Interpreting complex machine learning models is a critical challenge, especially for tabular data where model transparency is paramount. Local Interpretable Model-Agnostic Explanations (LIME) has been a very popular framework for interpretable machine learning, also inspiring many extensions. While traditional surrogate models used in LIME variants (e.g. linear regression and decision trees) offer a degree of stability, they can struggle to faithfully capture the complex non-linear decision boundaries that are inherent in many sophisticated black-box models. This work contributes toward bridging the gap between high predictive performance and interpretable decision-making. Specifically, we propose the NDT-LIME variant that integrates Neural Decision Trees (NDTs) as surrogate models. By leveraging the structured, hierarchical nature of NDTs, our approach aims at providing more accurate and meaningful local explanations. We evaluate its effectiveness on several benchmark tabular datasets, showing consistent improvements in explanation fidelity over traditional LIME surrogates.
Figures & tables
Figure 1: Transforming a DT to its equivalent NDT.
Figure 2: Local decision boundaries of surrogate models on the Wine dataset (a subset of 2 features and 2 classes). From left to right: a standard DT, an NDT initialized from the DT without fine-tuning, and a fine-tuned NDT. While the DT exhibits rigid, axis-aligned partitions, the non-fine-tuned NDT produces degraded and poorly structured decision regions due to soft splits applied without optimization. Fine-tuning the NDT restores coherent and smooth decision boundaries, yielding a faithful approximation of the black-box model’s local decision boundary.
Figure 3: Stability matrices of DT-LIME (left) and NDT-LIME (right) explanations for a fixed instance from the Digits dataset.
Figure 4: Visual summary of the results comparing NDT-LIME with other LIME variants in terms of stability-vs-fidelity, regularity-vs-fidelity, and stability-vs-regularity. The results in tabilar format are presented in Tabs. 5 - 7 .
Figure 5: Local fidelity comparison of LIME surrogates (LR-LIME, DT-LIME, and NDT-LIME) across increasing complexity of black-box models on benchmark datasets.
Figure 6: Regularity of LIME surrogates (LR-LIME, DT-LIME, and NDT-LIME) across increasing k -NNs on benchmark datasets.
LIME variant
Datasets used
GLIME
No tabular data
DLIME
Breast Cancer Wisconsin, Indian Liver Patient,
Hepatitis
MPS-LIME
No tabular data
KL-LIME
No tabular data
MeLIME
Iris
Table 4: A list of the datasets used in the LIME-related literature.