ChainInfer: A Joint Method for Inferring Missing AI Supply Chain Information

Runbang Hu, Mohammad Shahedur Rahman, Hang Liu, Bingzhe Li, Yuede Ji

Keywords: AI supply chain, graph neural network, graph transformer, link prediction, edge classification

TL;DR: We model the AI supply chain as an attributed graph and propose a hybrid GIN+GT encoder that unifies local and global reasoning for robust link prediction and edge classification.

Abstract: The modern AI ecosystem forms an intricate AI supply chain, where models, datasets, and software components are deeply interconnected. Incomplete or inconsistent metadata on platforms such as Hugging Face and Kaggle leaves critical gaps in provenance, hindering reproducibility, risk management, and governance. To address this, we formalize AI supply chain inference as a coupled graph learning problem: link prediction to recover missing dependencies and edge classification to determine their semantic types. We propose ChainInfer, a hybrid architecture that integrates graph neural networks for local structural reasoning with graph transformers for global context, trained end-to-end on attributed supply chain graphs. Using a benchmark of 200K models from Hugging Face, ChainInfer outperforms GNN-, Transformer-, and ensemble baselines, achieving 0.93 joint accuracy while remaining efficient. Moreover, ChainInfer generalizes inductively to Kaggle, retaining 0.90 accuracy without retraining. These results demonstrate ChainInfer as a practical framework for scalable, accurate, and transferable AI supply chain provenance inference.

Primary Area: learning on graphs and other geometries & topologies

Submission Number: 8151