Pith. sign in

REVIEW 3 major objections 5 minor 30 references

Enhancing the Interpretability of Rule-based Explanations through Information Retrieval

T0 review · 3 major / 5 minor · reviewed 2026-08-06 · deepseek-v4-flash

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗ Review JSON

Pith's one-line read A tf-idf-based post-processing of rule-based AI output gives clinicians and patients factor-level rankings of what drove a lymphedema risk prediction, and a user study reports that the ranked visualizations are rated more interpretable…

desk verdict A lightweight, clearly-presented post-hoc method for ranking clinical factors from rule-based XAI, whose headline attribution claim rests on an unvalidated frequency heuristic and whose Eq. 7 has a fixable but real scaling error. read the letter →

arxiv 2507.05976 v1 pith:Y2JNRLA2 submitted 2025-07-08 cs.AI cs.IR

Alessandro Umbrico , Guido Bologna , Luca Coraci , Francesca Fracasso , Silvia Gola , Gabriella Cortellessa This is my paper · ORCID
classification cs.AIcs.IR
keywords explainableAItf-idfrule-basedexplanationsattributionmethodinterpretabilitylymphedemariskpredictionuserstudyinformationretrieval
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

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

The reading

The paper proposes a post-hoc attribution method that makes an already rule-based AI model's predictions easier for non-AI users to grasp. It borrows the tf-idf metric from information retrieval and applies it to the rules activated for a prediction, treating each rule as a document and each clinically grouped risk factor as a token. The claim is that a factor's frequency in the activated rule conditions proxies its contribution to the risk prediction, so reversing tf-idf yields a per-patient factor ranking. A user study with 30 participants reports that radar-chart and factor-list presentations are rated significantly more interpretable and useful than the raw rules, by both AI experts and non-experts. If this holds, the approach offers a lightweight, retraining-free route to more readable explainable-AI output in clinical risk communication.

What carries the argument

The central object is an adapted tf-idf score computed over the rules activated by a prediction. Attributes are first clustered into contributing factors, so each factor may appear zero or multiple times in a rule's preconditions. For each activated rule the method computes a term frequency weighted by the rule's coverage, and across the rule set an inverse document frequency that is high when a factor appears in few rules; the per-rule scores are averaged to give each factor a relevance value. Because the raw average is low for factors that appear everywhere, the paper applies a logistic transform and inverts the score, so factors appearing in many activated rules end up with relevance close to 1. That inversion is what converts a standard information-retrieval metric into an attribution score for rule-based predictions.

What would settle it

Construct or locate a patient case where a known decisive factor appears in only one activated rule while a neutral factor appears in many; if the proposed score ranks the neutral factor higher and removing the decisive factor from the rules changes the predicted risk, the frequency-proxy assumption is falsified.

Watch

Extended reading notes

Core claim

The paper argues that raw rule output from an explainable prediction model is not enough for clinicians and patients, and that the model can be treated as a black box and post-processed without retraining. It groups the model's attributes into semantically coherent factors, treats each activated rule as a document and each factor as a token, and computes a coverage-weighted term frequency and an inverse rule frequency. The standard tf-idf meaning is reversed: factors that recur across many rules receive high relevance, because they are the ones the paper takes to shape the prediction. Applied globally over all 301 rules it ranks familiar lymphedema risk factors such as BMI, involved nodes, and radiotherapy technique at the top, and applied locally it yields per-patient radar and list visualizations. In the user study, repeated-measures ANOVA found significant effects of visualization type on both interpretability and usefulness, with the processed visualizations scoring significantly higher than raw rules and no difference between AI-experts and non-experts in their judgments.

Load-bearing premise

The load-bearing premise is that a factor's frequency in the rule preconditions activated for a prediction measures its actual contribution to that prediction; the paper states this assumption but does not test it against clinical outcomes, model ablations, or alternative attribution methods.

