Statistical Impact Analysis Machine for Priority-Based PLS Prediction
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Current Partial Least Squares (PLS) algorithms, such as Latent Variable PLS (LV-PLS), lack the ability to specify strategic priorities for performance measures, leading to suboptimal prediction of higher-priority performance metrics and are limited by high correlation among predictor variables and ineffective coefficient constraints.
Innovation Solution
A computer-implemented apparatus employing a value-based weighting partial least squares process with an inside approximation scheme to optimize latent variable weights according to prediction priorities, and a patient PLS regression process to calculate path coefficients between predictor and dependent latent variables, allowing for improved prediction and control of industrial or commercial processes.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Adaptability or versatility
If traditional LV-PLS algorithm is used to estimate latent variable case values as linear combinations of manifest variables, then the model structure best fitting the given data set is obtained, but the user cannot specify strategic priorities of performance measures to tune predictions for higher-priority measures
Solution Approach 1:
The patent introduces dynamic weighting factors that allow the PLS model to adaptively prioritize different performance measures based on user-specified strategic priorities. The weighting scheme modifies the traditional PLS algorithm to give different importance levels to different dependent variables, enabling the model to dynamically adjust predictions according to organizational priorities rather than treating all measures equally.
Solution Approach 2:
The patent changes the parameter structure by introducing priority-based weighting parameters into the PLS algorithm. These parameters allow users to specify the relative importance of different performance measures, fundamentally altering how the model processes and prioritizes information when making predictions about high-priority performance measures.
2Productivity
If traditional PLS regression methods based on NIPALS are used, then the algorithm is computationally simple, but when predictor variables are highly correlated, only one or two components explain most variation and subsequent components are limited
Solution Approach 1:
The patent applies local quality by treating different predictor variables and components differently based on their specific characteristics. Instead of applying a uniform approach to all components, the method identifies and prioritizes components that are most relevant to high-priority performance measures, giving them greater weight in the prediction process while still maintaining computational efficiency.
Solution Approach 2:
The patent introduces dynamic component selection and weighting that adapts to the correlation structure of the data. Rather than being limited to fixed one or two components, the algorithm dynamically determines which components are most valuable for predicting specific high-priority measures, allowing more components to be utilized when they provide predictive value while maintaining computational tractability.
3Ease of manufacture
If traditional PLS models are used, then the implementation is straightforward, but there is no effective way of constraining the coefficients used in the regression
Solution Approach 1:
The patent applies preliminary action by pre-specifying constraint parameters and priority weights before running the regression analysis. Users can define the desired properties of coefficients (such as sparsity, magnitude limits, or relationships between coefficients) in advance, and the algorithm incorporates these constraints during the estimation process, making coefficient control an integral part of the model specification rather than a post-hoc adjustment.
Data Source
AI summary
A computer-implemented initial run module processes manifest variable data using computer-defined model specification parameters stored in a database to provide initial estimates of weights that are associated with latent variables. The initial run module employs a unique value-based weighting partial least squares computer-implemented process. A final run module then operates upon the manifest variable data to determine the importance of the predictor values that are then used to control the industrial, manufacturing or commercial process. The final run module implements a unique patient partial least squares regression model utilizing a boosting learning technique.