Transparent Machine Learning Recommendation Engine
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Machine-learning based recommendation systems operate as black-boxes, making it difficult to determine the contributing factors to recommendations and potential biases, hindering directed feature engineering and transparency in healthcare settings like post-acute care recommendations.
Innovation Solution
A computer-implemented method that receives patient data, executes machine-learning models to generate recommendations, and generates graphical user interfaces showing the relative importance of data entries contributing to these recommendations, using multiple models to identify relevant features and alternative options, and facilitates prior-authorization processes.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If machine-learning models are used to generate recommendations, then recommendation accuracy is improved, but transparency of the recommendation process deteriorates
Solution Approach 1:
The patent introduces an intermediary component that reverse-engineers the machine-learning model's decision process. This intermediary analyzes the black-box model's inputs and outputs to generate explanatory information about which features contributed most to each recommendation, thereby mediating between the accurate but opaque ML model and the users who need transparency.
Solution Approach 2:
The patent extracts key contributing features from the machine-learning model's internal processing and presents them separately to users. By taking out the most influential features and displaying them in a human-understandable format, the system provides transparency without requiring users to understand the complex internal workings of the ML model.
2Reliability
If multiple machine-learning models are executed to improve recommendation quality, then recommendation reliability is improved, but system complexity increases
Solution Approach 1:
The patent combines multiple machine-learning models into a unified recommendation system. Instead of presenting separate outputs from multiple models, the system integrates their results and uses a reverse-engineering approach to explain the combined decision, thereby maintaining reliability while managing complexity through unified processing.
Solution Approach 2:
The reverse-engineering component serves multiple functions: it analyzes inputs, determines feature importance, generates explanations, and presents results to users. This multi-functional approach reduces overall system complexity by consolidating several tasks into a single versatile module rather than requiring separate components for each function.
3Loss of information
If detailed feature analysis is provided to users, then user understanding is improved, but information processing time increases
Solution Approach 1:
Instead of providing complete analysis of all features, the patent applies local quality by identifying and displaying only the most influential features for each specific recommendation. This selective approach provides users with sufficient understanding while avoiding the time cost of processing and displaying information about less relevant features.
Solution Approach 2:
The system performs partial analysis by focusing on the top contributing features rather than analyzing all input features in detail. This partial action provides adequate user understanding for most recommendations without the excessive time investment required for complete feature analysis, striking a practical balance between comprehension and efficiency.
Data Source
AI summary
Machine-learning based recommendation engines are configured to execute machine-learning models to generate recommendations as output for a user based at least in part on functional data received at the recommendation engine. The recommendation engine is further configured to automatically determine the relative importance of one or more functional data entries in generating the recommendation. Moreover, the recommendation engine executes additional machine-learning models, including a machine-learning model trained to avoid negative outcomes, and an opportunity-based machine-learning model to identify alternative recommendation options based on alternative training logic.