Visualizing Dimensionality Reduction Artifacts: An Evaluation

Multidimensional scaling allows visualizing high-dimensional data as 2D maps with the premise that insights in2D reveal valid information in high-dimensions, but the resulting projections always suffer from artifacts such asfalse neighborhoods and tears. These artifacts can be revealed by interactively coloring the projection accordingto the original dissimilarities relative to a reference item. However, it is not clear if conveying these dissimilaritiesusing color and displaying only local information really helps to overcome the projections artifacts. We conducteda controlled experiment to investigate the relevance of this interactive technique using several datasets. We comparedthe bare projection with the interactive coloring of the original dissimilarities on different visual analysistasks involving outliers and clusters. Results indicate that the interactive coloring is effective for local tasks as itis robust to projection artifacts whereas using the bare projection alone is error prone