Improving Local Fidelity and Interpretability of LIME by Replacing Only the Sampling Process With CVAE

Knowing the basis of decisions is essential for using Machine Learning (ML) in various regions, such as medical diagnosis, finance, and organizational decision-making. The Local Interpretable Model agnostic Explanation (LIME) algorithm is an Explainable AI (XAI) method that can be applied to many black -box models. However, there is a problem in that the local fidelity of the interpretable model decreases. Therefore, local fidelity must be increased, and the more interpretability of the local surrogate model, the better. Moreover, the fewer the changes, the easier it is to apply them to other studies and the more significant the contribution to the XAI area. Therefore, there are studies that improve fidelity by changing only the sampling. However, no studies have been identified that improve local fidelity and interpretability by modifying only the sampling. To this end, we use a conditional variational autoencoder (CVAE) to modify the sampling of LIME in the classification task to improve local fidelity and interpretability. We conducted a study to improve local fidelity and interpretability by modifying only the sampling part and clarifying the mechanism behind this improvement. Results show that the proposed method improves local fidelity while increasing interpretability compared to existing methods and the factors contributing to this improvement. We also conducted experiments on stability, stability against noise, and execution time, and confirmed that the proposed method performs well, although not as well as methods specific to these issues. Our code is publicly released on GitHub (<uri xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">https://github.com/YasuiDaisuke/CVAE-LIME.git</uri>) under the Apache 2.0 License.