Instance elimination strategy for non-convex multiple-instance support vector machine

Abstract

Multiple-instance support vector machine (SVM) is a classic and effective algorithm to address the classification of bags. However, training multiple-instance SVM is expensive. To deal with this issue, the safe screening rule is introduced to solve optimization problem, which can significantly mitigate the storage burden by identifying inactive instances. Specifically, an instance elimination strategy is designed for the inner solver of a signal concave–convex procedure (CCCP). Then, between the iterations of CCCP, due to the label of instance may change, to cope with this difficulty, a dual screening method with variational inequalities (DVI) is used to identify the part of inactive instances. To further speed up the solution efficiency, a smart dual coordinate descent method (SDCDM) is introduced for the inner solver, which skips over many updates that result in no change to the current iteration. Experiments on 35 benchmark datasets significantly exhibit the outstanding performance of the proposed strategy.