Hybrid Variable Selection for ML Data Conditioning
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing methods for conditioning datasets for machine learning models are inefficient in capturing interaction effects between variables, leading to less effective models due to excessive computational latency when examining all possible interactions.
Innovation Solution
A method for selecting hybrid variables by sampling interaction effect structures, calculating lift values, labeling and training machine learning models, and retaining variables with high discriminatory strength, which involves sampling, calculating moment statistics, and using a GINI coefficient for filtering.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Reliability
If all possible interactions between variables are examined to capture comprehensive information about the target, then model effectiveness is improved, but computational latency becomes excessive
Solution Approach 1:
The patent applies partial action by examining only a subset of possible variable interactions rather than all combinations. It uses statistical sampling to select representative interaction terms that capture the essential relationships in the data, thereby achieving sufficient model effectiveness without the prohibitive computational cost of exhaustive analysis
Solution Approach 2:
The patent segments the vast space of all possible variable interactions into manageable subsets based on statistical criteria. By dividing the comprehensive interaction space into sampled segments, the method enables parallel processing and reduces the computational burden while maintaining model effectiveness
2Reliability
If the number of hybrid variables is increased to capture more interaction effects, then predictive power is improved, but model complexity increases
Solution Approach 1:
The patent changes the parameters used to select hybrid variables by introducing statistical criteria such as p-values, effect sizes, and information gain thresholds. These parameter-based filters enable the model to include only those interaction terms that meet predefined significance levels, thereby improving predictive power while controlling complexity through quantitative parameter constraints
Data Source
AI summary
Embodiments generally relate to a method for selecting hybrid variables. The method comprises sampling at least one interaction effect structure of at least one multivariable dataset, sampling at least one hybrid variable for each sampled interaction effect structure, calculating a lift value for each sampled hybrid variable, and comparing the lift value to a threshold lift criteria, labeling each sampled hybrid variable based on determining that the lift value of the sample hybrid variable exceeds the threshold lift criteria, training a machine learning model to predict the likelihood of a hybrid variable having a lift which exceeds the threshold lift criteria, applying the trained machine learning model to each hybrid variable within each sampled interaction effect structure to determine a value corresponding to the likelihood of each hybrid variable having a lift which exceeds the threshold lift criteria, and retaining only hybrid variables with a likelihood value that exceeds a decision criteria. The training of the machine learning model is performed using the labeled sampled hybrid variables.