Pith. sign in

REVIEW 4 major objections 6 minor 20 references

Models that gain the most accuracy from country and behaviour features also show the strongest direct dependence on country in AML flagging.

Reviewed by Pith at T0; open to challenge. T0 means a machine referee read the full paper against a public rubric. the ladder, T0–T4 →

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

T0 review · grok-4.5

2026-07-11 08:48 UTC pith:6DE744FX

load-bearing objection Solid first application of path-specific counterfactual fairness to AML; the accuracy–fairness ranking is real on the synthetic data, with the mediator construction as the main (already-flagged) soft spot. the 4 major comments →

arxiv 2607.05101 v1 pith:6DE744FX submitted 2026-07-06 cs.LG

Counterfactual Methods for Detecting Unfairness in Anti-Money Laundering Algorithms

Lea Multerer , Michele Inchingolo , David Kletz , Adrian Cosma , Alessandro Antonucci , Martina Gogova This is my paper
classification cs.LG
keywords Counterfactual Fairness AnalysisMediation AnalysisStructural Causal ModelsGraph Neural NetworksAnti-Money LaunderingKnow Your Customer DataPath-specific effectsAccuracy-fairness trade-off
verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved

The pith

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

Banks train machine-learning systems on transaction data plus sensitive Know-Your-Customer attributes to flag possible money laundering. Those attributes are often required by regulation, yet can produce unfair decisions if the model leans on them beyond legitimate behavioural differences. This paper builds a structural causal model on synthetic IBM AMLSim data, adding a country attribute and a scalar transaction-behaviour mediator, and shows that the extra features raise detection scores for decision-tree and graph-neural-network classifiers. Using path-specific counterfactual effects it then decomposes each model’s total country effect into an indirect (mediated, treated as fair) part and a direct (unmediated, treated as unfair) part. The architecture that improves most when the new features are added also exhibits the largest relative direct effects, giving a concrete accuracy–fairness trade-off and arguing that AML systems need systematic causal fairness audits.

Core claim

On the extended synthetic AML data, fairness violations measured by non-negligible natural direct effects of the protected country attribute tend to be more pronounced precisely for the models whose predictive performance benefits most from those country and behaviour features, furnishing a concrete accuracy–fairness trade-off.

What carries the argument

Path-specific counterfactual effects (natural direct effect NDE and natural indirect effect NIE) inside a structural causal model that treats country as protected attribute, transaction behaviour as mediator, and the trained classifier as the deterministic structural equation for the prediction.

Load-bearing premise

The hand-crafted transaction-behaviour score plus a tree regression is assumed to capture the legitimate causal path from country to the model’s flag; if that mediator is incomplete or misspecified, the split into fair and unfair effects no longer holds.

What would settle it

Re-run the same NDE/NIE decomposition after replacing the constructed mediator with an alternative behavioural summary (or with real KYC country labels) and check whether the ranking of models by direct-effect size still matches their accuracy gains.

Watch this falsifier — get emailed when new claim-graph text bears on it.

If this is right

  • AML pipelines that add KYC country features for accuracy gains should be required to report path-specific direct effects, not only aggregate performance.
  • Graph architectures that harvest large lifts from country node features are the first candidates for fairness remediation or for restricting which paths may use those features.
  • Regulators can treat a large relative natural direct effect as evidence that a model is relying on protected attributes beyond task-relevant behaviour.
  • The same causal template can be applied to other prescribed risk indicators (account age, residency status) once they appear in production data.

Where Pith is reading between the lines

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

  • Because real AML labels are noisier and more imbalanced than the synthetic set, the same direct-effect inflation may appear earlier and at lower accuracy levels in live systems.
  • Privacy-preserving multi-bank training could raise baseline graph performance without country features, potentially relaxing the accuracy–fairness tension the paper observes.
  • If the mediator model itself is biased by country, the “fair” indirect path may silently absorb unfairness that the current audit would miss.

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

4 major / 6 minor

Summary. The paper applies a structural-causal, path-specific fairness audit (flip rates, total effect, natural direct/indirect effects) to AML transaction classifiers on the synthetic IBM AMLSim HI-Small dataset. The authors construct a pseudo-KYC protected attribute (account country, recovered by bank-name matching and currency fallback) and a scalar transaction-behaviour mediator, train XGBoost, GNN-GIN and GNN-PNA on basic vs extended feature sets, and show that models whose F1 gains most from the extended features (especially GNN-GIN) also exhibit larger relative natural direct effects of country. They interpret this as a concrete accuracy–fairness trade-off and argue for systematic counterfactual fairness analysis in AML.

