Explainable Artificial Intelligence (XAI) methods like LIME provide local explanations for black-box models by approximating their behavior with interpretable surrogate models. However, LIME’s reliance on random perturbations and sampling can cause instability and locality issues, especially when training data is scarce. This scarcity leads to unrealistic sample generation that deviates from the true data manifold, reducing the surrogate model’s ability to accurately capture the original model’s decision boundary. To overcome these challenges, the authors propose ITL-LIME, which integrates instance-based transfer learning into the LIME framework to improve explanation quality in data-constrained environments. ITL-LIME leverages relevant real instances from a related source domain, identified through clustering and prototype selection, to supplement the target domain data. Instead of generating random perturbations, the method retrieves source instances from clusters whose prototypes closely resemble the target instance, combining them with neighboring target instances. A contrastive learning-based encoder is then used to assign weights to these combined instances based on their proximity to the target instance, defining a compact locality for the surrogate model. This weighted set of source and target instances is used to train the surrogate model, enhancing explanation fidelity and stability. The approach mitigates the risks of unrealistic sample generation and instability inherent in LIME under low-resource conditions. ITL-LIME’s methodology suggests potential improvements in interpretability for applications with limited data, offering a pathway to more reliable local explanations through transfer learning and advanced weighting mechanisms.
👉 Pročitaj original: arXiv AI Papers
