Pith. sign in

REVIEW 5 major objections 7 minor 24 references

Reinforcement Learning for Autonomous Warehouse Orchestration in SAP Logistics Execution: Redefining Supply Chain Agility

T0 review · 5 major / 7 minor · reviewed 2026-08-07 · deepseek-v4-flash

Pith's one-line read A Deep Q-Network trained on synthetic SAP Logistics Execution data can, the paper claims, autonomously allocate warehouse tasks, reroute picks, and absorb disruptions with 95% accuracy and 60% faster processing than rule-based methods.

desk verdict A plausible DQN application to SAP LE orchestration, but the evaluation conflates disruption detection with optimization quality and leaves the central claims unmeasured. read the letter →

arxiv 2506.06523 v1 pith:CHB7N2P2 submitted 2025-06-06 cs.AI

classification cs.AI
keywords reinforcementlearningDeepQ-NetworkSAPLogisticsExecutionwarehouseorchestrationtaskoptimizationsupplychainagilitysyntheticdatasetdisruptionhandling
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

This paper sets out to show that a standard deep reinforcement learning algorithm, a Deep Q-Network (DQN), can take over the job of orchestrating warehouse work inside SAP Logistics Execution (LE), the enterprise module that runs warehouses, transportation, and deliveries. The idea is to recast everyday decisions — which worker gets which order, how to reroute a pick during equipment downtime, what to prioritize during an order surge — as actions in a learning environment, with rewards tied to cutting delays and raising throughput. As evidence, it builds a synthetic 300,000-transaction SAP LE dataset with Python's faker library, injects 15,000 disruptions as ground truth, and reports that the DQN correctly optimizes 14,250 of them (95% accuracy), outscoring random forests (88%) and rule-based tools (80%) while cutting processing time by 60%. If the claim holds up, it offers a privacy-safe template for autonomous warehouse orchestration that can be developed without exposing real SAP data.

What carries the argument

The load-bearing object is the Deep Q-Network (DQN), a Q-learning agent whose value function is approximated by a deep neural network (five layers, 0.01 learning rate), which the paper pairs with an encoded warehouse state: task priorities, inventory locations, and equipment status vectorized as embeddings, one-hot encodings, and normalized processing times. The action space is the set of task assignments — allocating a worker to an order, rerouting a picking task — and the reward function is shaped around efficiency gains such as reducing delays and improving throughput. The supporting fixture is the synthetic dataset: 300,000 SAP LE transactions generated with Python's faker library from SAP LE schema templates, with 5% injected disruptions (equipment downtime, order surges), 10% multilingual fields, and 3% missing values, where the injected disruptions serve as the ground-truth labels that define whether an optimization is correct.

What would settle it

Run the same DQN pipeline on real SAP LE transaction logs, or on an independently built warehouse simulation whose ground truth is measured operational cost, and check two things: whether accuracy on disrupted tasks stays near 95%, and whether end-to-end warehouse cycle time actually drops by 60% when the agent dispatches work. If the agent merely reproduces injected labels, or the speedup turns out to be a comparison of model compute time rather than warehouse processing time, the paper's central claim is refuted.

Watch

Extended reading notes

Core claim

The central claim is that a DQN agent embedded in the SAP LE workflow can autonomously orchestrate warehouse tasks in real time: it encodes warehouse state (task priority, inventory location, equipment status) as vectors, chooses task assignments from an action space, and learns a policy whose reward reflects efficiency gains. Tested on the 300,000-transaction synthetic dataset with 5% injected disruptions, the agent achieves 95% task-optimization accuracy — meaning it correctly optimized 14,250 of the 15,000 disrupted tasks — with 0.96 precision, 0.94 recall, 0.95 F1-score, and a 0.98 ROC AUC. The paper reports that this beats random forests (88% accuracy, 1,800 disruptions missed) and SAP's rule-based tools (80%), reduces processing time by 60% (70 minutes on TensorFlow versus 20 hours for rule-based processing), and saves $1 million in simulated operational costs. Accuracy holds at 95% even at full 900-field schema complexity, though Spanish-language fields cost 4 points of recall and task-priority features account for 60% of errors.

Load-bearing premise

The entire result depends on the assumption that the synthetic dataset generated by Python's faker library from SAP LE schema templates — and especially its injected disruption labels, which are treated as ground truth for what a correct optimization is — faithfully represents how real SAP Logistics Execution operations behave, since the paper provides no comparison to real SAP data and no production validation.

