Software metrics have been developed to measure the quality of software systems. A proper use of metrics requires thresholds to determine whether the value of a metric is acceptable or not. Many approaches propose to define thresholds based on large analyses of software systems. However it has been shown that thresholds depend greatly on the context: the programming language or the application domain of the project for example. Thus there is a need for an approach that computes thresholds by taking into account this context. In this paper we propose such approach with the objective to reach a trade-off between representativeness of the threshold and computation cost. Our approach is based on an unbiased selection of software entities and makes no assumptions on the statistical properties of the software metrics values. It can therefore be used by any one, ranging from developer to manager, for computing a representative metric threshold tailored to their context.