TRIDENT-Seq: Profile-Matched Prefix-Future Contrast for Label-Efficient Payment Risk Modeling

Weiyu Han; Haoyu Jiang; Jie Sheng; Ruobing Zuo; Huilin He; Ze Zhao; Xiao Tan; Nairong Zhang

cs.LGOct 27, 2026

TRIDENT-Seq: Profile-Matched Prefix-Future Contrast for Label-Efficient Payment Risk Modeling

Authors: Weiyu Han, Haoyu Jiang, Jie Sheng, Ruobing Zuo, Huilin He, Ze Zhao, Xiao Tan, Nairong Zhang,

Organizations: Ant International · East China Normal University · Tongji University · Shanghai Jiao Tong University

Abstract

Ant International accumulates large volumes of transaction history across users and payment businesses. Confirmed fraud and chargeback outcomes, by contrast, are scarce, delayed, and specific to each business. A single encoder pretrained on unlabeled histories could serve all businesses, but only if pretraining captures the behavior that separates risk. Coarse activity differences can make contrastive pretraining on transaction logs too easy: transaction volume, amount, and timing can distinguish a history's continuation from arbitrary candidates without modeling how payment resources are used. We present TRIDENT-Seq, a transaction foundation model trained with Prefix-Future Contrast (PFC). PFC matches each prefix to a summary of its subsequent behavior among candidates with similar activity profiles; activity prediction and dispersion regularization act on the encoder output. Across three production payment businesses, the pretrained encoder improves average precision (AP) and recall at a 0.5% intervention budget over architecture-matched scratch training in all six business-budget settings. With a fixed validation set, it exceeds full-label scratch training in mean AP and recall using 10% of the training labels on Wallet and 25% on Shopping and SME. Pretraining only on another business's histories still helps. Replacing profile-matched candidates with random ones cuts recall from 24.62% to 5.64%. In Ant International's deployment evaluation, recorded amount-based chargeback and overall loss rates are 37.7% and 37.9% lower than those of the previous XGBoost-based system under the same intervention budget.

Figures & tables

Figure 1. TRIDENT-Seq overview. (a) PFC learns behavioral representations without risk labels. (b) The shared foundation model adapts to each business through supervised fine-tuning (SFT) with limited labels. (c) At Ant International, online A/B evaluation across deployed payment businesses shows a 37.9% reduction in the overall loss rate under the same intervention budget. Three stages: behavior pretraining from prefixes and future summaries; reuse of the same pretrained encoder for business-specific supervised fine-tuning; and production risk evaluation.

Figure 2. TRIDENT-Seq architecture. (a) Event tokenization. (b) Shared history encoder. (c) Pretraining with PFC, activity prediction, and dispersion regularization, and business-specific supervised adaptation. Event attributes are encoded into tokens. A Transformer produces a history representation. Pretraining uses prefix and summary projection heads; downstream adaptation uses a business risk head.

Data	Scored events	History events	Mean length	History sparsity	Train positives
Wallet	127,375	16.03M	125.8	0.50%	732
Shopping	501,898	14.78M	29.5	30.63%	183,398
SME	371,600	2.91M	7.8	76.91%	46,740

Table 1. Dataset statistics.

	Wallet		Shopping		SME
Model	AP $\uparrow$	[email protected]% $\uparrow$	AP $\uparrow$	[email protected]% $\uparrow$	AP $\uparrow$	[email protected]% $\uparrow$
10% training labels
Scratch	0.0342 $\pm$ 0.0082	3.31 $\pm$ 0.77	0.3067 $\pm$ 0.0375	4.79 $\pm$ 0.51	0.3476 $\pm$ 0.0798	4.67 $\pm$ 0.51
CoLES (adapted)	0.0274 $\pm$ 0.0124	2.92 $\pm$ 0.55	0.2879 $\pm$ 0.0182	5.18 $\pm$ 0.84	0.3514 $\pm$ 0.0691	5.04 $\pm$ 0.52
BYB (adapted)	0.0475 $\pm$ 0.0160	2.86 $\pm$ 0.56	0.2447 $\pm$ 0.0574	3.56 $\pm$ 1.80	0.3624 $\pm$ 0.0488	5.91 $\pm$ 2.17
CPC (adapted)	0.0279 $\pm$ 0.0061	3.34 $\pm$ 2.14	0.3569 $\pm$ 0.0094	5.79 $\pm$ 0.70	0.3874 $\pm$ 0.0283	5.60 $\pm$ 0.68

Table 2. Performance comparison (%). Best results are bold; second-best results are underlined.

