Editable Rules-Based Model for Transparent Machine Learning
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Machine learning models, such as recommender systems based on Collaborative Filtering or Matrix Decomposition, are not transparent and difficult to edit without retraining, making it challenging to understand and modify their decision-making processes.
Innovation Solution
A computerized method for creating and managing an editable rules-based model that aggregates confidence values from trained models, allowing users to view, edit, and update rules through a model management interface without requiring additional training, enabling transparency and bias identification.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If machine learning models are used to achieve high precision output, then prediction accuracy is improved, but model transparency deteriorates
Solution Approach 1:
The patent creates a rules-based model that copies the behavior and predictions of the trained machine learning model. By extracting rules from the ML model's decision-making process, the system generates a transparent representation that mimics the original model's output while making the decision logic visible and interpretable to users.
2Manufacturing precision
If machine learning models are trained to achieve high precision, then output quality is improved, but model editability deteriorates
Solution Approach 1:
The patent segments the monolithic machine learning model into discrete, editable rules. Each rule represents a specific decision logic extracted from the trained model, allowing individual rules to be modified, added, or removed without affecting the entire model. This segmentation enables granular editing while preserving the overall predictive capability.
Solution Approach 2:
The patent transforms the static, fixed machine learning model into a dynamic system where rules can be continuously updated and refined. The model evolution from initial training through rule extraction to iterative editing creates a flexible framework that adapts to new requirements without requiring complete retraining.
3Measurement precision
If machine learning models are used to achieve high precision output, then prediction accuracy is improved, but model editing complexity increases
Solution Approach 1:
The patent creates simplified rule representations that serve as lightweight proxies for the complex machine learning model. These rules are easier to manipulate and understand, acting as an intermediate layer between the complex ML model and human users. The rules can be independently edited without dealing with the full complexity of the underlying model architecture.
4Manufacturing precision
If machine learning models are trained to achieve high precision, then output quality is improved, but retraining time increases
Solution Approach 1:
The patent performs rule extraction from the trained machine learning model in advance, creating a rules-based representation before any editing is needed. This preliminary transformation allows subsequent rule modifications to be made directly without triggering time-consuming retraining processes, as the rules can be updated independently of the original training data and model architecture.
Data Source
AI summary
The disclosure herein describes creating and managing an editable rules-based model. A model is trained to generate output values for data sources based on input feature data of the data sources. Using the trained model, results are generated that include feature-output value pairs and associated confidence values for each data source. The confidence values are aggregated for each feature-output value pair of the generated results, and rules of a rules-based model are built using the feature-output value pairs and the respective aggregated confidence values of the generated results. A user is provided access to the rules of the rules-based model via a model management interface. The rules-based model is updated based on rule change instructions received from the user via the model management interface.