Graph Query Embeddings for Topology-Aware List Expansion
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing semantic search engines struggle to leverage the topological and taxonomic data present in graph databases, limiting their ability to provide comprehensive and relevant search results beyond string-based property fields.
Innovation Solution
A graph query platform that generates property-level and node-level vector embeddings based on graph data, allowing for list expansion by determining similar entities through graph embeddings, incorporating both topology and taxonomy in search results.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If semantic search engines use only string-based property fields for searching, then the search process is simple and fast, but the search results lack comprehensiveness and relevance to topological and taxonomic data in graph databases
Solution Approach 1:
The patent replaces traditional string-based mechanical text matching with vector embedding-based semantic similarity computation. Vector embeddings transform textual and structural data into numerical vectors that capture semantic meaning and topological relationships, enabling the search engine to understand context and relationships rather than merely matching keywords. This substitution fundamentally changes the search mechanism from surface-level string comparison to deep semantic understanding.
Solution Approach 2:
The patent creates a composite search approach that integrates multiple data types and representation methods: combining vector embeddings of node properties with graph topology structures, merging semantic similarity scores with structural relationship scores. This composite approach synthesizes information from both the content (taxonomy) and the relationships (topology) to produce comprehensive search results that leverage the full power of graph database structures.
2Loss of information
If semantic search engines incorporate both topology and taxonomy data from graph databases, then search result comprehensiveness and relevance improve, but the computational complexity and processing time increase
Solution Approach 1:
The patent pre-computes and stores vector embeddings for graph node properties before actual search operations. These embeddings capture the semantic meaning of node attributes and are stored in an efficient retrieval structure. During search operations, the system only needs to compute embeddings for query terms and perform vector similarity comparisons, rather than processing entire graph structures in real-time. This preliminary preparation significantly reduces online search latency while maintaining comprehensive information retrieval.
Solution Approach 2:
The patent divides the search process into distinct modular components: (1) query embedding generation, (2) candidate node filtering based on embedding similarity, (3) topology-based relationship verification, and (4) result ranking. This segmentation allows each component to be optimized independently and enables parallel processing of multiple search operations. The modular architecture processes only relevant portions of the graph data rather than the entire structure, reducing overall processing time while preserving information completeness.
Data Source
AI summary
In some aspects, the techniques described herein relate to a method including: receiving a first list entity vector embedding, wherein the first list entity vector embedding is of a categorical type and is associated with a list entity; generating a first similarity score between the first list entity vector embedding and a stored vector embedding of a plurality of stored vector embeddings, wherein the stored vector embedding of the plurality of stored vector embeddings is of the categorical type and is associated with a stored entity; adding the first similarity score to a second similarity score, wherein the sum of the first similarity score and the second similarity score is an overall similarity score associated with the stored entity; querying a datastore to retrieve data associated with the stored entity; and returning, as output of the list expansion module, the data associated with the stored entity.