Paper: arXiv 2507.08835
Authors: Harold Guéneau, Alain Celisse, Pascal Delange
Abstract
The present work tackles the money laundering detection problem. A new procedure is introduced which exploits structured time series of both qualitative and quantitative data by means of a transformer neural network. The first step of this procedure aims at learning representations of time series through contrastive learning (without any labels). The second step leverages these representations to generate a money laundering scoring of all observations. A two-thresholds approach is then introduced, which ensures a controlled false-positive rate by means of the Benjamini-Hochberg (BH) procedure. Experiments confirm that the transformer is able to produce general representations that succeed in exploiting money laundering patterns with minimal supervision from domain experts. It also illustrates the higher ability of the new procedure for detecting nonfraudsters as well as fraudsters, while keeping the false positive rate under control. This greatly contrasts with rule-based procedures or the ones based on LSTM architectures.
Complexity vs Empirical Score
- Math Complexity: 6.5/10
- Empirical Rigor: 4.0/10
- Quadrant: Lab Rats — high math complexity, limited empirical testing
Why this score: The paper introduces advanced ML techniques (transformers with contrastive learning, Benjamini-Hochberg procedure) requiring significant mathematical infrastructure, but it lacks concrete backtesting details, specific datasets, or implementation code, focusing instead on a conceptual/procedural framework.
Research Flowchart
flowchart TD
A["Research Goal"] --> B["Input Data: Cash/Transaction Time Series"]
B --> C{"Step 1: Contrastive Learning"}
C --> D["Unsupervised Representation Learning via Transformer"]
D --> E{"Step 2: Supervised Scoring"}
E --> F["Money Laundering Score Generation"]
F --> G["Benjamini-Hochberg Thresholding"]
G --> H["Key Outcomes"]
subgraph H ["Key Outcomes"]
H1["High Detection Accuracy"]
H2["Controlled False-Positive Rate"]
H3["Superior to LSTM/Baselines"]
end
Evidence Card
What the authors claim, on what data, and who has checked it. Read from the paper's own text by a language model: every quote is verbatim, nothing was re-run. How to read this card
Proposes
The paper proposes a two‑step methodology: a transformer pretrained by contrastive learning on raw transaction time series, followed by a two‑threshold classification calibrated with the Benjamini‑Hochberg procedure to control false positives.
Source sentence
... learning representations of time series through contrastive learning (without any labels). The second step leverages these representations to generate a money laundering scoring of all observations. A two‑thresholds approach is t
Assumes 6
- The dataset is strongly imbalanced with a very low proportion of fraudsters. data
Source sentence
One important specificity of our dataset is its strong imbalance regarding the (expected) very low proportion of fraudsters among all observations.
- Gaussian noise added in latent space is small enough to preserve the local structure of the data. model
Source sentence
This mechanism helps in creating ’new observations’ from the original ones by Gaussian perturbation [DT17], which yields realistic observations provided σ²
- Descriptors of normal time series do not vary over time, allowing train/test split by disjoint periods. data
Source sentence
Assuming that the relevant descriptors of normal time series (as opposed to time series of abnormal accounts) should not vary along the time, the training and test sets have been created by considering two disjoint time periods.
- Anonymization via differential privacy with ε=1 preserves enough information for modeling. data
Source sentence
For this procedure, the privacy parameter ϵ is equal to 1. After such an anonymization process, the data only reflect the generic financial activity of a set of accounts...
Show 2 more
- The transformer can learn useful representations without any labeled data. model
Source sentence
... stems from the use of contrastive learning for (pre‑)training the transformer without labeled data.
- The two‑threshold BH calibration effectively controls the false discovery rate. statistical
Source sentence
These two thresholds are calibrated by means of the Benjamini‑Hochberg multiple testing procedure, which allows for controlling the False Discovery rate (FDR) criterion and therefore avoids too many false positives.
