PaperlayerView exact revision on arXiv

Source-linked AI summary

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, Shuai Chen, Tianyi Zhang

arXiv:2610.37365v1cs.LG

TL;DR

Payment histories are plentiful, but confirmed risk labels are scarce, delayed, and business-specific. TRIDENT-Seq pretrains a shared encoder with profile-matched Prefix-Future Contrast, then adapts it to each business. Across three businesses, it improves AP and recall over scratch training and exceeds full-label scratch results with a fraction of the labels.

  • Problem

    Payment risk modeling has abundant transaction histories but scarce, delayed, business-specific outcomes, creating a need for useful representations across businesses with limited labels.

    Source: §1 · p. 1

  • Method

    TRIDENT-Seq pretrains an encoder by matching history prefixes to future-behavior summaries among candidates with comparable activity, then fine-tunes it for each business.

    Source: §1 · p. 1

  • Results

    Across three businesses, pretraining improves AP and recall over architecture-matched scratch training in all six business–budget settings and surpasses full-label scratch in mean AP and recall with 10%–25% of training labels.

    Source: §1 · p. 2, §6 · p. 8

  • Takeaways & Limitations

    One pretrained encoder supports label-efficient adaptation across businesses, including when the target business is excluded from pretraining.

    Source: §6 · p. 8

    On SME, 76.91% of traffic has no history; results also rest on one pretrained checkpoint, and Wallet AP varies considerably across fine-tuning runs.

    Source: §6 · p. 8

Abstract

from arXiv · show

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.

1 Introduction

Ant International has abundant transaction histories but scarce, delayed, business-specific risk labels, motivating a shared encoder that can adapt across businesses. TRIDENT-Seq uses profile-matched future contrast to learn behavioral representations, and its evaluation reports label-efficient improvements across three production businesses.

1 Introduction

  • Confirmed risk outcomes are scarce, delayed, and specific to each business, while a separate risk model for each business multiplies labeling and upkeep costs [32].The proposed shared encoder is pretrained on unlabeled histories and adapted using each business’s available labels.Source: §1 · p. 1
  • Random contrastive candidates allow activity level, amount, and timing to distinguish continuations, whereas comparable-activity candidates yield 24.62% recall versus 5.64% with random candidates at a 0.5% budget.The comparison is reported after fine-tuning with 10% of the labels.Source: §1 · p. 1
  • TRIDENT-Seq matches each prefix to a summary of subsequent behavior among candidates with comparable activity, while activity prediction and dispersion regularization shape the encoder output.Each business fine-tunes its own encoder copy with a risk head.Source: §1 · p. 1
  • Pretraining improves AP and recall over architecture-matched scratch training in all six business–budget settings; full-label scratch is surpassed in mean AP and recall using 10% of Wallet labels and 25% on Shopping and SME.On Wallet, the 10% setting uses 76 positive examples; pretraining only on another business’s histories still helps.Source: §1 · p. 2
  • The evaluation spans three production businesses with history coverage ranging from nearly complete to mostly empty, and deployment reports 37.7% lower chargeback and 37.9% lower overall loss rates than the previous XGBoost-based system.The deployment comparison uses the same intervention budget.Source: §1 · pp. 1–2
  • The stated contributions combine profile-matched pretraining, cross-business label-efficient evidence, and deployment evaluation within a 120 ms inference budget.The deployment evidence includes a portfolio-level A/B result and a two-stage SME evaluation.Source: §1 · p. 2

2 Related Work

The related work spans supervised transaction detectors, self-supervised event-sequence objectives, and transaction foundation models. TRIDENT-Seq differs by contrasting multi-event future summaries under observable activity matching.

2 Related Work

  • Feature-based, sequence, and graph detectors model transaction risk in different ways, but typically learn from labels for a single task [8] [14].Delayed feedback means each new business must accumulate its own labels [6].Source: §2 · p. 2
  • Event-sequence pretraining includes masked-field reconstruction, subsequence contrasts, and predictive objectives such as CPC and BYB [33, 2, 49, 43, 48].PFC instead targets a fixed statistical summary of subsequent events and restricts negatives to comparable activity.Source: §2 · p. 2
  • Transaction foundation models include generative, attribute-and-trajectory, structured-token, and masked-modeling approaches [40] [26] [32].TRIDENT-Seq contrasts multi-event summaries under activity matching and evaluates reuse across businesses through deployment.Source: §2 · pp. 2–3

3 Problem Setup

The setup scores a transaction from the user’s preceding event history, using delayed fraud or chargeback labels for supervised adaptation. Pretraining uses unlabeled histories, and deployment ranks scores under a fixed intervention budget.

