arxiv: 2605.12350 · v3 · submitted 2026-05-12 · 💻 cs.LG · cs.AI

Recognition: no theorem link

A New Technique for AI Explainability using Feature Association Map

Sayantani Ghosh , Amit Kumar Das , Amlan Chakrabarti

Authors on Pith no claims yet

Pith reviewed 2026-05-15 05:33 UTC · model grok-4.3

classification 💻 cs.LG cs.AI

keywords explainable AIfeature importancegraph theoryXAIclassificationfeature association mapFAMeXSHAP

0 comments

The pith

A graph linking features by their associations ranks input importance for AI classifications more accurately than SHAP or permutation importance.

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The paper proposes FAMeX, which first builds a Feature Association Map as a graph with features as nodes and statistical associations as edges. It then derives feature importance scores from this graph structure to explain a classifier's output. Tests across eight benchmark datasets show these scores align better with actual model behavior than scores from Permutation Feature Importance or SHAP. The approach matters for any setting where users need to trust or audit automated decisions, because it explicitly accounts for how features relate to one another rather than treating them in isolation. If the method generalizes, it offers a practical alternative for generating explanations without requiring model internals.

Core claim

The authors claim that modeling the feature set as a Feature Association Map graph, where edges represent measured associations, enables the FAMeX algorithm to produce feature importance rankings that are superior to those of PFI and SHAP when evaluated on eight standard classification benchmarks.

What carries the argument

The Feature Association Map (FAM), a graph whose nodes are input features and whose edges encode pairwise association strengths, from which importance values are extracted to explain classification predictions.

If this is right

Explanations for classification models can be generated by analyzing a graph of feature associations instead of relying solely on additive or permutation-based scores.
Feature importance derived from association graphs may better capture interactions that affect decision boundaries.
XAI pipelines could replace or augment SHAP and PFI with FAMeX for tasks where feature dependencies are strong.
Model auditing in regulated domains gains a graph-based tool that highlights which inputs matter in context.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

The same graph construction might be adapted to regression or clustering tasks by redefining the target association measure.
Different choices of association metric or graph pruning threshold could change the rankings, so users may need to test sensitivity on each new dataset.
Combining the association graph with causal discovery algorithms could move the explanations closer to identifying true causal drivers.

Load-bearing premise

The statistical associations used to build the graph must reliably reflect the features' actual influence on the model's decisions rather than incidental correlations or dataset artifacts.

What would settle it

Construct a synthetic classification dataset with known ground-truth important features and known interaction structure, then compare whether FAMeX recovers the ground-truth importance ranking more accurately than SHAP and PFI.

read the original abstract

Lack of transparency in AI systems poses challenges in critical real-life applications. It is important to be able to explain the decisions of an AI system to ensure trust on the system. Explainable AI (XAI) algorithms play a vital role in achieving this objective. In this paper, we are proposing a new algorithm for Explaining AI systems, FAMeX (Feature Association Map based eXplainability). The proposed algorithm is based on a graph-theoretic formulation of the feature set termed as Feature Association Map (FAM). The foundation of the modelling is based on association between features. The proposed FAMeX algorithm has been found to be better than the competing XAI algorithms - Permutation Feature Importance (PFI) and SHapley Additive exPlanations (SHAP). Experiments conducted with eight benchmark algorithms show that FAMeX is able to gauge feature importance in the context of classification better than the competing algorithms. This definitely shows that FAMeX is a promising algorithm in explaining the predictions from an AI system

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit. Tearing a paper down is the easy half of reading it; the pith above is the substance, this is the friction.

Desk Editor's Note private letter to a colleague

FAMeX offers a graph-based XAI approach for feature importance but its superiority claims over SHAP lack any supporting metrics or details.

read the letter

The one thing to know is that this paper introduces FAMeX, a method that builds a graph of feature associations to explain importance in classification models, and it claims this works better than SHAP or permutation importance across eight benchmarks. The second thing is that the abstract gives no numbers or construction details, so those claims are not yet convincing. The paper does something new by framing feature associations as a graph for explainability purposes. This is a reasonable extension of graph ideas from feature selection into the XAI space, and it could appeal to users who think in terms of networks rather than additive scores. The authors do a good job stating the problem of AI transparency and positioning their approach as practical. The main soft spot is the lack of evidence. The central claim of better performance is stated without any metrics, comparisons, or even basic description of the association function or graph building process. There is no ablation on how sensitive the results are to those choices, which means we cannot rule out that the reported wins are due to particular settings or datasets. The stress-test note hits this exactly. If the full manuscript has the experiments with proper stats, that would change things, but right now the soundness is low. This paper is for applied machine learning engineers and data scientists who want another option in their XAI toolkit for classification tasks. Someone who values simple, graph-based visualizations might get value from trying it out. A reader focused on theoretical XAI or reproducible benchmarks will not get much from it as is. It deserves a serious referee because the idea is clear and could be made useful with better validation, even though it will likely need substantial revisions to the experimental section. I recommend sending it to peer review so the authors can provide the missing details and let reviewers assess whether the method actually delivers on the claims.