Editorial extensions

If this is right

  • Global explanations over all rules yield a ranked list of risk factors, helping clinicians check the model against known lymphedema literature.
  • Local explanations per patient show which factors drove each prediction, with rankings and visualizations that differ across patient profiles.
  • The attribute-to-factor clustering can be changed to coarser or finer granularity without retraining the model, allowing explanations to be tailored to different user expertise levels.
  • In the reported user study, radar and factor-list presentations beat raw rule output on interpretability and usefulness, with no significant difference between AI-expert and non-expert ratings.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

  • An extension left implicit is that the same frequency-as-relevance score could be applied to any model that emits multiple activated rules per prediction, as long as rule conditions can be mapped to semantically grouped factors.
  • A stronger validation would compare the factor rankings against known ground-truth causal roles in controlled rule sets, or against perturbation-based attributions; the paper itself does not test the frequency proxy in that way.
  • The clustering flexibility suggests a testable follow-up: measuring whether coarse versus fine factor groupings change clinicians' judgments or trust using decision-based or calibration metrics rather than self-report alone.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Referee Report

3 major / 5 minor

Summary. The paper proposes an information-retrieval-based method to enhance the interpretability of rule-based explanations produced by an ensemble MLP model for lymphedema risk prediction. The method aggregates input attributes into semantic factors, applies tf-idf statistics to the set of activated rules for a given prediction, inverts the resulting scores, and visualizes factor relevance via radar charts or ranked lists. A user study with 30 participants compared the raw rule output against the two proposed visualizations and found statistically significant improvements in perceived interpretability and usefulness.

Significance. If the relevance scores are accepted as meaningful measures of factor contribution, the approach offers a lightweight, flexible, and model-agnostic way to contextualize rule-based XAI output for non-expert users, and the user study provides preliminary evidence of improved perceived interpretability. The use of IR metrics as a post-hoc explanation layer is a novel idea with potential applicability beyond this clinical use case. However, the paper's central claim that the method 'computes the relevance of each attribute to the prediction' is not supported by the evidence: the frequency-based proxy is unvalidated, and the user study only measures subjective preference. The internal inconsistency in Eq. (7) further undermines confidence in the reported scores.

major comments (3)
  1. [Section 3.2, Eq. (7)] For non-negative tf-idf values, the transformation tf_idf*_i = 1 - 1/(1 + exp(-tf_idf_i)) produces values in (0, 0.5], not in [0, 1]. The statement that 'Values close to 1 are assigned to factors i that are common' is therefore mathematically impossible. This affects the interpretation of Figures 2-4 and the logistic inversion description. Please correct the formula (e.g., use 2/(1+exp(-x)) - 1 or a suitable scaling) or revise the text and figures to reflect the actual range.
  2. [Section 3, paragraph after Eq. (5)] The underlying assumption that the frequency of an attribute in the preconditions of activated rules is proportional to its contribution to the prediction is stated but never validated. The user study in Section 4.3 measures only subjective interpretability and usefulness; it does not test whether the computed scores correspond to the model's actual behavior or to clinically established risk factors. Consequently, the abstract's claim that the method 'computes the relevance of each attribute to the prediction' is unsupported. Please either add a validation study (e.g., model ablations, comparison with SHAP or LIME, or expert assessment of score correctness) or temper the claims to describe the scores as frequency-based indicators of prominence rather than validated contributions.
  3. [Section 4.2] The local explanation demonstration uses 10 profiles with randomly assigned sets of 7 rules each, not actual patient inputs processed through the model. This shows the flexibility of the visualization but does not validate that the local scores reflect how the model would explain a real patient's prediction. Please clarify this limitation explicitly and consider including at least one case with a real input and the rules actually activated by the model for that input.
minor comments (5)
  1. [Section 3.2, Eq. (2)] The text uses 'call' while the equation uses c_all for total coverage; please unify the notation throughout the paper.
  2. [Section 4.3] The reported degrees of freedom F(2, 56) are inconsistent with the stated sample size of 30 participants, as a repeated-measures ANOVA with k=3 conditions and n=30 would have df_error = 29*2 = 58. Please verify the statistical analysis and correct any typographical errors.
  3. [Section 4.3, list of factors] The 'list of factors' visualization presents impact as a percentage, but the paper does not explain how these percentages are computed from the tf_idf* scores. Please specify the normalization used.
  4. [Section 2] The description of the rule extraction algorithm from reference [11] is brief; a few more details (e.g., the number of attributes, typical rule length, and how fidelity is thresholded) would improve reproducibility and help readers interpret the subsequent results.
  5. [Section 5] The phrase 'infers causal correlations' overstates what a frequency-based analysis can establish. Suggest replacing 'causal correlations' with 'associations' or 'prominence' to avoid implying causal inference.

