Seeing Graphs Like Humans: Benchmarking Computational Measures and MLLMs for Similarity Assessment

A research team from Seoul National University has published a groundbreaking study titled 'Seeing Graphs Like Humans,' which directly benchmarks automated graph similarity measures against human visual perception. The paper addresses a critical gap in visual analytics: traditional computational metrics often provide recommendations that conflict with analysts' intuitive judgments, potentially increasing confusion rather than reducing cognitive load. Through three interconnected experiments using a dataset of 1,881 node-link diagrams and judgments from 32 human participants, the study establishes a human baseline for graph similarity, revealing that people prioritize global shapes and edge densities over exact topological details.

The research then benchmarks 16 established computational measures against this human consensus, identifying Portrait divergence as the best-performing traditional metric, though with only moderate alignment. The most significant finding comes from evaluating three state-of-the-art Multimodal Large Language Models (MLLMs): GPT-5, Gemini 2.5 Pro, and Claude Sonnet 4.5. The results demonstrate that MLLMs, particularly GPT-5, significantly outperform all traditional measures in aligning with human perception of graph similarity. Furthermore, these models provide interpretable natural language rationales for their decisions, a key advantage over opaque algorithmic scores. Claude Sonnet 4.5 was noted for its superior computational efficiency. The findings suggest MLLMs hold significant promise as effective, explainable proxies for human perception and as intelligent guides that can uncover subtle visual nuances in complex graph data.