Significance. AML is a high-stakes, heavily regulated domain where KYC risk factors are both mandated and potentially sensitive; a causal mediation lens that separates legitimate behavioural pathways from direct dependence on protected attributes is therefore timely. The work is, to the authors’ knowledge, the first path-specific counterfactual fairness study in AML, spans tabular and state-of-the-art GNN detectors, reports careful experimental protocol (temporal splits, class-weighted losses, F1-tuned thresholds, mediator R²≈0.80), and releases code. If the accuracy–fairness ranking survives stronger mediator and GNN-intervention checks, it supplies a useful empirical benchmark for regulators and practitioners weighing KYC features in automated monitoring.

major comments (4)
  1. Section 3 and the mediator paragraph of Section 5: the central claim that larger F1 gains co-occur with larger ‘unfair’ natural direct effects rests on the scalar mediator M = log(N_tx+1)+log(vol_kUSD+1) and the tree regressor ˆf_M(A,X) (R²=0.80) adequately isolating the legitimate path A→M→Ŷ. Country fixed effects explain only R²=0.09 of M, so most variation is within-country; residual country-specific behavioural channels left outside M would be mislabelled as direct (unfair) effect. A sensitivity analysis with alternative or multi-dimensional mediators (or an explicit residual-country term) is needed before the accuracy–fairness ranking can be read as fairness violations rather than mediator incompleteness.
  2. Section 5 (GNN intervention scheme): for GNN-GIN/PNA the authors duplicate the graph, rewrite country on all nodes of the intervened country, and freeze the remaining structure and neighbourhood. This answers a global ‘what if all transactions from this country…’ question and severs potential feedback from A into neighbourhood context that message-passing models exploit. Under that intervention it is unclear that the classical NDE/NIE formulae (Eqs. 4–5) remain path-specific in the intended sense. Either a formal argument for identification under this graph intervention, or a comparison with a local (single-node / single-edge) intervention, is required for the GNN fairness numbers that drive the main ranking.
  3. Table 2 vs Table 3 / Figure 4: the headline claim is that ‘fairness violations tend to be more pronounced for models whose predictive performance benefits the most from the extended features.’ GNN-GIN’s F1 jump (0.20→0.42) and dispersed RDEs (e.g. Saudi Arabia RDE=5.90) support this; GNN-PNA’s tiny gain (0.56→0.58) and near-zero effects also fit. The claim is currently qualitative (kernel densities and selected extremes). A quantitative summary (e.g. correlation of ΔF1 with |RDE| or median |RDE| across countries, with uncertainty) would make the trade-off claim falsifiable rather than impressionistic.
  4. Section 5, relative-effect scaling: effects are divided by E[Ŷ_{A←US}], which differs by nearly an order of magnitude across models (0.248 XGBoost, 0.093 GNN-GIN, 0.035 GNN-PNA). Relative scaling aids cross-model readability but can amplify small absolute differences for low-flagging models and compress them for high-flagging ones. Absolute TE/NDE/NIE (or effects on the probability scale with confidence intervals) should be reported alongside the relative quantities so that the ranking is not an artefact of baseline flagging rates.
minor comments (6)
  1. Throughout: country-specific numerical results are repeatedly caveated as ‘not expected to reflect country-specific patterns,’ which is appropriate; consider moving the strongest caveats into the abstract/introduction so readers do not over-interpret Table 1 or Table 3.
  2. Figure 3: flip rates are shown in percent; the y-axis label and caption should state units explicitly and note that US factual flip rates are omitted (or shown separately) for XGBoost vs GNN.
  3. Section 2, structural equations: UA,X is written as a shared exogenous confounder of A and X; a brief remark on whether the fitted models allow residual dependence between A and X after conditioning on observed covariates would help readers assess unmeasured confounding.
  4. Table 1: only countries with >30k transactions are shown; state the total number of country categories (34 including crypto) and the share of volume covered by the displayed rows.
  5. References / related work: a short pointer to concurrent fairness-in-finance / transaction-fraud fairness papers (beyond the two correlational AML citations) would better situate the contribution.
  6. Typographical: ‘operationalsation’ → ‘operationalisation’; occasional spacing issues around en-dashes and equation references.