3 Problem Setup

  • A scored transaction’s history contains preceding events with amount, categorical payment, device, and network attributes, and age relative to the scoring time; histories can be empty.The binary label indicates recorded fraud or chargeback, which can arrive months later.Source: §3 · p. 3
  • Pretraining splits an observed history into an older prefix and a subsequent future segment, both preceding the scored transaction, so no post-decision events enter pretraining.Source: §3 · p. 3
  • Each business initializes an encoder from pretrained parameters and trains it with a business-specific risk head on its labeled transactions.The decision system ranks transactions by predicted risk and intervenes on a fixed fraction of traffic; the production budget is 0.5%.Source: §3 · p. 3

4 Methodology

TRIDENT-Seq uses a shared history encoder, pretrains it with Prefix–Future Contrast and auxiliary objectives, then adapts a copy for each business. Its event encoding and profile-matched future summaries are designed to represent transaction behavior for business-specific risk prediction.

4.1 Overview

  • The overview separates unlabeled pretraining, business-specific fine-tuning, and online scoring, with future summaries, contrastive candidates, and auxiliary heads used only during pretraining.Online scoring uses history retrieval and one pass through the adapted encoder and risk head.Source: §4.1 · p. 3

4.2 Event Encoding

  • Categorical attributes use fixed-width multi-probe hash vectors, while amounts and ages are transformed before learned projections form event tokens for the Transformer encoder.The Transformer reads a CLS token followed by event tokens; its CLS output represents the history and is shared across pretraining, adaptation, and serving.Source: §4.2 · p. 3

4.3 Prefix–Future Contrast

  • PFC pairs an older history prefix with a 24-dimensional summary of future behavior and selects candidate futures with similar coarse activity profiles.The summary covers amount and event-age distributions, interval regularity, and occupancy and change of hashed attribute values; profile matching directs contrast toward finer structure.Source: §4.3 · pp. 3–4

4.4 Auxiliary Objectives

  • Activity prediction regresses encoder outputs to activity targets, while dispersion regularization penalizes low average within-batch standard deviation of prefix representations.The activity target uses counts, amounts, ages, and hash-field statistics; the dispersion penalty vanishes once average spread reaches its threshold.Source: §4.4 · p. 4

4.5 Business-Specific Adaptation

  • Each business fine-tunes a copy of the pretrained encoder with a new risk head, while class-balanced cross-entropy gives both classes equal total weight.The whole encoder is fine-tuned because each business defines risk differently and its labels should be able to reweight the shared representation.Source: §4.5 · p. 4

5 Experiments

Across three payment businesses, the experiments test risk ranking under label scarcity, cross-business transfer, PFC components, learned representations, and online deployment. TRIDENT-Seq improves on scratch training across all six business-budget settings, while transfer, ablations, and deployment results examine the scope and behavior of those gains.

5.1 Datasets

  • Wallet, Shopping, and SME differ in history coverage and label counts; 30.63% of Shopping and 76.91% of SME test traffic has no history.The training splits contain 732 Wallet positives, 183,398 Shopping positives, and 46,740 SME positives; empty-history transactions remain in evaluation.Source: §5.1 · p. 5, Table 1 · p. 5

5.2 Experimental Setup

  • Evaluation uses weighted AP and R@0.5%, with fixed label subsets and a common validation set across budgets.Fine-tuning tests 1%, 5%, 10%, 25%, and 100% label budgets where specified; Wallet reports 10% and 100% because its smaller subsets contain only 7 and 34 positives.Source: §5.2 · p. 5

5.3 Overall Performance (RQ1)

  • TRIDENT-Seq improves both AP and R@0.5% over architecture-matched scratch training in all six business-budget settings.Wallet benefits most; Shopping and SME comparisons are more competitive, with some results within run-to-run variation.Source: §5.3 · pp. 5–6, Table 2 · p. 6

