🤖 AI

Explainable AI Through a Democratic Lens: DhondtXAI for D'Hondt-Projected Feature Attribution

This paper introduces DhondtXAI, a novel, SHAP-independent explainable AI framework for tabular data that adapts the D'Hondt apportionment method to allocate feature attribution through background-interventional effects, feature alliances, and thresholds, demonstrating high accuracy in synthetic tests and strong agreement with SHAP on healthcare datasets while preserving completeness by construction.

Original authors: Turker Berk Donmez

Published 2026-06-02

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Turker Berk Donmez

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, computer program (an AI) that helps doctors predict if a patient has breast cancer or diabetes. The problem is, this computer is a "black box." It gives an answer, but it doesn't easily explain why it made that choice. It looks at dozens of clues (like cell size, age, or symptoms), but we don't know which clues matter the most.

This paper introduces a new way to understand these clues called DhondtXAI. Instead of using complex math jargon, the author, Türker Berk Dönmez, decides to treat the AI's decision-making process like a democratic election.

Here is how it works, broken down into simple concepts:

1. The Features are Political Parties

In a normal election, different political parties run for office. In this AI system, every "clue" (or feature) the computer looks at is treated like a political party.

The Votes: The computer calculates how important each clue is. This importance score is turned into "votes." A clue that is very important gets millions of votes; a less important one gets fewer.
The Seats: The goal is to see how many "seats" in a parliament each clue wins. The total number of seats is fixed (like 600 seats in a parliament).

2. The D'Hondt Method: A Fair Way to Count

The paper uses a specific rule called the D'Hondt method to count the votes. This is a real-world rule used in countries like Turkey and Spain to make sure smaller parties get a fair chance at representation, not just the biggest ones.

How it works: Imagine the computer has 600 empty seats to fill. It looks at who has the most votes and gives them a seat. Then, it recalculates the odds for everyone else. It keeps handing out seats one by one until all 600 are filled.
The Result: The clues that get the most "seats" are the ones the AI relied on most heavily to make its decision. If "Cell Shape" gets 124 seats and "Age" only gets 26, you know the computer cared much more about the shape of the cell than the age of the patient.

3. Forming Alliances (Teamwork)

In real politics, small parties sometimes join forces to form a "coalition" or "alliance" so they can win more seats together. DhondtXAI lets you do the same thing with AI clues.

The Example: In the diabetes study, the author grouped related symptoms together. Instead of looking at "thirst," "frequent urination," and "weight loss" as three separate parties, he put them in one big "Diabetes Symptoms" alliance.
The Benefit: This alliance combined their votes. Suddenly, this "super-party" won a massive number of seats (361 out of 600), showing that symptoms as a group are the most powerful predictor, even if individual symptoms might look small on their own.

4. The Threshold Filter (The Minimum Bar)

Sometimes, there are too many small parties (or unimportant clues) cluttering up the parliament. The paper introduces a threshold, which is like a minimum bar a party must jump over to get into the election.

How it works: The user can say, "Only parties with at least 10% of the votes get a seat."
The Result: Any clue that didn't make the cut is kicked out. Its votes are then redistributed to the stronger clues that did make the cut. This cleans up the noise and leaves you with a clear list of only the most important factors.

What Did They Find?

The author tested this new "Election System" against a standard, well-known method called SHAP (which uses complex math to explain AI).

Breast Cancer Test: When they ran the test on a breast cancer model, the "Election" and the "Math" agreed almost perfectly. The clues that won the most seats in the parliament were the exact same clues that the math said were most important.
Diabetes Test: They did the same for diabetes. Again, the results matched up. The "Alliance" of diabetes symptoms won the election, confirming that these symptoms are the key drivers of the prediction.

Why Does This Matter?

The paper argues that this method is great because it turns boring, confusing data into a story we all understand: Politics.

Instead of looking at a spreadsheet of numbers, a doctor or a patient can look at a "Parliament" and see which "parties" (clues) are holding the most power.
It makes the AI feel more transparent and fair, like a democracy where every voice (clue) gets a chance to be heard, but the loudest voices (most important clues) get the most seats.

In short, DhondtXAI takes the scary, invisible logic of an AI and turns it into a visible, democratic election where you can clearly see who is in charge of the decision.

Technical Summary: Explainable AI Through a Democratic Lens: DhondtXAI for Proportional Feature Importance Using the D'Hondt Method

Problem Statement
The paper addresses the challenge of interpretability in complex machine learning models, particularly "black box" systems like neural networks and ensemble tree-based models (e.g., CatBoost, XGBoost). While existing Explainable AI (XAI) techniques such as SHAP (Shapley Additive Explanations) and LIME provide post-hoc explanations, the authors argue that there is a need for alternative frameworks that offer intuitive, structured, and transparent representations of feature importance. Specifically, the paper seeks to bridge the gap between technical model mechanics and human understanding by adapting principles from democratic electoral systems to the domain of feature attribution. The core problem is how to represent feature importance in a way that allows for proportional representation, alliance formation (grouping related features), and thresholding (filtering low-impact variables), thereby enhancing user trust and clarity in high-stakes fields like healthcare.