Circularity Check

0 steps flagged

No circularity: path-specific effects are standard post-hoc causal quantities on independently trained classifiers; accuracy gains and fairness metrics are separate empirical measurements.

full rationale

The paper trains three classifiers (XGBoost, GNN-GIN, GNN-PNA) on the basic (TB) and extended (TE) feature sets, reports ordinary classification metrics (F1, precision, recall) on a held-out temporal split, then evaluates the standard counterfactual quantities FR, TE, NDE and NIE (Eqs. 2–5) under interventions on the constructed country attribute. The mediator mechanism is approximated by a separately fitted tree regressor ˆf_M (R² = 0.80 on validation), which is the usual practical step in mediation analysis and does not algebraically force the subsequent effect sizes. The accuracy–fairness ranking is an empirical observation read off Tables 2–3, not a quantity derived by rearranging the training loss or the fitted mediator parameters. Citations are to external causal-fairness literature (Pearl, Chiappa, Kusner et al.); there are no load-bearing self-citations, uniqueness theorems imported from the authors, or ansatzes smuggled via prior work. The construction of the synthetic country and behaviour features is an experimental design choice, not a definitional loop that makes the reported effects true by construction. Hence the derivation chain is self-contained against the paper’s own equations and external benchmarks.

Axiom & Free-Parameter Ledger

4 free parameters · 4 axioms · 2 invented entities

The central claim rests on a standard causal fairness model plus several domain-specific modelling choices required because the public synthetic data lack real KYC fields. Free parameters are the usual ML hyper-parameters and the hand-crafted mediator formula. Invented entities are the pseudo-country labels and the scalar transaction-behaviour mediator. No new physical or mathematical objects are postulated beyond these operational constructions.

free parameters (4)
  • transaction-behaviour formula coefficients = additive logs with +1 offsets
    The mediator is defined as log(N_tx + 1) + log(volume_kUSD + 1); the additive combination and the unit scaling are chosen by the authors rather than derived.
  • mediator regression hyper-parameters = 50 trees, depth 10, leaf 10
    Tree ensemble used to approximate f_M (50 trees, max depth 10, min leaf 10) is selected by validation R².
  • classifier decision thresholds = per-model F1-maximising thresholds
    Thresholds maximising F1 on the validation split are chosen after training and directly affect the binary predictions whose counterfactual effects are reported.
  • XGBoost / GNN training hyper-parameters = as listed in Section 4
    Learning rates, depths, sampling rates, neighbour-sampling sizes and epoch counts are fixed by the authors or taken from the original AMLSim benchmark.
axioms (4)
  • domain assumption The standard fairness model causal graph (A → Ŷ, A → M → Ŷ, X covariates, possible UA,X confounding) correctly describes the data-generating process for AML flagging.
    Invoked throughout Section 2 and used to define NDE/NIE; domain expertise is cited as the source of the mediator/covariate partition.
  • domain assumption The natural direct effect of country on the prediction is the unfair component; the natural indirect effect mediated by transaction behaviour is the fair component.
    Stated explicitly after Equations (4)–(5) and in the introductory example of Section 1.
  • ad hoc to paper For GNN classifiers, intervening on country by rewriting the attribute on all nodes of that country while freezing the rest of the graph structure yields a valid counterfactual world for path-specific effects.
    Acknowledged simplification in Section 5; neighbourhood feedback is ignored for computational tractability.
  • ad hoc to paper Country labels recovered by string matching on bank names plus majority-currency fallback are a usable proxy for a protected KYC attribute.
    Construction procedure given in Section 3; the paper repeatedly notes that results are not expected to reflect real country patterns.
invented entities (2)
  • pseudo-country protected attribute no independent evidence
    purpose: Supplies a categorical protected feature absent from the original AMLSim release so that fairness analysis becomes possible.
    Obtained by rule-based mapping of bank names and currencies; no independent real-world KYC validation is provided.
  • scalar transaction-behaviour mediator no independent evidence
    purpose: Serves as the legitimate pathway through which country may influence AML risk without constituting direct discrimination.
    Hand-crafted log-sum of activity and volume; the paper reports that country explains only R² ≈ 0.09 of its variance, consistent with a soft mediator but still an author-defined construct.

pith-pipeline@v1.1.0-grok45 · 19254 in / 3408 out tokens · 30610 ms · 2026-07-11T08:48:11.219673+00:00 · methodology

0 comments
read the original abstract

The application of machine learning-based predictive algorithms to Anti-Money Laundering (AML) has grown rapidly, driven by the vast volume of financial transaction data available to banks. These algorithms are typically trained not only on transactional data but also on sensitive client information, which may raise fairness concerns. Despite this, AML detection systems remain largely underexplored from a fairness perspective, even though deeper analytical methods based on counterfactuals are now available. Such techniques enable the decomposition of the direct and indirect effects of potentially sensitive features on model predictions, thereby supporting the evaluation of whether their influence is acceptable from a fairness perspective. Closing this gap, we consider the synthetic IBM AMLSim transaction dataset and construct additional features of the country of an account and its average behaviour. This improves the predictive performance of diverse machine learning models, ranging from baseline decision trees to state-of-the-art graph neural networks. We assess the potential unfairness associated with these features through a counterfactual, path-specific effect analysis. This reveals that fairness violations tend to be more pronounced for models whose predictive performance benefits the most from the extended features. Such a finding highlights a concrete instance of the trade-off between predictive accuracy and fairness in AML applications, thus underscoring the urgency of a systematic fairness analysis in such critical domains.

