Knowledge Graph Embedding With Parallel Sampling for Cognitive Inference
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing knowledge graph technologies face challenges in efficiently managing big data and performing cognitive inferences, particularly in financial organizations, due to limitations in data structuring and machine learning integration.
Innovation Solution
A method and system for knowledge graph based embedding and multi-task learning, utilizing parallel processing with CPUs and GPUs to read, sample, and train graph data, incorporating frequency-based edge sampling and oversampling, and employing objective functions to generate embeddings for nodes and edges, with mechanisms for explainability and counterfactual link generation.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Productivity
If traditional knowledge graph technologies are used to manage big data, then data structuring is achieved, but efficiency and effectiveness in cognitive inferences are insufficient
Solution Approach 1:
The patent replaces traditional mechanical data processing systems with a parallel computing architecture that utilizes multiple GPUs to simultaneously process different portions of the knowledge graph. This substitution enables efficient handling of big data while maintaining inference accuracy through distributed computation and specialized processing units optimized for graph neural network operations.
Solution Approach 2:
The patent segments the knowledge graph processing into multiple independent threads that can be executed in parallel across different GPU devices. Each thread handles specific graph operations such as embedding generation, sampling, or training on particular data subsets, thereby improving overall productivity while maintaining reliability through modular error handling and result aggregation.
2Speed
If parallel processing with multiple threads is implemented, then processing speed is improved, but system complexity increases
Solution Approach 1:
The patent implements a universal parallel processing framework where the same GPU architecture and processing logic can handle multiple types of graph operations including embedding generation, link prediction, and classification tasks. This multi-functionality reduces system complexity by avoiding the need for separate specialized systems for each operation while maintaining high processing speeds through parallel execution.
Solution Approach 2:
The patent introduces intermediary data structures and communication mechanisms that facilitate efficient data exchange between parallel threads and between CPU and GPU systems. These intermediaries include optimized memory buffers, queue structures for task management, and standardized data formats that simplify the coordination of parallel operations without requiring complex point-to-point synchronization logic.
3Manufacturing precision
If frequency-based sampling and oversampling is applied, then data distribution is improved, but computational overhead increases
Solution Approach 1:
The patent performs frequency analysis and determines sampling strategies in advance during data preprocessing stages, before the main computational workload begins. By pre-calculating node and edge frequencies and pre-determining which edges require sampling versus oversampling, the system avoids repeated computational overhead during the main processing phases while still achieving accurate data distribution in the final model training.
Data Source
AI summary
Methods, systems, and computer program products for knowledge graph based embedding, explainability, and/or multi-task learning may connect task-specific inductive models with knowledge graph completion and enrichment processes.