Editorial extensions

If this is right

  • Adaptive learning beats static dispatch: if the 95%-versus-80% margin is real, rule-based scheduling in SAP LE is the bottleneck, and replacing it with a learned policy is the direct path to the claimed efficiency gain.
  • The 88%-to-95% gap over random forests implies the advantage comes from sequential, state-dependent decision-making rather than from better static feature classification.
  • Graceful degradation under schema complexity (97% accuracy at 100 fields, 95% at 900 fields) suggests the approach can scale to SAP LE's full 900-field data structures without reworking the core algorithm.
  • The two identified failure patterns — multilingual fields cutting recall by 4 points and task-priority features causing 60% of errors — give concrete next targets: NLP preprocessing for Spanish fields and re-encoding task priority in the state vector.
  • Integration with SAP BTP and HANA is the claimed deployment path, meaning the framework is designed to run where the data lives rather than exporting warehouse data to a separate ML stack.

Reading between the lines

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

  • The reported 95% accuracy measures agreement with the dataset's own injected disruption labels, so the metric is self-referential: it shows the agent learned the synthetic ground truth, not that it reduced real warehouse cost. A production-minded test would replace those labels with a cost-based reward built from labor minutes, delay penalties, and throughput.
  • The 60% processing-time reduction is ambiguous as reported: the paper's time comparison is between DQN training and inference (70 minutes) and rule-based processing (20 hours), which is compute time, not warehouse cycle time from order arrival to completion. These two quantities should not be conflated in a deployment decision.
  • A direct next experiment would re-run the identical pipeline on real SAP LE transaction logs or a validated warehouse simulator to see whether the accuracy ordering 95/88/80 survives when ground truth is actual operational cost rather than injected labels.
  • Nothing in the DQN formulation is specific to SAP LE, so the same state-action-reward template — task queues, resource availability, disruptions — should transfer to neighboring problems such as transportation dispatch, manufacturing job-shop scheduling, or any ERP module with task-allocation workflows.
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

5 major / 7 minor

Summary. The paper proposes a Deep Q-Network (DQN) framework for autonomous task orchestration in SAP Logistics Execution (LE). It constructs a synthetic dataset of 300,000 LE transactions with 5% injected disruptions, trains a DQN, and reports 95% task-optimization accuracy, 0.96 precision, 0.94 recall, 0.95 F1, an ROC AUC of 0.98, and a 60% processing-time reduction relative to rule-based methods. The central claim is that RL can autonomously orchestrate SAP LE warehouse tasks and substantially outperform existing rule-based approaches.

Significance. If the central claims were established, the paper would describe a practically important application: applying deep RL to SAP LE warehouse orchestration at scale, with tangible efficiency gains. The paper is transparent that its data are synthetic and it gives some implementation details (TensorFlow, a five-layer DQN, a 0.01 learning rate, and an 80/20 split). However, the significance cannot be assessed from the current manuscript because the evaluation does not define the prediction objective, the environment, the reward, or the comparison protocol. No machine-checked proofs, reproducible code, or parameter-free derivations are supplied; the reported numbers are therefore not verifiable.

