Pith. sign in

REVIEW 4 major objections 4 minor 1 references

Cooperative effects in feature importance of individual patterns: application to air pollutants and Alzheimer disease

T0 review · 4 major / 4 minor · reviewed 2026-08-06 · deepseek-v4-flash

Pith's one-line read Per-case feature scores reveal O3-NO2 synergy in Alzheimer's deaths

desk verdict A plausible per-instance extension of Hi-Fi that I cannot assess because the full text is unreadable; the abstract alone does not answer the key identifiability question about the local reference distribution. read the letter →

arxiv 2508.00930 v1 pith:ZNYMG67H submitted 2025-07-30 cs.LG physics.data-an

classification cs.LGphysics.data-an
keywords featureimportanceexplainableAIsynergyredundancypartialinformationdecompositionLeaveOneCovariateOutairpollutionAlzheimerdisease
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's aim is to turn feature importance for regression models from a single global score per feature into a per-pattern decomposition: for each individual data point, each feature gets a unique (two-body) contribution, plus redundant and synergistic higher-order contributions. The framework is built on an adaptive version of the Leave One Covariate Out (LOCO) measure, called Hi-Fi, and is compared with the Shapley effect. Applied to a One-Health question, the method predicts Alzheimer's disease mortality from air pollutants and urban green-area density; the central finding is a synergistic association between $O_3$ and $NO_2$ with mortality, strongest in the provinces of Bergamo and Brescia, and a synergistic influence of green-area density with pollutants. A sympathetic reader would care because standard importance tools report only isolated feature relevance, whereas this framework claims to expose cooperative effects that only appear when features act together, at the resolution of single predictions.

What carries the argument

The central object is the local (per-pattern) Hi-Fi decomposition, an adaptive extension of the Leave One Covariate Out (LOCO) measure. For each data point, it uses a variance-based partial information decomposition to split a feature's importance into a unique two-body term, a redundant term, and a synergistic higher-order term. This decomposition carries the argument because it is what turns 'this feature matters' into a statement about how features matter together, and it is the object whose values identify Bergamo and Brescia as sites of $O_3$-$NO_2$ synergy and green areas as synergistic contributors.

What would settle it

Recompute the local Hi-Fi decomposition on the same data with a different well-validated regression model (for example, a regularized linear model or a tree ensemble), and also on a control data set where $O_3$ and $NO_2$ values are independently shuffled. If the strong synergistic scores in Bergamo and Brescia disappear or appear in the shuffled control, the claimed $O_3$-$NO_2$ synergy is a property of the estimator rather than of the underlying relationship.

Watch

Extended reading notes

Core claim

The paper claims that cooperative effects in feature importance can be measured for individual patterns, not only averaged over a data set. In the authors' formulation, the predictive contribution of a feature in a regression model splits into three components, unique, redundant, and synergistic, and this split is computed locally, pattern by pattern. The empirical result in the Alzheimer's disease application is that ozone ($O_3$) and nitrogen dioxide ($NO_2$) act synergistically in predicting mortality, particularly in the provinces of Bergamo and Brescia, and that the density of urban green areas also participates synergistically with the pollutants. These claims are established by comparing the local Hi-Fi scores with the global Shapley effect and by inspecting which provinces carry the synergistic component.

Load-bearing premise

The load-bearing premise is that the regression model predicting Alzheimer's disease mortality is well specified, so that decomposing its per-pattern predictions tells us about real cooperative effects in the health and environmental data rather than about artifacts of the model or the LOCO baseline.

Editorial extensions

If this is right

  • Each feature in a regression model can be reported as three per-instance scores (unique, redundant, synergistic) rather than one global importance, giving explanations at the level of individual predictions.
  • $O_3$ and $NO_2$ should be treated as a pair in future analyses of air pollution and Alzheimer's disease mortality, since their joint contribution exceeds what either feature alone explains.
  • Urban green-area density enters the prediction of Alzheimer's mortality by modulating pollutant effects, not just as an independent protective factor.
  • Local Hi-Fi can be transferred to other regression settings that need high-order cooperative effects, such as biomarker panels or environmental exposure mixtures.

