This part looks alike this: identifying important parts of explained instances and prototypes

TL;DR

The paper tackles interpretability for prototype-based explanations on tabular data by aligning feature importance across instances and their nearest prototypes to identify alike parts. It uses SHAP-based per-feature importances, normalizes and squares them to produce $\hat{\phi}$, defines per-feature weight $w_l = \hat{\phi}(h,\mathbf{x}_i^l)\hat{\phi}(h,\mathbf{p}_j^l)$, and derives a binary mask for the shared informative features. A FI-informed objective $ f(\mathcal{P}) = \sum_{i=1}^{|\mathcal{S}|} \min_{\mathbf{p} \in \mathcal{P}} \left( d(\mathbf{x}_i, \mathbf{p}) + \beta \cdot fi(\mathbf{x}_i, \mathbf{p}) \right) $, with $ fi(\mathbf{x}_i, \mathbf{p}) = \sum_{l=1}^d \left( \hat{\phi}(h, \mathbf{x}_i^l) \right)^2 \left( \hat{\phi}(h, \mathbf{p}^l) \right)^2 $, balances distance and feature-importance to yield diverse prototypes. Experiments on six datasets show improved user comprehension and preserved or improved predictive accuracy, with ablation indicating controllability via $\beta$; results generalize to multiple prototype-selection algorithms and motivate user studies and extensions to non-tabular data.

Abstract

Although prototype-based explanations provide a human-understandable way of representing model predictions they often fail to direct user attention to the most relevant features. We propose a novel approach to identify the most informative features within prototypes, termed alike parts. Using feature importance scores derived from an agnostic explanation method, it emphasizes the most relevant overlapping features between an instance and its nearest prototype. Furthermore, the feature importance score is incorporated into the objective function of the prototype selection algorithms to promote global prototypes diversity. Through experiments on six benchmark datasets, we demonstrate that the proposed approach improves user comprehension while maintaining or even increasing predictive accuracy.

Figures (4)

• Figure 1: Comparison of prototypes (x-axis -- prototype index) and important features (y-axis -- feature index) for the Diabetes dataset. The top row displays prototypes generated using the original raw algorithm, while the bottom row incorporates an extended target function with feature importance (FI). The size of the inner circle represents feature importance, and pink highlights features identified as important for a given prototype.
• Figure 2: The comparison of the frequency of feature highlighting between the original (raw) and Feature Importance (FI)-informed strategies across different benchmark datasets. The results are shown for three prototype selection algorithms: A-Pete, G-KM, and SM-A.
• Figure 3: The comparison of accuracy (hue) achieved by A-Pete, G-KM, and SM-A algorithms across benchmarks against algorithm-specific hyperparameters (x-axis) and $\beta$ (y-axis). Note that the bottom line of each subfigure ($\beta = 0$) represents the original definition of the algorithms, where only the tree distance is minimized.
• Figure 4: The comparison of mean feature importance of the features included in alike parts (left y-axis) and the length of the vector identified as alike parts between the explained instance and the prototype (right y-axis). The plot illustrates these two values tested against different $\beta$ values (x-axis).