The success of machine learning approaches to solving real-world problems motivated the plethora of new algorithms. However, it raises the issue of algorithm selection, as there is no algorithm that performs better than all others. Approaches for predicting which algorithms provide the best results for a given problem become useful, especially in the context of building workflows with several algorithms. Domain knowledge (in the form of constraints, preferences) should also be considered and used to guide the process and improve results. In this work, we propose a meta-learning approach that characterizes sets of constraints to decide which constrained clustering algorithm should be employed. We present an empirical study over real datasets using three clustering algorithms (one unsupervised and two semi-supervised), which shows improvements in cluster quality when compared to existing semi-supervised methodolo-gies.