DimenFix: a Novel Meta-Dimensionality Reduction Strategy for Feature Preservation

Dimensionality Reduction (DR) methods have become essential tools for the data analysis toolbox. Typically, DR methods combine features of a multi-variate dataset to produce dimensions in a reduced space, preserving some data properties, usually pairwise distances or local neighborhoods. Preserving such properties makes DR methods attractive, but it is also one of their weaknesses. When calculating the embedded dimensions, through usually non-linear strategies, the original feature values are lost and not explicitly represented in the spatialization of the produced layouts, making it challenging to verify the features' contribution to the attained representations. Some strategies have been proposed to tackle this issue, such as coloring the DR layout or generating explanations. Still, they are post-processes, so specific features (values) are not guaranteed to be preserved or represented. This paper proposes DimenFix, a novel meta-DR strategy that explicitly preserves the values of a particular feature or external data (e.g., class, time, or ranking) in one of the embedded dimensions. DimenFix works with virtually any gradient-descent DR method and, in our results, has shown to be capable of representing features without heavily impacting distance or neighborhood preservation, allowing for creating hybrid layouts joining characteristics of scatter plots and DR methods.