Pith. sign in

REVIEW 4 major objections 5 minor 34 references

Amatriciana: Exploiting Temporal GNNs for Robust and Efficient Money Laundering Detection

T0 review · 4 major / 5 minor · reviewed 2026-08-07 · deepseek-v4-flash

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

Pith's one-line read Amatriciana is a graph neural network that detects money launderers from the full transaction graph, using temporal and edge information to cut false positives by 55 percent compared with state-of-the-art baselines.

desk verdict Solid temporal-GNN AML paper with a credible FP reduction, but the comparison is missing a feature-parity ablation that determines whether the architecture or the engineered inputs drive the gain. read the letter →

arxiv 2506.00654 v1 pith:YB4Q7OVM submitted 2025-05-31 cs.CR cs.LG

Marco Di Gennaro , Francesco Panebianco , Marco Pianta , Stefano Zanero , Michele Carminati This is my paper
classification cs.CRcs.LG
keywords anti-moneylaunderinggraphneuralnetworkstemporallearningnodeclassificationclassimbalanceMatthewscorrelationcoefficientLSTMaggregationfalsepositivereduction
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 aims to show that a GraphSAGE-style GNN, modified to consume the whole transaction graph rather than time-sliced subgraphs and to mix in edge features and an LSTM over time steps, can detect launderer accounts at higher precision than existing GNN baselines. On the IT-AML synthetic dataset it reports F1 0.76, precision 0.81, recall 0.72, and AUC 0.82, with false positives 55 percent lower than the comparison models. The authors argue this matters because money-laundering detection in practice is bottlenecked by false positives, and current temporal GNNs either distort topology or ignore time and edge information. They also propose a weighted Matthews-correlation-coefficient loss to rebalance training and a two-phase memory-efficient training scheme. If correct, the framework would be a practical inductive detector that improves precision without fragmenting the transaction graph.

What carries the argument

The carrying mechanism is an encoder pair. A Node Encoder averages sampled neighbor node and edge messages and applies a weight matrix to generate initial embeddings; a Temporal Encoder groups those messages by discrete hourly time steps, averages within each step, and passes the resulting sequence through an LSTM to produce time-aware embeddings. Skip connections concatenate intermediate embeddings into a fully connected classifier. Training alternates between updating the LSTM weights on batches and then updating the Node Encoder and classifier with frozen LSTM weights, a two-phase loop that keeps memory bounded and avoids splitting the graph into time-based subgraphs. A weighted MCC loss replaces cross-entropy to focus updates on minority-class mistakes.

What would settle it

Retrain GraphSAGE or the GCN baseline on the same engineered node-feature set Amatriciana uses, with comparable parameter counts, and measure precision and F1 on the same IT-AML test split; if it matches or beats Amatriciana, the architecture's claimed advantage is unsupported. Alternatively, ablate the engineered features from Amatriciana and check whether the false-positive reduction disappears.

Watch

Extended reading notes

Core claim

On the paper's own terms, the central discovery is that incorporating transaction-level edge features and temporal ordering into a GraphSAGE-style inductive GNN improves launderer detection enough to beat three baselines on IT-AML: Amatriciana reaches precision 0.8093, recall 0.7153, accuracy 0.7734, F1 0.76, and AUC 0.82, while EvolveGCN, GCN with Focal Loss, and GraphSAGE score lower F1 and precision. The architecture keeps the full graph intact by feeding time-step-grouped messages into an LSTM aggregator, and the custom MCC-derived loss weights false positives and false negatives to counter the 0.61 percent launderer imbalance. The authors interpret the precision gain as evidence that temporal and edge-aware embeddings plus the weighted loss are what reduce false positives.

Load-bearing premise

The comparison is fair: the paper credits the false-positive reduction to Amatriciana's architecture and loss, but the baselines are not shown to have received the same hand-built node features, so the extra inputs rather than the temporal or edge mechanism could explain the gains.

Editorial extensions

If this is right

  • If Amatriciana holds up, an inductive anti-money-laundering detector can track temporal laundering patterns without slicing the transaction graph, avoiding the topology distortion that time-subgraph methods like EvolveGCN incur.
  • The false-positive reduction (precision 0.81 versus 0.67 for GraphSAGE) would mean fewer innocent accounts flagged for investigation per true launderer caught, directly addressing the cost bottleneck of AML alerts.
  • The weighted MCC loss matches cross-entropy on F1 while letting operators tilt the model toward fewer false positives, giving a tunable control knob for regulatory priorities.
  • The two-phase training loop decouples LSTM updates from node-encoder updates, so training can run on the whole graph rather than fragments, which could scale to larger national-level transaction graphs.

