A Tabu Search Meta-Heuristic for Image Semi-Supervised Classification

We investigate the utility of tabu search (TS) meta-heuristics for semi-supervised image classification tasks. The proposed heuristic solves the integer programming transductive support vector machine (MIP-TSVM) formulation considered. Preliminary results, with a linear kernel show that our TS implementation can effectively find optimal global solutions for TSVM with relatively large problem dimensions and is competitive, in terms of generalization performance, with transductive SVM^light package on LIBSVM benchmarks. However on corel image database, TSVM^light demonstrates superior performance. As a result, the usefulness of such MIP-TSVM formulation may be application dependant