Abstract
Path-based counterfactual explanation is an explanatory method that proposes paths from input features to obtain desirable prediction outcomes from a trained classifier. The demand for such explanations is increasing in areas such as loan approval and health diagnostics. However, conventional methods face several challenges. Specifically, they propose zigzag paths that are hard to interpret, and find it challenging to incorporate sparsity. To address these issues, this study proposes a method that defines paths as solutions to a curve energy minimization problem, approximates this optimization problem using the finite difference method, and reduces it to a nonlinear programming problem. Theoretically, we demonstrate that this energy minimization problem can be evaluated using this approximation method as proportional to the -1/2 power of the number of partitions of the integration interval. In the experiments, we confirm the acquisition of smooth paths and the enhancement of sparsity, specifically reducing the necessary features of conventional methods to 73%.
