Interpretation of Machine Learning Models : Explainable Boosting Machines, Accumulated Local Effects Plots and Shapney Additive Explanations Plots
Häggblad, Thomas (2025)
2025
All rights reserved. This publication is copyrighted. You may download, display and print it for Your own personal use. Commercial use is prohibited.
Julkaisun pysyvä osoite on
https://urn.fi/URN:NBN:fi:amk-2025062523474
https://urn.fi/URN:NBN:fi:amk-2025062523474
Tiivistelmä
This study examines the interpretability of machine learning models. As the use of complex models increases, understanding how they make predictions has become important. The research task is to investigate how inherently interpretable models compare with black box models when explaining model behaviour.
Key concepts include model interpretability, glass box models, and post hoc explanation methods. The research uses the Explainable Boosting Machine as an inherently interpretable model and XGBoost as a black box model. Post hoc methods include Accumulated Local Effects and Shapley Additive Explanations. Publicly available datasets were used, and the models were applied to both classification and regression tasks.
The results show that Explainable Boosting Machines provide inherently interpretable predictions through additive feature contributions. Post hoc methods offered useful insights into black box predictions. The study concludes that different interpretability tools serve different purposes and combining them supports a more complete understanding of machine learning models.
Key concepts include model interpretability, glass box models, and post hoc explanation methods. The research uses the Explainable Boosting Machine as an inherently interpretable model and XGBoost as a black box model. Post hoc methods include Accumulated Local Effects and Shapley Additive Explanations. Publicly available datasets were used, and the models were applied to both classification and regression tasks.
The results show that Explainable Boosting Machines provide inherently interpretable predictions through additive feature contributions. Post hoc methods offered useful insights into black box predictions. The study concludes that different interpretability tools serve different purposes and combining them supports a more complete understanding of machine learning models.