major comments (5)
  1. [Section V, Table 1] The central metric 'accuracy (correct optimizations)' is never defined. The manuscript does not specify what constitutes a correct optimization, how the DQN's chosen actions are scored, or what ground-truth labels beyond the injected disruption flags are used. As written, 'optimizing 14,250 of 15,000 disrupted tasks' merely restates 95% of the injected disruptions and does not establish that the DQN improved warehouse operations. This is load-bearing because the abstract and conclusion rest on this number.
  2. [Section II; Section V] The reward function is not written down. The paper states only that it 'rewards actions that reduce delays or improve throughput,' and then Section V and Figure 5 report a 60% processing-time reduction. Measuring success with the same objective used to train the policy is not an independent evaluation: a policy trained to minimize a proxy will trivially appear to improve that proxy on its training distribution. The manuscript needs an explicit reward definition, a separate evaluation environment, and a precise account of how processing time is computed for both the DQN and the baselines.
  3. [Section IV-A] The synthetic dataset is generated with Python's faker library and SAP LE schema templates, but the paper provides no evidence that the generated records reproduce the temporal and operational dynamics of real SAP LE warehouses. It also does not define how disruptions are injected or why the injected labels should serve as ground truth. Without validation against real SAP data or a justified simulator, the reported accuracy and processing-time gains have unknown external validity for actual SAP deployments.
  4. [Section V, Figures 3-4] The evaluation conflates RL policy learning with binary classification. Precision, recall, F1, and the ROC curve treat the model as a classifier that flags disruptions, whereas a DQN selects actions in an MDP. There is no description of how action-level decisions are converted to binary 'optimized/disrupted' decisions, nor of the decision threshold used. Consequently, the reported precision/recall/AUC numbers do not measure task-orchestration quality.
  5. [Sections III and V] Several reported comparisons are internally inconsistent. Rule-based systems are given 85% accuracy in Section III but 80% in Section V; random forests are 90% in Section III but 88% in Section V; and the Introduction says the framework 'surpasses rule-based systems by 40%,' while the actual difference is 15 percentage points. The cross-validation protocol is also unclear: Section IV-A states an 80/20 train-test split with 5-fold cross-validation, but no fold-level results, error bars, or confidence intervals are reported.
minor comments (7)
  1. [Section IV-A and Section VII-C] The statements about data availability are inconsistent: Section IV-A cites [zenodo.org/sample-le], while Section VII-C says data will be made available on reasonable request; no valid repository identifier is provided.
  2. [Section IV-B] The hyperparameter search is described only as 'GridSearchCV' with final values of five layers and a 0.01 learning rate; the search range and selection criterion are missing.
  3. [Section V] Please provide the standard deviation or confidence interval for the 5-fold cross-validation results; Table 1 reports only point estimates.
  4. [References] Many references, especially those on credit-card fraud detection and blockchain, are not shown to be methodologically relevant to SAP LE warehouse orchestration; the related-work section should instead engage with RL-for-logistics and warehouse-management literature.
  5. [Section II] The state space, action space, and reward function are described only in prose; at least one equations block with formal definitions is needed.
  6. [Figure 2 caption] The caption says red zones indicate 'major contributors to model errors,' while the body text says the heatmap shows 'disruption frequencies'; the caption and the text should be aligned.
  7. [Section VII-B] The declaration contains a typo: 'no known competing for financial interests' should read 'no known competing financial interests.'

Circularity Check

3 steps flagged · score 7.0 of 10

The two central numbers restate their own inputs: '95% accuracy' is simply 14,250/15,000 disrupted tasks with no independent definition of correctness, and the '60% processing-time reduction' is the reward objective relabeled as an evaluation result.

  1. self definitional [Section V, 'Experimental Analysis', first paragraph (metrics and first result)]
    "Metrics included accuracy (correct optimizations), precision (true positives among flagged tasks), recall (disruption coverage), and F1-score [2]. ... The DQN model achieved 95% accuracy, optimizing 14,250 of 15,000 disrupted tasks [5]."

    Accuracy is defined only as 'correct optimizations', but no operational definition of a correct optimization (cost function, optimal assignment, or validated simulator) is given anywhere in the paper. The reported numerator and denominator, 14,250 of 15,000 disrupted tasks, equal exactly 95%. The headline accuracy is therefore the proportion of injected disruptions relabeled as 'optimized', not a measured quality of the DQN's orchestration decisions. The central claim reduces by construction to the rate at which the author counts disruptions as handled; there is no independent yardstick being predicted.

  2. fitted input called prediction [Section II (reward design) and Section V (main result)]
    "The action space is defined to include task assignments, such as allocating a specific worker to an order or rerouting a picking task, while the reward function is designed to reflect efficiency gains, rewarding actions that reduce delays or improve throughput [13]. ... DQN reduced processing times by 60%, saving $1 million [14]."

    The headline outcome 'reduced processing times by 60%' is the same quantity that the reward function was explicitly constructed to maximize ('reduce delays or improve throughput'). The evaluation is not an external benchmark of orchestration quality: training optimizes this reward, and the result reports improvement on it. No independent cost model, real warehouse data, or held-out operational objective is used to validate the 60% figure. The reported efficiency gain is the training objective renamed as a result, so the claim is fitted input called prediction.

