Paper deep dive
FinFraudBench: A Heterogeneous Graph Benchmark for Financial Fraud Detection
Yixuan Chen, Hongyu Zhan, Jie Sheng, Weiyu Han, Shuai Chen, Tianyi Zhang, Xiao Tan, Jun Xia
Intelligence
Status: succeeded | Model: Gemma-4-26B-A4B | Prompt: intel-v1 | Confidence: 93%
Last extracted: 8/18/2026, 5:52:20 AM
Summary
The paper introduces FinFraudBench, a heterogeneous graph benchmark for financial fraud detection, addressing gaps in existing public benchmarks that often simplify financial ecosystems into homogeneous or single-node-type graphs. FinFraudBench provides two large-scale datasets, CreditCard-Fraud and BankTrans-Fraud, preserving six financial entity types and fourteen directed edge types with realistic class imbalance and limited label availability. The authors establish a standardized evaluation protocol and demonstrate that heterogeneous graph learning methods outperform homogeneous and multi-relational approaches in detecting fraud.
Entities (20)
Relation Signals (19)
FinFraudBench → contains → CreditCard-Fraud
confidence 98% · FinFraudBench contains two heterogeneous graph datasets (CreditCard-Fraud and BankTrans-Fraud)
FinFraudBench → contains → BankTrans-Fraud
confidence 98% · FinFraudBench contains two heterogeneous graph datasets (CreditCard-Fraud and BankTrans-Fraud)
FinFraudBench → evaluates → Heterogeneous Graph
confidence 95% · FinFraudBench is a heterogeneous graph benchmark
CreditCard-Fraud → hasnodetype → Customer
confidence 95% · customers, cards, merchants, categories, and locations provide relational context
CreditCard-Fraud → hasnodetype → Card
confidence 95% · customers, cards, merchants, categories, and locations provide relational context
CreditCard-Fraud → hasnodetype → Merchant
confidence 95% · customers, cards, merchants, categories, and locations provide relational context
CreditCard-Fraud → hasnodetype → Category
confidence 95% · customers, cards, merchants, categories, and locations provide relational context
CreditCard-Fraud → hasnodetype → Location
confidence 95% · customers, cards, merchants, categories, and locations provide relational context
Cypher Suggestions (0)
No Cypher suggestions yet.
Abstract
Abstract:The increasing complexity of digital financial systems has reshaped financial fraud detection from isolated transaction classification into relational risk reasoning over interconnected financial entities. This shift has motivated graph-based fraud detection, where models identify fraudulent nodes by exploiting dependencies among customers, cards, merchants, categories, and locations. However, despite rapid progress in graph-based methods, existing public benchmarks remain misaligned with real-world financial systems in two important aspects. First, they often simplify financial ecosystems into homogeneous or single-node-type multi-relational graphs, failing to preserve the multi-entity and multi-relational nature of financial data. Second, they rarely provide large-scale heterogeneous financial graph datasets with realistic operating conditions such as extreme class imbalance and limited label availability, making it difficult to assess the practical effectiveness of current methods. To address these gaps, we present FinFraudBench, a heterogeneous graph benchmark for financial fraud detection. FinFraudBench contains two heterogeneous graph datasets (CreditCard-Fraud and BankTrans-Fraud) with up to 8.99M nodes and 89.23M directed typed edges. Each dataset preserves six financial entity types, fourteen directed edge types, and natural fraud rates that mirror deployment constraints. With these datasets, we establish a standardized evaluation protocol covering both ranking and imbalance-sensitive classification metrics, and evaluate representative baselines. Extensive experiments yield empirical insights into current methods' limitations and suggest promising avenues for future research. FinFraudBench is available at this https URL.
Tags
Links
- Source: https://arxiv.org/abs/2608.15177v1
- Canonical: https://arxiv.org/abs/2608.15177v1
Trouble viewing inline? Open PDF directly →
Full Text
55,611 characters extracted from source content.