Methodology
The study introduces DhondtXAI, a library and methodology that applies the D'Hondt method—a proportional representation system used in political elections—to allocate "seats" (representative units) to features or feature groups based on their calculated importance. The methodology proceeds through the following structured steps:

Feature Importance Calculation: The process begins with standard feature importance extraction from tree-based models (Random Forests, Gradient Boosted Trees). This is calculated by aggregating the reduction in impurity (Information Gain) across all nodes and trees where a feature is used as a split criterion.
User-Defined Parameters: The framework allows users to specify:
- Total Votes ( $V$ ): A scalar representing the total resource to be distributed.
- Total Seats ( $M$ ): The number of representative units to be allocated.
- Exclusions: Features to be removed from the analysis.
- Alliances: User-defined groupings of features treated as single voting entities.
- Threshold: A minimum percentage of votes required to be eligible for seat allocation.
Vote Distribution: Initial votes are allocated to features or alliances proportionally based on their importance scores relative to the total importance of the set.
Threshold Application and Redistribution: If a user-defined threshold is set, features or alliances falling below this vote count are excluded. Their votes are then redistributed proportionally among the remaining "above-threshold" entities based on their relative importance.
Seat Allocation (D'Hondt Method): The final seats are distributed iteratively. For each feature/alliance $i$ , a seat ratio $S_i$ is calculated as $S_i = \frac{\text{redistributed\_vote}_i}{k + 1}$ , where $k$ is the current number of seats held. The entity with the highest $S_i$ receives the next seat, $k$ is incremented, and the process repeats until all $M$ seats are allocated.

Key Contributions

Novel Framework: The paper presents DhondtXAI, a novel approach that translates feature importance into a "parliamentary" view, where features compete for seats based on proportional representation.
Alliance Mechanism: It introduces the concept of "feature alliances," allowing users to group related variables (e.g., symptoms of a disease) to evaluate their collective impact, analogous to political parties forming coalitions.
Thresholding: It incorporates an electoral-style threshold to filter out noise, ensuring that only features with significant influence are highlighted in the final interpretation.
Integration with Standard XAI: The method is designed to complement existing techniques like SHAP, offering a resource-allocation perspective alongside traditional mathematical attribution.

Results
The study validates DhondtXAI using two datasets:

Wisconsin Breast Cancer Dataset (CatBoost Model):
- The model achieved 96% accuracy and an AUC of 1.00.
- DhondtXAI allocated 600 seats to 10 features. "Mean concave points" received the most seats (124), followed by "mean texture" (121).
- A Spearman correlation analysis between DhondtXAI seat counts and global SHAP values yielded a coefficient of 0.83 ( $p=0.0029$ ), indicating a strong, statistically significant positive correlation.
- The parliamentary visualization successfully mirrored SHAP's identification of top features and their directional correlations (e.g., red for negative correlation with benign outcomes).
Early Stage Diabetes Risk Prediction Dataset (XGBoost Model):
- The model achieved 98% accuracy and an AUC of 1.00.
- Features were grouped into alliances (e.g., Diabetes_Symptoms, Metabolic_Body).
- The Diabetes_Symptoms alliance dominated the allocation with 361 out of 600 seats.
- A Spearman correlation between alliance seat counts and mean SHAP values was 0.40 ( $p=0.505$ ), which was not statistically significant. The authors note this discrepancy but argue that the alliance-based approach offers a different, more intuitive perspective for grouped features.
- Threshold Application: Applying a 10% vote threshold filtered out the Vision_Nervous and Psychological alliances (reducing their seats to 0), while increasing the seat counts for the top three alliances (Metabolic_Body, Diabetes_Symptoms, Skin_Infection), demonstrating the method's ability to refine focus on dominant drivers.

Significance and Claims
The paper claims that DhondtXAI offers a complementary and intuitive tool for Explainable AI. Its significance lies in:

Democratizing Interpretability: By using a familiar political metaphor (parliamentary seats), it makes complex model decisions accessible to non-technical stakeholders, fostering a sense of fair representation for different features.
Enhanced Flexibility: The ability to form alliances and apply thresholds allows for a more nuanced analysis of feature interactions and noise reduction, which standard point-based importance metrics may not explicitly support.
Ethical AI Alignment: The proportional representation framework is presented as a step toward ethical AI, ensuring that feature influence is distributed transparently and that minor but collectively significant factors (via alliances) are not overlooked.
Validation: The strong correlation with SHAP in the breast cancer dataset suggests that DhondtXAI provides consistent rankings of feature importance, validating its reliability as a secondary interpretability tool.

The authors conclude that while DhondtXAI may not replace granular methods like SHAP for individual prediction analysis, it provides a valuable, structured, and socially resonant framework for understanding the collective influence of features in high-stakes decision-making environments.