Subgraph Search over Neural-Symbolic Graphs

In this paper, we propose neural-symbolic graph databases (NSGDs) that extends traditional graph data with content and structural embeddings in every node. The content embeddings can represent unstructured data (e.g., images, videos, and texts), while structural embeddings can be used to deal with incomplete graphs. We can advocate machine learning models (e.g., deep learning) to transform unstructured data and graph nodes to these embeddings. NSGDs can support a wide range of applications (e.g., online recommendation and natural language question answering) in social-media networks, multi-modal knowledge graphs and etc. As a typical search over graphs, we study subgraph search over a large NSGD, called neural-symbolic subgraph matching (NSMatch) that includes a novel ranking search function. Specifically, we develop a general algorithmic framework to process NSMatch efficiently. Using real-life multi-modal graphs, we experimentally verify the effectiveness, scalability and efficiency of NSMatch.