AI Safety, Alignment & Interpretability
The machine learning model and feature collinearity highly affect SHAP and LIME explanations, a biomedical case study suggests
Researchers discussed two common explainable artificial intelligence tools, SHAP and LIME. In a case study that classified people with or without a myocardial infarction, a heart attack, the particular machine-learning model and collinearity among input features highly affected the explanation scores from both tools. The authors say this raises a note of caution about using and interpreting them.
What are SHAP and LIME?
SHAP is short for SHapley Additive exPlanations. LIME is short for Local Interpretable Model Agnostic Explanation. Both are explainable artificial intelligence methods. They try to make the behavior of a machine-learning model easier for people to understand. Researchers use them widely with tabular data, the kind of data arranged in rows and columns. The authors discuss how the two methods create their explanation scores and propose a way to interpret those scores, with attention to strengths and weaknesses.
What did the paper examine?
The paper is a perspective piece, not a fresh experiment. It discusses how SHAP and LIME explanation scores depend on the machine-learning model and on collinearity among features. The case study comes from the biomedical domain and classifies people with or without myocardial infarction.
FAQ
Why does model-dependence matter?
The authors report that SHAP and LIME explanations are highly affected by the machine-learning model that was adopted. The authors raise a note of caution on their usage and interpretation.
What is feature collinearity?
Collinearity generally refers to input features that are closely related to one another. The authors found that SHAP and LIME were highly affected by feature collinearity in the case study.
Does the finding apply outside this medical case?
The abstract describes only a biomedical case study, so the reported result is tied to that setting. The abstract does not say whether the authors tested other settings.
Limits and what's next
The abstract does not list limitations. It also does not state specific next steps. The paper is peer-reviewed, and Advanced Intelligent Systems published it in 2024.
Source: Ahmed Salih, Zahra Raisi-Estabragh, Ilaria Boscolo Galazzo, Petia Radeva, Steffen E. Petersen, Karim Lekadir, and Gloria Menegaz, "A Perspective on Explainable Artificial Intelligence Methods: SHAP and LIME," Advanced Intelligent Systems, 2024.