OpenRTAG Benchmark Assesses AI Robustness in Low-Quality Text-Attributed Graphs

A new benchmark named OpenRTAG has been introduced to systematically evaluate the robustness of machine learning models designed for Text-Attributed Graphs (TAGs). TAGs, which combine relational structures with rich textual information at nodes, are prevalent in real-world applications but often suffer from data quality issues such as sparsity, noise, and imbalance across text, graph structure, and labels. Existing research on mitigating these issues is fragmented, making it difficult to comprehensively understand model robustness. OpenRTAG addresses this by organizing TAG quality problems into a unified 3x3 taxonomy, covering nine distinct degradation scenarios. It supports standardized evaluation across nine datasets and three downstream tasks, allowing for a thorough comparison of traditional Graph Neural Networks (GNNs), LLM-GNNs, and Graph Foundation Models (GFMs). The benchmark investigates scenario validity, model sensitivity, and the effectiveness of matched baselines, even under composite degradation scenarios, providing a crucial tool for developing more resilient TAG learning algorithms.