EN
Matching payment histories helps risk models learn from fewer labels
TRIDENT-Seq: Profile-Matched Prefix-Future Contrast for Label-Efficient Payment Risk Modeling · arXiv:2610.37365
In plain words
Payment businesses have many transaction histories, but confirmed fraud and chargeback labels are scarce, delayed, and different from one business to another. TRIDENT-Seq is a model that first learns from unlabeled histories, then adapts to each business using its risk labels.
It learns to match an earlier part of a history with a summary of what happened later, comparing histories with similar overall activity.
In offline tests across three businesses, it scored better than the same model trained from scratch on both reported risk-ranking measures at each of two label budgets.
The research
Payment businesses can collect large transaction histories while waiting months for confirmed fraud or chargeback labels, and each business has its own risk criteria. Can a model learn useful behavior from unlabeled histories without relying on easy differences in transaction volume or timing?
TRIDENT-Seq pretrains a shared transaction-history encoder by matching future behavior among similarly active histories, then adapts it to each business’s risk labels.
The central question is whether unlabeled transaction histories can help when confirmed risk labels are scarce and business-specific, without letting coarse activity differences dominate pretraining.
A history may be easy to distinguish by volume or timing alone, even when those cues do not capture the behavior that separates risky from legitimate payment use.
TRIDENT-Seq matches an earlier history with a summary of what follows among candidates with similar activity. In offline tests, it scored higher than architecture-matched scratch training on average precision and risk recall in all six business–budget settings.
The Wallet result also showed higher risk recall with 10% of labels than scratch training with all labels, under the same 0.5% traffic intervention budget.
The offline comparison uses one pretrained checkpoint, and Wallet average precision varies widely across fine-tuning runs. The results cover three businesses and their evaluated history coverage; on SME, 76.91% of test traffic has no history for the encoder to use.
What the evidence shows
- Histories, few labels
TRIDENT-Seq is a transaction foundation model for payment risk. It addresses a practical data mismatch: businesses accumulate many transactions, but confirmed fraud and chargeback labels are scarce, delayed, and business-specific.
The paper also identifies a pretraining risk: a model might distinguish histories by transaction volume, amount, or timing without learning finer patterns of payment behavior.
- Make continuations harder
The method, Prefix–Future Contrast (PFC), trains an encoder to match an older part of a transaction history with a summary of the events that follow it. It selects candidate summaries with similar coarse activity profiles, making volume, amount, and event timing less able to distinguish the correct continuation.
Activity prediction and dispersion regularization also shape the encoder output that each business later reuses.
- Compare six settings
The main offline comparison covered three Ant International businesses at 10% and 100% label budgets. Against the same encoder and risk-head architecture trained from random initialization, TRIDENT-Seq had higher average precision and higher risk recall within the highest-scoring 0.5% of traffic in all six business–budget settings. Each result is a mean across three fine-tuning runs for that setting.
- Read the Wallet evidence
Wallet illustrates the label-efficiency result: with 10% of its training labels, TRIDENT-Seq captured 27.75% of weighted test risk within the highest-scoring 0.5% of traffic, versus 10.41% for scratch training with all Wallet labels.
The paper’s ablation also found lower Wallet recall when PFC used random candidates instead of candidates matched on coarse activity. The authors interpret this as evidence for an activity shortcut, while noting that recall does not directly reveal which cues the encoder used.
- Extend, then bound
The paper tested transfer by pretraining on one business’s histories while excluding the target business’s histories. For SME adapted with 10% of its labels, Shopping-only pretraining produced 8.67% risk recall at the 0.5% intervention budget, versus 4.67% without pretraining.
A separate online evaluation found lower recorded portfolio chargeback and overall loss rates for TRIDENT-Seq than for XGB-AutoRefit during the reported period. These findings support the approach in the tested settings, with limits from history coverage, checkpoint count, and evaluation scope.
Deep
- Original paper authors
- Weiyu Han, Haoyu Jiang, Jie Sheng, Ruobing Zuo, Huilin He
- Submitted to arXiv
- License
- CC BY 4.0