Figure 3. Label efficiency on Shopping and SME. Four plots show AP and recall at a 0.5 percent intervention budget for TRIDENT-Seq and scratch training across five training-label budgets on Shopping and SME, with a dashed line for full-label scratch.

	10% labels		25% labels
Pretraining source	AP	[email protected]%	AP	[email protected]%
Target: Shopping
No pretraining	0.3067	4.79	0.3591	5.68
Shopping only	0.3994	5.97	0.4361	7.85
SME only	0.3668	5.96	0.4663	8.70
Shopping + SME	0.3331	6.46	0.4511	7.98

Table 3. Effect of the pretraining source (%).

	10% labels			100% labels
Variant / budget	0.1%	0.5%	1.0%	0.1%	0.5%	1.0%
Full PFC	8.54	24.62	32.70	11.26	28.44	36.29
Random negatives	3.77	5.64	8.98	3.74	7.38	13.10
Shuffled pairs	1.32	14.62	24.79	2.45	16.22	20.04
PFC only	7.02	14.37	20.65	9.33	16.07	21.28
w/o activity	7.85	15.10	22.50	10.31	16.38	24.07

Table 4. Wallet ablation: recall (%) at three intervention budgets.

Figure 4. Pretrained representations of annotated histories. PCA and t-SNE projections of pretrained representations, with two behavior groups and an outlined t-SNE region used for case inspection.

Figure 5. Card and device switching after rejection. Three payment attempts show changes in anonymous card and device identities following rejections while retaining the IP identity, ending in a successful attempt.

Method	Traffic (M)	CB rate change $\downarrow$	Loss rate change $\downarrow$
XGB-AutoRefit	9.06	base	base
SAINT-style	10.56	$-$ 13.21%	$-$ 18.18%
TRIDENT-Seq	8.58	$-$ 37.74%	$-$ 37.88%

Table 5. Portfolio-level online A/B results at Ant International. Pooled arm-level rate changes relative to XGB.

Appendix figures & tables12 assets

Supplementary material from the paper’s appendix.

Appendix

Setting	Value
Layers / hidden size / heads	8 / 1,024 / 8
Encoder parameters / history cap	92.92M / 300 events
Hash bits / probes	128 / 4 per field
Eligible history / prefix	At least 2 events / older 60%
Future summary	24 dimensions (Table 7 )
Profile	Three coordinates (Equation 4 )

Appendix

Table 6. Encoder and pretraining configuration.

Coordinates	Definition
1	Constant 1
2	$\min(10,\operatorname{Std}(\Delta\delta)/\max(\operatorname{Mean}\|\Delta\delta\|,10^{-6}))/10$
3–5	$\operatorname{Mean}(u)/16$ , $\operatorname{Std}(u)/16$ , $\max(u)/16$
6–7	$\operatorname{Mean}(\delta)/92$ , $\operatorname{Std}(\delta)/92$
8	$\sum_{j,k}h_{jk}/(1536m)$
9	Field-level switching rate SW between adjacent events

Appendix

Table 7. Future-summary coordinates.

Setting	Value
Optimizer / weight decay	AdamW ( Loshchilov and Hutter, 2019 ) / 0.01
Encoder learning rates	$3{\times}10^{-5}$ , $10^{-4}$ , $3{\times}10^{-4}$
Risk-head learning rate	$10^{-3}$
Global batch size	128
Updates per LR / final run	1,024 / 2,048
Warmup updates	40

Appendix

Table 8. Downstream optimization settings.

