Localized pairwise constraint proximal support vector machine

Proximal support vector machine (PSVM) adopts the class labels as priori and performs the similar level of accuracy as the regular SVM and is significantly faster. However, PSVM dose not take the local structure of the data points into account. In this paper, by introducing the pairwise constraints as priori, we propose a Localized Pairwise Constraint Proximal Support Vector Machine (LPCPSVM) for classification learning. The central idea is to find a projection vector such that can assure the maximum margin of the SVM hyperplane, and considers improving the tightness among distances between the similar data pairs under the Must-link constraint, while expanding the distances between the dissimilar ones under the Cannot-link constraint. We also show that LPCPSVM can be extended to non-linear KLPCPSVM by using the standard kernel trick and demonstrate the practical usefulness and good performance of the proposed algorithms for classification through extensive simulations with benchmark datasets. Experimental results show that our method can select the good features and has comparable test correctness and faster computational time to that of PSVM classifiers.