Reduction schemes approximate the lower eigenvalues that represent the global behavior of the structures. But, they are not efficient to be applied to large-scaled problems because these schemes require considerable amount of computing time in constructing reduced one from the original large-scaled systems. In addition, the selection of the primary degrees of freedom might be localized to cause the excessive emphasis of the lower mode or lost of the important modes. In the present study, a new reduction method combined with the subdomain method is proposed. For the construction of the final reduced system the system of each domain subdivided into primary, slave and interface degrees of freedom. It is remarkably efficient and accurate comparable to full-scale system. Numerical examples demonstrate that the proposed method saves computational cost effectively and provides a reduced system which predicts accurate eigen-pairs of global system.