💻 computer science

Alike Parts: A Feature-Informed Approach to Local and Global Prototype Explanations

This paper introduces "Alike Parts," a framework that enhances prototype-based explanations by using feature importance scores to highlight shared relevant features for local interpretability and to promote feature diversity in global prototype selection without compromising prediction fidelity.

Original authors: Jacek Karolczak, Jerzy Stefanowski

Published 2026-05-22

📖 4 min read☕ Coffee break read

CC BY 4.0

Original authors: Jacek Karolczak, Jerzy Stefanowski

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine you have a very smart, but mysterious, "Black Box" computer that makes decisions—like approving a loan, diagnosing a disease, or sorting fruit. You trust it because it's accurate, but you can't see how it thinks. To fix this, scientists often use prototypes.

Think of a prototype as a "representative example." If the computer says, "This apple is ripe," it might point to a specific photo of a perfect red apple from its training data and say, "I'm saying this because it looks just like this one."

The Problem:
Sometimes, pointing to the whole example isn't helpful. If the prototype is a photo of a fruit basket with 20 different items, or a medical record with 50 different numbers, it's overwhelming. You don't know which specific part of that example actually mattered. Was it the color? The weight? The age? The computer is hiding the most important clues inside a sea of data.

The Solution: "Alike Parts"
The authors of this paper propose a new way to explain these decisions called "Alike Parts."

Instead of showing you the whole prototype, their method acts like a highlighter pen. It looks at the specific item you asked about (the "Instance") and the representative example it was compared to (the "Prototype"). Then, it uses a special scoring system to find the shared features that both the item and the prototype have in common and that the computer thinks are the most important for the decision.

The Analogy: Imagine you and a friend are both wearing red hats. If the computer decides you are "Team Red," it shouldn't just show you your friend's entire outfit (jeans, shoes, jacket). It should highlight only the red hats you both share, because those are the "Alike Parts" that actually caused the decision.

How They Did It (The Two-Step Magic):

Local Explanation (The Highlighter):
When explaining a single decision, the system calculates which features are "important" for both your item and the prototype. It then creates a mask that hides everything else, showing you only the critical, shared features. This stops you from getting distracted by irrelevant details.
Global Explanation (The Diverse Team):
Usually, when computers pick a set of prototypes to represent the whole system, they just look for the ones that are mathematically closest to the data. The authors changed the rules. They told the computer: "Don't just pick the closest examples; pick examples that cover a diverse range of important features."
- The Analogy: If you are building a team of experts to explain a complex topic, you don't want five experts who all know the exact same thing. You want one expert on the budget, one on the timeline, and one on the design. By forcing the computer to pick a "diverse team" of prototypes, they ensure that when you look at the whole set, you see a wider variety of reasons why decisions are made.

What They Found:
The researchers tested this on six different datasets (like predicting wine quality, diabetes risk, or passenger satisfaction).

It didn't break the computer: Adding these new rules to pick prototypes didn't make the computer less accurate. In fact, in many cases, the simplified "prototype" version of the computer performed just as well as the original complex one.
It found better clues: In a test with diabetes data, the old method only highlighted "Glucose" levels. The new "Alike Parts" method also highlighted "Age" and "Family History," which are actually more medically relevant.
It showed more variety: The new method ensured that the set of prototypes used to explain the system covered different types of important features, rather than just repeating the same few clues over and over.

In a Nutshell:
This paper introduces a way to stop showing users the "whole messy picture" and start showing them the "specific, shared clues" that actually drove the decision. It makes AI explanations clearer by highlighting the "Alike Parts" and ensures the AI's library of examples is diverse enough to explain different kinds of decisions without losing its accuracy.

Technical Summary: ALIKE PARTS: A FEATURE-INFORMED APPROACH TO LOCAL AND GLOBAL PROTOTYPE EXPLANATIONS

Problem Statement
Prototype-based explanations offer an intuitive, example-based method for interpreting black-box machine learning classifiers by linking predictions to representative training instances. However, a significant limitation exists: while prototypes provide a holistic view, they often lack feature-level granularity. Users struggle to identify which specific features within a prototype are most relevant to a prediction, leading to cognitive overload when dealing with high-dimensional tabular data. Furthermore, global prototype selection often prioritizes spatial distribution or distance minimization without ensuring diversity in the feature attributions of the selected prototypes, potentially missing critical, diverse decision boundaries.

Methodology
The authors propose a framework that integrates feature importance scores at two distinct levels to address these gaps: local explanation refinement and global prototype selection augmentation.