5.4 Label Efficiency (RQ1)

  • With 25% of training labels, TRIDENT-Seq exceeds full-label scratch in mean AP and recall on both Shopping and SME.On Shopping, the comparison is 0.5246% AP and 8.33% recall versus 0.4205% and 7.81%; on SME, it is 0.6928% and 10.16% versus 0.6370% and 8.11%.Source: §5.4 · p. 6, Figure 3 · p. 6

5.5 Cross-Business Transfer (RQ2)

  • Pretraining only on the other business improves AP and recall over no pretraining in all four target-budget settings.The benefit depends on direction and budget: SME histories match Shopping in recall at 10% and beat them in both metrics at 25%, while Shopping histories help SME more at 10% than SME-only histories.Source: §5.5 · p. 6, Table 3 · p. 7

5.6 Ablation Study (RQ3)

  • Random negatives yield 5.64% recall at the 0.5% budget, while combining activity and dispersion terms yields 24.62% with 10% labels.Shuffled pairs also reduce recall; the auxiliary-term gains concentrate at 0.5% and 1% budgets, near the production operating point. The ablation is on Wallet and measures downstream recall.Source: §5.6 · p. 7, Table 4 · p. 7

5.7 Case Study

  • In the 374-history t-SNE region, evasion-type histories dominate; annotated rapid retries there show 95.6% same-merchant pairs versus 70.7% outside.The median intervals are 4.8 minutes in the region and 22.8 hours outside; inspected histories include card and device changes after rejection.Source: Figure 4 · p. 7, Figure 5 · p. 7, §5.7 · p. 8

5.8 Online Deployment (RQ4)

  • Across the portfolio, TRIDENT-Seq records 37.74% lower amount-based chargeback and 37.88% lower overall loss rates than XGB-AutoRefit.The A/B evaluation assigned about 28.20 million transactions across three arms; SME Stage I recall within the 0.5% budget is 29.79% for TRIDENT-Seq versus 7.39% for XGB-AutoRefit.Source: §5.8 · p. 8, Table 5 · p. 8

6 Conclusion

TRIDENT-Seq uses Prefix–Future Contrast (PFC) to pretrain a shared transaction encoder for payment risk, matching future summaries with similar activity profiles. Across three payment businesses, it improves risk ranking with limited labels and transfers when the target business is excluded from pretraining.

6 Conclusion

  • The encoder exceeds full-label scratch training in mean AP and recall using 10% to 25% of training labels, and transfers when the target business is excluded from pretraining.The benefit depends on history coverage: on SME, 76.91% of traffic has no history, so the encoder receives no events for it.Source: §6 · p. 8
  • PFC matches prefixes to future summaries with similar activity profiles, narrowing coarse activity differences among contrastive candidates.The paper argues that candidate choice deserves as much attention as the prediction target.Source: §6 · p. 8

A Implementation Details

The appendix specifies event encoding, pretraining targets, and optimization settings. The implementation combines hashed event fields and a Transformer encoder with future-summary and activity-prediction targets, then fine-tunes the encoder and risk head.

A.1 Event Encoding and Network Configuration

  • For each nonempty field value, MurmurHash3 hashes the string field-name:value to 128 bits split into four 32-bit slices, setting one bit per block; missing values produce all-zero vectors.Source: §A.1 · p. 10

A.2 Pretraining Targets

  • A future summary uses 24 coordinates, including amount, timing, and attribute-change statistics, computed over events after a prefix of about 60% of the history.For a one-event future, adjacent-event statistics are zero; histogram counts are normalized by the future length.Source: §A.2 · p. 10, Table 7 · p. 10
  • The 25-dimensional activity target combines statistics from recent-history windows of 75, 150, and 300 events with selected differences between the shortest and longest windows.Empty-window and undefined statistics are set to zero; smooth L1 loss averages over the batch and 25 coordinates.Source: §A.2 · p. 10

A.3 Optimization and Protocol

  • Ablations use random negatives with the matched candidate count or cyclically shift future summaries and profiles before constructing the mask.Downstream adaptation jointly fine-tunes the encoder and risk head with class-balanced binary cross-entropy; weighted validation AP selects learning rates and checkpoints.Source: §A.3 · p. 10, Table 8 · p. 10

B Baselines

The baselines include contrastive and predictive sequence methods, separately designed transaction models, and an XGBoost reference using production risk indicators. Controlled-source results use matched pretraining pools and separate checkpoints for each source.

