GAN Counterfactual Explanations for Realistic ML Interpretability
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Complex machine learning models, particularly deep learning models, are challenging to interpret due to increased complexity, making it difficult to provide meaningful counterfactual explanations that are realistic and relevant to human understanding, especially when dealing with tabular datasets containing both numerical and categorical features.
Innovation Solution
Utilizing Generative Adversarial Networks (GANs) to generate realistic counterfactual explanations by optimizing a generator model to produce synthetic samples that are locally similar to the instance of interest while having different predictions, handling both numerical and categorical features and adhering to the training data distribution.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Adaptability or versatility
If deep learning models with tens to hundreds of layers are used to solve complex problems, then model capacity and generalization improve, but interpretability and ease of understanding deteriorate
Solution Approach 1:
The patent introduces counterfactual explanations as an intermediary mechanism between the complex deep learning model and human users. These explanations serve as a mediator that translates the model's complex internal reasoning into comprehensible causal relationships, allowing users to understand model predictions without needing to comprehend the underlying complex architecture
Solution Approach 2:
The patent transforms the complex model's internal parameters and operations into a different representation form - counterfactual explanations that describe causal relationships in natural language. This parameter transformation converts high-dimensional weight matrices and activation patterns into human-understandable statements about feature importance and causal mechanisms
2Loss of information
If counterfactual explanations are generated by perturbing original samples, then causal relationships are revealed, but realism and relevance to human understanding deteriorate
Solution Approach 1:
The patent applies parameter changes by transforming perturbed sample data into a different representation - counterfactual explanations that describe causal relationships in natural language. This transformation converts numerical perturbations into meaningful statements that humans can understand, maintaining both causal clarity and readability
Solution Approach 2:
The patent replaces the mechanical perturbation process with a linguistic description system. Instead of simply showing perturbed numerical values, the system substitutes this with natural language explanations that describe the causal relationships, making the information both causally accurate and human-comprehensible
3Manufacturing precision
If minimal perturbations are applied to generate counterfactuals, then local similarity is maintained, but the number of operations and computational complexity increase
Solution Approach 1:
The patent applies local quality by focusing the explanation generation process on local perturbations around the specific instance of interest. The counterfactual explanations are generated by making minimal, targeted changes to the original sample's features, concentrating computational effort on the most relevant local regions of the feature space rather than exhaustive global analysis
Data Source
AI summary
Herein are counterfactual explanations of machine learning (ML) inferencing provided by generative adversarial networks (GANs) that ensure realistic counterfactuals and use latent spaces to optimize perturbations. In an embodiment, a first computer trains a generator model in a GAN. A same or second computer hosts a classifier model that inferences an original label for original feature values respectively for many features. Runtime ML explainability (MLX) occurs on the first or second or a third computer as follows. The generator model from the GAN generates a sequence of revised feature values that are based on noise. The noise is iteratively optimized based on a distance between the original feature values and current revised feature values in the sequence of revised feature values. The classifier model inferences a current label respectively for each counterfactual in the sequence of revised feature values. Satisfactory discovered counterfactuals are promoted as explanations of behavior of the classifier model.