	AP			[email protected]%
Setting	XGB	Scratch	Ours	XGB	Scratch	Ours
Wallet 10%	0.0133 $\pm$ 0.0027	0.0342	0.8214	2.01 $\pm$ 0.37	3.31	27.75
Wallet 100%	0.1042 $\pm$ 0.0109	0.1148	0.8355	4.31 $\pm$ 1.16	10.41	28.33
Shopping 10%	0.2146 $\pm$ 0.0098	0.3067	0.4331	3.41 $\pm$ 0.94	4.79	6.61
Shopping 100%	0.3573 $\pm$ 0.0307	0.4205	0.5678	6.25 $\pm$ 0.41	7.81	9.61
SME 10%	0.2810 $\pm$ 0.0411	0.3476	0.4886	3.97 $\pm$ 0.33	4.67	9.14

Appendix

Table 9. XGBoost on Ant International’s production risk indicators (%).

		10% labels			25% labels
Target	Pretraining source	AP	AUROC	[email protected]%	AP	AUROC	[email protected]%
Shopping	No pretraining	0.3067 $\pm$ 0.0375	73.90 $\pm$ 0.66	4.79 $\pm$ 0.51	0.3591 $\pm$ 0.0597	74.91 $\pm$ 1.40	5.68 $\pm$ 1.73
	Shopping only	0.3994 $\pm$ 0.0650	75.10 $\pm$ 0.56	5.97 $\pm$ 0.78	0.4361 $\pm$ 0.1080	74.73 $\pm$ 2.00	7.85 $\pm$ 2.06
	SME only	0.3668 $\pm$ 0.1107	71.13 $\pm$ 6.86	5.96 $\pm$ 2.48	0.4663 $\pm$ 0.0435	75.57 $\pm$ 1.25	8.70 $\pm$ 0.76
	Shopping + SME	0.3331 $\pm$ 0.1150	71.71 $\pm$ 5.84	6.46 $\pm$ 1.81	0.4511 $\pm$ 0.1033	74.93 $\pm$ 0.46	7.98 $\pm$ 1.99
SME	No pretraining	0.3476 $\pm$ 0.0798	67.84 $\pm$ 13.67	4.67 $\pm$ 0.51	0.4103 $\pm$ 0.0494	69.52 $\pm$ 13.10	5.04 $\pm$ 1.25

Appendix

Table 10. Cross-business transfer results (%).

Data	Split	Events	Positives	Positive rate
Wallet	Train	30,232	732	2.42%
	Validation	46,807	807	1.72%
	Test	50,336	3,336	6.63%
Shopping	Train	283,040	183,398	64.80%
	Validation	20,583	373	1.81%
	Test	198,275	2,949	1.49%

Appendix

Table 11. Sampled split sizes and positives.

Data	1%	5%	10%	25%	100%
Wallet	7	34	76	172	732
Shopping	1,881	9,265	18,423	45,859	183,398
SME	469	2,382	4,671	11,608	46,740

Appendix

Table 12. Training positives at each label budget.

	AP		[email protected]%
Labels	TRIDENT-Seq	Scratch	TRIDENT-Seq	Scratch
Shopping
1%	0.2344 $\pm$ 0.1087	0.2609 $\pm$ 0.0496	3.31 $\pm$ 2.18	4.76 $\pm$ 0.38
5%	0.3546 $\pm$ 0.0939	0.3049 $\pm$ 0.0703	5.87 $\pm$ 1.62	5.11 $\pm$ 1.46
10%	0.4331 $\pm$ 0.0735	0.3067 $\pm$ 0.0375	6.61 $\pm$ 1.16	4.79 $\pm$ 0.51
25%	0.5246 $\pm$ 0.0566	0.3591 $\pm$ 0.0597	8.33 $\pm$ 1.01	5.68 $\pm$ 1.73

Appendix

Table 13. Shopping and SME results across label budgets (%).

	10% labels			100% labels
Variant	AP	AUROC	[email protected]%	AP	AUROC	[email protected]%
Full PFC	0.8373	91.46	24.62	1.2725	91.63	28.44
PFC only	0.5913	90.33	14.37	0.3703	91.05	16.07
w/o activity	0.6508	91.41	15.10	0.4399	92.12	16.38
w/o dispersion	0.7921	92.05	16.14	0.4153	93.35	17.21

Appendix

Table 14. Wallet ablation: AP, AUROC, and [email protected]% (%).

