Conventional wisdom considers that multi-dimensional scaling (projection) is good at representing the big picture, whereas hierarchical clustering handles local details more faithfully.
Ideally, one would want a low-dimensional configuration rendering both the global and local properties; such that, for example, if one drew between the final points a hierarchical tree obtained from the original data, this tree would appear simple and related branches would stay next to each other.
Our research adapts existing scaling algorithms to create such tree-friendly configurations, able to capture conceptual structure in a visually striking way." - Tree-Expansion Strategy & Conceptual Structures Visialization.
зы: спасибо urbansheep-е за "символическую" поддержку.