Reading between the lines

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

  • The reported false-positive reduction may owe as much to the engineered node features (transaction statistics, centralities, PageRank, clustering) as to the temporal and edge-aware architecture; a feature-ablation study would separate the two.
  • The temporal LSTM with hourly discretization is only as meaningful as the discretization granularity, which the paper itself flags as future work, so the architecture may be sensitive to bin size on other data.
  • A stricter test would be cross-generator transfer or real-world bank data, since synthetic IT-AML labels may embed generator-specific patterns that inflate precision estimates.
  • The weighted MCC loss is not graph-specific and could be transferred directly to other imbalanced tabular or sequence detection tasks.
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

4 major / 5 minor

Summary. The paper presents Amatriciana, a Graph Neural Network-based detector for money laundering that augments GraphSAGE with edge features, an LSTM-based temporal encoder operating on the full transaction graph, and a weighted Matthews Correlation Coefficient (MCC) loss. The authors evaluate on the synthetic IT-AML HI_SMALL dataset, reporting that Amatriciana achieves an F1 score of 0.76, an AUC of 0.82, and a 55% reduction in false positives relative to EvolveGCN, GCN with Focal Loss, and GraphSAGE. A second experiment compares Amatriciana trained with different loss functions, showing that the weighted MCC loss performs comparably to cross-entropy while allowing tunable control over false positives.

Significance. If the reported results hold under controlled comparison, Amatriciana would be a useful contribution to AML detection: it avoids graph fragmentation by processing the full temporal graph, incorporates edge features and temporal information in an inductive framework, and proposes a loss function adapted to class imbalance. The use of a public dataset (IT-AML) and multiple baselines is a strength, and the arithmetic behind the 55% false-positive reduction is plausible from the precision values in Table I. However, the central comparison is currently not sufficiently controlled to attribute the performance gain to the architecture: the paper does not establish that the baselines received the same engineered node features, and it lacks ablations isolating the contributions of the temporal encoder, edge features, and enriched inputs. Unsupported claims about learning from limited data and memory efficiency also weaken the presentation.

major comments (4)
  1. [Section V-A and Table I] The paper does not establish feature parity between Amatriciana and the baselines. Section V-A describes four categories of engineered node features (transactions-related, topological, clustering, and PageRank) that are added to the training and validation graphs, but Section VI-B only states that models are 'configured comparably' and that GraphSAGE's adjacency matrix sums transaction values. It is never stated whether EvolveGCN, GCN with Focal Loss, and GraphSAGE received these same enriched node features. Because the main claim of a 55% false-positive reduction is derived from the precision advantage in Table I, an uncontrolled difference in input features could explain part or all of the gap. The authors should either explicitly state that all baselines used the identical feature vectors, or run an ablation in which the baselines are given the same engineered features as Amatriciana.
  2. [Section VI-B and Table I] There is no ablation isolating the architectural components that are claimed to cause the improvement. Table I compares the full Amatriciana model against baselines, and Table II varies the loss function while keeping the rest of the architecture fixed, but no experiment removes the temporal encoder or the edge-feature aggregation from Amatriciana, and no experiment adds these components to a baseline. Consequently, the precision gain cannot be attributed specifically to the temporal/edge-aware machinery as opposed to the extra feature engineering or the loss function. An ablation study that progressively adds edge features, the LSTM temporal encoder, and the engineered node features would be needed to support the attribution in the abstract and conclusion.
  3. [Section VI-A and Table I] The evaluation protocol mixes different adjustments across models, which makes the comparison non-standard. The text says 'All metrics are adjusted for class imbalance' and that metrics for EvolveGCN are 'scaled by averaging True Positive, False Positive, True Negative, and False Negative values across test subgraph, then normalizing them to match the adjusted sample sizes used for other models.' This means the numbers in Table I may not be directly comparable. The authors should report the raw confusion-matrix counts or standard metrics computed identically on the same test graph for all models, along with variance over multiple seeds or runs, so that the claimed precision advantage can be verified.
  4. [Abstract, Section V-C, and Section VII] The claims that the model 'can learn from a limited amount of data' and that the training procedure is 'memory-efficient' are not supported by any experiment in the paper. No experiment varies the size of the training set, and no runtime or memory measurements are reported despite the word 'efficient' in the title and the detailed description of the batched training algorithm in Section V-C. These claims should either be substantiated with dedicated experiments or removed from the abstract and introduction.
minor comments (5)
  1. [Section VI-C] The sentence 'As shown in Table I, the Cross-Entropy and Weighted MCC losses achieve the best results' appears to refer to Table II, which contains the loss-function comparison.
  2. [Section V-D, Equation (1)] Equation (1) is described as a weighted MCC loss, but the weights appear only in the denominator and the equation is not explicitly defined as a loss to be minimized. Please clarify the objective (e.g., loss = 1 - MCC or loss = -MCC) and report the specific weight values used for the experiments.
  3. [Section V-A and Section VI-A] The terminology for the data splits is inconsistent: Section V-A describes 'training and validation subgraphs', while Section VI-A first mentions a 'test graph' and then splits the remaining data into training and validation. Please use consistent naming and clearly state which split is used for model selection and which for final evaluation.
  4. [Section VI-B] No random seeds, number of runs, or confidence intervals are reported for the metrics in Tables I and II, so it is impossible to assess the stability of the reported precision and F1 values.
  5. [Section VII] The conclusion contains 'outperformedSOTA' with a missing space, and the sentence 'We have also shown that our custom loss function matches the performance of existing standard loss functions while empowering the tradeoff between FPs and FNs' uses 'empowering' where 'enabling' or 'controlling' would be clearer.