1 more flagged steps
  1. renaming known result [Section V, 'Experimental Analysis', metrics and ROC paragraph]
    "Metrics included accuracy (correct optimizations), precision (true positives among flagged tasks), recall (disruption coverage), and F1-score [2]. ... The ROC curve (Figure 4) achieved an AUC of 0.98, indicating strong discrimination [13]."

    Precision, recall, and ROC are binary-classification metrics over flagged disruptions, but the method is a DQN choosing task-assignment actions. No mapping from Q-values or actions to predicted binary classes or to a 'correct optimization' label is specified. The evaluation therefore relabels a binary disruption-detection exercise as 'task optimization accuracy' and 'optimized vs disrupted' discrimination. The claimed result is numerically equivalent to a classification label count, not to the RL orchestration policy's action quality; the rename supplies the appearance of a measured outcome without a derivation.

full rationale

The paper's proposed derivation chain — an RL agent trained on a synthetic SAP LE dataset yielding autonomous warehouse orchestration — is not supported by an external benchmark. The circularity is concentrated in the evaluation. In Section V, 'accuracy (correct optimizations)' is immediately instantiated as '14,250 of 15,000 disrupted tasks', which is exactly 95%; without a cost function or ground-truth optimal assignments, accuracy is just the rate at which injected disruptions are relabeled as optimized. Similarly, Section II defines the reward to reduce delays and improve throughput, and Section V reports a 60% processing-time reduction as the main success metric; this is the training objective reported as an independent outcome. The confusion-matrix and ROC metrics additionally presuppose binary labels, while the claimed method is a DQN over task-assignment actions, so the reported numbers cannot be measuring orchestration quality. The synthetic faker-generated dataset is a validity concern, but the internal circularity is independent of data realism: even with real SAP data, the two headline quantities would still reduce to definitions. Self-citations to prior preprints are numerous but not load-bearing for the numeric derivation, so no score of 8 or higher from a self-citation chain is warranted. The central quantitative claims, however, are substantially self-referential, giving a circularity score of 7.

Assumptions & free parameters 5 free parameters · 4 assumptions · 0 invented entities

The central claim rests on a synthetic dataset, author-defined disruption labels, and an unverified integration story. The free parameters are either tuned on the same data used for evaluation or chosen by hand with no robustness checks. The axioms are domain assumptions that the simulated environment is a valid stand-in for real SAP LE operations. No new entities are introduced, but the 'efficiency_score' metric is defined without external validation.

free parameters (5)
  • DQN learning rate = 0.01
    Tuned via GridSearchCV on the synthetic training set; the paper does not report sensitivity to this value.
  • DQN hidden layer count = 5
    Chosen by hyperparameter search; no ablation or robustness analysis is provided.
  • Disruption injection rate = 5% of 300,000 transactions
    The 15,000 disrupted tasks are the ground truth used to compute the 95% accuracy, so the evaluation target is a modeling choice made by the author.
  • Outlier cap percentile = 99th percentile
    Used to cap 1,000-unit order batches; the choice of cap can change the reported performance and is not justified.
  • Correlation redundancy threshold = r > 0.8
    Used to remove redundant fields; no justification or robustness check is given for the threshold.
assumptions (4)
  • domain assumption Python faker-generated synthetic transactions capture real SAP LE distributions and disruption patterns.
    Section IV-A states the dataset was created with faker and SAP LE schema templates, but no comparison to real SAP LE logs or production data is presented.
  • domain assumption The DQN converges to a policy whose learned rewards correspond to real operational efficiency.
    Sections II and IV-B assert that DQN learns optimal task sequences, but there is no convergence proof, real deployment, or human evaluation.
  • ad hoc to paper Ground-truth labels for 'correct optimizations' are well-defined.
    Section V says 14,250 of 15,000 disrupted tasks were optimized, but the paper never defines what makes a task optimization correct; the labels come from the synthetic setup.
  • domain assumption SAP BTP and SAP HANA integration works as described without a real system test.
    Sections II and IV claim seamless integration and real-time processing, but no API traces, pilot deployment, or performance measurements on SAP infrastructure are provided.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Reinforcement Learning for Autonomous Warehouse Orchestration in SAP Logistics Execution: Redefining Supply Chain Agility." pith.science (2026). https://pith.science/paper/CHB7N2P2

