Temporal Contrastive Transformer for Financial Crime Detection: Self-Supervised Sequence Embeddings via Predictive Contrastive Coding
Researchers have developed a Temporal Contrastive Transformer (TCT) model for financial crime detection, using self-supervised learning to create embeddings from transaction sequences. While the TCT embeddings alone showed promising predictive performance with an AUC of 0.8644, they did not significantly improve upon existing domain-engineered features when combined, achieving a similar AUC of 0.9205 versus the baseline's 0.9245. This suggests the model captures relevant temporal signals but faces challenges in adding value beyond established feature engineering methods, indicating a promising but intermediate step in automated financial crime analysis. AI
IMPACT This research offers a new self-supervised approach for financial crime detection, potentially reducing reliance on manual feature engineering.