Publication Date
2022
Conference Title
Modeling, Simulation and Visualization Student Capstone Conference 2022
Conference Track
General Sciences & Engineering
Document Type
Paper
Abstract
The explainability of a model has been a topic of debate. Some research states explainability is unnecessary, and some ”white-box” models, such as regression models or decision trees, are inherently explainable. This paper conducts a multiple regression model analysis with highly correlated features to illustrate how the model’s explainability fails when dealing with complex data. In this case, trusting the model explanations can be problematic. The Shapley net effect technique, which helps determine the marginal contribution of the features, is employed to improve the model explainability and reveal more information about the prediction. The work concludes that explainability is necessary to avoid biased and erroneous conclusions in all circumstances, including simple models or even more apparent cases.
Keywords:
Machine learning model, Explainability, Explainable AI, Feature importance
Start Date
4-14-2022
End Date
4-14-2022
Recommended Citation
Grigoryan, Gayane and Collins, Andrew J., "Is Explainability Always Necessary? Discussion on Explainable AI" (2022). Modeling, Simulation and Visualization Student Capstone Conference. 2. DOI:10.25776/2ta8-8058 https://digitalcommons.odu.edu/msvcapstone/2022/scienceengineering/2
DOI
10.25776/2ta8-8058
Is Explainability Always Necessary? Discussion on Explainable AI
The explainability of a model has been a topic of debate. Some research states explainability is unnecessary, and some ”white-box” models, such as regression models or decision trees, are inherently explainable. This paper conducts a multiple regression model analysis with highly correlated features to illustrate how the model’s explainability fails when dealing with complex data. In this case, trusting the model explanations can be problematic. The Shapley net effect technique, which helps determine the marginal contribution of the features, is employed to improve the model explainability and reveal more information about the prediction. The work concludes that explainability is necessary to avoid biased and erroneous conclusions in all circumstances, including simple models or even more apparent cases.