Graph Learning Explainability With Cumulative Denoising

Resolve Bottlenecks,
Find Innovative Solutions
Generate Solutions

Solution Overview

Problem

Existing explainable graph machine learning models generate inconclusive and localized explanations for predictions, leading to transparency issues and confirmation bias, particularly in applications like finance and fraud detection.

Innovation Solution

A method involving cumulative denoising and clustered search is employed to process explanation scores, identifying and isolating noise, and determining actual prediction-driving graph elements through global statistical analysis and clustered search, enhancing explanation accuracy and transparency.

Engineering Contradictions & Design Principles

VSEngineering Contradiction Analysis

1Productivity

If post-hoc explanation generation is used for graph neural network predictions, then explanation generation is simple and fast, but the explanations become inconclusive and subject to confirmation bias

Engineering Contradiction:
Improveexplanation generation speedVSAvoidexplanation accuracy
Core Design Contradiction:
ProductivityVSMeasurement precision

Solution Approach 1:

The patent applies preliminary action by integrating explanation generation into the training phase of the graph neural network. The explainable GNN model is trained simultaneously to perform both prediction and explanation generation, rather than generating explanations after predictions are made. This allows the model to learn explanatory patterns during training, improving explanation accuracy while maintaining efficiency through unified model architecture.

Inventive Principle:
Principle #10Preliminary action

2Loss of information

If localized explanation generation is performed for each prediction, then individual prediction transparency is improved, but global consistency and trustworthiness deteriorate due to confirmation bias

Engineering Contradiction:
Improveprediction transparencyVSAvoidglobal explanation consistency
Core Design Contradiction:
Loss of informationVSReliability

Solution Approach 1:

The patent merges localized and global explanation generation by designing a unified explainable GNN model that produces both instance-specific explanations and global explanation patterns. The model integrates explanation generation at multiple levels: local explanations for individual predictions and global explanations that capture overall decision-making patterns, ensuring consistency across all predictions while maintaining individual transparency.

Inventive Principle:
Principle #5Merging (Combining)

3Device complexity

If graph elements are filtered based on explain scores alone, then explanation generation is simple and fast, but noise in the graph data affects explanation accuracy

Engineering Contradiction:
Improveexplanation processing complexityVSAvoidgraph element identification accuracy
Core Design Contradiction:
Device complexityVSMeasurement precision

Solution Approach 1:

The patent introduces an intermediary denoising mechanism that processes graph data before explanation generation. The system applies denoising operations to remove noise from graph structures and features, then uses the cleaned data as input for explanation generation. This intermediary step improves the accuracy of graph element identification without significantly increasing overall system complexity, as the denoising is integrated into the model pipeline.

Inventive Principle:
Principle #24Intermediary (Mediator)

Data Source

PatentEP4614353A1Global interpretation of graph learning explainability based on cumulative denoising and clustered search
Publication Date: 2025.09.10 FUJITSU LTD
  • EP4614353A1 patent drawingFigure 1
  • EP4614353A1 patent drawingFigure 2
  • EP4614353A1 patent drawingFigure 3

AI summary

In an embodiment, operations include receiving a dataset including a set of graphs, applying a graph explainer model on the dataset to determine explain scores, and filtering graph elements of each graph based on the explain scores to determine a first set of graph elements. Further, another dataset of randomized graphs is generated. The graph explainer model is applied on the other dataset to determine noise scores, which are used determine denoised explain scores from the explain scores. Graph elements of each graph are filtered based on the denoised explain scores to determine a second set of graph elements. Thereafter, feature vectors are determined for the first set of graph elements and the second set of graph elements, and prediction driving graph elements are determined based on the feature vectors. Explainability information associated with the set of graphs is rendered, based on the prediction driving graph elements.