SADIRE: a context-preserving sampling technique for dimensionality reduction visualizations

Sampling techniques are widely used in the effort to reduce complexity and improve interpretability of datasets. Given the enormous availability of data, these techniques try to select representative data points that inherently reflect the data structure. In this work, we propose a novel sampling technique that preserves the structures imposed by dimensionality reduction techniques when visualized as scatter plots. In the experiments, we demonstrate how our technique is able to reflect the class boundaries and layout structures, besides decreasing redundancy of the datasets visualized as scatter plots. We also provide an user experiment regarding the perception of sampling from scatter plot visualizations. Graphic abstract