A Generic Framework and Library for Exploration of Small Multiples through Interactive Piling

Fritz Lekschas, Xinyi Zhou, Wei Chen, Nils Gehlenborg, Benjamin Bach, Hanspeter Pfister

Introduction

"Small multiples"—grids of miniature visualizations like sparklines or maps—are a cornerstone of data visualization. However, they struggle to scale. When the number of items reaches the hundreds or thousands, screen real estate disappears, and the cognitive load of scanning for patterns becomes overwhelming. Traditional solutions like filtering or global summarization often obscure important outliers or local trends.

This paper proposes interactive piling as a middle ground. By treating digital items like physical stacks of paper, users can flexibly organize, explore, and compare large numbers of small multiples by interactively aggregating them into piles, mimicking natural physical organization.

The Piling Framework and PILING.JS

The authors contribute both a theoretical design space and a practical implementation called PILING.JS.

The Design Space

The framework identifies five core analytical tasks for piling:

  1. Grouping (T1): Putting items together manually or automatically by similarity.
  2. Arrangement (T2): Positioning piles using grids or data-driven layouts.
  3. Previewing (T3): Hinting at pile contents via galleries or "foreshortened" stacks.
  4. Browsing (T4): Peeking inside piles by "leafing" through items.
  5. Aggregation (T5): Summarizing a pile into a single representative visual, such as an average image.
The visual piling workflow: from small multiples to aggregated piles for comparison.

Implementation

PILING.JS is a data-agnostic JavaScript library built on PixiJS (WebGL) for high-performance rendering. It uses a declarative approach, allowing developers to define item appearance and behavior without low-level code. The library maintains a smooth 60 FPS for up to 1,000 items, making it suitable for real-time exploration.

Diverse Use Cases and Results

The framework's versatility was tested across eight diverse datasets, including Google Quickdraw sketches, microscopy cell images, and global temperature anomalies.

Examples of piling applied to COVID-19 infection rates and Worldbank development indicators.

In the "Necklace" sketch example, piling revealed four distinct sketching styles that were invisible in a standard grid. The library successfully handled various data types—images, SVGs, and WebGL programs—demonstrating its generic nature. For instance, global area charts of COVID-19 infection rates were arranged geographically and then piled to identify regional trends.

Key Insights

  • Piling vs. Filing: Borrowing from cognitive science, the authors note that piling is cognitively easier than filing because it involves loose, temporary classification rather than rigid categorization.
  • The Power of Previews: Previews act as a bridge between individual items and aggregates. "Leafing" allows rapid inspection without breaking the spatial layout.
  • Dynamic State: Piling is a transient process; piles are dynamic entities that users can disperse or merge on the fly to test different hypotheses.

Limitations and Future Work

While the library handles up to 2,000 items, initialization time for custom WebGL renderers remains a bottleneck. Future work includes "texture bundling" for better scalability and perceptual studies to determine which preview and browsing techniques best support specific cognitive tasks. Expanding to touch and pen-based gestures could also make the "piling" metaphor feel even more natural.

Summary updated

BibTeX
@article{lekschas2021generic,
  author = {Fritz Lekschas and Xinyi Zhou and Wei Chen and Nils Gehlenborg and Benjamin Bach and Hanspeter Pfister},
  title = {{A Generic Framework and Library for Exploration of Small Multiples through Interactive Piling}},
  journal = {IEEE Trans. Vis. Comput. Graph.},
  year = {2021},
  doi = {10.1109/tvcg.2020.3028948},
}
Figure