A Faithful and Stable Explanation Method for Image Classification

Deep learning models, despite their exceptional predictive performance, suffer from opaque decision-making processes and require explainability to build trust, ensure accountability, and validate model decisions, especially in critical application domains. A prominent explanation method is Local Interpretable Model-agnostic Explanation (LIME), which explains machine learning models' decision by fitting a simple interpretable surrogate model in the vicinity of the explained sample. However, LIME is limited by its deficiencies in fidelity, stability, and comprehensibility, which stem from LIME's random perturbation and super-pixel representation. Recent improvements, such as GLIME-BINOMIAL and ConceptLIME, partially address these problems by introducing weighted sampling and concept representation, respectively. However, they each have their own limitations, and none simultaneously achieves high stability, fidelity, and comprehensibility. To address these challenges, this work introduces Stable Concept-based LIME (SCLIME), which enhances LIME's fidelity, stability, and comprehensibility simultaneously. SCLIME formulates its explanations using semantic concepts, which are extracted independently of annotated datasets, thereby enhancing comprehensibility. Subsequently, the method perturbs the explained sample with concepts through a rejection sampling process to generate samples nearby for surrogate model construction, which improves fidelity and stability. Comprehensive experiments across multiple datasets demonstrate SCLIME's effectiveness compared to a wide range of relevant methods. SCLIME significantly outperforms LIME in all evaluated metrics, achieving state-of-the-art fidelity and stability while providing enhanced comprehensibility. Module comparisons and ablation studies further validate each module's contribution to SCLIME's overall improvements.