Circularity Check

0 steps flagged · score 0.0 of 10

No circularity: the reported metrics are measured on a disjoint test graph with independent evaluation metrics, and the weighted-MCC loss is a tuned training objective rather than a fitted prediction target.

full rationale

The paper's derivation chain is empirical rather than definitional: raw transactions are preprocessed into a graph, node and edge features are computed separately for each split, a GraphSAGE-inspired model with an LSTM temporal encoder is trained with a weighted-MCC loss, and predictions are evaluated on a test graph that is disjoint from training and validation. The paper explicitly states: 'the test graph is disjoint from both the training and validation graphs.' Final metrics (F1, Precision, Recall, Accuracy, AUC) are computed from predictions on that held-out test graph, not from the loss function or its hyperparameters. The weighted-MCC loss weights are treated as tunable hyperparameters and are selected using validation performance, which is standard practice rather than circularity. The comparison to EvolveGCN, GCN with Focal Loss, and GraphSAGE is performed on the same public IT-AML HI_SMALL dataset, providing an external benchmark. The only overlap with prior author work is the citation of Amaretto [13] in the related-work section, and that citation is not used to justify the central architecture, loss, or experimental claims; it is not load-bearing. The skeptic concern about feature parity across baselines is a legitimate comparison-control limitation, but it does not make the derivation circular: even if the baselines lacked engineered features, the reported numbers are still measured, not derived from the model's inputs by construction. No equation in the paper reduces to its own inputs, and no fitted parameter is renamed as a prediction. Therefore the paper exhibits no significant circularity.

Assumptions & free parameters 2 free parameters · 3 assumptions · 0 invented entities

The central result is an empirical benchmark claim, so the ledger is dominated by dataset and evaluation assumptions rather than mathematical axioms. The free parameters are the MCC loss weights and the hourly time discretization; both are chosen by the authors and both can move the reported numbers. The main domain axioms are that IT-AML synthetic transactions approximate real money laundering, that the adjusted and scaled metrics are comparable across models, and that the GraphSAGE sum-weighted adjacency is a fair baseline against Amatriciana's multidimensional temporal adjacency. There are no invented entities.

