CONSISTENCY OF LIME EXPLANATIONS: EVALUATING EXPLAINABLE AI FOR PLANT CLASSIFICATION

With the increasing adoption of machine learning techniques, the demand for explaining decisions has grown. Explainability is the ability of a classification model to provide insights about its decisions, and it has been investigated in the field of explainable artificial intelligence (XAI). Explanation techniques have been proposed recently, and one such technique is Local Interpretable Model-Agnostic Explanations (LIME). This paper investigates whether LIME produces consistent explanations through a case study on plant classification. A convolutional neural network (CNN) classifier is trained with images of leaves, and LIME is applied to provide explanations of the decisions. LIME is evaluated with different parameter values to determine which produced stable explanations. Results show that LIME can consistently provide explanations, but its consistency depends on the number of samples used.