🤖 machine learning

A New Technique for AI Explainability using Feature Association Map

This paper introduces FAMeX, a novel graph-theoretic explainable AI algorithm based on Feature Association Maps that demonstrates superior performance in gauging feature importance for classification compared to existing methods like PFI and SHAP.

Original authors: Sayantani Ghosh, Amit Kumar Das, Amlan Chakrabarti

Published 2026-05-13

📖 4 min read☕ Coffee break read

CC BY 4.0

Original authors: Sayantani Ghosh, Amit Kumar Das, Amlan Chakrabarti

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 super-smart robot that can predict things, like whether a patient has a disease or if a stock price will go up. This robot is incredibly accurate, but it's a "black box." You ask it, "Why did you make that decision?" and it just says, "Because I said so." That's scary, especially when the decision affects real lives. We need to know why the robot thinks what it thinks.

This paper introduces a new tool called FAMeX (Feature Association Map based eXplainability) to solve that problem. Here is how it works, using simple analogies:

The Problem: The "Crowded Room" of Data

Imagine the robot is trying to make a decision based on a room full of people (the data features). Some people are shouting important facts, while others are just repeating what their neighbors are saying.

Relevance: How much useful information a person brings to the conversation.
Redundancy: How much a person is just echoing someone else.

Old methods of explaining AI (like PFI and SHAP) are like asking everyone in the room to raise their hand one by one to see who is important. They might miss the fact that three people are just repeating the same thing, making it look like there are three important voices when there is really only one.

The Solution: The "FAMeX" Party Planner

The authors propose FAMeX, which acts like a smart party planner who draws a map of the room before the party starts.

Drawing the Map (The Feature Association Map):
FAMeX looks at every pair of people (features) and draws a line between them if they are talking about the same thing.
- If two people are whispering the exact same secret (high similarity), FAMeX draws a thick line between them.
- If someone is telling a unique story that no one else is (low similarity), they stand alone.
Grading the Guests:
The algorithm gives everyone a "grade" based on how much they are repeating others:
- Grade 1 (The Originals): People who aren't repeating anyone else. They get a high score for being unique.
- Grade 2 (The Echoes): People repeating one or two others. They get a medium score.
- Grade 3 (The Chorus): People repeating many others or repeating someone very loudly. They get a low score for being redundant.
The Final Score:
FAMeX calculates a final "Importance Score" for each person.
- If you bring unique, valuable information (High Relevance) AND you aren't just repeating others (Low Redundancy), you get a super high score.
- If you are just repeating others, your score goes down, even if what you are saying is true.

Did It Work?

The authors tested this new "Party Planner" against two other famous methods (PFI and SHAP) using eight different real-world datasets (like medical records, wine quality, and vehicle data).

They played a game:

They asked each method to pick the top 30% of the most important people (features).
They fed only those top people into a standard classifier (a simple decision-maker) to see how well it could predict the outcome.
They also tested the bottom 30% (the least important people) to see if the prediction failed.

The Results:

FAMeX won. When they used the top 30% of features picked by FAMeX, the prediction accuracy was consistently higher than when using the top picks from PFI or SHAP.
The Gap was clear. For FAMeX, the difference between using the "best" features and the "worst" features was huge (often over 10% better accuracy). This proves FAMeX actually knows who the important people are.
The Competitors stumbled. Sometimes, PFI and SHAP picked "important" features that actually made the prediction worse than picking the least important ones.

The Tool

The authors didn't just write code; they built a visual tool (a GUI) that anyone can use. You can upload your data, and it will show you a colorful map (like the one in Figure 3 of the paper) where the most important features light up, helping you understand exactly what drives the AI's decision.

In a Nutshell

FAMeX is a new way to explain AI that doesn't just ask "What is important?" but also asks "Who is just copying others?" By filtering out the noise (redundancy) and highlighting the unique, valuable signals (relevance), it gives a clearer, more accurate explanation of why an AI made a specific decision. The paper claims this method is more reliable than the current top competitors.

