We consider the problem of learning representations of relational data in spaces of constant sectional curvature, i.e., Euclidean, Hyperbolic, and Spherical space. In this context, we explore how to identify a suitable embedding curvature for a given relational dataset. For this task, we investigate the use of a scalable heuristic based on local graph neighborhoods and evaluate it on classic benchmark graphs.