Social network systems rely on very large underlying graphs. Consequently, to achieve scalability, most data analytics and data mining algorithms are distributed and graphs are partitioned over a set of servers. In most real-world graphs, the edges and/or vertices have different semantics and queries largely consider this semantics. But while several works focus on efficient graph computations on these "multi-semantic" graphs, few ones are dedicated to their partitioning. In this work, we propose a novel approach to achieve edge partitioning for multi-layer graphs, which considers both structural and edge-types (labels) localities. Our experiments on real life datasets with benchmark graph applications confirm that the execution time and the inter-partition communication can be significantly reduced with our approach.