Feature Importance Modeling for Interpretable Supply Chain PGMs
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing supply chain machine learning systems generate convoluted probabilistic graphical models that require significant time and resources to interpret and do not effectively assess the relative importance of features and feature classes, hindering efficient decision-making.
Innovation Solution
A probabilistic graphical model feature importance system using LASSO regression analysis to prune and bin features, generating simplified models that highlight important features and classes, enabling rapid inference and planning decisions.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Reliability
If a Bayesian network PGM is used to model the flow of all variables through a supply chain network, then the model comprehensively represents all variables and their conditional dependencies, but the model becomes extremely dense and convoluted, requiring significant time and resources to interpret and derive meaningful inferences
Solution Approach 1:
The patent extracts and removes less important features and edges from the Bayesian network PGM based on calculated importance scores. This extraction process eliminates unnecessary complexity while retaining the essential relationships between variables, producing a simplified model that maintains reliability for decision-making purposes.
Solution Approach 2:
The patent segments the features into different importance levels through binning, allowing the model to be analyzed and interpreted in manageable segments. By categorizing features into high, medium, and low importance groups, the system enables focused interpretation on critical segments without being overwhelmed by the entire complex model.
2Reliability
If a Bayesian network PGM is used to model the flow of all variables through a supply chain network, then the model represents all variables and their dependencies, but the system does not assess and emphasize the relative importance of some features and feature classes, which is undesirable
Solution Approach 1:
The patent introduces a new parameter - feature importance score - to the Bayesian network PGM. By calculating and assigning importance scores to features based on their impact on target variables, the system transforms the uniform treatment of all features into a differentiated assessment, highlighting which features matter most for supply chain decisions.
Solution Approach 2:
The patent introduces an intermediary analysis layer that calculates feature importance scores and binning information. This intermediary process acts as a mediator between the raw Bayesian network model and the decision-making process, translating the complex model into actionable insights about which features drive supply chain outcomes.
3Reliability
If all features are included in the PGM model, then the model is comprehensive, but the time and resources required to interpret and derive meaningful inferences increases significantly
Solution Approach 1:
The patent applies partial action by including only the most important features in the final simplified model, rather than all possible features. By using feature importance thresholds and binning to select a subset of critical features, the system achieves sufficient model completeness for decision-making while dramatically reducing interpretation time and computational resources required.
Data Source
AI summary
Embodiments of the following disclosure provide a feature importance system and method to identify features relevant to determining whether a target variable will achieve a particular value. One example method includes generating a probabilistic graphical model to represent the performance of one or more entities in a supply chain and selecting or more target variables. The method further includes collating a list of features pertaining to the one or more selected target variables, pruning at least one of the one or more features from the list and generating one or more bins in which to distribute the one or more features in the list. The method further includes modeling a network graph incorporating the one or more features in the list and bins and determining one or more inferences pertaining to the one or more supply chain entity target variables.