Semiconductor Data Modeling with Iterative Feature Selection
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing data processing methods for semiconductor manufacturing face a trade-off between preventing over-learning and reducing calculation load when creating model formulas, with current evaluation indices like AIC not always effective in preventing over-learning and cross-validation methods being time-consuming.
Innovation Solution
A data processing apparatus and method that rearranges and evaluates feature amounts using evaluation index values to select significant features, creating data groups and updating their order iteratively to optimize prediction performance, allowing for the deletion of non-contributory feature amounts and improving model generalization without excessive calculation.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If many feature amounts are extracted from time-series signals to represent signal features comprehensively, then the representation accuracy of signal features is improved, but the calculation time increases and over-learning occurs in model creation
Solution Approach 1:
The patent extracts only the significant feature amounts from the time-series signals that contribute to prediction accuracy, rather than using all extracted features. This selective extraction removes meaningless features that would increase calculation time and cause over-learning, while maintaining comprehensive representation of signal characteristics through the significant features alone
Solution Approach 2:
The patent changes the parameter of feature selection by introducing an evaluation index that quantifies the significance of each feature amount. This parameter change enables automatic selection of optimal features based on their contribution to prediction, transforming the feature set from exhaustive to optimized, thereby reducing calculation time while maintaining representation accuracy
2Reliability
If cross-validation method is used to effectively prevent over-learning, then the generalization performance is improved, but the calculation time increases greatly
Solution Approach 1:
The patent extracts and uses only the significant feature amounts identified by the evaluation index, rather than performing cross-validation on all possible feature combinations. This extraction of essential features reduces the search space dramatically, enabling effective over-learning prevention with much lower calculation time compared to full cross-validation
Solution Approach 2:
The patent applies partial cross-validation or simplified validation on the selected significant features rather than exhaustive validation on all features. This partial action provides sufficient over-learning prevention for the optimized feature set while avoiding the excessive calculation time required for complete cross-validation of all possible feature combinations
Data Source
AI summary
A data processing apparatus in which a trade-off between over-learning prevention and calculation load prevention is eliminated when creating a model formula is provided. The data processing apparatus includes: a recording unit that records electronic data; and a computing unit that performs computing using the electronic data, in which the computing unit includes a feature amount selection unit used for computing, and the feature amount selection unit performs feature amount selection including: a first step (S101) of ranking feature amounts and rearranging the feature amounts from top; a second step (S103) of creating a plurality of data groups using only a part of the feature amounts according to the order; a third step (S104) of calculating a value that is an index for evaluating prediction performance of a regression or classification problem using each of the data groups using only a part of the feature amounts; a fourth step (S105) of deleting feature amounts based on the calculated prediction performance index; and a fifth step (S106) of updating the order of the feature amounts, which are feature amounts other than the deleted feature amount, using the prediction performance index, in which the second step to the fifth steps are iterated (S102) until an optimal value of the prediction performance index calculated in the third step is no longer updated.