The Role of (non-)Topological Features as Drivers of Systemic Risk: a machine learning approach