Figures

Figures reproduced from arXiv: 2607.05101 by Adrian Cosma, Alessandro Antonucci, David Kletz, Lea Multerer, Martina Gogova, Michele Inchingolo.

Figure 1
Figure 1. Figure 1: Causal mediation example for AML flagging decisions. The protected demographic attribute influences the probability of [PITH_FULL_IMAGE:figures/full_fig_p003_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: Structural causal graph with protected attribute [PITH_FULL_IMAGE:figures/full_fig_p005_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: Counterfactual flip rates FRYˆ(US) under the intervention A ← US, stratified by factual sending country and model (XGBoost, GNN-GIN, GNN-PNA). Results are based on simulated data and are not expected to reflect country-specific patterns. Germany China Switzerland India United Kingdom France Japan Israel Italy Australia Canada Spain Russia Mexico Saudi Arabia Brazil Netherlands Belgium Austria Finland Portu… view at source ↗
Figure 4
Figure 4. Figure 4: Estimated relative total (RTE), direct (RDE), and indirect (RIE) effects with respect to the US, shown as kernel density [PITH_FULL_IMAGE:figures/full_fig_p012_4.png] view at source ↗

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Reference graph

Works this paper leans on

20 extracted references · 10 canonical work pages · 1 internal anchor

  1. [1]

    Chirag Agarwal, Himabindu Lakkaraju, and Marinka Zitnik. 2021. Towards a unified framework for fair and stable graph representation learning. InProceedings of Machine Learning Research, V ol. 161. PMLR, 2114–2124. doi:10.48550/arXiv.2102.13186

    work page internal anchor Pith review Pith/arXiv arXiv doi:10.48550/arxiv.2102.13186 2021

  2. [2]

    Erik Altman, Jovan Blanuša, Luc von Niederhäusern, Béni Egressy, Andreea Anghel, and Kubilay Atasu. 2023. Realistic synthetic financial transactions for anti-money laundering models. InAdvances in Neural Information Processing Systems 36. Curran Associates Inc., 29851 – 29874. doi:10.5555/3666122.3667422

    work page doi:10.5555/3666122.3667422 2023

  3. [3]

    April Chen, Ryan Rossi, Namyong Park, Puja Trivedi, Yu Wang, Tong Yu, Sungchul Kim, Franck Dernoncourt, and Nesreen Ahmed

  4. [4]

    doi:10.1145/3649142

    Fairness-aware graph neural networks: a survey.ACM Transactions on Knowledge Discovery from Data18, 6 (2024), 1–23. doi:10.1145/3649142

    work page doi:10.1145/3649142 2024

  5. [5]

    Silvia Chiappa. 2019. Path-specific counterfactual fairness. InProceedings of the AAAI Conference on Artificial Intelligence, V ol. 33. AAAI Press, 7801–7808. doi:10.1609/aaai.v33i01.33017801

    doi:10.1609/aaai.v33i01.33017801 2019

  6. [6]

    Béni Egressy, Luc V on Niederhäusern, Jovan Blanuša, Erik Altman, Roger Wattenhofer, and Kubilay Atasu. 2024. Provably powerful graph neural networks for directed multigraphs. InProceedings of the AAAI Conference on Artificial Intelligence, V ol. 38. AAAI Press, 11838–11846. doi:10.1609/aaai.v38i10.29069

    doi:10.1609/aaai.v38i10.29069 2024

  7. [7]

    Xiaoxiong Gu, Jingwen Yang, and Min Liu. 2025. Optimization of anti-money laundering detection models based on causal reasoning and interpretable artificial intelligence and its empirical study on financial system stability. InProceedings of the 2025 2nd International Conference on Digital Economy and Computer Science. ACM, 304–308. doi:10.1145/3785706.3785754

    work page doi:10.1145/3785706.3785754 2025

  8. [8]

    Zhimeng Guo, Jialiang Li, Teng Xiao, Yao Ma, and Suhang Wang. 2023. Towards fair graph neural networks via graph counterfactual. In Proceedings of the 32nd ACM International Conference on Information and Knowledge Management. ACM, 669–678. doi:10.1145/ 3583780.3615092

    arXiv 2023

  9. [9]