Technical Summary: FAMeX – A New Technique for AI Explainability using Feature Association Map

Problem Statement
The increasing adoption of Artificial Intelligence (AI) in critical real-life applications—ranging from disease detection and weather forecasting to stock price prediction—has highlighted a significant challenge: the opacity of "black-box" systems. While complex learning algorithms often yield high predictive accuracy, their decision-making processes frequently lack transparency. This lack of interpretability makes it difficult to determine if decisions are based on scientific facts or spurious coincidences, thereby hindering trust, accountability, and reliability in high-stakes environments. Existing Explainable AI (XAI) methods, such as Permutation Feature Importance (PFI) and SHapley Additive exPlanations (SHAP), attempt to address this by ranking feature importance. However, the authors argue that these methods often fail to conduct a holistic analysis of critical feature aspects, specifically relevance (information contribution to the target) and redundancy (similarity/overlap with other features).

Methodology: The FAMeX Algorithm
The paper proposes FAMeX (Feature Association Map based eXplainability), a novel algorithm designed to evaluate feature importance by simultaneously considering relevance and redundancy through a graph-theoretic formulation.

Feature Association Map (FAM): The core of the methodology is the FAM, a graph where features are modeled as vertices and associations between them are represented as edges.
Feature Similarity (Redundancy): Redundancy is measured using Pearson's correlation coefficient. Features are graded based on their correlation with other features:
- Grade 1: Low redundancy (correlation < 0.67 with all other features).
- Grade 2: Moderate redundancy (correlation > 0.67 but < 0.9 with at least one feature, but fewer than three).
- Grade 3: High redundancy (correlation > 0.9 with at least one feature OR > 0.67 with three or more features).
  A Similarity Score is calculated based on these grades, where higher grades (indicating higher redundancy) result in a higher score.
Feature Relevance: Relevance is measured using Mutual Information (MI) between a feature and the target class variable, normalized to calculate a Relevance Score.
Feature Importance Score: The final importance score is derived by integrating the two metrics. The algorithm posits that a feature is more important if it has high relevance and low redundancy. Consequently, the score is calculated as:
$\text{Feature Importance Score} = \frac{\text{Relevance Score}}{\text{Similarity Score}}$
This inverse relationship ensures that highly redundant features (high similarity score) are penalized, while highly relevant features are rewarded.

The algorithm operates with a time complexity of $O(n^2)$ , where $n$ is the number of features.

Key Contributions

Holistic Feature Analysis: Unlike competing methods that may focus primarily on marginal contribution, FAMeX explicitly models the trade-off between feature relevance and feature redundancy.
Graph-Theoretic Formulation: The introduction of the Feature Association Map provides a visual and mathematical framework to understand feature interdependencies.
Algorithmic Implementation: The authors provide a complete algorithm (Algorithm 1) and a platform-independent Graphical User Interface (GUI) tool to facilitate the application of FAMeX on various datasets.

Experimental Results
The authors evaluated FAMeX against PFI and SHAP using eight benchmark datasets from the UCI Machine Learning repository (e.g., Wisconsin, ILPD, WineQuality, WBDC). The evaluation involved:

Setup: 10-fold cross-validation repeated for 100 iterations.
Metrics: Classification accuracy was measured using four standard classifiers: Support Vector Machine (SVM), Random Forest (RF), Naive Bayes (NB), and Decision Tree (DT).
Validation Strategy: The top 30% and bottom 30% of features identified by each algorithm were used to train the classifiers. A superior XAI algorithm should yield significantly higher accuracy when using its top-ranked features compared to its bottom-ranked features.

Findings:

Superior Performance: Across all four classifiers and eight datasets, FAMeX consistently generated higher classification accuracy using the top 30% of features compared to PFI and SHAP. For instance, with the Random Forest classifier, FAMeX achieved an average accuracy of 80.16%, compared to 71.47% for PFI and 71.82% for SHAP.
Consistency: FAMeX demonstrated a clear distinction between important and unimportant features. In most cases, the accuracy drop when using the bottom 30% of features was significant (often >10% difference).
Competitor Limitations: In contrast, PFI and SHAP showed inconsistent results. In several instances (e.g., WBDC dataset with SVM), the "bottom 30%" features selected by PFI or SHAP yielded higher accuracy than the "top 30%" features, indicating a failure to correctly identify critical features in those contexts.

Significance and Claims
The paper claims that FAMeX is a promising algorithm for explaining AI predictions because it effectively gauges feature importance in the context of classification by accounting for both relevance and redundancy. The authors state that the experimental results establish the efficacy of FAMeX ahead of PFI and SHAP. The work is presented as a step toward fostering trust in AI systems by providing a transparent, holistic explanation of decision-making processes. The authors conclude modestly, suggesting that future work should involve applying FAMeX to more complex, large-scale datasets and comparing it with additional state-of-the-art XAI algorithms.