free parameters (2)
  • MCC loss weights (w_TP, w_FP, w_TN, w_FN) = Not reported
    Equation 1 defines a weighted MCC loss; the paper calls the weights tunable hyperparameters but does not give the values used in Tables I and II, so the false-positive and precision results are conditional on an undisclosed setting.
  • Time-step discretization = Hourly, 384 steps for 10 days
    Section V-A converts dates into discrete hourly time steps; the conclusion lists studying how timestep discretization influences results as future work, so the temporal encoding and all results depend on this choice.
assumptions (3)
  • domain assumption The IT-AML synthetic generator produces transaction graphs whose laundering labels and patterns are realistic enough to support conclusions about real AML detection.
    Section VI-A adopts the generator; Section VII acknowledges that synthetic data is a limitation, making this an explicit load-bearing assumption.
  • ad hoc to paper Performance metrics adjusted for class imbalance and the EvolveGCN scaling and normalization are comparable across models.
    Section VI-A states metrics are adjusted and EvolveGCN values are scaled by averaging and normalizing true positives, false positives, true negatives, and false negatives; no derivation of comparability is provided.
  • ad hoc to paper Baseline models received inputs comparable to Amatriciana's, including the engineered node features.
    Section VI-B says models are configured comparably but never confirms feature parity; the claim of fewer false positives depends on this.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Amatriciana: Exploiting Temporal GNNs for Robust and Efficient Money Laundering Detection." pith.science (2026). https://pith.science/paper/YB4Q7OVM

@misc{pith2026250600654,
  author       = {Pith},
  title        = {Pith review of: Amatriciana: Exploiting Temporal GNNs for Robust and Efficient Money Laundering Detection},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/YB4Q7OVM}},
  note         = {Machine review of arXiv:2506.00654}
}
read the original abstract

Money laundering is a financial crime that poses a serious threat to financial integrity and social security. The growing number of transactions makes it necessary to use automatic tools that help law enforcement agencies detect such criminal activity. In this work, we present Amatriciana, a novel approach based on Graph Neural Networks to detect money launderers inside a graph of transactions by considering temporal information. Amatriciana uses the whole graph of transactions without splitting it into several time-based subgraphs, exploiting all relational information in the dataset. Our experiments on a public dataset reveal that the model can learn from a limited amount of data. Furthermore, when more data is available, the model outperforms other State-of-the-art approaches; in particular, Amatriciana decreases the number of False Positives (FPs) while detecting many launderers. In summary, Amatriciana achieves an F1 score of 0.76. In addition, it lowers the FPs by 55% with respect to other State-of-the-art models.

Figures

Figures reproduced from arXiv: 2506.00654 by the authors.

Figure 1
Figure 1. High-level Overview of Amatriciana Architecture [PITH_FULL_IMAGE:figures/full_fig_p004_1.png] view at source ↗
Figure 2
Figure 2. ROC Curves of Amatriciana vs State-of the-art [PITH_FULL_IMAGE:figures/full_fig_p006_2.png] view at source ↗
Figure 3
Figure 3. Training and Validation Loss Comparison. [PITH_FULL_IMAGE:figures/full_fig_p007_3.png] view at source ↗

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

34 extracted references · 28 canonical work pages

  1. [1]

    Money laundering—a global obstacle,

    B. Buchanan, “Money laundering—a global obstacle,” Research in International Business and Finance , vol. 18, no. 1, pp. 115–127, 2004. [Online]. Available: https://www.sciencedirect.com/science/article/pii/ S0275531904000029

    work page 2004

  2. [2]

    Payments statistics: 2020,

    “Payments statistics: 2020,” https://www.ecb.europa.eu/press/pr/stats/ paysec/html/ecb.pis2020~5d0ea9dfa5.en.html, accessed: 2023-08-07

    work page 2020

  3. [3]

    Application of data mining for anti- money laundering detection: A case study,

    N. A. Le Khac and M.-T. Kechadi, “Application of data mining for anti- money laundering detection: A case study,” in 2010 IEEE International Conference on Data Mining Workshops , 2010, pp. 577–584

    work page 2010

  4. [4]

    Fraud detection decision support system for indonesian financial institution,

    C. Lawrencia and W. Ce, “Fraud detection decision support system for indonesian financial institution,” in 2019 International Conference on Information Management and Technology (ICIMTech), vol. 1, 2019, pp. 389–394

    work page 2019

  5. [5]

    A new algorithm for money laundering detection based on structural similarity,

    R. Soltani, U. T. Nguyen, Y . Yang, M. Faghani, A. Yagoub, and A. An, “A new algorithm for money laundering detection based on structural similarity,” in 2016 IEEE 7th Annual Ubiquitous Computing, Electronics & Mobile Communication Conference (UEMCON) , 2016, pp. 1–7

    work page 2016

  6. [6]

    Detecting money laundering using filtering techniques: a multiple-criteria index,

