Prediction Combiner Using Subset Segmentation and Selection
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing machine learning and prediction systems vary significantly in their ability to make accurate predictions, and existing techniques for combining multiple systems, such as pipelining, rescoring, and average ensembling, have limitations in improving overall performance.
Innovation Solution
A method that involves obtaining optimal selection values to maximize a predefined performance measure by splitting and combining predictions from multiple systems into intersection, fully-disjoint, and partially-disjoint subsets, selecting predictions based on these values, and merging them to generate a combined prediction that outperforms individual systems.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Reliability
If multiple prediction systems are combined using traditional techniques (pipelining, rescoring, average ensembling), then the ability to utilize diverse system strengths is improved, but the overall performance improvement is limited and the system complexity increases
Solution Approach 1:
The patent segments predictions into different subsets (intersection subset containing predictions from all systems, fully-disjoint subsets containing predictions from only one system, and partially-disjoint subsets) and applies different selection strategies to each subset. This segmentation allows the system to handle different prediction types optimally without uniformly increasing complexity across all predictions.
Solution Approach 2:
The patent dynamically selects which prediction systems to use and how to combine their predictions based on the specific input sample and the characteristics of each prediction subset. The selection values determine the degree to which predictions from different systems are combined, allowing the system to adapt its complexity level depending on the prediction task at hand.
2Ease of operation
If predictions from multiple systems are combined using fixed ratio methods, then the processing simplicity is improved, but the ability to optimize for different prediction types is worsened
Solution Approach 1:
The patent applies different combination strategies and selection values to different subsets of predictions based on their characteristics. The intersection subset may use one combination approach while fully-disjoint subsets use another, allowing each prediction type to be handled with the most appropriate method rather than applying a uniform simple ratio across all predictions.
Solution Approach 2:
The patent changes the selection values (parameters) based on the specific prediction type and subset being processed. These selection values control the weighting and combination of predictions from different systems, allowing the system to optimize performance for different prediction types by adjusting these parameters rather than using fixed ratios.
3Productivity
If all predictions from multiple systems are merged equally, then the computational efficiency is improved, but the precision of the combined prediction is worsened
Solution Approach 1:
The patent applies partial action by selectively combining predictions rather than merging all predictions equally. The selection values determine the proportion of predictions from each system that are included in the final combined prediction, allowing the system to include only the most valuable predictions while maintaining computational efficiency.
Solution Approach 2:
The patent introduces selection values as intermediary parameters that mediate between the raw predictions from multiple systems and the final combined prediction. These selection values act as a filtering mechanism that preserves precision by controlling which predictions are included in the final output while maintaining the efficiency of automated processing.
Data Source
AI summary
Combining predictions made by multiple different prediction systems, including: obtaining a new input sample; obtaining a combiner which comprises optimal selection values that are configured to maximize a predefined performance measure; automatically applying multiple different prediction systems to the new input sample, to generate predictions; and automatically selectively combining the generated predictions based on the optimal selection values, to generate a combined prediction whose predefined performance measure is improved compared to individual usage of each of the prediction systems. The selective combination includes splitting the generated predictions into an intersection subset, fully-disjoint subsets, and optionally one or more partially-disjoint subsets; selecting predictions, from the subsets, at a ratio corresponding to the optimal selection values associated with the subsets, and modifying the subsets to remove all non-selected predictions; and merging all predictions in the modified subsets, to produce the combined prediction.