Mochi: Aligning Pre-training and Inference for Efficient Graph Foundation Models via Meta-Learning
A new graph AI aligns training with inference, cutting time drastically.
Researchers João Mattos and Arlei Silva have introduced Mochi, a Graph Foundation Model that tackles two persistent challenges in graph machine learning: task unification and training efficiency. Traditional graph models pre-train with reconstruction-based objectives like link prediction, then rely on a separate post-hoc unification step (e.g., class prototypes) to align representations with downstream tasks. The authors demonstrate through synthetic and real-world experiments that this approach has inherent limitations that degrade downstream performance.
Mochi addresses these limitations by adopting a meta-learning framework where pre-training occurs on few-shot episodes that directly mirror the downstream evaluation protocol. This aligns the training objective with inference from the start, eliminating the need for a separate unification step. The model, along with its more powerful variant Mochi++, achieves competitive or superior performance across 25 real-world graph datasets covering node classification, link prediction, and graph classification. Critically, Mochi requires 8 to 27 times less training time than the strongest baseline, making it a highly efficient solution for graph foundation models.
- Mochi uses meta-learning to align pre-training with inference, avoiding post-hoc unification steps.
- It achieves competitive performance on 25 real-world datasets for node classification, link prediction, and graph classification.
- Training time is 8–27x faster than the strongest baseline, significantly improving efficiency.
Why It Matters
Mochi's meta-learning approach could revolutionize graph AI by slashing training costs and improving task alignment.