Neighbor-Based Counterfactual Explanations: A Method for Balancing Similarity and Plausibility

As machine learning models become increasingly intricate, the need for explainability has emerged as a pivotal area of research. While the emphasis has been predominantly on examining the internal architecture, substantial endeavors have been made to investigate the potential for modifying outputs of AI models. In this regard, Counterfactual Explanations (CFEs) have emerged as a pivotal tool, facilitating comprehension of the impact of features on instances and identifying the minimal alterations required to achieve different model outcomes. Applications of CFEs are extensive, spanning various industries. In the financial sector, for instance, banks can utilize CFEs to suggest actions that enhance their clients’ creditworthiness. Plausibility, a paramount factor in CFEs, ensures that they remain coherent and realistic within the confines of the provided data. In this study, we propose an instance-based CFE generation method that strikes a balance between similarity and plausibility. The proposed approach involves the identification of nearest reverse neighbors (NRNs) and the subsequent construction of a search space between the original instance and NRNs, based on differing features. Bayesian Search is then applied to identify the most similar CFE within this search space. While Gower and Hamming distances are utilized in this study, the method is versatile and can incorporate different distance metrics. Furthermore, it enables users to regulate the trade-off between similarity and plausibility by adjusting the number of NRNs considered. These configurations of the proposed method are tested across multiple datasets and compared with other CFE generation methods in the literature, demonstrating competitive results.