Balance Weighted Voting for Faster Ensemble Model Scoring
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Conventional ensemble modeling approaches for predictive analytics face challenges in balancing evaluation-measure based weighting, which focuses on model overall performance, and prediction-measure based weighting, which focuses on single record prediction effectiveness, leading to inefficiencies in resource usage and longer computation times.
Innovation Solution
Implementing balance weighted voting methods that calculate evaluation-based and prediction-based weights using normalized measures and hyperparameters to derive a balanced weighting predictor, allowing for hybrid nested weight filtering to select a preselected number of component models.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Reliability
If evaluation-measure based weighting is used to focus on model overall performance, then prediction accuracy is improved, but computation time increases
Solution Approach 1:
The patent segments the ensemble modeling process into distinct phases: an training phase where comprehensive evaluation-measure based weighting is performed to establish optimal weights, and a scoring phase where pre-calculated weights are applied without re-computation. This segmentation allows accurate predictions during scoring while avoiding repeated computation overhead.
Solution Approach 2:
The patent performs preliminary action by pre-calculating and storing optimal weighting factors during the training phase using comprehensive evaluation measures. These pre-computed weights are then reused during the scoring phase, eliminating the need for repeated computation and significantly reducing computation time while maintaining prediction accuracy.
2Reliability
If multiple component models are used in the ensemble, then prediction effectiveness is improved, but resource demand increases
Solution Approach 1:
The patent changes parameters by dynamically adjusting the number and type of component models based on performance metrics. The system monitors prediction effectiveness and resource consumption, modifying the ensemble composition to maintain optimal performance while reducing resource demand when multiple models are not necessary.
Solution Approach 2:
The patent implements discarding and recovering by selectively removing underperforming or redundant component models from the ensemble while preserving those that contribute most to prediction effectiveness. This reduces the number of active models and thus resource demand, while maintaining or improving prediction performance through optimized model selection.
3Reliability
If comprehensive weighting measures are calculated for all models, then ensemble effectiveness is improved, but computation time increases
Solution Approach 1:
The patent extracts and applies only the most critical weighting measures and performance metrics rather than calculating all possible comprehensive measures. By identifying and focusing on the most influential factors for ensemble effectiveness, the system achieves high effectiveness while avoiding unnecessary computation and improving overall computation efficiency.
Data Source
AI summary
A method, system, and computer for balance weighted voting. The method may comprise receiving, by a network interface, a scoring request. The method may further comprise, by a processing unit in response to the scoring request, generating a plurality of scores using a plurality of models. normalizing the plurality of scores, calculating an evaluation-based weighting factor from a first subset of the normalized scores, calculating a prediction-based based weighting factor from a second subset of the normalized scores, and calculating a balanced weighting predictor from the evaluation-based weighting factor and the prediction-based weighting factor. The method may further comprise returning, by the network interface, the balanced weighting predictor as an ensemble score for the scoring request.