Knowledge Graph Tensor Decomposition for Reliable Triple Prediction

Resolve Bottlenecks,
Find Innovative SolutionsGenerate Solutions

Solution Overview

Problem

Statistical machine learning in knowledge graphs faces challenges with incorrect prediction determinations, particularly for famous or characteristic cases, leading to reduced reliability and accuracy.

Innovation Solution

A machine learning program that generates a tensor representing triple data and performs tensor decomposition with fixed values for certain elements, using a modified Tucker decomposition to ensure accurate prediction by generating negative examples and allocating core tensor elements based on reserved triples.

Engineering Contradictions & Design Principles

VSEngineering Contradiction Analysis

1Adaptability or versatility

If statistical machine learning is used for knowledge graph embedding, then general prediction capability is improved, but prediction accuracy for specific famous or characteristic cases deteriorates

Engineering Contradiction:

Improvegeneral prediction capabilityVSAvoidprediction accuracy for specific cases

Core Design Contradiction:

Adaptability or versatilityVSMeasurement precision

Solution Approach 1:

The patent segments the prediction process into two distinct phases: (1) general statistical machine learning for overall prediction capability, and (2) rule-based verification for specific famous or characteristic cases. This segmentation allows each method to operate in its optimal domain, resolving the contradiction between general adaptability and specific precision.

Inventive Principle:

Principle #1Segmentation

Solution Approach 2:

The patent introduces an intermediary verification mechanism that acts as a bridge between statistical machine learning outputs and final predictions. This intermediary layer checks predictions against predefined rules and constraints, ensuring accuracy for specific cases while preserving the general capability of statistical learning.

Inventive Principle:

Principle #24Intermediary (Mediator)

2Productivity

If machine learning models are trained with standard loss functions, then overall model convergence is improved, but prediction reliability for designated triples deteriorates

Engineering Contradiction:

Improvemodel convergence speedVSAvoidprediction reliability for designated triples

Core Design Contradiction:

ProductivityVSReliability

Solution Approach 1:

The patent applies local quality by differentiating the treatment of different triples during training. Designated triples (famous or characteristic cases) are assigned special attention mechanisms and constrained loss functions, while other triples use standard loss functions. This localized differentiation ensures high reliability for important predictions without sacrificing overall convergence efficiency.

Inventive Principle:

Principle #3Local quality

Solution Approach 2:

The patent performs preliminary identification and marking of designated triples before the main training process. This preliminary action allows the model to prioritize learning these specific triples with appropriate constraints and verification mechanisms in place from the beginning, ensuring their prediction reliability is established during the convergence process.

Inventive Principle:

Principle #10Preliminary action

Data Source

PatentUS12481922B2Non-transitory computer-readable storage medium for storing machine learning program, machine learning method, and machine learning apparatus of improving prediction accuracy in knowledge graph embedding

Publication Date: 2025.11.25 FUJITSU LTD

AI summary

A non-transitory computer-readable recording medium storing a machine learning program for causing a computer to executes processing, the processing including: generating a tensor representing a plurality of pieces of triple data; and executing tensor decomposition, when performing the tensor decomposition of the tensor into a core tensor and factor matrices, under a condition that a value of a first element of the factor matrix corresponding to first triple data among the plurality of pieces of triple data and a value of an element of the core tensor are fixed.