Introducing Fairness in Graph Visualization (Poster Abstract)

Seok-Hee Hong, Giuseppe Liotta, Fabrizio Montecchiani, Martin Nöllenburg, Tommaso PiselliGD 2024 · DOI

TL;DR

This paper introduces the concept of "fairness" in graph visualizations, arguing that existing methods can inadvertently favor certain groups of nodes. They propose a novel research direction and a formalization of fairness for straight-line graph drawings, aiming to balance visual quality across different node subsets.

What Problem Does This Paper Address?

Information visualization tools are crucial for decision-making, especially when human feedback is involved. While fairness has been extensively studied in artificial intelligence, its application to information visualization, particularly graph drawing, has been largely overlooked. The authors highlight a critical gap: standard graph layout algorithms optimize for global quality metrics (like minimizing edge crossings or vertex overlap), but this optimization doesn't guarantee that the visual clarity is consistent across different groups of nodes within the graph. For instance, in a graph representing relationships between two competing parties (e.g., "reds" and "blues"), a layout optimized for overall readability might make the connections and positions of "red" nodes much clearer than those of "blue" nodes, or vice-versa. This disparity can lead to biased interpretations and hinder effective analysis. The paper aims to address this by proposing a framework for creating "fair" graph visualizations where visual quality is equitably distributed among predefined node groups.

Methods

The core contribution of this paper is the conceptualization and formalization of fairness in graph visualization. They draw an analogy from fairness in machine learning, defining it as the absence of prejudice or favoritism. For straight-line graph drawings, they propose to refine the concept of "stress," a common metric for layout quality that measures the difference between the desired distances between nodes in the graph (based on shortest paths) and their actual Euclidean distances in the drawing.

They define two group-specific stress measures: stressR(Γ) for red vertices and stressB(Γ) for blue vertices. The key idea is to minimize the difference between these group-specific stresses, normalized by the size of the groups. This leads to a formal definition of "unfairness" λ(Γ), which they aim to minimize. While the paper doesn't detail a specific algorithm for achieving this fairness, it mentions implementing a gradient-descent based algorithm capable of optimizing multiple drawing criteria simultaneously. This suggests an approach where fairness is treated as an additional objective alongside traditional quality metrics like stress.

Datasets

Not discussed in the paper. The paper focuses on the theoretical framework and conceptualization of fairness in graph visualization, using a hypothetical example of "red" and "blue" competing parties.

Key Insights and Ideas

  • Fairness is Crucial for Graph Visualization: Just as fairness is vital in AI, it's equally important for visualizations to avoid inherent biases that can skew interpretation.
  • Global Optimization Can Lead to Local Unfairness: Standard graph layout algorithms, by optimizing global metrics, can inadvertently create disparities in visual quality between different node groups.
  • Balancing Readability Across Groups: A fair visualization aims to distribute visual complexity and readability equitably, even if it means a slight compromise in overall global quality.
  • Extensibility of the Fairness Concept: The proposed notion of fairness, rooted in stress minimization, can likely be extended to other quality criteria and graph drawing paradigms.

Key Results

The paper presents a conceptual contribution by formalizing the notion of fair straight-line graph drawings. They also report on empirical results concerning the "price of fairness" in terms of additional stress. Their findings suggest that multi-objective optimization functions that balance fairness and stress can produce drawings with good fairness at the cost of only a relatively small increment in global stress. The paper includes a figure illustrating how optimizing for fairness can resolve ambiguities around specific nodes, improving their local readability.

Limitations and Future Work

The authors acknowledge that due to space limitations, they present their fairness model for straight-line drawings and refer readers to an extended abstract for more details. The paper focuses on straight-line drawings and the stress metric, implying that future work could explore fairness in other drawing paradigms (e.g., hierarchical layouts, force-directed layouts) and with different quality criteria. The empirical results are presented as preliminary, suggesting a need for more extensive investigations into the trade-offs between fairness and other visualization quality metrics. The paper also implicitly points to the need for practical algorithms that can efficiently compute fair graph layouts.

Summary updated

BibTeX
@inproceedings{hong2024introducing,
  author = {Seok-Hee Hong and Giuseppe Liotta and Fabrizio Montecchiani and Martin Nöllenburg and Tommaso Piselli},
  title = {{Introducing Fairness in Graph Visualization (Poster Abstract)}},
  booktitle = {GD},
  year = {2024},
  doi = {10.4230/lipics.gd.2024.49},
}
Figure