    Parameswaran Kamalaruban, Yulu Pi, Stuart Burrell, Eleanor Drage, Piotr Skalski, Jason Wong, and David Sutton. 2024. Evaluating fairness in transaction fraud models: fairness metrics, bias audits, and challenges. InProceedings of the 5th ACM International Conference on AI in Finance. ACM, 555–563. doi:10.1145/3677052.3698666

    doi:10.1145/3677052.3698666 2024

  10. [10]

    Matt Kusner, Joshua Loftus, Chris Russell, and Ricardo Silva. 2017. Counterfactual fairness. InProceedings of the 31st International Conference on Neural Information Processing Systems. Curran Associates Inc., 4069–4079. doi:10.5555/3294996.3295162

    work page doi:10.5555/3294996.3295162 2017

  11. [11]

    Shuhan Liu, Zheyun Qin, Xuan Hou, Yining Wang, Ziwen Wang, and Zhaohui Peng. 2026. Fairness-aware graph representation learning through bias disentanglement.Information and Software Technology193 (2026), 108034. doi:10.1016/j.infsof.2026.108034

    work page doi:10.1016/j.infsof.2026.108034 2026

  12. [12]

    Jing Ma, Ruocheng Guo, Mengting Wan, Longqi Yang, Aidong Zhang, and Jundong Li. 2022. Learning fair node representations with graph counterfactual fairness. InProceedings of the Fifteenth ACM International Conference on Web Search and Data Mining. ACM, 695–703. doi:10.1145/3488560.3498391

    doi:10.1145/3488560.3498391 2022

  13. [13]

    Pristly Turjo Mazumder. 2026. Explainable and fair anti-money laundering models using a reproducible SHAP framework for financial institutions.Discover Artificial Intelligence6, 1 (2026), 262. doi:10.1007/s44163-026-00944-7

    work page doi:10.1007/s44163-026-00944-7 2026

  14. [14]

    Brendan McMahan, Eider Moore, Daniel Ramage, Seth Hampson, and Blaise Agüera Arcas. 2017. Communication-efficient learning of deep networks from decentralized data. InProceedings of the 20th International Conference on Artificial Intelligence and Statistics, V ol. 54. Proceedings of Machine Learning Research, 1273–1282. doi:10.48550/arXiv.1602.05629 Proce...

    Pith/arXiv arXiv doi:10.48550/arxiv.1602.05629 2017

  15. [15]

    Soroor Motie and Bijan Raahemi. 2024. Financial fraud detection using graph neural networks: a systematic review.Expert Systems with Applications240 (2024), 122156. doi:10.1016/j.eswa.2023.122156

    doi:10.1016/j.eswa.2023.122156 2024

  16. [16]

    Sadia Nazar, Abdul Raheman, and Muhammad Anwar Ul Haq. 2024. The magnitude and consequences of money laundering.Journal of Money Laundering Control27, 5 (2024), 808–824. doi:10.1108/jmlc-09-2022-0139

    work page doi:10.1108/jmlc-09-2022-0139 2024

  17. [17]

    2009.Causality: models, reasoning, and inference(2 ed.)

    Judea Pearl. 2009.Causality: models, reasoning, and inference(2 ed.). Cambridge University Press, Cambridge

    2009

  18. [18]

    Drago Pleˇcko and Elias Bareinboim. 2024. Causal fairness analysis: a causal toolkit for fair machine learning.Foundations and Trends in Machine Learning17, 3 (2024), 304–589. doi:10.1561/2200000106

    doi:10.1561/2200000106 2024

  19. [19]

    Drago Pleˇcko and Elias Bareinboim. 2025. Fairness-accuracy trade-offs: a causal perspective. InProceedings of the AAAI Conference on Artificial Intelligence, V ol. 39. AAAI Press, 26344–26353. doi:10.1609/aaai.v39i25.34833

    work page doi:10.1609/aaai.v39i25.34833 2025

  20. [20]

    Zichong Wang, Zhibo Chu, Ronald Blanco, Zhong Chen, Shu-Ching Chen, and Wenbin Zhang. 2024. Advancing graph counterfactual fairness through fair representation learning. InMachine Learning and Knowledge Discovery in Databases. Research Track. V ol. 14947. Springer Nature Switzerland, 40–58. doi:10.1007/978-3-031-70368-3_3 Proceedings of ECAF’26. September...

    work page doi:10.1007/978-3-031-70368-3_3 2024