The final degree project aims to enhance the Local Interpretable Model-agnostic Explanations (LIME) method, which is widely used in machine learning to provide interpretable explanations for the predictions of complex, opaque models. LIME typically generates explanations by approximating the local decision boundary of a black-box model with a simpler, interpretable surrogate model based on perturbed input data and the corresponding predictions. This project proposes an innovative approach by incorporating partial derivatives into the LIME framework. Specifically, it examines how the partial derivatives of the model’s output with respect to its inputs can be utilized to refine the construction of the surrogate model. By using these derivatives, the project seeks to achieve a more precise representation of the model's behavior in the vicinity of the instance being explained, capturing not just the output variations but how these outputs change in response to small changes in inputs. This approach hypothesizes that integrating gradient information will allow the surrogate model to more accurately mimic the local decision surface of the original model, thus providing more faithful and insightful explanations. The project involves theoretical development, algorithmic enhancements, and extensive experimental validation using various datasets to evaluate the effectiveness of the modified LIME method in delivering clearer, more accurate interpretative insights into the functioning of complex models.
