Knowledge Graph Embedding via Grid Search Tuning
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing lightweight knowledge graph embedding techniques primarily focus on reducing model size, leading to issues such as increased memory requirements for decoding, complexity in knowledge distillation, and inefficient query processing.
Innovation Solution
A knowledge graph embedding device and method that employs hyperparameter tuning using grid search to generate a low-dimensional embedding model, and optionally includes quantization and reordering of entity embedding vectors to reduce model size and improve query processing efficiency.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Weight of stationary object
If codebook-based methodology is used to compress knowledge graph embedding vectors, then model size is reduced, but additional memory space is required for decoding and all encoded data must be decoded for repetitive tasks such as clustering
Solution Approach 1:
The patent extracts only the necessary embedding vectors for entities and relations, eliminating the need to store and decode all encoded data. By taking out only the essential components (embedding vectors) rather than all encoded information, the system reduces memory requirements while maintaining functionality for tasks like clustering without needing to decode everything.
Solution Approach 2:
The patent applies partial action by only decoding and storing the embedding vectors that are actually needed for the knowledge graph, rather than decoding all encoded data. This partial approach maintains sufficient functionality for clustering and other tasks while significantly reducing memory space requirements.
2Weight of stationary object
If knowledge distillation is used to reduce embedding vector size, then model size is reduced, but the method becomes complex and requires well-trained teacher models
Solution Approach 1:
The patent uses a simple and straightforward embedding vector storage approach rather than complex knowledge distillation. The method employs basic data structures to store embedding vectors, avoiding the need for well-trained teacher models and complex distillation processes. This simple approach is sufficient for reducing model size without introducing significant complexity.
3Weight of stationary object
If embedding vectors are reduced to low-dimensional space, then model size is reduced, but linear scanning of all entity embedding vectors is required upon query processing
Solution Approach 1:
The patent performs preliminary action by pre-computing and storing embedding vectors in an optimized structure that enables efficient query processing. By preparing the data in advance in a format that supports fast retrieval, the system avoids the need for linear scanning during query processing, thereby improving productivity while maintaining reduced model size.
Data Source
AI summary
Provided are a device and method for embedding a knowledge graph. The device includes an acquisition module configured to acquire a knowledge graph embedding model and a tuning module configured to generate a low-dimensional embedding model by performing hyperparameter tuning on the acquired knowledge graph embedding model on the basis of grid search.