Machine learning models are used in various areas. In addition to the accuracy of predictive models, their comprehensibility is also important. Understanding the machine learning model provides confidence in it. By understanding the predictive model, we can determine its bias and causes of errors. Complex models such as random forests, neural networks and support vector machines are not easy to understand and act as black box models; therefore, for their explanations we use post-hoc explanation methods that are model-independent and use perturbation sampling to explain each instance. The robustness of perturbation explanation methods has so far been poorly researched. Recent research has shown that due to poor perturbation sampling, these methods can be manipulated so that they do not recognize a biased classifier. In this work, we propose the use of better sampling, which prevents such manipulations. The proposed sampling uses data generators that better capture the training set distribution. We show that improved sampling increases the robustness of the LIME and SHAP explanation methods and speeds up the convergence of the IME explanation method.
|
