Preserving Vector Space Properties in Dimensionality Reduction: A Relationship Preserving Loss Framework