Data
- time-series-of-transactions-money-laundering alternative public
Source sentence
The dataset has been made freely available on Kaggle dataset: https://www.kaggle.com/datasets/haroldg/time-series-of-transactions-money-laundering/ data
Costs
Not applicable
Finds 3
- The transformer‑based procedure reduces false positives dramatically compared to rule‑based strategies while maintaining detection of both fraudsters and non‑fraudsters. performance
Source sentence
It also illustrates the higher ability of the new procedure for detecting non‑fraudsters as well as fraudsters, while keeping the false positive rate under control. This greatly contrasts with rule‑based procedures...
- Experimental results on real‑life data show a consistent behavior and a large reduction in false positives relative to rule‑based systems. performance
Source sentence
Despite this limitation, the reported empirical results demonstrate a consistent behavior, reducing a lot the number of false positives compared to rule‑based strategies.
- The approach suggests even better results could be achieved on larger datasets. other
Source sentence
All of this even suggests better results on larger datasets.
Checked by others
No other paper in the archive has been found to support or contradict these findings (compared with 27 related papers). Coverage is partial, so this is neither replication nor refutation.
Code
No repository by the authors found in the paper.
Extracted 2026-11-14 · same card as JSON · A language model chose these items: check the source sentence before relying on one.
Cited References (78)
- Large-scale machine learning with stochastic gradient descent (2010)
- Understanding and improving the role of projection head in self-supervised learning (2022) — link
- A survey on self-supervised learning: Algorithms, applications, and future trends (2024)
- A comprehensive survey on contrastive learning (2024)
- A natural language processing approach for financial fraud detection (2022)
- Anti-Money Laundering in the EU: Time to get serious. CEPS Task Force Report 28 Jan 2021. (2021) — link
- Money Laundering, Proceeds of Crime and the Financing of Terrorism — link
- Proposal for a directive on the mechanisms to be put in place by the Member States for the prevention of the use of the financial system for the purposes of money laundering or terrorist financing (2024) — link
and 70 more
- Financial Fraud and Cybercrime in Wartime: An Overview of the Scientific Landscape and Insights from Countries Engaged in Military Conflict (2024) — link
- The evolution of trade-based money laundering schemes: a regulatory dialectic perspective (2023) — link
- Design of a Monitor for Detecting Money Laundering and Terrorist Financing (2016)
- Rule-based Anti-Money Laundering in Financial Intelligence Units: Experience and Vision (2020) — link
- Deep Learning and Explainable Artificial Intelligence Techniques Applied for Detecting Money Laundering–A Critical Review (2021) — link
- Statistical Methods for Fighting Financial Crimes (2010) — link
- Anti-Money Laundering Alert Optimization Using Machine Learning with Graphs (2022) — link
- DELATOR: Money Laundering Detection via Multi-Task Learning on Large Transaction Graphs (2022) — link
- A Spatio-Temporal Attention-Based GCN for Anti-money Laundering Transaction Detection (2023) — link
- Graph-Based LSTM for Anti-money Laundering: Experimenting Temporal Graph Convolutional Network with Bitcoin Data (2022) — link
- HAMLET: A Transformer Based Approach for Money Laundering Detection (2023)
- Scalable and Imbalance-Resistant Machine Learning Models for Anti-money Laundering: A Two-Layered Approach (2020) — link
- The control of the false discovery rate in multiple testing under dependency (2001) — link
- A direct approach to false discovery rates (2002)
- A cross-validation based estimation of the proportion of true null hypotheses (2010)
- A Preprocessing Scheme for High‐Cardinality Categorical Attributes in Classification and Prediction Problems (2001) — link
- Comparative Study on the Performance of Categorical Variable Encoders in Classification and Regression Tasks (2024) — link
- Finding Structure in Time (1990) — link
- A learning algorithm for continually running fully recurrent neural networks (1989)