B Baselines

  • CoLES contrasts two contiguous crops from one history, BYB predicts a teacher representation of a recent window, and CPC predicts future event representations against cross-user negatives.Each uses the shared Transformer and 24,000 pretraining updates at batch size 128.Source: B Baselines · pp. 10–11
  • TransactionGPT uses next-event prediction, while PANTHER uses behavior and field vocabularies with next-event contrastive prediction.Both are pretrained from scratch on the same history pool for 24,000 updates, with calendar features and user-profile modules disabled.Source: B Baselines · p. 11
  • XGBoost trails scratch training in all six settings; on full-label Wallet, recall is 4.31% versus 10.41% for scratch and 28.33% for TRIDENT-Seq.XGBoost uses production risk indicators rather than event histories and is reported as a reference.Source: B Baselines · p. 11, Table 9 · p. 11

C Pretraining Source Composition

  • Controlled-source results use one checkpoint and three fine-tuning runs per source, with means and sample standard deviations reported.The checkpoints use matched pretraining pools and are separate from the main pretraining run; the no-pretraining baseline shares the downstream training and test sets.Source: C Pretraining Source Composition · p. 11

D Data Splits and Label Budgets

The split statistics account for negative sampling and set the scope of the label-budget evaluations. Wallet is evaluated at 10% and 100% of training labels because its smaller subsets contain few positives.

D Data Splits and Label Budgets

  • Positive rates differ across splits because of negative sampling, so validation and test metrics reweight examples by inverse sampling probability.Label budgets subsample the training split.Source: D Data Splits and Label Budgets · p. 11
  • Wallet is evaluated at 10% and 100% label budgets because its 1% and 5% subsets contain only 7 and 34 positives.The passage says Table 12 gives the resulting positive counts.Source: D Data Splits and Label Budgets · p. 11
  • Table 11 lists sampled split sizes and positive counts, while Table 12 gives positive counts after training-label subsampling.Source: D Data Splits and Label Budgets · p. 11

E Results at All Label Budgets

Table 13 reports results across label budgets as three-run means and sample standard deviations.

E Results at All Label Budgets

  • Table 13 reports three-run means and sample standard deviations on the full weighted test population, using identical training-label subsets across methods.The passage does not provide the table's individual values.Source: E Results at All Label Budgets · p. 11

F Ablation: AP and AUROC

The ablation compares component variants on Wallet, while separate passages specify the longer-adaptation schedule and the construction of Figure 4's inspected region.

F Ablation: AP and AUROC

  • Full PFC has the highest AP at both label budgets, while the three partial variants show decreasing AP but increasing R@0.5% from 10% to 100% labels.AUROC varies by less than 2.5 points across variants and does not favor Full PFC; the components mainly change the top of the ranking where interventions act.Source: F Ablation: AP and AUROC · p. 11

G Longer Adaptation

  • Full-label Wallet adaptation uses 2,048 updates per learning-rate trial and 8,192 for the final run, totaling 14,336 selection-and-training updates per method and run.Table 15 scales results so the scratch mean under this schedule equals 100 for each metric.Source: G Longer Adaptation · p. 11

H Case Inspection Details

  • Figure 4's inspected region is a post-hoc 55%-peak class-density contour on a fixed 140×140 Gaussian KDE grid over t-SNE.It includes 374 histories; the comparison set excludes the surrounding island and contains 2,312 histories.Source: H Case Inspection Details · p. 11
  • Feedback statistics use 361 and 916 annotated event pairs, and merchant links and outcomes come from event inspection rather than encoder inputs.Source: H Case Inspection Details · p. 11

I Reporting Delays and SME CB Estimates

The supplied table descriptions identify the scope and calculation conventions for reporting delays, transfer and label-budget results, and SME chargeback estimates.

I Reporting Delays and SME CB Estimates

Related papers

Loading 2610.37365v1…