Referee Report

3 major / 1 minor

Summary. The manuscript proposes FAMeX, a new XAI algorithm that constructs a graph-theoretic Feature Association Map (FAM) from pairwise feature associations and derives feature importance rankings from this structure. It claims that FAMeX outperforms Permutation Feature Importance (PFI) and SHAP when evaluated on eight benchmark datasets for classification tasks.

Significance. If the empirical claims are substantiated with quantitative metrics and ablations, the graph-based formulation could offer a distinct perspective on feature interactions that additive or permutation-based methods do not explicitly capture, potentially aiding interpretability in domains where feature dependencies matter.

major comments (3)

[Abstract] Abstract: the assertion that FAMeX is 'better than' PFI and SHAP on eight benchmarks is presented without any numerical results, tables, fidelity scores, consistency metrics, error bars, or statistical tests, so the central empirical claim cannot be evaluated from the provided text.
[Abstract] Abstract: the construction of the Feature Association Map is described only at a high level ('association between features'); no specific association measure, edge-weighting scheme, thresholding rule, or procedure for extracting importance scores from the graph is given, which are load-bearing details for reproducibility and for distinguishing the method from existing graph-based XAI approaches.
[Abstract] Abstract: no ablation or sensitivity analysis is mentioned regarding the choice of association function or graph-construction hyperparameters, leaving open the possibility that reported gains are artifacts of a particular measure or dataset rather than a general advantage of the FAM formulation.

minor comments (1)

[Abstract] Abstract: the phrasing 'eight benchmark algorithms' is ambiguous (datasets or models?) and should be clarified; likewise, 'gauge feature importance ... better' would benefit from a precise definition of the comparison criterion.

Simulated Author's Rebuttal

3 responses · 0 unresolved

We thank the referee for the constructive feedback. We address each major comment below and will revise the manuscript to improve the abstract's clarity, specificity, and support for the empirical claims.

read point-by-point responses

Referee: [Abstract] Abstract: the assertion that FAMeX is 'better than' PFI and SHAP on eight benchmarks is presented without any numerical results, tables, fidelity scores, consistency metrics, error bars, or statistical tests, so the central empirical claim cannot be evaluated from the provided text.

Authors: We agree that the abstract lacks specific numerical results and statistical details, making the central claim difficult to evaluate at a glance. The full manuscript includes tables and metrics (fidelity, consistency) comparing FAMeX to PFI and SHAP across the eight datasets. We will revise the abstract to include key quantitative highlights, such as average improvements and references to the experimental tables, while keeping it concise. revision: yes
Referee: [Abstract] Abstract: the construction of the Feature Association Map is described only at a high level ('association between features'); no specific association measure, edge-weighting scheme, thresholding rule, or procedure for extracting importance scores from the graph is given, which are load-bearing details for reproducibility and for distinguishing the method from existing graph-based XAI approaches.

Authors: The comment is accurate; the abstract is high-level by design. The main text specifies the association measure (e.g., pairwise correlation), edge-weighting, thresholding, and importance extraction via graph centrality. We will update the abstract with brief but concrete descriptions of these components to aid reproducibility and differentiation from prior graph-based XAI methods. revision: yes
Referee: [Abstract] Abstract: no ablation or sensitivity analysis is mentioned regarding the choice of association function or graph-construction hyperparameters, leaving open the possibility that reported gains are artifacts of a particular measure or dataset rather than a general advantage of the FAM formulation.

Authors: We recognize that the abstract does not reference ablations or sensitivity analysis. The manuscript evaluates FAMeX across multiple datasets and association measures to support generalizability, but explicit ablation studies on hyperparameters are limited. We will revise the abstract to note the robustness checks performed and expand the main text or supplementary material with additional sensitivity results where feasible. revision: partial

Circularity Check

0 steps flagged

No circularity: empirical claim with no derivations or self-referential reductions

full rationale

The paper introduces FAMeX as a graph-theoretic method based on feature associations and supports its superiority claim solely through experimental comparisons against PFI and SHAP on eight benchmark algorithms. No equations, derivations, fitted parameters renamed as predictions, or self-citations appear in the provided abstract or described content. The central claim reduces to reported experimental outperformance rather than any self-definitional loop, ansatz smuggled via citation, or uniqueness theorem imported from prior author work. This is a standard empirical XAI proposal with no detectable circular reduction in its derivation chain.

