Vector Estimating Program for Dynamic Knowledge Graph Updates
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing knowledge graph embedding models face difficulties in updating and re-learning when new knowledge is introduced, as they often require re-training on the entire graph, which is impractical and challenging, especially when the model does not support dynamic updating or when machine-learned vectors are inaccessible during the calculation of new entity vectors.
Innovation Solution
A vector estimating program that generates a second model based on existing entity information and graph structure data, allowing the estimation of new entity vectors by integrating peripheral nodes or performing machine re-learning on updated graph sections, thereby reducing the need for full graph re-training and enabling vector calculation for unknown entities.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Reliability
If full graph re-training is performed to update knowledge graph embeddings with new entities, then embedding quality is maintained, but computational cost and time consumption increase significantly
Solution Approach 1:
The patent segments the re-training process into two parts: (1) pre-trained embedding models are trained on historical knowledge graphs in advance, and (2) when new entities are added, only the necessary model parts are fine-tuned or adapted rather than re-training the entire model from scratch. This segmentation reduces the time loss while maintaining embedding quality.
Solution Approach 2:
The patent performs preliminary action by pre-training embedding models on historical knowledge graphs before new entities are added. The pre-trained models are stored and can be quickly adapted to new entities without requiring full re-training, thus reducing the time consumption when updates are needed.
2Productivity
If pre-trained embedding models are stored for quick adaptation, then update speed increases, but model accuracy may deteriorate when handling new entities
Solution Approach 1:
The patent implements a dynamic approach where the system determines whether to use pre-trained models directly or to perform fine-tuning based on the characteristics of new entities. This dynamic adaptation mechanism ensures both fast updates and maintained accuracy by selecting the appropriate strategy for each situation.
Solution Approach 2:
The patent introduces an intermediary mechanism (the determination unit) that decides whether to apply pre-trained models or perform fine-tuning. This intermediary ensures that the system can quickly use pre-trained models when appropriate while switching to fine-tuning when accuracy is critical, balancing speed and reliability.
3Loss of information
If machine learning is performed on the entire graph structure, then comprehensive knowledge is captured, but the complexity of the system increases
Solution Approach 1:
The patent extracts and stores pre-trained embedding models separately from the main system. When new entities are added, the system extracts only the necessary information from these pre-trained models rather than processing the entire graph structure, thus reducing system complexity while maintaining knowledge completeness.
Data Source
AI summary
A non-transitory computer-readable recording medium has stored therein a vector estimating program that causes a computer to execute a process including: obtaining a first vector and second entity information, the first vector being generated by using a first model with reference to graph structure data representing a relationship of a first entity group and being obtained by using first entity information related to the first entity group, the second entity information obtained by updating the first entity information and related to the first entity group and a second entity not being included in the first entity group; generating a second model based on the first vector and information on the first entity group included in the second entity information, the second model being used for obtaining vector data from the second entity information; and estimating a second vector corresponding to the second entity by using the generated second model.