Reading between the lines

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

  • If the local synergy scores survive model resampling and cross-validation, one testable extension is to use them as a screening map: provinces with consistently high per-pattern synergy would be candidates for targeted joint-pollutant monitoring.
  • The same per-pattern decomposition could be applied with a different importance baseline, checking whether the reported $O_3$-$NO_2$ synergy is robust to the choice of LOCO or is a property of the data-generating process.
  • A natural next experiment is to feed the local scores into a spatial model and ask whether high-synergy provinces cluster geographically or follow pollution transport patterns; the paper does not do this.
Share X Bluesky LinkedIn Reddit HN

Signed reviews

No signed human review yet.

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

4 major / 4 minor

Summary. The paper proposes a per-pattern extension of the Hi-Fi (high-order feature importance) framework, in which each input feature is assigned unique, redundant, and synergistic importance scores for individual data points, and it compares these scores with Shapley-effect feature importance. The method is applied to an ecological One-Health dataset linking air pollutants (O3, NO2), urban green density, and Alzheimer's disease mortality, with the reported main finding being a synergistic association between O3 and NO2 with mortality, especially in the Italian provinces of Bergamo and Brescia, and a synergistic influence of urban green density with pollutants. The abstract frames local Hi-Fi as a promising, widely applicable XAI tool.

Significance. If the per-pattern variance-based decomposition is theoretically justified and empirically validated, the framework could be a useful addition to explainable AI, particularly for uncovering higher-order interactions at the level of individual predictions. The One-Health application is topical and could be of broad interest. However, the current manuscript does not make this case accessible: the full text after the abstract is unreadable due to encoding corruption, and the abstract alone supplies no model specification, no local reference distribution, no out-of-sample or synthetic validation, and no uncertainty quantification. The significance is therefore entirely conditional on details that the manuscript does not presently make verifiable.

major comments (4)
  1. [Full text (after abstract)] The submitted full text is unreadable: after the abstract the characters are garbled, so the mathematical definitions of the per-pattern Hi-Fi scores, the LOCO estimator, and the empirical analysis cannot be checked. This must be corrected in a resubmission; without a readable manuscript no technical claim can be verified.
  2. [Abstract] The central claim of synergistic O3/NO2 and green-area effects is model-relative: the scores are computed from a fitted regression model for AD mortality, but the abstract reports no model class, no hyperparameters, no validation, and no uncertainty quantification. Because a model with interaction terms can manufacture synergy even under an additive data-generating process, the authors must report the model specification and an out-of-sample or synthetic benchmark before interpreting the reported associations.
  3. [Abstract] For a single pattern (one observation), a variance-based decomposition requires a distribution over feature values conditioned on that pattern; the abstract never specifies this local reference distribution or the LOCO baseline (estimator, resampling scheme). Without a stated and justified local distribution, the unique/redundant/synergistic split is not identifiable. The authors should define the local distribution explicitly and show that the reported synergy is robust to its choice.
  4. [Abstract] The application is an ecological, observational analysis at province level, but the abstract gives no adjustment for potential confounders (e.g., socioeconomic status, healthcare access, smoking) and no measure of uncertainty for the reported province-specific results. The statement 'especially in the provinces of Bergamo and Brescia' requires at least interval estimates or a multiple-comparison-aware analysis.
minor comments (4)
  1. [Abstract] The phrase 'Bergamo e Brescia' should be 'Bergamo and Brescia' in an English-language manuscript, and the disease name should be used consistently as 'Alzheimer's disease'.
  2. [Abstract] The terms 'Hi-Fi' and 'Shapley effect' are used without citations; the manuscript should give references for the baseline method and the comparison metric.
  3. [Abstract] The manuscript does not state data sources or availability; since the application is a key selling point, a reproducibility statement is needed.
  4. [Full text] The abstract refers to a recently proposed adaptive version of LOCO, but the citation is not visible in the readable portion; ensure the reference is explicit in the resubmission.