Axiom & Free-Parameter Ledger

0 free parameters · 1 axioms · 1 invented entities

The central claim rests on the unstated assumption that feature associations can be meaningfully encoded as a graph whose properties yield reliable importance rankings; no free parameters or invented entities beyond the FAM itself are described in the abstract.

axioms (1)

domain assumption Feature associations can be modeled as edges in a graph that capture decision-relevant structure
Foundation of the FAM construction stated in the abstract

invented entities (1)

Feature Association Map (FAM) no independent evidence
purpose: Graph representation of feature relationships for explainability
New construct introduced to support the FAMeX algorithm

pith-pipeline@v0.9.0 · 5476 in / 1227 out tokens · 36560 ms · 2026-05-15T05:33:07.098848+00:00 · methodology

Review history (3 revisions) →

discussion (0)

Reference graph

Works this paper leans on

26 extracted references · 26 canonical work pages · 3 internal anchors

[1]

Information fusion58, 82–115 (2020), https://doi.org/10.1016/j.inffus.2019.12.012

Arrieta, A., D´ ıaz-Rodr´ ıguez, N., Ser, J., Bennetot, A., Tabik, S., Barbado, A., Garc´ ıa, S., Gil-L´ opez, S., Molina, D., Benjamins, R., Chatila, R., and Her- rera, F. (2020).Explainable Artificial Intelligence (XAI): Concepts, Taxonomies, Opportunities and Challenges toward Responsible AI. ArXiv, 1910.10045

work page arXiv 2020
[3]

What Does Explainable AI Really Mean? A New Conceptualization of Perspectives

Doran, D., Schulz, S., and Besold, T.R. (2017).What Does Explainable AI Really Mean? A New Conceptualization of Perspectives. ArXiv,abs/1710.00794

work page internal anchor Pith review Pith/arXiv arXiv 2017
[4]

(2020).Hide-and-Seek: A Template for Explain- able AI

Tagaris, T., and Stafylopatis, A. (2020).Hide-and-Seek: A Template for Explain- able AI. ArXiv,abs/2005.00130

work page arXiv 2020
[5]

(2017).Towards A Rigorous Science of Interpretable Machine Learning

Doshi-Velez, F., and Kim, B. (2017).Towards A Rigorous Science of Interpretable Machine Learning. arXiv: Machine Learning. 14

work page 2017
[6]

Dosilovic, F.K., Brˇ ciˇ c, M., and Hlupic, N. (2018).Explainable Artificial Intelli- gence: A Survey.2018 41st International Convention on Information and Com- munication Technology, Electronics and Microelectronics (MIPRO), 0210-0215

work page 2018
[7]

arXiv preprint arXiv:2006.11371 (2020)

Das, A., and Rad, P. (2020).Opportunities and Challenges in Explainable Artificial Intelligence (XAI): A Survey. ArXiv,abs/2006.11371

work page arXiv 2020
[8]

(2016).”Why Should I Trust You?”: Explaining the Predictions of Any Classifier

Ribeiro, M.T., Singh, S., and Guestrin, C. (2016).”Why Should I Trust You?”: Explaining the Predictions of Any Classifier. Proceedings of the 22nd ACM SIGKDD International Conference on Knowledge Discovery and Data Mining

work page 2016
[9]

Learning to Explain: An Information-Theoretic Perspective on Model Interpretation

Chen, J., Song, L., Wainwright, M., and Jordan, M.I. (2018).Learning to Explain: An Information-Theoretic Perspective on Model Interpretation. ArXiv, abs/1802.07814

work page internal anchor Pith review Pith/arXiv arXiv 2018
[10]

Peeking Inside the Black Box: Visualizing Statistical Learning with Plots of Individual Conditional Expectation

Goldstein, A., Kapelner, A., Bleich, J., and Pitkin, E. (2014).“Peeking Inside the Black Box: Visualizing Statistical Learning with Plots of Individual Conditional Expectation.” arXiv,1309.6392

work page internal anchor Pith review Pith/arXiv arXiv 2014
[11]

(2016).A Permutation Importance-Based Fea- ture Selection Method for Short-Term Electricity Load Forecasting Using Random Forest

Huang, N., Lu, G., and Xu, D. (2016).A Permutation Importance-Based Fea- ture Selection Method for Short-Term Electricity Load Forecasting Using Random Forest. Energies,9, 767

work page 2016
[12]

