One of the most effective and successful approaches to construct effective Euclidean representation of graphs, i.e., embedding, are based on random walks. This solution allows to learn a latent representation by capturing nodes similarities from a series of walk contexts while optimizing the likelihood of preserving their neighborhood in the Euclidean space. In this paper, we address the question of enhancing the existing random walks based methods by making the walk generation process aware of the global graph structure. To this end, we propose a node embedding method based on random walks guided by weights calculated considering a macro observation of the graph structure rather than a micro one. Experimental results on both synthetic and real-world networks show that our computed embedding allows to reach better accuracy rates on two tasks: node classification and link prediction.