- Long Short-Term Memory (1997)
- Unsupervised Learning of Video Representations using LSTMs (2016) — link
- Deep learning-based time series forecasting (2024) — link
- Attention Is All You Need (2017)
- Sentence-BERT: Sentence Embeddings using Siamese BERT-Networks (2019) — link
- BERT: Pre-training of Deep Bidirectional Transformers for Language Understanding (2019)
- Improving Language Understanding by Generative Pre-Training (2018) — link
- Language Models are Unsupervised Multitask Learners (2019) — link
- Position Information in Transformers: An Overview (2021) — link
- What Do Position Embeddings Learn? An Empirical Study of Pre-Trained Language Model Positional Encoding (2020) — link
- Nucleotide Transformer: building and evaluating robust foundation models for human genomics (2024) — link
- Portfolio Transformer for Attention-Based Asset Allocation (2022) — link
- An Image is Worth 16x16 Words: Transformers for Image Recognition at Scale (2021) — link
- Transformers in time series: A survey (2023)
- A comprehensive survey on applications of transformers for deep learning tasks (2024) — link
- Self-supervised Learning is More Robust to Dataset Imbalance (2022) — link
- Contrastive Representation Learning: A Framework and Review (2020) — link
- Representation Learning with Contrastive Predictive Coding (2019) — link
- Supervised Contrastive Learning (2021) — link
- A Simple Framework for Contrastive Learning of Visual Representations (2020) — link
- Momentum Contrast for Unsupervised Visual Representation Learning (2020) — link
- Investigating the Benefits of Projection Head for Representation Learning (2024) — link
- Understanding and Improving the Role of Projection Head in Self-Supervised Learning (2022) — link
- Universal Language Model Fine-tuning for Text Classification (2018) — link
- With a Little Help from My Friends: Nearest-Neighbor Contrastive Learning of Visual Representations (2021) — link
- Contrast to Divide: Self-Supervised Pre-Training for Learning with Noisy Labels (2022) — link
- Boosting Contrastive Self-Supervised Learning with False Negative Cancellation (2022) — link
- Dataset Augmentation in Feature Space (2017) — link
- Augmenting Imbalanced Time-series Data via Adversarial Perturbation in Latent Space (2021) — link
- Modern Information Retrieval: A Brief Overview (2001)
- Cross-Entropy Loss Functions: Theoretical Analysis and Applications (2023) — link
- Convex Optimization (2004)
- Transfer learning for medical image classification: a literature review (2022) — link
- RankMe: Assessing the downstream performance of pretrained self-supervised representations by their rank (2023) — link
- A Cluster Separation Measure (1979) — link
- A dendrite method for cluster analysis (1974) — link
- Information Theoretic Measures for Clusterings Comparison: Variants, Properties, Normalization and Correction for Chance (2010) — link
- Visualizing Data using t-SNE (2008) — link
- K-means clustering algorithms: A comprehensive review, variants analysis, and advances in the era of big data (2023) — link
- DELATOR: Money Laundering Detection via Multi-Task Learning on Large Transaction Graphs (2022) — link
- Symmetric matrix perturbation for differentially-private principal component analysis (2016) — link
- Diffprivlib: the IBM differential privacy library (2019)
- Controlling the False Discovery Rate: A Practical and Powerful Approach to Multiple Testing (1995) — link
- Training with Noise is Equivalent to Tikhonov Regularization (1995) — link
- Regularizing Deep Neural Networks by Noise: Its Interpretation and Optimization (2017) — link
- Learning Deep Representations Using Convolutional Auto-encoders with Symmetric Skip Connections (2017) — link
- Logistic Regression, a review (1988) — link
- Silhouettes: A graphical aid to the interpretation and validation of cluster analysis (1987) — link
- Temporal Ensembling for Semi-Supervised Learning (2017) — link
- An optimal algorithm for approximate nearest neighbor searching fixed dimensions (1998) — link
- XGBoost: A Scalable Tree Boosting System (2016) — link
0 of 78 cited works are in the corpus; resolved via arXiv source bundles, Semantic Scholar and OpenAlex. Corpus-wide: 2596 linked papers.