SHAP-Based Contribution Analysis for Process Condition Changes
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing machine learning models struggle to interpret the contribution degree of individual conditions to the change in predicted values when process conditions are altered, particularly in semiconductor manufacturing processes.
Innovation Solution
An analysis device and method that utilize the SHAP algorithm to calculate the contribution degree of individual conditions by subtracting SHAP values for different process condition sets, leveraging the properties of total group rationality and additivity.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Reliability
If machine learning models are used to predict complex phenomena, then prediction capability is improved, but interpretability of contribution degree of individual conditions deteriorates
Solution Approach 1:
The patent segments the overall prediction model into individual condition contributions by calculating SHAP values for each explanatory variable separately. This allows the model to maintain its predictive power while providing segmented interpretability showing how each individual condition contributes to the final prediction, thus resolving the contradiction between prediction capability and interpretability.
Solution Approach 2:
The patent introduces SHAP values as an intermediary mechanism that bridges the black-box prediction model and human interpretation. The SHAP values serve as a mediator that translates the complex model outputs into understandable contribution degrees for each condition, allowing both accurate prediction and clear interpretation to coexist.
2Loss of information
If SHAP values are calculated for each explanatory variable, then contribution degree information is obtained, but calculation complexity increases
Solution Approach 1:
The patent performs preliminary calculations by pre-computing SHAP values for all possible subsets of explanatory variables. This preliminary action stores the contribution information in advance, allowing the final contribution degree of each individual condition to be obtained through simple aggregation of pre-computed values, thereby reducing the complexity of real-time calculations.
Solution Approach 2:
The patent calculates SHAP values for all possible combinations of explanatory variables (excessive action), which provides more information than strictly necessary but simplifies the final computation. By computing contributions for all subsets, the system can efficiently derive individual condition contributions without complex real-time calculations.
Data Source
AI summary
According to one embodiment, an analysis device includes a processor. The processor calculates a first contribution degree, which is a contribution degree of a first explanatory variable for a first set, based on a training data set used for generating a prediction model. The prediction model receives explanatory variables and outputs a predicted value corresponding to the explanatory variables. The first set is a set of values of the explanatory variables. The first explanatory variable is one of the explanatory variables. The processor calculates, based on the training data set, a second contribution degree of the first explanatory variable for a second set that is a set of values of the explanatory variables and is different from the first set. The processor obtains a third contribution degree by calculating a difference between the first contribution degree and the second contribution degree and outputs the third contribution degree.