Message Tuning Outshines Graph Prompt Tuning: A Prismatic Space Perspective
Researchers have introduced Prismatic Space Theory (PS-Theory) to quantify the adaptation capacity of methods used for Graph Foundation Models (GFMs). This framework establishes an upper bound for graph prompt tuning, a common adaptation technique for GNN-based GFMs. Building on this theory, they developed Message Tuning for GFMs (MTG), a lightweight approach that enhances adaptation by injecting learnable message prototypes into GNN layers. Experiments show MTG surpasses graph prompt tuning baselines, validating the theoretical findings. AI
IMPACT Introduces a new theoretical framework for understanding and improving adaptation methods in graph foundation models.