Contrastive Explanation Generation via Latent Embeddings
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Conventional explainable artificial intelligence (XAI) and interpretable machine learning (iML) systems face challenges in providing contrastive explanations for machine-generated decisions and predictions, including the 'plateau effect' and the generation of meaningless or unrealistic pertinent negatives, while lacking extensibility for incorporating semantics.
Innovation Solution
A multi-label joint autoencoder generates latent embeddings of machine learning model predictions within an embedding space, enabling the identification of a nearest flipped neighbor and incorporating domain-specific semantics to provide realistic and meaningful contrastive explanations through interpolation.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Loss of information
If conventional XAI and iML systems provide contrastive explanations by identifying pertinent negatives, then they can explain what was lacking in a prediction, but they suffer from the plateau effect and generate meaningless or unrealistic pertinent negatives
Solution Approach 1:
The patent introduces an embedding space as an intermediary representation layer between the original data and the contrastive explanation. This embedding space transforms high-dimensional data into a lower-dimensional semantic space where meaningful contrasts can be identified, avoiding the plateau effect and meaningless pertinent negatives that occur in conventional direct approaches.
Solution Approach 2:
The patent changes the parameter space by projecting data into an embedding space with different dimensional characteristics. This transformation allows the system to identify pertinent negatives in a transformed space that correspond to meaningful, realistic changes in the original space, thereby improving reliability while maintaining explanation quality.
2Adaptability or versatility
If conventional systems search for contrastive explanations without semantic constraints, then they can find any differing features, but they lack extensibility for incorporating domain-specific semantics
Solution Approach 1:
The embedding space serves as a universal representation that can incorporate multiple types of domain-specific semantics through configurable constraints. The same embedding framework can be adapted to different domains (e.g., medical, financial, legal) by adjusting the semantic constraints, providing both extensibility and preservation of semantic meaning.
3Loss of information
If the system provides detailed contrastive explanations with domain-specific semantics, then explanation quality improves, but computational complexity and search difficulty increase
Solution Approach 1:
The patent extracts the semantic structure into a separate embedding space that can be pre-computed and stored. This extraction allows the complex semantic relationships to be captured once in the embedding space, simplifying subsequent contrastive explanation searches while maintaining high explanation quality.
Data Source
AI summary
Generating a contrastive explanation for a machine learning model prediction includes generating, by a multi-label joint autoencoder, latent embeddings of a plurality of machine learning model predictions. The latent embeddings are datapoints representing the plurality of predictions within an embedding space. The datapoints are positioned by the multi-label joint autoencoder within the embedding space based on a semantic labeling of each datapoint. Based on the latent embeddings, a plurality of computer-searchable data structures is generated. The computer-searchable data structures can identify a nearest flipped neighbor of datapoints within the embedding space. A nearest flipped neighbor determiner is output, the nearest flipped neighbor determiner constructed with the plurality of computer-searchable data structures.