Monotonic Constraints for Interpretable Machine Learning Models
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing machine learning models face a trade-off between accuracy and interpretability, particularly in credit rating systems, where the impact of individual input variables on the model is obscured, making it difficult to provide transparency and record-level variable importance.
Innovation Solution
Implementing monotonic relationships as constraints in machine learning algorithms to generate more transparent and interpretable models, allowing for the determination of record-level variable importance by altering input values based on these relationships and comparing predicted values to identify the impact of each factor.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If machine learning algorithms are used to generate predictive models, then model accuracy is improved, but interpretability and transparency deteriorate
Solution Approach 1:
The patent introduces monotonic relationships as an intermediary constraint between input variables and model predictions. These monotonic constraints act as a mediator that preserves the predictive power of complex machine learning models while imposing a structured relationship that enables interpretability. The monotonic relationships serve as a bridge between the black-box nature of ML algorithms and the need for transparent, explainable predictions.
Solution Approach 2:
The patent transforms the model structure by changing the parameter constraints from unconstrained to monotonic constrained. By imposing monotonicity constraints on the model parameters, the patent modifies the solution space of the machine learning algorithm to produce models that are both accurate and interpretable. This parameter transformation allows the model to maintain predictive performance while becoming transparent about variable impacts.
2Measurement precision
If the number of input variables is increased, then model accuracy is improved, but model complexity and interpretability worsen
Solution Approach 1:
Monotonic relationships serve as an intermediary structure that organizes the relationship between numerous input variables and predictions. This intermediary framework allows the model to handle many variables without becoming incomprehensible, as the monotonic constraints provide a clear structure for understanding variable impacts.
Solution Approach 2:
The patent segments the complex model interpretation task into individual variable impacts. By evaluating each input variable's effect separately under monotonic constraints, the patent breaks down the complexity of multi-variable interactions into manageable, interpretable components that can be understood individually.
3Loss of information
If traditional modeling techniques are used, then interpretability is maintained, but model accuracy deteriorates
Solution Approach 1:
The patent merges the advantages of traditional interpretable models with the predictive power of advanced machine learning algorithms. By combining monotonic constraint structures (from traditional models) with flexible ML algorithms, the patent creates a hybrid approach that achieves both high accuracy and interpretability simultaneously.
Solution Approach 2:
The patent creates a composite modeling approach that combines the structural discipline of traditional models with the predictive capabilities of modern machine learning. This composite model structure integrates the interpretability framework of traditional methods with the accuracy of ML algorithms, producing a model that exhibits properties of both parent approaches.
Data Source
AI summary
Embodiments herein provide for a machine learning algorithm that generates models that are more interpretable and transparent than existing machine learning approaches. These embodiments identify, at a record level, the effect of individual input variables on the machine learning model. To provide those improvements, a reason code generator assigns monotonic relationships to a series of input variables, which are then incorporated into the machine learning algorithm as metadata. In some embodiments, the reason code generator creates records based on the monotonic relationships, which are used by the machine learning algorithm to generate predicted values. The reason code generator compares an original predicted value from the machine learning model to the predicted values from the machine learning model.