Abstract

The explanation of a predictive model is as important as its prediction accuracy. The existing Local Interpretable Model-Agnostic Explanations (LIME) has major limitations, including the necessity of tuning multiple hyperparameters and a lack of consistent explanation accuracy across various scenarios. This presentation introduces the development of the Multi-Objective Genetic Algorithm Explainer (MOGAE), a novel, model-agnostic local explainer specifically designed for image data. MOGAE significantly advances eXplainable Artificial Intelligence (XAI) by employing the Non- dominated Sorting Genetic Algorithm II (NSGA-II). A key innovation lies in its use of an adaptive Bit Flip Mutation (BFM), which incorporates both densify and sparsify operators. This allows MOGAE to automatically and dynamically adjust superpixel granularity, thereby streamlining the explanation process. Crucially, this innovative approach simplifies model interpretability by eliminating several critical hyperparameters traditionally required by established methods like Local Interpretable Model-Agnostic Explanations (LIME). MOGAE has been evaluated with the citrus disease dataset (an agricultural domain dataset, serving as the main testbed) and the melanoma detection dataset (a medical domain dataset, detailed in the appendix). Experimental results revealed that MOGAE consistently showed superior accuracy of explanation in comparison with the existing LIME library and the similar evolutionary-based EGAE explainer.

About the Speaker

Dr. Hossein Nematzadeh earned his Ph.D. in Computer Science in 2014. His academic career includes serving as an Assistant Professor at Islamic Azad University in Iran and working as a Research Staff member at the University of Malaga in Spain. He is currently an Assistant Professor at the Modern College of Business and Science in Oman. His primary research interests cover various aspects of data mining and data analytics, specifically focusing on areas such as feature selection in small-sample, high-dimensional datasets, eXplainable AI (XAI), and Training Set Selection (TSS).