Local Explanations: "Alike Parts"
The paper introduces a method to identify "alike parts," defined as the most relevant, shared feature subsets between a classified instance ( $x_i$ ) and its nearest prototype ( $p_j$ ).
- Process: Feature importance scores ( $\phi$ ) are computed for both the instance and the prototype using model-agnostic or model-specific estimators (e.g., Tree SHAP, Tree Interpreter).
- Alignment: The method calculates a weight vector ( $w$ ) representing the similarity of importance between the instance and prototype. This is achieved via a similarity operator ( $\circ$ ), such as the Hadamard product (element-wise multiplication), element-wise $l_1$ , or $l_2$ distances.
- Masking: A binary mask is generated to highlight features where the importance alignment exceeds a threshold. Strategies include mean thresholding or adaptive Top- $k$ selection (e.g., $k = \lceil\sqrt{d}\rceil$ or $k = \lceil\log(d)\rceil$ ).
- Output: This directs user attention to a limited, high-relevance subset of features shared by the instance and the prototype, rather than presenting the entire feature vector.
Global Explanations: Augmented Objective Function
The authors modify the standard prototype selection objective function to actively promote diversity in feature attributions.
- Standard Objective: Traditional $k$ -medoids approaches minimize the distance $d(x_i, p_j)$ between training instances and their nearest prototypes.
- Augmented Objective: The authors introduce a feature importance term ( $f_i$ ) derived from the alignment scores used in the local method. The new objective function is:
  $f(P) = \sum_{i=1}^{|S|} \min_{p_j \in P} (d(x_i, p_j) + \beta \cdot f_i(x_i, p_j))$
  Here, $\beta$ acts as a trade-off parameter between metric distance and feature importance alignment. This encourages the selection of prototypes that not only cover the data space but also represent diverse subsets of important features.

Key Contributions

Alike Parts Identification: A novel method to generate binary masks highlighting shared, high-importance features between an instance and its nearest prototype, enhancing local interpretability.
Feature-Diverse Prototype Selection: An extension of prototype generation algorithms (specifically G-KM, SM-A, and A-PETE) that incorporates feature importance into the optimization loop to ensure global diversity in feature attributions.
Comprehensive Evaluation: The study evaluates multiple prototype generation algorithms alongside new evaluation functions, explores a wider set of feature importance algorithms (Tree SHAP, Tree Interpreter), and conducts extensive sensitivity analysis on hyperparameters (e.g., $\beta$ , similarity operators, masking strategies).
Model Agnosticism: The framework is designed to be compatible with various underlying components, though the experiments focus on Random Forest ensembles and tree-based distance metrics.

Experimental Results
Experiments were conducted on six benchmark datasets (Apple Quality, Australia Rain, Breast Cancer, Diabetes, Passenger Satisfaction, Wine Quality) using Random Forest as the black-box model.

Fidelity: The augmented selection process maintained or, in several cases, increased the prediction fidelity of the surrogate 1-Nearest Neighbor (1-NN) model compared to the original (raw) prototype selection. Statistical analysis (Wilcoxon signed-rank test) indicated a positive trend toward improved fidelity ( $p \approx 0.068$ ) and significant improvement over standard interpretable baselines like Naive Bayes, Logistic Regression, and Decision Trees ( $p < 0.05$ ).
Feature Diversity: The FI-informed approach successfully identified prototypes with greater feature diversification. For instance, in the Diabetes dataset, the FI-informed method highlighted clinically relevant features (Age, Diabetes Pedigree Function) that the raw distance-only method missed.
Hyperparameter Sensitivity: The Hadamard product was found to be the most effective similarity metric across datasets. The study noted that while some trends were global, tailored tuning of hyperparameters (like ignoring direction or normalization) remained necessary for specific datasets.
Activation Frequency: Analysis of feature activation frequencies confirmed that the FI-informed objective reduces the variance in feature usage, forcing the selection of a more diverse set of prototypes.

Significance and Claims
The paper claims that integrating feature importance into prototype selection does not compromise model fidelity; rather, seeking diverse feature attributions can be complementary to distance-based coverage. The primary significance lies in bridging the gap between high-level prototype examples and low-level feature reasoning. By identifying "alike parts," the method guides user attention to the specific rationale shared between an instance and its prototype, addressing the interpretability challenges of high-dimensional tabular data. The authors emphasize that this approach balances interpretability and diversity, enhancing both local explanations and the global prototype selection process without requiring modifications to the underlying black-box model's internal logic.

The authors modestly note that while the method improves the quality of the explanation (via proxy metrics like fidelity and feature diversity), the central hypothesis—that "alike parts" improve actual user comprehension—requires future validation through formal human-computer interaction studies. Additionally, while the current implementation focuses on tree ensembles, the modular design allows for adaptation to other architectures (e.g., neural networks) by substituting the distance and importance estimation components.

More like this