Shape analysis aims to infer precise structural properties of imperative memory states and has been applied heavily to verify safety properties on imperative code over pointer-based data structures. Recent advances in shape analysis based on separation logic has leveraged sum-marization predicates that describe unbounded heap regions like lists or trees using inductive definitions. Unfortunately, data structures with un-structured sharing, such as graphs, are challenging to describe and reason about in such frameworks. In particular, when the sharing is unstructured, it cannot be described inductively in a local manner. In this paper, we propose a global abstraction of sharing based on set-valued variables that when integrated with inductive definitions enables the specification and shape analysis of structures with unstructured sharing.