Ontology-Driven Neural Link Prediction for Knowledge Graphs
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Conventional neural link prediction systems are over-parameterized, leading to excessive memory usage and reduced scalability, and often require computationally expensive heuristics that degrade predictive power, especially when applied to large knowledge graphs.
Innovation Solution
The ontology-driven neural link prediction system uses memory and parameter-efficient ontology embeddings learned through an ontology lookup table and a subgraph encoder, which operates on a k-hop neighboring subgraph, reducing the need for expensive node distance computations and supporting both transductive and inductive link prediction tasks.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Reliability
If conventional neural link prediction systems use node embeddings for all nodes in the knowledge graph, then predictive power is maintained, but memory usage becomes excessive and scalability is reduced
Solution Approach 1:
The patent segments the knowledge graph into multiple partitions or blocks, where each partition contains a subset of nodes and their local connectivity information. Instead of maintaining embeddings for all nodes globally, the system processes and stores embeddings in distributed partitions, reducing peak memory requirements while preserving local predictive accuracy through localized embedding computations.
Solution Approach 2:
The patent introduces hierarchical or multi-resolution embedding representations that operate at different dimensions. Coarse-grained embeddings capture global structure with fewer parameters, while fine-grained embeddings provide local detail only where needed. This dimensional hierarchy allows the system to maintain predictive power through selective use of detailed embeddings rather than uniformly high-dimensional representations across all nodes.
2Quantity of substance
If conventional systems reduce embedding dimensionality to reduce memory usage, then memory footprint decreases, but predictive power is severely impacted
Solution Approach 1:
The patent applies local quality by using low-dimensional or compressed embedding representations for nodes that require less predictive detail, while maintaining high-dimensional embeddings only for critical nodes or regions of the knowledge graph where predictive accuracy is most important. This selective dimensionality allocation reduces overall memory footprint while preserving predictive power where it matters most.
Solution Approach 2:
The patent dynamically adjusts embedding dimensionality parameters based on local graph properties, node importance metrics, or query-specific requirements. Rather than using fixed low-dimensional embeddings everywhere, the system adapts parameter dimensions locally, allowing memory-efficient representations in less critical regions while maintaining high-dimensional embeddings for nodes requiring greater predictive fidelity.
3Reliability
If conventional neural link prediction systems use computationally expensive heuristics for node distance computations, then link prediction accuracy is maintained, but computational cost and time increase significantly
Solution Approach 1:
The patent pre-computes and stores approximate node distances, path lengths, or connectivity metrics during an offline preprocessing phase or caching mechanism. These pre-computed distance heuristics are then reused during online link prediction queries, eliminating the need for repeated expensive shortest-path computations while maintaining sufficient accuracy for prediction tasks.
Solution Approach 2:
The patent uses sampled or approximated distance metrics that replicate the essential information of exact shortest-path computations without requiring full graph traversal. By copying or estimating distance relationships through sampling techniques or simplified graph representations, the system achieves comparable predictive accuracy with significantly reduced computational overhead.
Data Source
AI summary
Methods, systems and apparatus, including computer programs encoded on computer storage medium, for training a neural link predictor. In one aspect a method includes obtaining triples that represent a knowledge graph, where each triple comprises specifies a subject and object node in the knowledge graph and a relation type between the subject node and object node; obtaining data that specifies entity types of nodes in the knowledge graph; and for each triple: retrieving, from an ontology lookup table, ontology embeddings for the knowledge graph, the ontology embeddings comprising embeddings for each entity type in the set of entity types, generating, using the retrieved ontology embeddings for the knowledge graph, node embeddings for the subject node and the object node included in the triple, scoring the triple using the generated node embeddings, and updating, using a loss of the scored triple, the ontology embeddings stored in the ontology lookup table.