TransactionGPT: Toward Foundational Transaction Modeling

Le résumé fourni par la source

We present TransactionGPT (TGPT), a foundation model for consumer transaction data within one of the world's largest payment networks. TGPT is designed to understand and generate transaction trajectories while simultaneously supporting a variety of downstream predictive tasks. We introduce a novel 3D-Transformer architecture specifically tailored for capturing the complex dynamics in payment transaction data. This architecture incorporates design innovations that enhance modality fusion and computational efficiency, while seamlessly enabling joint optimization with downstream objectives. Trained on billion-scale real-world transactions, TGPT significantly improves anomaly detection performance against a competitive production model and exhibits advantage over baselines in generating future transactions. We also validate the effectiveness of LLM-derived transaction field embeddings by using them as part of TGPT inputs. Given the prevalence of transaction-like data in industry, which consist of event-based records containing numerous categorical and numerical fields with highly diverse cardinalities, we anticipate that the architectural innovations and practical lessons introduced in this work will advance the development of foundation models for transaction-like data and stimulate future research in this emerging topic.

Ce résumé expose les affirmations des auteurs. BNTIC ne l’interprète pas comme une validation indépendante des résultats.