(2020).Shapley-Lorenz Decompositions in eXplain- able Artificial Intelligence

Giudici, P., and Raffinetti, E. (2020).Shapley-Lorenz Decompositions in eXplain- able Artificial Intelligence. Econometrics: Econometric Model Construction

work page 2020
[13]

(2020).The Explanation Game: Explaining Machine Learning Models Using Shapley Values

Merrick, L., and Taly, A. (2020).The Explanation Game: Explaining Machine Learning Models Using Shapley Values. CD-MAKE

work page 2020
[14]

(2020).Explaining Predictive Models with Mixed Features Using Shapley Values and Conditional Inference Trees

Redelmeier, A., Jullum, M., and Aas, K. (2020).Explaining Predictive Models with Mixed Features Using Shapley Values and Conditional Inference Trees. ArXiv, abs/2007.01027

work page arXiv 2020
[15]

(2020).Generalized SHAP: Generating Multiple Types of Explanations in Machine Learning

Bowen, D., and Ungar, L. (2020).Generalized SHAP: Generating Multiple Types of Explanations in Machine Learning. ArXiv,abs/2006.07155

work page arXiv 2020
[16]

(2019).Toward Safer Highways: Application of XGBoost and SHAP for Real- Time Accident Detection and Feature Analysis

Parsa, A., Movahedi, A., Taghipour, H., Derrible, S., and Mohammadian, A. (2019).Toward Safer Highways: Application of XGBoost and SHAP for Real- Time Accident Detection and Feature Analysis. Accident Analysis and Prevention, 136, 105405

work page 2019
[17]

TheApplication of Support Vector Machines to Medical Decision Support: A Case Study

Veropoulos, K., Cristianini, N., and Campbell, C. TheApplication of Support Vector Machines to Medical Decision Support: A Case Study. ACAI99

work page
[18]

Alty, S.R., Millasseaut, S.C., Chowienczyk, P.J., and Jakobsson, A. (2004). Cardiovascular Disease Prediction Using Support Vector Machines. IEEE, 0-7803- 8294-3/04/$20.00. 15

work page 2004
[19]

Pavlopoulos, S., Stasis, A., and Loukis, E. (2004). A Decision Tree–Based Method for the Differential Diagnosis of Aortic Stenosis from Mitral Regurgitation Using Heart Sounds.BioMedical Engineering OnLine,3, 21-21

work page 2004
[20]

Nahar, N., and Ara, F. (2018). Liver Disease Prediction by Using Different Decision Tree Techniques.International Journal of Data Mining and Knowledge Management Process,8, 01-09

work page 2018
[21]

(2017).Diagnosis of Chronic Kidney Disease by Using Random Forest

Subasi, A., Alickovic, E., and Kevric, J. (2017).Diagnosis of Chronic Kidney Disease by Using Random Forest

work page 2017
[22]

Parthiban, G., Srivatsa, S., and Rajesh, A. (2011). Diagnosis of Heart Disease for Diabetic Patients Using Naive Bayes Method.International Journal of Computer Applications,24, 7-11

work page 2011
[24]

(2020).Forensic Malware Identification Using Naive Bayes Method

Ramadhan, B., Purwanto, Y., and Ruriawan, M.F. (2020).Forensic Malware Identification Using Naive Bayes Method. 2020 International Conference on Information Technology Systems and Innovation (ICITSI), 1-7

work page 2020
[25]

(2019).Explaining a Black-Box Using Deep Variational Information Bottleneck Approach.ArXiv,abs/1902.06918

Bang, S., Xie, P., Wu, W., and Xing, E. (2019).Explaining a Black-Box Using Deep Variational Information Bottleneck Approach.ArXiv,abs/1902.06918

work page arXiv 2019
[26]

(2016).A Graph- Theoretic Approach for Visualization of Data Set Feature Association

Das, A., Goswami, S., Chakraborty, B., and Chakrabarti, A. (2016).A Graph- Theoretic Approach for Visualization of Data Set Feature Association. ACSS

work page 2016
[27]

(2017).A New Hybrid Feature Selection Approach Using Feature Association Map for Supervised and Unsupervised Classification

Das, A., Goswami, S., Chakrabarti, A., and Chakraborty, B. (2017).A New Hybrid Feature Selection Approach Using Feature Association Map for Supervised and Unsupervised Classification. Expert Systems with Applications,88, 81-94. [28]https://archive.ics.uci.edu/ml/datasets

work page 2017
[28]

(2017).A Study on Liver Disease Diagnosis Based on Assessing the Importance of Attributes

Akyol, K., and G¨ ultepe, Y. (2017).A Study on Liver Disease Diagnosis Based on Assessing the Importance of Attributes. International Journal of Intelligent Systems and Applications,9(11), 1-9. 16

work page 2017