Circularity Check

0 steps flagged · score 2.0 of 10

No significant circularity: the relevance scores are explicit formula-transformations of rule frequencies, with the frequency-as-contribution link stated as an assumption rather than derived; the only self-citation supplies the input rule set, not the conclusion.

full rationale

The paper's relevance scores are defined directly by Eqs. 1-7 from rule statistics (tf, idf, tf-idf, logistic inversion). No fitted parameter, target response, or external outcome is used to construct the scores, so the computation does not secretly reuse the quantity it claims to explain. The statement in Section 3 that 'the more a set of attributes is present in the precondition of the activated rules, the greater the relevance (or contribution)' is an explicitly stated modeling assumption, not a derived result presented as a prediction; the paper does not pretend to test or derive it from first principles. The self-citation [11] (Fidex) is used to obtain the rule-based model and the set of rules; this is an input dependency, not a load-bearing circular step, because the IR post-processing would apply to any rule set and the relevance computation does not depend on the internal derivation of [11]. The user study measures perceived interpretability and usefulness of visualization formats, not the accuracy of the relevance scores, so it cannot validate the proxy assumption; this is a validity limitation, not circularity. Separately, Eq. 7 as written maps all non-negative tf-idf values to (0, 0.5], making the stated range [0,1] and 'values close to 1' impossible; this is a technical correctness concern but not a circularity. Overall, the derivation chain is self-contained and the central claim rests on an explicit, unvalidated assumption rather than on circular reasoning.

Assumptions & free parameters 1 free parameters · 3 assumptions · 0 invented entities

The central claim rests on a small set of design choices: the factor clustering, the tf-idf analogy, and the coverage weighting. None of these are empirically validated. There are no fitted scalar parameters beyond the hand-crafted clustering, so the circularity burden is low. The paper does not introduce new physical or conceptual entities beyond abstract factor groupings.

free parameters (1)
  • Semantic clustering mapping = hand-crafted taxonomy (e.g., SMOKER, FORMER SMOKER, CURRENT SMOKER -> SMOKER)
    The aggregation of attributes into factors is designed by the authors and determines all relevance scores; no validation or sensitivity analysis is provided.
assumptions (3)
  • domain assumption Frequency of a factor in activated rules is proportional to its causal influence on the prediction.
    Stated in Section 3 as 'The underlying assumption is that the more a set of attributes is present in the precondition of the activated rules, the greater the relevance (or contribution) of the associated factor on the considered prediction.' Never tested against ground truth.
  • ad hoc to paper Rules can be treated as documents and factors as tokens for tf-idf analysis.
    The analogy between rule preconditions and document token counts is a modeling choice; no formal justification links tf-idf to explanation quality.
  • ad hoc to paper Coverage weighting w_j = c_j / c_all is an appropriate aggregation.
    Introduced in Eq. 2 without analysis of its interaction with the final score; it is one of several possible weightings and its effect on interpretability is untested.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Enhancing the Interpretability of Rule-based Explanations through Information Retrieval." pith.science (2026). https://pith.science/paper/Y2JNRLA2

@misc{pith2026250705976,
  author       = {Pith},
  title        = {Pith review of: Enhancing the Interpretability of Rule-based Explanations through Information Retrieval},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/Y2JNRLA2}},
  note         = {Machine review of arXiv:2507.05976}
}
read the original abstract

The lack of transparency of data-driven Artificial Intelligence techniques limits their interpretability and acceptance into healthcare decision-making processes. We propose an attribution-based approach to improve the interpretability of Explainable AI-based predictions in the specific context of arm lymphedema's risk assessment after lymph nodal radiotherapy in breast cancer. The proposed method performs a statistical analysis of the attributes in the rule-based prediction model using standard metrics from Information Retrieval techniques. This analysis computes the relevance of each attribute to the prediction and provides users with interpretable information about the impact of risk factors. The results of a user study that compared the output generated by the proposed approach with the raw output of the Explainable AI model suggested higher levels of interpretability and usefulness in the context of predicting lymphedema risk.

