Dataset-Free Feature Attribution via Parametric Distributions
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Complex deep learning models are challenging to interpret due to their increased complexity, which hinders the explanation and understanding of their predictions and behavior, necessitating alternative techniques for machine learning explainability (MLX).
Innovation Solution
The approach involves fitting parametric distributions to each feature's empirical marginal distribution in a machine learning model, allowing for perturbation-based feature attribution without requiring a reference dataset, thereby generating more accurate and reliable explanations while reducing computational resources.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Productivity
If deep learning models increase complexity and depth to improve performance, then model capacity and generalization improve, but interpretability and ease of understanding deteriorate
Solution Approach 1:
The patent segments the complex deep learning model into individual features and their contributions, analyzing each feature's impact separately through perturbation methods. This allows the complex model to be understood in terms of discrete, interpretable feature contributions rather than as an opaque whole.
Solution Approach 2:
The patent introduces feature attribution scores as an intermediary between the complex model and human understanding. These scores act as a mediator that translates the model's internal complex computations into interpretable measurements of feature importance, enabling explanation without simplifying the underlying model architecture.
2Measurement precision
If perturbation-based feature attribution uses empirical marginal distributions from reference datasets, then explanation accuracy improves, but computational resources and data requirements increase
Solution Approach 1:
The patent replaces the need for large reference datasets with simple parametric distributions (such as uniform or normal distributions) that can be generated computationally cheaply. These lightweight distributional assumptions substitute for expensive data collection and processing, providing sufficient perturbation coverage without requiring substantial datasets.
Solution Approach 2:
The patent changes the approach from using empirical data distributions to using parametric distribution models. By parameterizing the marginal distributions (e.g., specifying mean and variance for normal distributions), the method achieves the same perturbation goals with minimal data, transforming a data-intensive problem into a parameter-based solution.
3Reliability
If perturbation-based feature attribution uses empirical marginal distributions from reference datasets, then explanation reliability improves, but computational time and resources increase
Solution Approach 1:
The patent performs preliminary action by establishing parametric distribution models that can be reused across multiple feature attribution computations. Once the simple parametric distributions are defined, they can be applied repeatedly without requiring reprocessing of large reference datasets, significantly reducing computational time for subsequent explanations.
Data Source
AI summary
In an embodiment, a computer assigns a respective probability distribution to each of many features that include a first feature and a second feature that are assigned different probability distributions. For each original tuple that are based on the features, a machine learning (ML) model infers a respective original inference. For each feature, and for each original tuple, the computer: a) generates perturbed values based on the probability distribution of the feature, b) generates perturbed tuples that are based on the original tuple and a respective perturbed value, c) causes the ML model to infer a respective perturbed inference for each perturbed tuple, and d) measures a respective difference between each perturbed inference and the original inference. A respective importance of each feature is calculated based on the differences measured for the feature. Feature importances may be used to rank features by influence and/or generate a global or local ML explainability (MLX) explanation.