Stabilizing and Simplifying Sharpened Dimensionality Reduction using Deep Learning

Abstract

Dimensionality reduction (DR) methods create 2D scatterplots of high-dimensional data for visual exploration. As such scatterplots are often used to reason about the cluster structure of the data, this requires DR methods with good cluster preservation abilities. Recently, Sharpened DR (SDR) was proposed to enhance the ability of existing DR methods to create scatterplots with good cluster structure. Following this, SDR-NNP was proposed to speed the computation of SDR by deep learning. However, both SDR and SDR-NNP require careful tuning of four parameters to control the final projection quality. In this work, we extend SDR-NNP to simplify its parameter settings. Our new method retains all the desirable properties of SDR and SDR-NNP. In addition, our method is stable vs setting all its parameters, making it practically a parameter-free method, and also increases the quality of the produced projections. We support our claims by extensive evaluations involving multiple datasets, parameter values, and quality metrics.