A tree-structured framework for purifying "complex" clusters with structural roles of individual data