Authors:
Jaakko Peltonen
and
Ziyuan Lin
Affiliation:
Aalto University and University of Tampere, Finland
Keyword(s):
Parallel Coordinates, Visualization, Machine Learning, Dimensionality Reduction.
Related
Ontology
Subjects/Areas/Topics:
Abstract Data Visualization
;
Computer Vision, Visualization and Computer Graphics
;
General Data Visualization
;
High-Dimensional Data and Dimensionality Reduction
;
Information and Scientific Visualization
;
Visual Data Analysis and Knowledge Discovery
Abstract:
Parallel Coordinate Plots (PCPs) are a prominent approach to visualize the full feature set of high-dimensional
vectorial data, either standalone or complementing other visualizations like scatter plots. Optimization of
PCPs has concentrated on ordering and positioning of the coordinate axes based on various statistical criteria.
We introduce a new method to construct PCPs that are directly optimized to support a common data analysis
task: analyzing neighborhood relationships of data items within each coordinate axis and across the axes. We
optimize PCPs on 1D lines or 2D planes for accurate viewing of neighborhood relationships among data items,
measured as an information retrieval task. Both the similarity measurement between axes and the axis positions
are directly optimized for accurate neighbor retrieval. The resulting method, called Parallel Coordinate Plots
for Neighbor Retrieval (PCP-NR), achieves better information retrieval performance than traditional PCPs in
experiments.