Artificial Intelligence 7 min read

How Dynamically Pruned Message Passing Networks Revolutionize Large‑Scale Knowledge Graph Reasoning

The Hulu AI team’s ICLR‑2020 paper introduces a consciousness‑prior‑driven graph neural network that dynamically prunes message‑passing subgraphs, achieving state‑of‑the‑art results on large‑scale knowledge‑graph completion tasks while improving interpretability and computational efficiency.

Hulu Beijing

Jan 3, 2020

How Dynamically Pruned Message Passing Networks Revolutionize Large‑Scale Knowledge Graph Reasoning

Recently, the Hulu AI research team’s paper Dynamically Pruned Message Passing Networks for Large‑scale Knowledge Graph Reasoning (authors Xiaoran Xu, Wei Feng, Yunsheng Jiang, Xiaohui Xie, Zhiqing Sun, Zhi‑Hong Deng) was accepted for oral presentation at the top‑tier deep‑learning conference ICLR 2020.

ICLR (International Conference on Learning Representations) was founded in 2013 by leading deep‑learning pioneers such as Yoshua Bengio and Yann LeCun, and is renowned for its open‑review process and high impact on the field.

The first author, Xiaoran Xu, joined Hulu as a researcher in 2017, focusing on deep learning and machine reasoning, and has published multiple top‑conference papers.

Paper link: https://openreview.net/forum?id=rkeuAhVKvB Code repository: https://github.com/anonymousauthor123/DPMPN

Knowledge graphs, though not new, remain a hot research topic as many internet companies aim to integrate domain knowledge into recommendation and advertising systems, especially when training data are scarce. Natural‑language‑processing applications also frequently rely on knowledge graphs.

The proposed Dynamic Pruned Message Passing Network (DPMPN) combines explicit reasoning with graph neural networks by constructing dynamically pruned subgraphs guided by a graph‑based attention mechanism, embodying the “consciousness prior”. The architecture consists of two GNN layers: a lower layer using full‑graph random sampling and an upper layer processing batches of subgraphs, linked by an attention‑transition mechanism.

Experiments on standard KG completion benchmarks (FB15K‑237, WN18RR) show that DPMPN achieves superior HITS@1 and HITS@3 scores compared to state‑of‑the‑art methods, while HITS@10 improvements are modest. The approach offers enhanced interpretability and controllable computational complexity.

In the emerging “deep learning 2.0” era, the synergy between graph neural networks and reasoning is expected to drive progress toward general artificial intelligence.