Semantic alignment can diverge from local neighborhood preservation in dimensionality reduction visualizations

Abstract Visualizations produced by dimensionality reduction are routinely interpreted as if spatial proximity reflects semantic similarity. This assumes that two properties—semantic alignment (whether pairwise distances in the layout track a semantic reference space) and structural preservation (whether the original relational geometry is maintained)—move together, which is rarely examined directly. Here we show that semantic alignment and local neighborhood preservation are dissociable: they respond differently to controlled manipulations and can diverge in empirical dimensionality-reduction demonstrations. We separate the two with the SSC-SP framework—Semantic-Spatial Correlation (SSC), a global distance-rank alignment measure, and Structural Preservation (SP), decomposed into local adjacency, global order, and cluster components. Under coordinate perturbations, SP degrades while SSC remains near zero; when semantic information is injected into a layout, the two trade off by construction. Applied to human similarity judgments (THINGS dataset), the method with the highest semantic alignment (MDS) showed the lowest local neighborhood preservation—a divergence not captured by single-metric evaluation. A component-wise analysis shows that this divergence is carried by local adjacency preservation, whereas the global order component is near-redundant with SSC. SSC-SP is best understood as a lightweight, reproducible reporting framework for decomposing visualization quality, rather than as a formal probabilistic test.