We investigate the utility of Reactive Tabu Search (RTS) meta-heuristic for semi-supervised classification tasks. We use RTS to solve the primal Mixed Integer Programming Transductive Support Vector Machine (MIP-TSVM) formulation considered in [7]. The proposed heuristic is an extension of the classical Tabu Search (TS) and can automatically adjust the generic parameters of TS and somehow learn during the search process. Preliminary results, with a linear kernel show that our RTS implementation can effectively find optimal global solutions for TSVM with relatively large problem dimension and is competitive, in term of generalization performance with Transductive SVMlight package on some LIBSVM benchmarks. Key Words: Reactive Tabu Search, Support Vector Machines, Mixed Integer Programming, Semi-Supervised Learning, Transductive inference.