Clustering-Guided Counterfactual Design Evolution

Counterfactuals provide valuable insights into how changes in input variables can affect outcomes. However, generating them in high-dimensional spaces presents significant challenges due to the data sparsity and increased complexity. Traditional methods often yield unrealistic or misleading explanations. This study tackles the specific challenge of generating reliable and actionable counterfactuals in vehicle chassis design problem by introducing clustering-guided counterfactual explanations (CCF). In the proposed approach, clustering helps to focus on more relevant regions of the search space, and a single-objective genetic algorithm (GA) searches for counterfactual designs. Experiments demonstrate that the proposed approach not only enhances the validity of counterfactuals but also improves the overall understanding of design decision spaces, in contrast to earlier methods that struggled with high-dimensional data. This research contributes to a broader framework for explainable AI and has the potential to address diverse problems. Ultimately, this work underscores the need for robust methodologies in counterfactual generation, which can advance knowledge across various scientific and engineering fields.