## Zhaocheng Zhu

*Knowledge Graph Reasoning with Graph Neural Networks*

Knowledge graphs encode real-world facts and are critical in various applications and domains such as natural language understanding, recommender systems, drug discovery. A fundamental problem on knowledge graphs is to predict the answer to some queries by reasoning over existing facts, a.k.a. knowledge graph reasoning. Such a problem is commonly solved by embedding methods. In this talk, we take another approach, and solves knowledge graph reasoning with graph neural networks. First, we propose Neural Bellman-Ford Networks (NBFNet) that solves single-hop reasoning with the generalized Bellman-Ford algorithm and learned operators in the path formulation. Then we introduce Graph Neural Network Query Executor (GNN-QE) that combines GNNs with fuzzy sets for answering multi-hop queries on knowledge graphs. Both models are highly related to traditional symbolic methods, such as personalized PageRank and subgraph matching, but can additionally deal with the missing links in knowledge graphs. Meanwhile, both models can visualize their intermediate steps, which may help us better understand the reasoning process.

When: **4th of August 2022 - 11am (EST)**

Where: **Pavillon André-Aisenstadt (Université de Montréal), 5th floor, Cirrelt Conference Room(5441)**

More: *Zhaocheng's Website *