Moment-matching is a state of the art projection-based model order reduction method. Using moment-matching, the reduced subspace becomes very large when dealing with systems with many inputs or outputs. For mechanical systems, this can for example be the case when subsystems are reduced individually and interconnected to an overall system after the reduction or when dynamic loads may occur on the entire surface of the structure. Typical methods to overcome this issue and to reduce the number of inputs and outputs are tangential directions, that interpolate the inputs or outputs along a specific direction. This contribution aims at comparing different methods to reduce the influence of the number of inputs and outputs on the dimension of the reduced order model. Therefore an error estimator that is able to estimate the error from every input to every output is used. This knowledge can be used with little additional computational effort in each iteration of the Greedy procedure to identify the inputs that have the largest transfer function error at the next frequency shift. For the next reduction step, only those inputs where the error is larger than a specific threshold are taken into account. This method is than compared to a combination of the error estimator and tangential directions.