A human trafficking detection model finds bank customers who may be laundering money from human trafficking (here, sex trafficking). It uses the bank's own data and combines NLP with other ML techniques.

Scenario

Human trafficking is estimated to be a 150 million USD business each year, and some of this money is laundered through major banks. In the country where the model was built, people commonly send money by email address, called an Email Money Transfer (EMT), and each transfer can carry a short transaction memo.

Structure

The model has 3 branches that each produce a score, and the scores are combined into one final risk score:

graph TD
    A[Email address, memo, KYC info] -->|BERT embedding| B[Client segmentation: Isolation Forest]
    C[Sequence of transactions] -->|input embedding| D[Transformer model]
    D -->|output embedding| E[Transaction segmentation: UMAP + HDBSCAN]
    F[Network of transactions] -->|graph creation| G[Graph score: graph algorithms]
    B --> H[Final risk score]
    E --> H
    G --> H
BranchInputMethod
Client segmentationEmail address, transaction memo, KYC information (for example the occupation of sender and receiver)All text is combined into 1 embedding by pre-trained BERT (768 dimensions), then grouped with Isolation Forest
Transaction segmentationA sequence of transactionsA Transformer turns the sequence into an output embedding, then UMAP + HDBSCAN cluster it into transaction segments
Graph scoreThe network of transactions between accountsGraph algorithms on the transaction graph

Key idea

A Transformer usually reads a sequence of words. Here it reads a sequence of transactions instead, and its output is a transformed embedding of that sequence.

Wording to expect

The course says Isolation Forest is used to cluster the client embeddings. Isolation Forest is normally an anomaly detection method, so read "client segmentation" here as separating unusual clients from normal ones.

Full names to recognise:

  • UMAP: Uniform Manifold Approximation and Projection
  • HDBSCAN: Hierarchical Density-Based Spatial Clustering of Applications with Noise
  • KYC: Know Your Customer

Downstream process

graph TD
    A[Risk score] -->|high risk customers| B[Level 1 reviewer]
    B -->|second review| C[Level 2 reviewer]
    C -->|highly suspicious customers| D[Off-boarding, authorities notified]
  • both levels of reviewers come from the bank's special investigations unit
  • highly suspicious customers can be off-boarded, and extreme cases are reported to the authorities

Remembering tip

Text, Time, Ties: text goes to BERT and Isolation Forest, time-ordered transactions go to the Transformer and UMAP + HDBSCAN, and ties between accounts go to the graph score.

The next application uses ChatGPT and a chain of prompts instead of a trained model, in Content Moderation with ChatGPT.