acceptodds
Under review as a conference paper at ICLR 2027

Multi-Identity Transaction Graph Pretraining for Fraud Detection

Abstract

Real-world payment risk modeling is constrained by scarce fraud labels and fragmented behavioral context across multiple payment identities. A physical transaction may appear simultaneously in multiple behavioral histories defined by different identities, while each identity induces distinct temporal dependencies. Existing sequence models typically encode these histories independently, duplicating the same physical event across trajectories and leaving cross-identity dependencies implicit. We propose Multi-Identity Transaction Graph Pretraining (MITraG), a self-supervised framework that organizes overlapping identity-conditioned transaction histories as a shared-event, multi-relational temporal graph. Each transaction is represented once as a node, while edges preserve identity-specific temporal relations and irregular time gaps. MITraG further uses Masked Transaction Node Reconstruction (MTNR) to learn contextual transaction representations from unlabeled histories before supervised adaptation to fraud detection. On real-world payment data under a strict out-of-time evaluation protocol, MITraG consistently outperforms transaction-level, sequence-pretraining, and temporal-graph baselines. These results suggest that modeling transactions as shared events connected by identity-specific temporal relations provides a general basis for self-supervised payment behavior representation learning.

Primary area:
unsupervised, self-supervised, semi-supervised, and supervised representation learning
Keywords:
self-supervised learning, temporal graph learning, transaction representation learning, multi-identity modeling, fraud detection

Reviews

Among:
Shape:
1234this paper, Bottom 10%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.67 over 3 reviews, above 10% of the others.
Reviewer_We2dOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper proposes MITraG, a self-supervised framework for fraud detection that models overlapping identity-conditioned transaction histories as a shared-event temporal graph. Each physical transaction is represented as a shared node, while identity-specific edges preserve temporal relations and time gaps. The framework uses Masked Transaction Node Reconstruction (MTNR) for pretraining and Target & Graph Interaction Fusion (TGIF) for downstream fraud prediction.

Critical strengths
  1. The paper addresses a practically meaningful fraud detection task by explicitly modeling overlapping transaction histories over multiple identities. Representing shared physical transactions as common graph nodes is well motivated, and ablation results support its effectiveness.

  2. The experimental evaluation is relatively comprehensive. The authors use three out-of-time test folds, multiple training seeds, component ablations and significance tests. MITraG delivers consistent improvements over the baselines.

Critical weaknesses
  1. Although the paper motivates self-supervised pretraining by highlighting the scarcity of fraud labels, it does not evaluate model performance under varying labeled-data regimes.

  2. The evaluation is limited to a single proprietary payment dataset and one downstream fraud detection task.

  3. The baseline comparison is somewhat outdated. The most recent baseline in Table 1 is GTAN (2023), while several more recent fraud detection and graph representation learning methods from 2025–2026 are discussed in the Related Work but not evaluated.

  4. The benefit of historical graph modeling varies considerably across OOT partitions. Compared with Matched Target-Only, MITraG shows only marginal improvements in January and February, while most of the gain comes from March. This raises concerns about whether its effectiveness depends on specific temporal data distributions.

  5. The efficiency evaluation is limited. Although the paper reports training cost and batched inference latency, it lacks latency comparisons with the baselines and an end-to-end analysis including graph construction.

  6. The presentation quality of the figures needs improvement. Figure 1 is overly crowded, with small labels, legends, and mathematical annotations that are difficult to read at normal paper size. Several experimental plots also contain small axis labels and numerical annotations, reducing readability.

Questions for authors
  1. Evaluate MTNR under different labeled-data fractions to verify its effectiveness in low-label settings.

  2. Explain the absence of recent fraud detection and graph learning methods from the main baseline comparisons.

  3. Explain why the improvement over Matched Target-Only is concentrated in March, with only marginal gains in January and February.

  4. Provide end-to-end inference latency comparisons with the baselines, including graph construction costs.

Reviewer_BNjfOpenReview ↗
Rating
1
Confidence
4
Informativeness
1
Presentation
2
Soundness
1
Read review
Paper summary

The authors consider the task of identification of fraudulent transactions. They solve this problem on the basis of a private, industrial dataset, which represents several million transactions coming from different identities (card view, account view, ...). The main contribution of the authors is treating such data not as isolated transactions or identity-specific sequences but as graph data, which use shared transactions across identities to construct cross-identity relations. The resulting structure serves as the basis for conducting self-supervised pretraining to get the representations of the history preceding the target transaction. The encoded multi-identity history along with the additional features of the target transaction produces the input representation to the classification head for detecting a fraudulent entity.

