Neural Logical Index for Fast Knowledge Graph Complex Query Answering
Keywords: complex logical query, knowledge graph, query embedding, neural link predictor
TL;DR: We present an efficient symbolic search framework.
Abstract: Complex query answering (CQA) over knowledge graphs is a crucial multi-hop reasoning task aimed at addressing first-order logical queries within large and incomplete knowledge graphs. Direct traversal search methods rely solely on graph topology and often miss answers due to the incompleteness of the graph, thus neural models have been proposed to generalize the neglected answers from observed facts. There are primarily two lines of research tackling the challenges of CQA. Query embedding models learn representations for complex queries, offering fast speed but often providing only generic performance. In contrast, neural symbolic search methods deliver better performance, although they tend to be computationally more expensive. In this paper, we propose an efficient and scalable search framework that combines the precision of symbolic methods with the speed of embedding techniques. Our model utilizes embedding methods to compute Neural Logical Indices (NLI) to reduce the search domain for each variable in advance, followed by an approximate symbolic search for fine ranking. The search is precise for tree-structured queries and approximates cyclic queries (which are NP-complete) in quadratic complexity concerning the search domain, matching the complexity of tree-form queries. Experiments on various CQA benchmarks show that our framework reduces computation by 90% with a minimal performance loss, alleviating both efficiency and scalability issues.
Primary Area: learning on graphs and other geometries & topologies
Code Of Ethics: I acknowledge that I and all co-authors of this work have read and commit to adhering to the ICLR Code of Ethics.
Submission Guidelines: I certify that this submission complies with the submission instructions as described on https://iclr.cc/Conferences/2025/AuthorGuide.
Anonymous Url: I certify that there is no URL (e.g., github page) that could be used to find authors’ identity.
No Acknowledgement Section: I certify that there is no acknowledgement section in this submission for double blind review.
Submission Number: 7276
Loading