Hyperbolic Embedding Link Prediction for Knowledge Graphs
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing methods for predicting links in knowledge graphs face challenges in handling unseen triples, under-represented entities, and under-represented relations, especially in low-dimensional embedding spaces, with limited ability to extrapolate effectively.
Innovation Solution
A computer-implemented method that uses multi-dimensional vectors in hyperbolic spaces for entity embeddings, involving relation-specific translation and rotation, and learnable parameters for curvature and rotation extent, to improve link prediction by decomposing embeddings into sub-vectors and selecting representations based on distance calculations in these spaces.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Adaptability or versatility
If traditional link prediction methods are used in knowledge graphs, then existing models can handle seen triples, but they fail to effectively extrapolate to unseen triples, under-represented entities, and under-represented relations
Solution Approach 1:
The patent transitions from standard Euclidean embedding spaces to hyperbolic embedding spaces, utilizing the curved geometry of hyperbolic space to better represent the hierarchical and compositional structure of knowledge graphs. This dimensional change enables the model to capture relational patterns and perform effective extrapolation to unseen triples, entities, and relations while maintaining high prediction accuracy.
2Reliability
If high-dimensional embedding spaces are used, then link prediction performance improves, but computational complexity and memory requirements increase
Solution Approach 1:
The patent changes the geometric parameters of the embedding space by adopting hyperbolic geometry with specific curvature values. This parameter change allows the model to achieve high-dimensional capacity and expressiveness in a more efficient manner, improving link prediction performance while reducing the actual computational burden compared to traditional high-dimensional Euclidean spaces.
3Reliability
If relation-specific translation and rotation are applied to sub-vectors, then compositionality and prediction accuracy improve, but processing time increases
Solution Approach 1:
The patent segments the entity embedding into multiple sub-vectors, where each sub-vector can be independently processed through relation-specific translation and rotation operations. This segmentation allows for parallel computation and optimized processing, maintaining high prediction accuracy while managing computational time through efficient sub-vector operations in hyperbolic space.
Data Source
AI summary
A device and computer implemented method for determining a link in a knowledge graph, wherein the link comprises a first entity, a second entity, and a relation. The method includes determining a first representation which represents an embedding of the first entity; selecting a second representation, from a set of representations of embeddings of entities of the knowledge graph, wherein the second representation represents an embedding of the second entity, including determining a prediction for the second representation and selecting the second representation depending on the prediction for the second representation; and determining the link including the first entity, the second entity, and the relation.