EPLIME: An Edge-Driven, Proxy-Free Framework for Local Model-Agnostic Explanations

To mitigate the dual challenges of instability and redundancy in local model-agnostic explanations, this paper introduces EPLIME, an Edge-Driven, Proxy-free Local Interpretation Method. EPLIME tackles instability by leveraging intrinsic edge information to guide a deterministic perturbation process, which eradicates the variance caused by random sampling. To enhance the explanation's sparsity, it refines an initial high-confidence result by iteratively eliminating redundant superpixels based on a robust, multi-operator feature importance evaluation. This refinement process yields a final explanation that is significantly more concise while maintaining high fidelity. Experimental results demonstrate that EPLIME produces stable and substantially sparser explanations while maintaining strong fidelity.