Figures

Figures reproduced from arXiv: 2507.05976 by the authors.

Figure 1
Figure 1. Excerpt of the mapping between attributes and factors. [PITH_FULL_IMAGE:figures/full_fig_p004_1.png] view at source ↗
Figure 2
Figure 2. Interpretation of the global explanation generated on the whole set of rules (i.e., the predictive model). [PITH_FULL_IMAGE:figures/full_fig_p005_2.png] view at source ↗
Figure 3
Figure 3. Interpretations of the local explanations generated for the defined patient profiles. [PITH_FULL_IMAGE:figures/full_fig_p007_3.png] view at source ↗
Figures from the paper (3 more)
Figure 4
Figure 4. Figure 4: Heatmap showing the distribution of factor relevance over the profiles. [PITH_FULL_IMAGE:figures/full_fig_p007_4.png]
Figure 5
Figure 5. Figure 5: The three different output visualizations for a single patient profile. From left to right: Raw output containing [PITH_FULL_IMAGE:figures/full_fig_p008_5.png]
Figure 6
Figure 6. Figure 6: The mean scores obtained for both Interpretability and Usefulness of the three visualization options. is not directly applicable to our case since each prediction is made of multiple rules, and a single factor could occur multiple times. So, it is not easy to find a di…

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

30 extracted references · 29 canonical work pages

  1. [1]

    The importance of interpretability and visualization in machine learning for applications in medicine and health care

    Alfredo Vellido. The importance of interpretability and visualization in machine learning for applications in medicine and health care. Neural Computing and Applications, 32(24):18069–18083, 2020

    work page 2020

  2. [2]

    Marzyeh Ghassemi, Luke Oakden-Rayner, and Andrew L. Beam. The false hope of current approaches to explainable artificial intelligence in health care. The Lancet Digital Health, 3(11):e745–e750, 2021

    work page 2021

  3. [3]

    Markus, Jan A

    Aniek F. Markus, Jan A. Kors, and Peter R. Rijnbeek. The role of explainability in creating trustworthy artificial intelligence for health care: A comprehensive survey of the terminology, design choices, and evaluation strategies. Journal of Biomedical Informatics, 113:103655, 2021

    work page 2021

  4. [4]

    Causability and explainability of artificial intelligence in medicine

    Andreas Holzinger, Georg Langs, Helmut Denk, Kurt Zatloukal, and Heimo Müller. Causability and explainability of artificial intelligence in medicine. WIREs Data Mining and Knowledge Discovery, 9(4), 2019

    work page 2019

  5. [5]

    What do we want from explainable artificial intelligence (xai)? – a stakeholder perspective on xai and a conceptual model guiding interdisciplinary xai research

    Markus Langer, Daniel Oster, Timo Speith, Holger Hermanns, Lena Kästner, Eva Schmidt, Andreas Sesing, and Kevin Baum. What do we want from explainable artificial intelligence (xai)? – a stakeholder perspective on xai and a conceptual model guiding interdisciplinary xai research. Artificial Intelligence, 296:103473, 2021

    work page 2021

  6. [6]

    black box

    Fei Wang, Rainu Kaushal, and Dhruv Khullar. Should health care demand interpretable artificial intelligence or accept "black box" medicine? Annals of internal medicine, 172(1):59—60, January 2020

    work page 2020

  7. [7]