@misc{pith2026250606523,
  author       = {Pith},
  title        = {Pith review of: Reinforcement Learning for Autonomous Warehouse Orchestration in SAP Logistics Execution: Redefining Supply Chain Agility},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/CHB7N2P2}},
  note         = {Machine review of arXiv:2506.06523}
}
read the original abstract

In an era of escalating supply chain demands, SAP Logistics Execution (LE) is pivotal for managing warehouse operations, transportation, and delivery. This research introduces a pioneering framework leveraging reinforcement learning (RL) to autonomously orchestrate warehouse tasks in SAP LE, enhancing operational agility and efficiency. By modeling warehouse processes as dynamic environments, the framework optimizes task allocation, inventory movement, and order picking in real-time. A synthetic dataset of 300,000 LE transactions simulates real-world warehouse scenarios, including multilingual data and operational disruptions. The analysis achieves 95% task optimization accuracy, reducing processing times by 60% compared to traditional methods. Visualizations, including efficiency heatmaps and performance graphs, guide agile warehouse strategies. This approach tackles data privacy, scalability, and SAP integration, offering a transformative solution for modern supply chains.

Figures

Figures reproduced from arXiv: 2506.06523 by the authors.

Figure 1
Figure 1. Workflow Diagram This diagram illustrates the end-to-end RL pipeline used for task orchestration. It begins with raw Logistics Execution (LE) transactions, followed by preprocessing and state encoding. A Deep Q-Network (DQN) is then applied for policy learning and action optimization. The framework integrates with SAP BTP for orchestration and SAP HANA for real-time execution, supporting intelligent and adaptive war… view at source ↗
Figure 2
Figure 2. ’s heatmap [9]. Task priority features caused 60% of errors, addressed via state embeddings [10] [PITH_FULL_IMAGE:figures/full_fig_p004_2.png] view at source ↗
Figure 4
Figure 4. ROC Curve This color ROC curve evaluates the DQN model’s performance, achieving an AUC of 0.98 (blue), compared to random forests (green) and rule-based tools (red). The curve plots true positive rates against false positive rates, demonstrating DQN’s superior discrimination power in task optimization. The high AUC indicates excellent ability to distinguish between optimized and disrupted tasks, even with disruption… view at source ↗
Figures from the paper (1 more)
Figure 5
Figure 5. Figure 5: Processing Time Reduction Bar Graph [PITH_FULL_IMAGE:figures/full_fig_p005_5.png]

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

24 extracted references · 19 canonical work pages

  1. [1]

    SAP Logistics Execution Guide,

    SAP SE, “SAP Logistics Execution Guide,” SAP Press, 2023

  2. [2]

    SmartSync: Machine Learning for Seamless SAP RAR Data Migration,

    G. Bandarupalli and V. Kanaparthi, “SmartSync: Machine Learning for Seamless SAP RAR Data Migration,” Research Square, Apr. 2025, doi: 10.21203/rs.3.rs-6459008/v1. 6

  3. [3]

    Warehouse Management Costs,

    Gartner, “Warehouse Management Costs,” Gartner Research, 2022

  4. [4]

    Efficient Deep Neural Network for Intrusion Detection Using CIC-IDS-2017 Dataset,

    G. Bandarupalli, “Efficient Deep Neural Network for Intrusion Detection Using CIC-IDS-2017 Dataset,” Nov. 2024, doi: 10.21203/RS.3.RS-5424062/V1

  5. [6]

    Industry 4.0 Standards,

    UN, “Industry 4.0 Standards,” 2022, doi: 10.1002/9781119376897

  6. [7]

    Advancing Smart Transportation via AI for Sustainable Traffic Solutions in Saudi Arabia,

    G. Bandarupalli, “Advancing Smart Transportation via AI for Sustainable Traffic Solutions in Saudi Arabia,” Nov. 2024, doi: 10.21203/RS.3.RS-5389235/V1

  7. [8]

    Enhancing Microservices Performance with AI-Based Load Balancing: A Deep Learning Perspective,

    G. Bandarupalli, “Enhancing Microservices Performance with AI-Based Load Balancing: A Deep Learning Perspective,” Apr. 2025, https://doi.org/10.21203/rs.3.rs- 6396660/v1

  8. [9]

    Enhancing Sentiment Analysis in Multilingual Social Media Data Using Transformer-Based NLP Models: A Synthetic Computational Study,