    S. Yang and L. Wei, “Detecting money laundering using filtering techniques: a multiple-criteria index,” Journal of Economic Policy Reform, vol. 13, no. 2, pp. 159–178, 2010. [Online]. Available: https://doi.org/10.1080/17487871003700796

    work page doi:10.1080/17487871003700796 2010

  7. [7]

    Deep learning anomaly detection as support fraud investigation in brazilian exports and anti-money laundering,

    E. L. Paula, M. Ladeira, R. N. Carvalho, and T. Marzagão, “Deep learning anomaly detection as support fraud investigation in brazilian exports and anti-money laundering,” in 2016 15th IEEE International Conference on Machine Learning and Applications (ICMLA) , 2016, pp. 954–960

    work page 2016

  8. [8]

    Analyzing and detecting money-laundering accounts in online social networks,

    Y . Zhou, X. Wang, J. Zhang, P. Zhang, L. Liu, H. Jin, and H. Jin, “Analyzing and detecting money-laundering accounts in online social networks,” IEEE Network, vol. 32, no. 3, pp. 115–121, 2018

    work page 2018

Show all 34 references
  1. [9]

    Machine learning and sampling scheme: An empirical study of money laundering detection,

    Z. Yan and T. Peter, “Machine learning and sampling scheme: An empirical study of money laundering detection,” Computational Economics, vol. 54, no. 3, pp. 1043–1063, Oct 2019. [Online]. Available: https://doi.org/10.1007/s10614-018-9864-z

    2019 doi

  2. [10]

    Detecting money laundering transactions with machine learning,

    M. Jullum, A. Løland, R. B. Huseby, G. Ånonsen, and J. Lorentzen, “Detecting money laundering transactions with machine learning,” Journal of Money Laundering Control , vol. 23, no. 1, pp. 173–186, Jan

  3. [11]

    Intelligent anti-money laundering solution based upon novel community detection in massive transaction networks on spark,

    X. Li, X. Cao, X. Qiu, J. Zhao, and J. Zheng, “Intelligent anti-money laundering solution based upon novel community detection in massive transaction networks on spark,” in 2017 Fifth International Conference on Advanced Cloud and Big Data (CBD) , 2017, pp. 176–181

    2017

  4. [12]

    Autoregressive-based outlier algorithm to detect money laundering activities,

    K. S. and S. K., “Autoregressive-based outlier algorithm to detect money laundering activities,” Journal of Money Laundering Control , vol. 20, no. 2, pp. 190–202, Jan 2017. [Online]. Available: https://doi.org/10.1108/JMLC-07-2016-0031

    2017 doi

  5. [13]

    Amaretto: An active learning framework for money laundering detection,

    D. Labanca, L. Primerano, M. Markland-Montgomery, M. Polino, M. Carminati, and S. Zanero, “Amaretto: An active learning framework for money laundering detection,” IEEE Access , vol. 10, pp. 41 720– 41 739, 2022

    2022

  6. [14]

    Detection of money laundering groups using supervised learning in networks,

    D. Savage, Q. Wang, P. Chou, X. Zhang, and X. Yu, “Detection of money laundering groups using supervised learning in networks,” 2016

    2016

  7. [15]

    A study on gcn using focal loss on class- imbalanced bitcoin transaction for anti-money laundering detection,

    P. Humranan and S. Supratid, “A study on gcn using focal loss on class- imbalanced bitcoin transaction for anti-money laundering detection,” in 2023 International Electrical Engineering Congress (iEECON) , 2023, pp. 101–104

    2023

  8. [16]

    Semi-supervised classification with graph convolutional networks,

    T. N. Kipf and M. Welling, “Semi-supervised classification with graph convolutional networks,” 2017

    2017

  9. [17]

    Anti-money laundering in bitcoin: Experimenting with graph convolutional networks for financial foren- sics,

    M. Weber, G. Domeniconi, J. Chen, D. K. I. Weidele, C. Bellei, T. Robinson, and C. E. Leiserson, “Anti-money laundering in bitcoin: Experimenting with graph convolutional networks for financial foren- sics,” 2019

    2019

  10. [18]

    Catch me if you can: Semi-supervised graph learning for spotting money laundering,

    M. R. Karim, F. Hermsen, S. A. Chala, P. de Perthuis, and A. Mandal, “Catch me if you can: Semi-supervised graph learning for spotting money laundering,” 2023

    2023

  11. [19]

    Laundrograph: Self-supervised graph representation learning for anti-money laundering,

