Neural Network Segmentation for Explainable Latent Feature Extraction
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Machine learning models, particularly neural networks, are often 'black box' in nature, making it difficult to explain their complex relationships and predictions, which is a challenge for business applications where interpretability is required due to regulatory constraints.
Innovation Solution
The development of a neural network architecture that allows for the extraction of interpretable relationships among predictive input variables by simplifying hidden nodes through regularization, enabling the creation of explainable latent features that can be used in other models, such as scorecards, while maintaining predictive power.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If a fully connected neural network model is used to capture complex relationships among predictive input variables, then prediction accuracy is improved, but the model becomes a black box that is not explainable to humans
Solution Approach 1:
The patent segments the complex neural network model into multiple simplified neural network models, each handling a specific subset of relationships. This segmentation allows the complex prediction task to be divided into interpretable components, where each simplified model can be understood by domain experts while collectively maintaining high prediction accuracy.
Solution Approach 2:
The patent introduces simplified neural network models as intermediaries between the complex fully connected neural network and human interpreters. These intermediary models translate the complex latent relationships into simplified representations that preserve predictive power while being explainable to domain experts.
2Loss of information
If an interpretable model is used to ensure explainability of predictions, then model transparency is improved, but prediction accuracy decreases compared to complex machine learning models
Solution Approach 1:
The patent employs dynamic model selection where the complexity of the neural network model is adjusted based on the specific application requirements. For applications requiring high explainability, simplified models are used; for applications prioritizing accuracy, more complex models are deployed. This dynamic approach allows optimization of both explainability and accuracy based on contextual needs.
Solution Approach 2:
The patent changes key parameters of the neural network models, specifically controlling the number of input variables each hidden node receives (limiting to N variables). This parameter change transforms the model from a complex black box to an interpretable model while retaining sufficient predictive capability for many applications.
3Adaptability or versatility
If the number of input variables to each hidden node is increased to capture more complex relationships, then the model's ability to learn complex patterns is improved, but the model becomes less interpretable to domain experts
Solution Approach 1:
The patent segments the input variables and hidden nodes such that each hidden node receives a limited set of N input variables. This segmentation creates a many-to-one relationship where multiple simplified nodes process different subsets of relationships, collectively capturing complex patterns while maintaining individual node interpretability.
Solution Approach 2:
The patent transitions from a single-dimension view (one complex node handling all relationships) to a multi-dimensional architecture where multiple simplified nodes each handle specific relationships. This dimensional change allows the system to maintain both complexity handling capability and interpretability through the distributed architecture.
Data Source
AI summary
Systems and methods that use a neural network architecture for extracting interpretable relationships among predictive input variables. This leads to neural network models that are interpretable and explainable. More importantly, these systems and methods lead to discovering new interpretable variables that are functions of predictive input variables, which in turn can be extracted as new features and utilized in other types of interpretable models, like scorecards (fraud score, etc.), but with higher predictive power than conventional systems and methods.