Data Visualization with Structural Control of Global Cohort and Local Data Neighborhoods

In Abstract

A mathematical solution to help humans see through high volume data by trusting their eyes

In data analytics, it has always been of high interest to study ways to create effective visual representations for large volumes of high-dimensional measurements, so that human viewers can acquire insight and understand underlying data properties and intrinsic structures based on their cognitive and perceptual skills. One popular strategy is to preserve the local relationships and similarities between neighbouring individual patterns in a visible, low-dimensional (e.g., 2D or 3D) space. A great deal of effort has been put to developing such algorithms in the past two decades.

Now, computer scientists at the Universities of Manchester and Liverpool have discovered that this type of “local” strategy can result in incorrect approximation of the global data structure and introduce substantial noise to the positioning of data groupings and cohorts in the visualising space. In addition to analysing this phenomenon, the researchers propose a mathematical solution to effectively “correct” such noise, together with a theoretical learning framework and nine algorithmic variations. The empirical performance of the proposed solution is extensively studied for five application problems and the researchers demonstrate more reliable visualisation output. An important target application of this work is to visualise high-dimensional data produced by various industries, such as ones related to biology, healthcare, media, marketing, economics, and research fields across the clinical, social and natural sciences and engineering.

Click here to read the full article - DOI 10.1109/TPAMI.2017.2715806