Circularity Check

0 steps flagged · score 1.0 of 10

No circularity demonstrated: the per-pattern Hi-Fi scores are a variance-based decomposition of fitted model predictions, with an external Shapley-effect comparison; the model-relativity concerns are correctness risks, not circular steps, and the unreadable full text prevents exhibiting any equation-level reduction.

full rationale

The only quotable text is the abstract; the supplied full text is corrupted mojibake and cannot support equation-level evidence. On the abstract alone, the claimed derivation is a per-pattern decomposition (unique, redundant, synergistic) of predictions from a regression model fitted to air-pollutant and AD-mortality data. Decomposing a fitted model's predictions into variance components is a derived, model-relative operation of the same kind as LOCO and Shapley effects; it is not definitional circularity because the synergy scores are outputs, not re-labeled inputs, of the decomposition. The abstract states the framework is 'while comparing it with the well-known measure of feature importance named Shapley effect', which is an external reference rather than a self-citation chain. No uniqueness theorem, no load-bearing self-citation, and no fitted-parameter-renamed-as-prediction step can be exhibited from the readable text. The skeptic's points (variance undefined for a single pattern without a local reference distribution; synergy scores model-relative; possible misspecification) are genuine validation and identifiability concerns, which fall under correctness risk rather than circularity per the analysis rules. Proportionately, with no quotable reduction, the honest finding is no significant circularity; score 1 reflects only the residual uncertainty from the unreadable full text, not any identified circular step.

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

Only the abstract is available, so the ledger is limited to the modeling assumptions that are explicitly visible: the PID/LOCO framework and the quality of the ecological data. No new physical entities are introduced.

free parameters (1)
  • Regression model parameters for AD mortality = not reported in abstract
    The framework derives feature importance from a predictive model fitted to the data; the fitted values of those parameters are not disclosed in the abstract.
assumptions (3)
  • domain assumption The partial information decomposition of model predictions correctly separates unique, redundant, and synergistic contributions
    The whole method rests on this decomposition being meaningful; no proof or external benchmark is visible in the abstract.
  • domain assumption The dataset used (air pollutants, green areas, AD mortality in Italian provinces) contains valid measurements with no major confounding at the province level
    The main applied finding is an ecological association; the abstract does not discuss confounding, spatial autocorrelation, or measurement error.
  • domain assumption The Leave One Covariate Out (LOCO) based Hi-Fi scores are an appropriate baseline for feature importance in this regression setting
    The framework adapts LOCO; the validity of LOCO as a baseline is assumed without visible justification in the abstract.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Cooperative effects in feature importance of individual patterns: application to air pollutants and Alzheimer disease." pith.science (2026). https://pith.science/paper/ZNYMG67H

@misc{pith2026250800930,
  author       = {Pith},
  title        = {Pith review of: Cooperative effects in feature importance of individual patterns: application to air pollutants and Alzheimer disease},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/ZNYMG67H}},
  note         = {Machine review of arXiv:2508.00930}
}
abstract

Leveraging recent advances in the analysis of synergy and redundancy in systems of random variables, an adaptive version of the widely used metric Leave One Covariate Out (LOCO) has been recently proposed to quantify cooperative effects in feature importance (Hi-Fi), a key technique in explainable artificial intelligence (XAI), so as to disentangle high-order effects involving a particular input feature in regression problems. Differently from standard feature importance tools, where a single score measures the relevance of each feature, each feature is here characterized by three scores, a two-body (unique) score and higher-order scores (redundant and synergistic). This paper presents a framework to assign those three scores (unique, redundant, and synergistic) to each individual pattern of the data set, while comparing it with the well-known measure of feature importance named {\it Shapley effect}. To illustrate the potential of the proposed framework, we focus on a One-Health application: the relation between air pollutants and Alzheimer's disease mortality rate. Our main result is the synergistic association between features related to $O_3$ and $NO_2$ with mortality, especially in the provinces of Bergamo e Brescia; notably also the density of urban green areas displays synergistic influence with pollutants for the prediction of AD mortality. Our results place local Hi-Fi as a promising tool of wide applicability, which opens new perspectives for XAI as well as to analyze high-order relationships in complex systems.

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

1 extracted references · 1 canonical work pages

  1. [1]

    �������� ��������� �������� ����� ������� ������������� ����� ��� �� ����� ��� �������� ������������� �������� ������������� �������� ������� ������� ��� �������� ����������� ����� ���������� �� ����������� ��� ����������� �������������������� ��������� ��������� �������� ���� ����� ��������� �� ������ �� ���� ������� ������ ��� ������� ������ ���� ���� �...

Pith tools

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