Use of Fitness Sharing in the Local Rule-Based Explanations Method

Abstract

Recent machine learning algorithms present remarkable results in many problems. However, the decisions made by some of these algorithms are very often difficult for human experts to interpret. Some recent works in the literature try to minimize this disadvantage, proposing algorithms that explain the decisions taken by any black-box model. One of these models is the Local Rule Based Explanations (LORE), that generates local explanations by using a Decision Tree (DT) that locally reproduces the decision boundaries of the black-box model. In LORE, the DT is trained by using an artificial dataset generated by a standard Genetic Algorithm (GA). In this paper, we show that the GA employed in LORE does not necessarily preserve the diversity of solutions in the final population. The diversity of the population is important to generate DTs that can more accurately reproduce the decision boundaries of the black-box model close to the instance to be explained. We then propose the use of fitness sharing in the GA in order to preserve the diversity of the population and generate local decision boundaries of the DT more similar to those of the black-box model. Experimental results show that LORE with fitness sharing generally produces better local explanations.