Clustering methods are not restricted to a particular topological type of their input and output data. In the simplest and fastest case vertices are clustered by discrete gridding and coordinate truncation. All vertices in one cluster are replaced by a single vertex. Edges and triangles that have become degenerate are removed. This procedure results in a simpler mesh, which might have a different topological type.
An interesting approach is described in [4]. It combines techniques used for Progressive Meshes and clustering methods. The authors report very good results for objects consisting of many separate parts.