Mochi: Meta-Learning Framework for Graph Foundation Models

A team of researchers has introduced Mochi, a Graph Foundation Model that employs a meta-learning training framework aimed at enhancing task unification and training efficiency. Unlike earlier models that utilize reconstruction-based objectives like link prediction and require a separate unification phase with class prototypes, Mochi focuses on few-shot episodes that replicate the downstream evaluation process during pre-training. This approach aligns the training goals with inference, thus removing the necessity for post-hoc unification. Tests conducted on both synthetic and real-world datasets reveal limitations in traditional methods that hinder downstream performance. Mochi, along with its advanced variant Mochi++, demonstrates competitive or even superior results compared to existing Graph Foundation Models across 25 real-world datasets.