    McCradden, and Anna Goldenberg

    Sana Tonekaboni, Shalmali Joshi, Melissa D. McCradden, and Anna Goldenberg. What clinicians want: Contex- tualizing explainable machine learning for clinical end use. In Finale Doshi-Velez, Jim Fackler, Ken Jung, David Kale, Rajesh Ranganath, Byron Wallace, and Jenna Wiens, editors, Proceedings of the 4th Machine Learning for Healthcare Conference, volume...

    work page 2019

  8. [8]

    Explainable ai (xai): A systematic meta-survey of current challenges and future opportunities

    Waddah Saeed and Christian Omlin. Explainable ai (xai): A systematic meta-survey of current challenges and future opportunities. Knowledge-Based Systems, 263:110273, 2023

    work page 2023

Show all 30 references
  1. [9]

    Methods for interpreting and understanding deep neural networks

    Grégoire Montavon, Wojciech Samek, and Klaus-Robert Müller. Methods for interpreting and understanding deep neural networks. Digital Signal Processing, 73:1–15, 2018

    2018

  2. [10]

    Moore, Marinka Zitnik, and John H

    Carlo Combi, Beatrice Amico, Riccardo Bellazzi, Andreas Holzinger, Jason H. Moore, Marinka Zitnik, and John H. Holmes. A manifesto on explainability for artificial intelligence in medicine. Artificial Intelligence in Medicine, 133:102423, 2022

    2022

  3. [11]

    Fidex: An algorithm for the explainability of ensembles and svms

    Guido Bologna, Jean-Marc Boutay, Quentin Leblanc, and Damian Boquete. Fidex: An algorithm for the explainability of ensembles and svms. In José Manuel Ferrández Vicente, Mikel Val Calvo, and Hojjat Adeli, editors, Bioinspired Systems for Translational Applications: From Roboti...

    2024

  4. [12]

    Research challenges in trustworthy artificial intelligence and computing for health: The case of the pre-act project

    Foivos Charalampakos, Thomas Tsouparopoulos, Yiannis Papageorgiou, Guido Bologna, André Panisson, Alan Perotti, and Iordanis Koutsopoulos. Research challenges in trustworthy artificial intelligence and computing for health: The case of the pre-act project. In 2023 Joint Europe...

    2023

  5. [13]

    Madai, and the Precise4Q consortium

    Julia Amann, Alessandro Blasimme, Effy Vayena, Dietmar Frey, Vince I. Madai, and the Precise4Q consortium. Explainability for artificial intelligence in healthcare: a multidisciplinary perspective. BMC Medical Informatics and Decision Making, 20(1):310, 2020

    2020

  6. [14]

    Incidence and risk factors of lymphedema after breast cancer treatment: 10 years of follow-up

    Ana Carolina Padula Ribeiro Pereira, Rosalina Jorge Koifman, and Anke Bergmann. Incidence and risk factors of lymphedema after breast cancer treatment: 10 years of follow-up. The Breast, 36:67–73, 2017

    2017

  7. [15]

    Incidence and risk factors of breast cancer lymphoedema

    Jin Hee Park, Won Hee Lee, and Hae Soo Chung. Incidence and risk factors of breast cancer lymphoedema. Journal of Clinical Nursing, 17(11):1450–1459, 2008

    2008

  8. [16]

    Statistical power analyses using g*power 3.1: Tests for correlation and regression analyses

    Franz Faul, Edgar Erdfelder, Axel Buchner, and Albert-Georg Lang. Statistical power analyses using g*power 3.1: Tests for correlation and regression analyses. Behavior Research Methods, 41:1149–1160, 2009

    2009

  9. [17]

    European Union. Regulation (EU) 2016/679 of the European Parliament and of the Council of 27 April 2016 on the protection of natural persons with regard to the processing of personal data and on the free movement of such data (General Data Protection Regulation). https://eur-l...

    2016

  10. [18]

    Exploring multi-task learning for explainability

    Foivos Charalampakos and Iordanis Koutsopoulos. Exploring multi-task learning for explainability. In Sła- womir Nowaczyk, Przemysław Biecek, Neo Christopher Chung, Mauro Vallati, Paweł Skruch, Joanna Jaworek- Korjakowska, Simon Parkinson, Alexandros Nikitas, Martin Atzmüller, ...