    G. Bandarupalli, “Enhancing Sentiment Analysis in Multilingual Social Media Data Using Transformer-Based NLP Models: A Synthetic Computational Study,” TechRxiv, Apr. 11, 2025. [Online]. Available: https://doi.org/10.36227/techrxiv.174440282.23013172/v1

Show all 24 references
  1. [10]

    AI-Driven Code Refactoring: Using Graph Neural Networks to Enhance Software Maintainability,

    G. Bandarupalli, “AI-Driven Code Refactoring: Using Graph Neural Networks to Enhance Software Maintainability,” arXiv, 2025. [Online]. Available: https://arxiv.org/abs/2504.10412

  2. [11]

    Code Reborn: AI-Driven Legacy Systems Modernization from COBOL to Java,

    G. Bandarupalli, “Code Reborn: AI-Driven Legacy Systems Modernization from COBOL to Java,” arXiv, 2025. [Online]. Available: https://arxiv.org/abs/2504.11335

  3. [12]

    Machine Learning-Driven Analysis of the Economic Impact of Current U.S. Trade Tariffs on Global Supply Chains,

    G. Bandarupalli, “Machine Learning-Driven Analysis of the Economic Impact of Current U.S. Trade Tariffs on Global Supply Chains,” TechRxiv, Apr. 17, 2025. [Online]. Available: https://doi.org/10.36227/techrxiv.174490701.17481632/v1

  4. [13]

    Warehouse Optimization Challenges,

    Deloitte, “Warehouse Optimization Challenges,” Deloitte Insights, 2023

  5. [14]

    Enhancing Predictive Models,

    S. Naz and G. S. Kashyap, “Enhancing Predictive Models,” Int. J. Inf. Technol., 2024, doi: 10.1007/S41870-023-01721- W

  6. [15]

    AI in Agriculture,

    G. S. Kashyap et al., “AI in Agriculture,” Research Square, Feb. 2024, doi: 10.21203/RS.3.RS-3984385/V1

  7. [16]

    Large Language Models’ Versatility,

    P. Kaur et al., “Large Language Models’ Versatility,” arXiv, Feb. 2024, doi: 10.48550/arXiv.2402.16142

  8. [17]

    Multi-Robot Exploration,

    G. S. Kashyap et al., “Multi-Robot Exploration,” arXiv, Nov. 2023, doi: 10.48550/arXiv.2311.16958

  9. [18]

    AI in Digital Finance,

    V. Kanaparthi, “AI in Digital Finance,” arXiv, Jan. 2024, doi: 10.48550/arXiv.2401.15700

  10. [19]

    Blockchain in Accounting,

    T. Garanina et al., “Blockchain in Accounting,” Emerald Group Holdings Ltd., Aug. 2022, doi: 10.1108/AAAJ-10- 2020-4991

  11. [20]

    Credit Card Fraud Detection,

    A. Dal Pozzolo et al., “Credit Card Fraud Detection,” in Proc. Int. Joint Conf. Neural Networks, 2015, doi: 10.1109/IJCNN.2015.7280527

  12. [21]

    Credit Card Fraud Detection,

    D. Varmedja et al., “Credit Card Fraud Detection,” in Proc. 18th Int. Symp. INFOTEH-JAHORINA, 2019, doi: 10.1109/INFOTEH.2019.8717766

  13. [22]

    Credit Card Fraud Detection with GA,

    E. Ileberi et al., “Credit Card Fraud Detection with GA,” J. Big Data, vol. 9, no. 1, pp. 1–17, 2022, doi: 10.1186/s40537- 022-00573-8

  14. [23]

    Neural Network Ensemble for Fraud Detection,

    E. Esenogho et al., “Neural Network Ensemble for Fraud Detection,” IEEE Access, vol. 10, pp. 16400–16407, 2022, doi: 10.1109/ACCESS.2022.3148298

  15. [24]

    Streaming Credit Card Fraud Detection,

    F. Carcillo et al., “Streaming Credit Card Fraud Detection,” Inf. Fusion, vol. 41, pp. 182–194, 2018, doi: 10.1016/j.inffus.2017.09.005

  16. [25]

    Credit Card Fraud Detection,

    V. N. Dornadula and S. Geetha, “Credit Card Fraud Detection,” Procedia Computer Science, 2019, doi: 10.1016/j.procs.2020.01.057

Pith tools

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