Recursive Temporal Knowledge Graph Completion via Embedding Updates
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Current temporal knowledge graph completion methods are limited by the static nature of discrete event models and the need to retrain models with new timestamps, lacking extensibility and efficiency in handling dynamic changes and incomplete data.
Innovation Solution
A recursive temporal knowledge graph completion method that uses embedding learning to update features and parameters across timestamps, allowing for continuous fact prediction without retraining the entire model, leveraging static knowledge graphs and embedding algorithms like TransE, TransD, RotatE, RGCN, and RDGCN to extend knowledge graph information naturally.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Adaptability or versatility
If discrete event completion model is used to complement missing facts in temporal knowledge graph, then completion can be performed at certain timestamps, but the model lacks extensibility for completion of new static knowledge graphs and requires retraining when new facts appear
Solution Approach 1:
The patent segments the temporal knowledge graph into multiple static knowledge graphs corresponding to different timestamps. Each static knowledge graph can be processed independently through embedding learning, allowing incremental updates without retraining the entire model. This segmentation enables the system to handle new facts by only processing the affected timestamp segments.
Solution Approach 2:
The patent introduces dynamic embedding learning parameters that can be updated recursively as new facts appear. The embedding parameters are not fixed but can adapt to new information through incremental learning, allowing the model to extend to new static knowledge graphs without complete retraining. The dynamic nature of the parameters enables the system to accommodate temporal changes efficiently.
2Reliability
If continuous fact completion model trains facts of all timestamps at the same time, then comprehensive completion is achieved, but the model cannot efficiently handle dynamic changes when new facts continue to appear
Solution Approach 1:
The patent divides the continuous training problem into discrete timestamp segments, where each static knowledge graph is processed separately through embedding learning. This segmentation maintains comprehensive coverage of all timestamps while enabling efficient incremental processing when new facts appear, as only the affected segments need to be updated.
Solution Approach 2:
The patent performs embedding learning to obtain updated features and parameters in advance for each timestamp segment. By pre-computing the embedding parameters for each static knowledge graph, the system prepares the data structures needed for fact prediction, enabling efficient handling of dynamic changes without reprocessing all timestamps when new facts appear.
3Measurement precision
If embedding learning is performed on static knowledge graph features, then updated features are obtained for fact prediction, but the model needs to retrain when new timestamps are added to maintain accuracy
Solution Approach 1:
The patent creates a universal embedding learning framework that can process multiple static knowledge graphs across different timestamps using the same embedding parameters. The embedding parameters learned from one timestamp can be transferred and adapted to other timestamps, reducing the need for retraining when new timestamps are added while maintaining prediction accuracy.
Solution Approach 2:
The patent uses parameter updates through embedding learning to adapt the model to new timestamps. Instead of retraining the entire model, the system updates the embedding parameters incrementally based on new factual information, maintaining accuracy while reducing complexity. The parameters are adjusted to reflect temporal changes without requiring complete retraining.
Data Source
AI summary
The disclosure provides a temporal knowledge graph completion method, including: obtaining a temporal knowledge graph; obtaining a corresponding static knowledge graph, and obtaining an updated feature through performing embedding learning on a feature of the static knowledge graph and the static knowledge graph; starting from the sub knowledge graph with the first timestamp, obtaining, based on recursion, an updated embedding learning parameter and an updated feature by taking a sub knowledge graph with a current timestamp, and a feature and an embedding learning parameter of the sub knowledge graph with the current timestamp as input of embedding learning; determining the updated embedding learning parameter and the updated feature as an embedding learning parameter and a feature of a sub knowledge graph with an adjacent next timestamp, until all the sub knowledge graph sequences in the temporal knowledge graph are traversed; and performing fact prediction on each sub knowledge graph.