    2023

  11. [19]

    Lundberg and Su-In Lee

    Scott M. Lundberg and Su-In Lee. A unified approach to interpreting model predictions. In Proceedings of the 31st International Conference on Neural Information Processing Systems, NIPS’17, page 4768–4777, Red Hook, NY , USA, 2017. Curran Associates Inc

    2017

  12. [20]

    Why Should I Trust You?

    Marco Tulio Ribeiro, Sameer Singh, and Carlos Guestrin. “Why Should I Trust You?”: Explaining the Predictions of Any Classifier. In Proceedings of the 22nd ACM SIGKDD International Conference on Knowledge Discovery and Data Mining, KDD ’16, page 1135–1144, New York, NY , USA, ...

    2016

  13. [21]

    Analysis of regression in game theory approach

    Stan Lipovetsky and Michael Conklin. Analysis of regression in game theory approach. Applied Stochastic Models in Business and Industry, 17(4):319–330, 2001

    2001

  14. [22]

    Explaining prediction models and individual predictions with feature contributions

    Erik Štrumbelj and Igor Kononenko. Explaining prediction models and individual predictions with feature contributions. Knowledge and Information Systems, 41(3):647–665, Dec 2014

    2014

  15. [23]

    Covert, Scott M

    Hugh Chen, Ian C. Covert, Scott M. Lundberg, and Su-In Lee. Algorithms to estimate shapley value feature attributions. Nature Machine Intelligence, 5(6):590–601, Jun 2023

    2023

  16. [24]

    Learning important features through propagating activation differences

    Avanti Shrikumar, Peyton Greenside, and Anshul Kundaje. Learning important features through propagating activation differences. In Proceedings of the 34th International Conference on Machine Learning - Volume 70, ICML’17, page 3145–3153. JMLR.org, 2017

    2017

  17. [25]

    On pixel-wise explanations for non-linear classifier decisions by layer-wise relevance propagation

    Sebastian Bach, Alexander Binder, Grégoire Montavon, Frederick Klauschen, Klaus-Robert Müller, and Wojciech Samek. On pixel-wise explanations for non-linear classifier decisions by layer-wise relevance propagation. PLOS ONE, 10(7):1–46, 07 2015

    2015

  18. [26]

    A holistic approach to interpretability in financial lending: Models, visualizations, and summary-explanations

    Chaofan Chen, Kangcheng Lin, Cynthia Rudin, Yaron Shaposhnik, Sijia Wang, and Tong Wang. A holistic approach to interpretability in financial lending: Models, visualizations, and summary-explanations. Decision Support Systems, 152:113647, 2022

    2022

  19. [27]

    Developing a catalogue of explainability methods to support expert and non-expert users

    Kyle Martin, Anne Liret, Nirmalie Wiratunga, Gilbert Owusu, and Mathias Kern. Developing a catalogue of explainability methods to support expert and non-expert users. In Max Bramer and Miltos Petridis, editors, Artificial Intelligence XXXVI, pages 309–324, Cham, 2019. Springer...

    2019

  20. [28]

    Explainability in nlp model: Detection of covid-19 twitter fake news

    Wan Yit Yong, Rajesh Jaiswal, and Fernando Perez Tellez. Explainability in nlp model: Detection of covid-19 twitter fake news. In Proceedings of the 2023 Conference on Human Centered Artificial Intelligence: Education and Practice, HCAIep ’23, page 1, New York, NY , USA, 2023....

    2023

  21. [29]

    Sparkly: A simple yet surprisingly strong tf/idf blocker for entity matching

    Derek Paulsen, Yash Govind, and AnHai Doan. Sparkly: A simple yet surprisingly strong tf/idf blocker for entity matching. Proc. VLDB Endow., 16(6):1507–1519, 2023. 11

    2023

  22. [2016]

    Official Journal of the European Union, L119, 1–88

Pith tools

Reviewed August 6, 2026 · model on record in the stance chip above.