Knowledge Graph Embedding for Out-of-Sample Entity Prediction
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing methods struggle with out-of-sample reasoning in knowledge graphs, where predictions need to be made for entities unseen during training, leading to inconsistencies between training and testing.
Innovation Solution
A computer-implemented method and system for out-of-sample representation learning using knowledge graphs, which involves generating random numbers to select different knowledge graph embedding model processors, retrieving and calculating embeddings, and using aggregation functions like oDistMult-ERAvg and oDistMult-LS to ensure consistent training and testing.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If traditional transductive training methods are used for knowledge graph embedding, then in-sample prediction accuracy is maintained, but out-of-sample prediction accuracy deteriorates due to training-testing inconsistency
Solution Approach 1:
The patent applies preliminary action by pre-processing training triples to identify and separate in-sample entities from out-of-sample entities before the actual training process. This allows the model to be trained specifically on in-sample entities while preparing for out-of-sample predictions, resolving the inconsistency between training and testing phases. The method pre-calculates which entities should be treated as out-of-sample and adjusts the training objective accordingly.
Solution Approach 2:
The patent implements local quality by applying different treatment to different entities within the knowledge graph. In-sample entities receive standard embedding training, while out-of-sample entities are handled through a separate aggregation mechanism that computes their embeddings by aggregating information from their neighboring in-sample entities. This localized differentiation resolves the training-testing inconsistency.
2Reliability
If multiple knowledge graph embedding model processors are selected randomly for different training data entries, then model robustness and out-of-sample performance improve, but computational complexity increases
Solution Approach 1:
The patent applies dynamics by making the selection of knowledge graph embedding model processors dynamic rather than static. Different processors are randomly selected for different training data entries, allowing the model to learn from multiple embedding perspectives and improve robustness. This dynamic selection process helps the model generalize better to out-of-sample entities while managing computational complexity through efficient implementation.
3Measurement precision
If aggregation functions like oDistMult-ERAvg and oDistMult-LS are used for out-of-sample entities, then prediction accuracy for unseen entities improves, but computational overhead increases
Solution Approach 1:
The patent applies preliminary action by pre-computing and storing embeddings for all in-sample entities during the training phase. When predicting for out-of-sample entities, the system only needs to aggregate information from these pre-computed embeddings rather than performing full training computations. This significantly reduces the computational overhead at inference time while maintaining high prediction accuracy.
Data Source
AI summary
Disclosed are systems, methods, and devices for out-of-sample representation learning using knowledge graphs. An embedding data structure reflective of a knowledge graph embedding model is received. A training data set including a plurality of training data entries, each of the training data entries reflective of a head entity, a tail entity, and a relation therebetween, wherein at least one of the head entities or the tail entities includes an out-of-sample entity, is received. A plurality of knowledge graph embedding model processors is provided. A random number is generated and compared to at least one criterion. A knowledge graph embedding model processor is selected from among the plurality of knowledge graph embedding model processors based at least in part on the comparing. The embedding data structure is processed with the selected knowledge graph embedding model processor.