Top-Frequency Parallel Coordinates Plots

TL;DR

This paper introduces a frequency-based parallel coordinates plotting method that effectively visualizes large multivariate datasets, including discrete variables and missing values, by discretizing continuous variables and selecting high-count patterns.

Contribution

It extends existing frequency-based visualization techniques to handle discrete variables and missing data, improving pattern detection in large multivariate datasets.

Findings

Effective visualization of large datasets with discrete variables.

Novel approach to handling missing values in parallel coordinates.

Enhanced pattern detection in high-dimensional data.

Abstract

Parallel coordinates plotting is one of the most popular methods for multivariate visualization. However, when applied to larger data sets, there tends to be a "black screen problem," with the screen becoming so cluttered and full that patterns are difficult or impossible to discern. Xie and Matloff (2014) proposed remedying this problem by plotting only the most frequently-appearing patterns, with frequency defined in terms of nonparametrically estimated multivariate density. This approach displays "typical" patterns, which may reveal important insights for the data. However, this remedy does not cover variables that are discrete or categorical. An alternate method, still frequency-based, is presented here for such cases. We discretize all continuous variables, retaining the discrete/categorical ones, and plot the patterns having the highest counts in the dataset. In addition, we propose some novel approaches to handling missing values in parallel coordinates settings.