Multilayer networks are a widely used model that provide a more realistic representation of the heterogeneous relationships that can characterise a complex system. Nevertheless, taking into account this complex information is a major challenge due to the noise contained in the data and the choice of entities and relationships to be taken into account in the analysis. For this reason, multi-layer simplification techniques have been proposed in order to select important information, reduce the computational burden and ameliorate the quality of the analysis as well as the visualisation of the information contained in these multilayer networks. Unfortunately, all the techniques proposed so far are task-agnostic and rely on unsupervised heuristics. In this work, we propose a framework to simplify multi-layer networks according to the final downstream task. We rely on two main components: i) a node relationship simplification module and ii) a (multi-layer) graph neural network to generate node embeddings for the purpose of solving a specific task. Here, we tackle the node classification task as goal but the method is directly transposable to other (supervised) multilayer network analysis tasks. Experimental results on different real multilayer networks prove the quality of our approach which provides a suitable simplification for the node classification task.