Graph Visualization Design Guidelines as Learnable Predicates

Abstract

Graphs are widely used to represent complex, interconnected data across domains, yet choosing an effective visualization remains difficult because existing design knowledge is fragmented and inconsistent. This lack of a unified foundation prevents researchers from integrating findings into a cumulative body of knowledge, leaving valuable results isolated. It also hinders designers and practitioners, who cannot readily translate such findings into actionable strategies for their own goals and contexts. We propose a predicate-based representation that formalizes visualization guidelines as bounded conditions over descriptive graph statistics. Predicates directly mirror the qualitative structure of design guidelines. For example, a rule might specify that if graph density is low, a node-link diagram is appropriate, whereas if density is high, an adjacency matrix should be used. Unlike static handcrafted rules, they can also be learned, optimized, and adapted as new findings or usage contexts emerge. As a result, fragmented knowledge is consolidated into a formal and extensible foundation for graph visualization design and recommendation. We evaluate this approach by testing its ability to (i) recover expert rules and (ii) adapt to user-specific preferences while generalizing to unseen graphs. The results show that the learned predicates closely reproduce expert-derived guidelines, accommodate diverse preference patterns, and achieve strong performance on held-out data, demonstrating a promising path toward more systematic and cumulative graph visualization design knowledge.