Fraud-label acquisition for payment transactions is both expensive and inconsistent across tasks, while high-cardinality identifiers such as merchant names make traditional one-hot feature pipelines impractical at scale. The new CARTA model addresses these problems by learning self-supervised transaction embeddings directly from raw payment data.
CARTA combines heterogeneous tokenization with a cross-attention mechanism, allowing it to represent diverse transaction fields in a unified way. Because the embeddings are reusable, they can be applied to multiple fraud-related tasks without retraining from scratch or requiring task-specific labels.
The arXiv abstract describes the motivation and architecture but does not include experimental results or comparisons with existing methods. As a single source, there are no conflicting findings to weigh; the contribution is presented as a design solution to known labeling and scalability bottlenecks.