    M. Cardoso, P. Saleiro, and P. Bizarro, “Laundrograph: Self-supervised graph representation learning for anti-money laundering,” 2022

    2022

  12. [20]

    Inductive representation learning on large graphs,

    W. Hamilton, Z. Ying, and J. Leskovec, “Inductive representation learning on large graphs,” Advances in neural information processing systems, vol. 30, 2017

    2017

  13. [21]

    Evolvegcn: Evolving graph convolutional networks for dynamic graphs,

    A. Pareja, G. Domeniconi, J. Chen, T. Ma, T. Suzumura, H. Kanezashi, T. Kaler, T. B. Schardl, and C. E. Leiserson, “Evolvegcn: Evolving graph convolutional networks for dynamic graphs,” 2019

    2019

  14. [22]

    Long short-term memory,

    S. Hochreiter, “Long short-term memory,” Neural Computation MIT- Press, 1997

    1997

  15. [23]

    Graph neural networks: A review of methods and applications,

    J. Zhou, G. Cui, S. Hu, Z. Zhang, C. Yang, Z. Liu, L. Wang, C. Li, and M. Sun, “Graph neural networks: A review of methods and applications,” AI open, vol. 1, pp. 57–81, 2020

    2020

  16. [24]

    Graph convolutional networks: a comprehensive review,

    S. Zhang, H. Tong, J. Xu, and R. Maciejewski, “Graph convolutional networks: a comprehensive review,” Computational Social Networks , vol. 6, no. 1, pp. 1–23, 2019

    2019

  17. [25]

    Graph attention networks,

    P. Veli ˇckovi´c, G. Cucurull, A. Casanova, A. Romero, P. Lio, and Y . Ben- gio, “Graph attention networks,” arXiv preprint arXiv:1710.10903, 2017

    2017 arXiv

  18. [26]

    Scalable graph learning for anti-money laundering: A first look,

    M. Weber, J. Chen, T. Suzumura, A. Pareja, T. Ma, H. Kanezashi, T. Kaler, C. E. Leiserson, and T. B. Schardl, “Scalable graph learning for anti-money laundering: A first look,” 2018

    2018

  19. [27]

    The anatomy of a large-scale hypertextual web search engine,

    L. P. Sergey Brin, “The anatomy of a large-scale hypertextual web search engine,” 1999

    1999

  20. [28]

    Optimization of graph neural networks: Implicit acceleration by skip connections and more depth,

    K. Xu, M. Zhang, S. Jegelka, and K. Kawaguchi, “Optimization of graph neural networks: Implicit acceleration by skip connections and more depth,” in Proceedings of the 38th International Conference on Machine Learning, ser. Proceedings of Machine Learning Research, M. Meila an...

    2021

  21. [29]

    It-aml dataset, kaggle url,

    “It-aml dataset, kaggle url,” https://www.kaggle.com/datasets/ ealtman2019/ibm-transactions-for-anti-money-laundering-aml, accessed: 2023-08-07

    2023

  22. [30]

    Realistic synthetic financial transactions for anti-money laundering models,

    E. Altman, B. Egressy, J. Blanuša, and K. Atasu, “Realistic synthetic financial transactions for anti-money laundering models,” 2023

    2023

  23. [31]

    Anti-Money Laundering Datasets: InPlusLab anti-money laundering datadatasets,

    T. Suzumura and H. Kanezashi, “Anti-Money Laundering Datasets: InPlusLab anti-money laundering datadatasets,” http://github.com/IBM/ AMLSim/, 2021

    2021

  24. [32]

    M. J. Zaki and W. Meira Jr, Data mining and machine learning: fundamental concepts and algorithms . Cambridge University Press, 2020

    2020

  25. [33]

    A study of cross-validation and bootstrap for accuracy estimation and model selection,

    R. Kohavi, “A study of cross-validation and bootstrap for accuracy estimation and model selection,” inProceedings of the 14th International Joint Conference on Artificial Intelligence - Volume 2 , ser. IJCAI’95. San Francisco, CA, USA: Morgan Kaufmann Publishers Inc., 1995, p....

    1995

  26. [2020]

    Available: https://doi.org/10.1108/JMLC-07-2019-0055

    [Online]. Available: https://doi.org/10.1108/JMLC-07-2019-0055

    2019 doi

Pith tools

Reviewed August 7, 2026 · model on record in the stance chip above.