Adaptive Random Trees Integer Non-Linear Programming Scorecard Optimization
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Traditional manual techniques for segmenting a population into risk pools overlook valuable information and require significant time to improve predictive power, as they make branch-by-branch decisions without considering the entire tree structure.
Innovation Solution
The Adaptive Random Trees (ART) and Integer Non-Linear Programming (INLP) system generates optimized scorecards by evolving populations of fully formed trees using genetic algorithms, enabling the determination of optimal score weights for each segment, thereby enhancing predictive power and reducing development time.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Reliability
If traditional manual techniques make branch-by-branch decisions to segment population, then the process is simple to implement, but valuable information is overlooked and predictive power is insufficient
Solution Approach 1:
The system performs preliminary actions by building complete decision trees first, then applying INLP optimization to determine optimal score weights across the entire tree structure before final segmentation. This preliminary full-tree construction allows the system to capture all valuable information from the population data, which is then optimized globally rather than through sequential branch-by-branch decisions.
Solution Approach 2:
The system implements feedback by using INLP optimization to evaluate and adjust score weights based on the complete tree structure and all available data. The optimization process provides feedback on how different weight combinations affect overall predictive power, allowing the system to iteratively improve the segmentation model by considering information from all parts of the tree simultaneously.
2Reliability
If traditional techniques sequentially determine best splits down through the growing tree, then the process is fast to execute, but significant time is spent evaluating all possible splits to improve predictive power
Solution Approach 1:
The system changes parameters by transitioning from sequential branch-by-branch split evaluation to a global optimization approach using INLP. Instead of evaluating splits one branch at a time, the system formulates the entire segmentation problem as an optimization problem where score weights and segment assignments are determined simultaneously, dramatically reducing the time needed to achieve high predictive power.
Solution Approach 2:
The system substitutes the mechanical sequential evaluation process with a mathematical optimization framework. Rather than physically traversing and evaluating each possible split in sequence, the INLP formulation allows the system to compute optimal score weights and segment assignments through mathematical optimization, replacing the step-by-step mechanical process with a more efficient computational approach.
3Quantity of substance
If traditional methods consume data until none is left in leaf nodes, then all data is utilized, but analysts must spend significant time evaluating all possible splits for each subpopulation
Solution Approach 1:
The system achieves universality by creating a unified optimization framework that handles the entire segmentation problem across all subpopulations simultaneously. The INLP formulation allows score weights and segment assignments to be determined in a single comprehensive optimization process, rather than requiring separate evaluations for each subpopulation, thereby utilizing all data efficiently while minimizing evaluation time.
Data Source
AI summary
A system and method for building segmented scorecards for a population is presented. A model of the population is built using a model builder computer, and one or more variables used by the model builder to build the model is stored in a repository. A scorecard is generated for each segment of the population based on the model and using an adaptive random tree computer program. Next, the scorecard for each segment is enhanced using a integer non-linear programming computer program to determine optimal score weights associated with the variables used by the model builder to build the model, and to generate an enhanced segmented scorecard for the population.