Critical strengths
  1. The considered task of fraud detection in the financial domain is an urgent and significant problem.
  2. The experiments are conducted on a large, industrial dataset that allows establishing the method’s performance in a real-world setup.
Critical weaknesses

There are two critical problems with the paper:

  1. There are missing important details of the problem setup, which hinders the understanding of where the proposed pipeline can be leveraged elsewhere. The key concept of the paper is the shared events across identities, but neither the nature, conditions, nor examples of the shared events within the considered data are provided. Moreover, the proposed pipeline classifies the target transaction, which must be shared across several identities, without discussing the case when the non-shared event must be classified. Therefore, the absence of understanding of the cases in which and circumstances under which I can leverage the proposed method makes me lean toward rejection.

  2. There are concerns about the generalization of the approach to different datasets. The method proposes quite specific architectural solutions, but there is no evidence that they will work effectively on different data, which undermines the usefulness of the paper for the community.

Less significant issues:

  1. The formulas in the text are hard to follow since some details are missing (e.g., how the field tokenizer is working, the design of the view-specific decoder and target-transaction encoder, and the MLP for time in formula 7).

  2. The information in Figure 1 is not aligned with the paper text. The device view, hash buckets in field-aware encoding, and the concept of base-field sequences are not discussed in the text.

Questions for authors
  1. Why do we need additional features of the target transaction before entering the classification head? Why can we not use them in the graph pretraining stage for all transactions? Do the transactions from different identities have the same feature sets?

  2. The performance of the transaction-level method (LightGBM) is already comparable to that of MITraG. Does it mean that there are few transactions that need cross-identity context and the majority of them can be identified with analysis of the transaction in isolation, maybe even in some rule-based manner for the feature (e.g., the strange time of a transaction at midnight or too large an amount for a transaction that was not encountered before)?

  3. Which standard deviation values do all methods from Table 1 have? Without such information, it is hard to understand the effect size from MITraG. Also, the discussion on the statistical significance of the proposed method should be added to the main part of the paper.

  4. How is your method applied if I want to infer whether a transaction, which is not shared and does not have a history beforehand, is fraudulent?

  5. What are the benefits of the suggested specialized architecture with mask-based self-supervised learning tasks compared to other established architectures (e.g. Heterogeneous Graph Transformer) and pretraining strategies for (temporal) heterogeneous graphs? There is only one graph-based baseline (GTAN), which does not adopt either pretraining or a heterogeneous graph nature.

Reviewer_avw7OpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper represents a payment transaction once, as a node shared across the account- and card-conditioned histories, and learns a local multi-relational temporal graph with masked node reconstruction before fine-tuning for fraud detection. The problem is real and the out-of-time protocol is carefully described. The evidence does not isolate the graph, though. On the zero-history samples, where no graph can exist, the model still beats the feature-only baseline by 0.0765, more than any ablation gain, and the one input-matched control gives 0.0146. The sequence baseline is also untuned and weaker than using no history. My decision is weak reject.

Critical strengths
  1. the shared-event formulation is the right abstraction for this problem. Merging occurrences of one physical transaction into a single node while keeping identity-specific edge types fixes the duplication without discarding either view.

  2. the out-of-time protocol is unusually careful for an industrial paper: three monthly partitions, causal histories, and an overlap-stratified analysis that tries to link the gain to the mechanism.

Critical weaknesses
  1. the headline gain is not attributable to the graph. MITraG receives base fields that the FT-Transformer baseline does not, and on the zero-history samples, where no graph is possible, it still leads FT-Transformer by 0.0765 (Appendix C). That is larger than any ablation effect. The only input-matched comparison is the internal target-only model, where the gain is 0.0146. Without an external baseline that receives the same fields, the shared-event claim is not isolated.

  2. the sequence baseline cannot carry the headline gap. The dual-sequence TabBERT is the authors' own adaptation, with no architecture, corpus, hyperparameters or budget reported, and it scores 0.7623, below the no-history FT-Transformer at 0.7670 despite strictly more inputs. A comparator that loses to using no history cannot show that the shared-event representation is what helps. The ablation story has a similar hole: the variant without MTNR and the matched target-only control both print 0.8028, so the claimed MTNR gain is a rounding artifact.

Questions for authors
  1. How is the physical-transaction identifier obtained at inference? Is it always available and consistent across the account and card views, and how does performance degrade when it is noisy?
  2. The variant without MTNR and the matched target-only control both print 0.8028. What does MTNR add beyond the target features?
  3. Can you report the total number of transactions and identities, the exact fraud rate, and either a sanitized sample or a public benchmark result?
  4. No table reports dispersion, and the single adjusted p-value is 0.003998, which is the bootstrap floor for 2,000 replicates. Can you report seed variance and test Recall@1%?
  5. Histories contain up to 35 transactions. Can a test graph contain training or validation nodes, and how is transductive contamination ruled out?

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.