Visual Identification and Extraction of Intrinsic Axes in High-Dimensional Data

Jiazhi Xia, Fenjin Ye, Fangfang Zhou, Yi Chen, Xiaoyan KuiIEEE Access 2019 · DOI

Introduction

Visualizing high-dimensional data often involves projecting it onto a 2D plane. However, standard linear techniques like Principal Component Analysis (PCA) struggle when data lies on a curved "manifold." Researchers have developed a new visual analytics system that allows users to interactively "unfold" high-dimensional data by extracting its intrinsic, non-linear axes. By combining topological skeleton extraction with a graph-based diffusion algorithm, the system enables users to define meaningful axes that follow the actual shape of the data manifold, even when it contains complex branches and holes.

The Challenge of High-Dimensional Visualization

Standard projection techniques often act as "black boxes," where the resulting axes lack clear semantic meaning and global structures can be distorted. Real-world datasets frequently possess complex topologies with multiple branches or holes. Projecting a data point from one branch onto an axis defined by a completely different branch is mathematically possible but semantically nonsensical. There is a clear need for tools that respect both local and global data structures while providing interactive control over axis definition.

Proposed Methodology

The proposed workflow centers around a k-Nearest Neighbors (kNN) graph, which serves as a proxy for the data's intrinsic structure. The approach involves several novel steps:

  1. Topology Extraction: Inspired by the "Mapper" method, the system builds a skeleton of the dataset. It uses 1D Multi-Dimensional Scaling (MDS) to color-code points by their geodesic distance and clusters them into a simplified graph of nodes and edges.
  2. Interactive Axis Definition: Users define an "intrinsic axis" by either selecting a sequence of data points (finding the shortest path on the kNN graph) or sketching a freeform line over a force-directed layout.
  3. Intrinsic Projection Algorithm: To map points onto these non-linear axes, a diffusion-based algorithm is used. Coordinates "flow" from the defined axis to neighboring points through the kNN graph, ensuring the projection respects the data's "curviness."
  4. Multi-View Interface: The system includes a Topology View (skeleton), Detail View (force-directed layout), Intrinsic Axes View (visualizing trends), and a Projection View (final 2D scatterplot).
The topology extraction process: (a) kNN graph, (b) color-coding by filter value, (c) binning/clustering, and (d) the resulting topological skeleton.

Key Insights: Topology and Diffusion

A central tenet of this work is that understanding the data's skeleton must precede projection. By identifying branches first, users avoid "cross-branch projection" errors.

  • The Power of Diffusion: Unlike traditional "explicit" projection using straight lines, the diffusion process treats the manifold like a heat-conducting surface, allowing coordinates to wrap around curves naturally.
  • Semantic Interpretability: Unlike t-SNE, the axes in this system represent specific user-defined trends, such as moving from high-protein to high-carb foods in a nutritional dataset.
Abstraction of data structure: (a) 3D point cloud, (b) the underlying kNN graph, and (c) the extracted topological skeleton.

Evaluation and Results

The system was evaluated using the ISOMAP Face dataset, NBA player statistics, and USDA food data.

  • Face Dataset: The system successfully extracted independent axes for pose and lighting, whereas t-SNE often created disjointed "clumps."
  • NBA Case Study: The tool revealed a bifurcation in player stats, clearly separating "Centers/Power Forwards" from "Guards/Small Forwards" and identifying a "tail" of substitute players.
  • Comparison with AxisSketcher: While AxisSketcher operates in 2D projection space, this method operates in the intrinsic space of the kNN graph, leading to more stable and accurate axis extraction.

Limitations and Future Work

Scalability remains a challenge; as dimensionality and point counts increase, kNN calculations and visual complexity become bottlenecks. Force-directed layouts can also introduce distortions. Future work includes integrating hierarchical layout algorithms and exploring automated ways to suggest interesting topological branches to users.

Summary updated

BibTeX
@article{xia2019visual,
  author = {Jiazhi Xia and Fenjin Ye and Fangfang Zhou and Yi Chen and Xiaoyan Kui},
  title = {{Visual Identification and Extraction of Intrinsic Axes in High-Dimensional Data}},
  journal = {IEEE Access},
  year = {2019},
  doi = {10.1109/access.2019.2922997},
}
Figure