteriorating effectiveness too much. That means a more desirable classifier in this context should run faster but still retain high accuracy. In this paper, we show a simple but fast method for modeling a linear granular support vector machine by splitting the feature space to two smaller subspaces and then building a SVM for each of them. The hyperplane used to halve the feature space is searched by applying the extended statistical margin maximization principle along the direction orthogonal to the first principle component. One public medical dataset is used to compare the resulting linear GSVM to one optimized SVM with the radial basis function (RBF) kernel in the whole feature space. The experimental results show that finding the splitting hyperplane is not a trivial task and the linear GSVM is even a little better than the optimized RBF-SVM in terms of testing accuracy, but the linear GSVM is more robust against noises, more stable to model parameters, and runs much faster. In general, GSVM provides an interesting new mechanism, which is competitive to kernel mapping methods, to address complex classification problems effectively and efficiently