Model	AP $\uparrow$	AUROC $\uparrow$	[email protected]% $\uparrow$
Scratch	100.0 $\pm$ 156.1	100.0 $\pm$ 25.4	100.0 $\pm$ 36.1
TRIDENT-Seq	178.8 $\pm$ 94.8	115.4 $\pm$ 18.2	252.7 $\pm$ 48.4

Appendix

Table 15. Longer adaptation on full-label Wallet, relative to scratch.

Cohort	Label	30d	60d	90d	120d	D95	D99
April	Union	67.16	85.29	93.54	97.90	99	148
	CB	65.41	85.31	94.34	99.37	93	119
May	Union	64.09	86.06	94.28	98.70	92	124
	CB	63.02	85.34	94.51	99.43	92	117
Apr–May	Union	65.53	85.70	93.93	98.32	95	135
	CB	64.14	85.32	94.43	99.40	92	119

Appendix

Table 16. Observed label-reporting delays (%); D95/D99 in days.

Method	Pooled	Date-weighted
SAINT-style	12.66 [2.58, 22.75]	12.49 [1.43, 23.16]
TRIDENT-Seq	28.28 [10.69, 42.38]	26.09 [9.24, 41.34]

Appendix

Table 17. SME transaction-date CB reduction versus XGB (%); approximate 95% CIs in brackets.

Financial Fraud Detection

Financial fraud detection identifies suspicious transactions, accounts, claims, and behavior to prevent losses and guide investigation.

See all 137 papers

CARTA: Self-Supervised Reusable Transaction Embeddings via Heterogeneous Tokenization and Cross-Attention
Oct 17, 2026Shreyas K. Chandrahas, Jose Munoz, Satish Babu BellapuRepresentation Learning Tabular Representation Learning
Visa Inc. Foster City, CA, USA · Massachusetts Institute of Technology Cambridge, MA, USA
SeqLLM: Augmenting LLMs with Behavioral-Sequence Modeling for High-Stakes Decisions at WeChat Pay
Aug 4, 2026Guilin Li, Jiaxing Zhang, Matthias Hwai Yong Tan +2Sequential Recommendation Financial Fraud Detection
Shanghai Jiao Tong University China
Causal Label Recovery in Payment Networks
May 28, 2026Gaurav DhamaFinancial Fraud Detection Learning with Noisy Labels
Mastercard

Oct 17, 2026J/K move · Enter open · S save

CARTA: Self-Supervised Reusable Transaction Embeddings via Heterogeneous Tokenization and Cross-Attention

Shreyas K. Chandrahas, Jose Munoz, Satish Babu Bellapu

Visa Inc. Foster City, CA, USA · Massachusetts Institute of Technology Cambridge, MA, USA

Representation Learning Tabular Representation Learning Self-Supervised Learning Financial Fraud Detection

Open paper AskPDF arXiv

Acquiring fraud labels for payment transactions is expensive and inconsistent across tasks, and high-cardinality identifiers such as merchant names make one-hot feature pipelines infeasible at scale. We present CARTA, a self-supervised encoder that learns a compact 16-dim embedding from raw transactions without labels; the frozen embedding is then reused for label-efficient fraud probing and exploratory unsupervised risk stratification without retraining the encoder. CARTA uses type-specific projections with learned embedding tables for high-cardinality identifiers, trained with token excision and InfoNCE contrastive loss. On IEEE-CIS, its linear probe at 1% of labels exceeds a raw-feature probe at 100%; on IBM TabFormer (2M rows), its frozen embedding reaches GBDT AUPRC 0.508, above frequency encoding (0.411) and the merchant-dropping baseline (0.022), though fully-supervised CatBoost is stronger. CARTA is a label-free, reusable representation that complements task-specific supervised models.

Abstract

Financial Fraud Detection

Explore similar work

CARTA: Self-Supervised Reusable Transaction Embeddings via Heterogeneous Tokenization and Cross-Attention

SeqLLM: Augmenting LLMs with Behavioral-Sequence Modeling for High-Stakes Decisions at WeChat Pay

Causal Label Recovery in Payment Networks