Graph Explainer Denoising for Global Prediction Driver Search
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing explainable graph machine learning models generate localized and often inconclusive explanations for predictions, leading to confirmation bias and inaccuracies due to noise in explanation scores, which affects transparency and trustworthiness.
Innovation Solution
A method involving cumulative denoising and clustered search is employed to process explanation scores globally, identifying and isolating noise through randomized graph datasets and statistical analysis, followed by clustering to determine actual prediction-driving graph elements.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Ease of manufacture
If explanations are generated as a post-processing step after prediction, then the model can generate explanations for predictions, but the explanations become inconclusive and may incorrectly identify graph elements driving predictions
Solution Approach 1:
The patent applies preliminary action by integrating explanation generation into the prediction process itself rather than as a post-processing step. The graph explainer model operates concurrently with the prediction model to identify prediction-driving graph elements in real-time, ensuring accurate identification while maintaining ease of explanation generation.
2Adaptability or versatility
If explanations are generated on a localized basis for each prediction in silo, then each prediction can have its own explanation, but confirmation bias affects transparency and explanations differ significantly even for same input data
Solution Approach 1:
The patent merges localized explanation generation with global statistical analysis. The system combines individual prediction explanations across multiple predictions to identify consistent prediction-driving graph elements, eliminating confirmation bias and ensuring transparency while maintaining the ability to generate explanations for each prediction.
3Loss of information
If a graph explainer model is applied to determine explanation scores, then explanations can be generated for predictions, but noise in the explanation scores leads to incorrect identification of graph elements
Solution Approach 1:
The patent extracts noise from explanation scores through statistical analysis. The system separates signal from noise by analyzing explanation scores across multiple predictions and identifying consistent patterns, thereby improving the accuracy of graph element identification while preserving information about prediction-driving elements.
4Measurement precision
If cumulative explanation scores are analyzed globally through statistical analysis and clustered search, then noise can be isolated and denoised scores determined, but the processing complexity increases
Solution Approach 1:
The patent uses copying by creating randomized graph datasets that replicate the structure and properties of original graphs. By analyzing explanation scores on these copied randomized graphs, the system can isolate noise patterns and determine denoised scores for original predictions, improving accuracy while managing processing complexity through efficient statistical methods.
Data Source
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.