Expand or collapse full text
FinFraudBench: A Heterogeneous Graph Benchmark for Financial Fraud Detection Yixuan Chen Hongyu Zhan11footnotemark: 1 Jie Sheng Weiyu Han Thanks: Equal contribution. Affiliation: HKUST-GZ Affiliation: HKUST-GZ Affiliation: Jilin University Affiliation: Ant Group Affiliation: Ant Group Shuai Chen Tianyi Zhang Xiao Tan Jun Xia22footnotemark: 2 Thanks: Corresponding authors. Affiliation: HKUST-GZ Affiliation: Ant Group Affiliation: Ant Group Affiliation: Ant Group Affiliation: HKUSTchenyx2123@mails.jlu.edu.cn, hzhan701@connect.hkust-gz.edu.cnshengjie.sheng, hanweiyu.hwy, shuai.cs@ant-intl.comzty113091, alex.tx@ant-intl.comjunxia@hkust-gz.edu.cn Abstract The increasing complexity of digital financial systems has reshaped financial fraud detection from isolated transaction classification into relational risk reasoning over interconnected financial entities. This shift has motivated graph-based fraud detection, where models identify fraudulent nodes by exploiting dependencies among customers, cards, merchants, categories, and locations. However, despite rapid progress in graph-based methods, existing public benchmarks remain misaligned with real-world financial systems in two important aspects. First, they often simplify financial ecosystems into homogeneous or single-node-type multi-relational graphs, failing to preserve the multi-entity and multi-relational nature of financial data. Second, they rarely provide large-scale heterogeneous financial graph datasets with realistic operating conditions such as extreme class imbalance and limited label availability, making it difficult to assess the practical effectiveness of current methods. To address these gaps, we present FinFraudBench, a heterogeneous graph benchmark for financial fraud detection. FinFraudBench contains two heterogeneous graph datasets (CreditCard-Fraud and BankTrans-Fraud) with up to 8.99M nodes and 89.23M directed typed edges. Each dataset preserves six financial entity types, fourteen directed edge types, and natural fraud rates that mirror deployment constraints. With these datasets, we establish a standardized evaluation protocol covering both ranking and imbalance-sensitive classification metrics, and evaluate representative baselines. Extensive experiments yield empirical insights into current methods’ limitations and suggest promising avenues for future research. FinFraudBench is available at https://anonymous.4open.science/r/FinFraudBench-B002. 1 Introduction Financial fraud detection is a central task in modern financial systems, where fraud can cause substantial economic loss and erode trust in digital payment services 2; 33. Fraud is rarely an isolated property of a single record. A suspicious transaction often becomes easier to identify when considered together with related financial entities and their interactions, such as the customer behind the transaction, the payment instrument, the receiving merchant, and the associated context. This relational nature has motivated graph-based fraud detection, where models can capture dependencies that are difficult to represent from tabular records alone in fraud analysis 1; 30; 26. Figure 1: Comparison of three graph formulations for financial fraud detection. Homogeneous graphs use one node type and one relation type; multi-relational graphs keep a single target node type while distinguishing multiple relation types; heterogeneous graphs preserve multiple entity types, multiple relation types, and type-specific attributes. Existing public graph benchmarks for fraud detection, however, often simplify this financial ecosystem before learning begins. We distinguish three graph formulations in this paper. Here, each node represents an entity instance, such as a transaction, user, account, card, merchant, or location. In a homogeneous graph, transactions, users, or accounts are usually selected as the only node type, and all edges describe one broad relation, such as a transaction between users or a payment flow between transactions 42; 20. A multi-relational graph still keeps a single target node type, but its relation types are often induced by typed meta paths in the original heterogeneous ecosystem, such as two target nodes connected through a shared device, shared account, shared card, or shared merchant 39; 10; 34; 31; 27; 11. A heterogeneous graph contains multiple node types and multiple relation types, allowing financial entities and their interactions to be represented directly 39; 19. Homogeneous and multi-relational formulations are useful abstractions and have supported important progress in graph-based fraud detection, but both lose part of the fraud entity-relation structure. This representation gap matters for benchmark design. Financial transactions naturally involve multiple entity types, multiple relation semantics, and type-specific attribute spaces. When typed entities are folded into features, converted into target–target edges, or omitted, benchmarks no longer directly test entity- and relation-aware fraud modeling. Figure 1 clearly illustrates this distinction among homogeneous graphs, multi-relational graphs, and heterogeneous graphs considered in this work. To address this gap, we present FinFraudBench, a benchmark consisting of two heterogeneous graph datasets constructed from public financial fraud data sources 38; 8; the original sources are transaction-level tables rather than pre-existing graphs. To the best of our knowledge, FinFraudBench is the first benchmark to construct comparatively large heterogeneous graphs specifically for financial fraud detection from public transaction-level sources. Both datasets use a unified schema with transaction nodes as prediction targets and additional typed financial entities as contextual information. FinFraudBench releases the heterogeneous graph as the canonical representation and supports deterministic projections for feature-only, homogeneous, or multi-relational methods. FinFraudBench also provides a reproducible evaluation protocol. Each dataset is organized into labeled training transactions, unlabeled training transactions, validation transactions, a full test split, and balanced mini-test subsets, supporting controlled comparison under limited labels and extreme class imbalance in realistic deployment settings. We further reproduce representative baselines across non-GNN, homogeneous GNN, multi-relation, fraud-oriented, and heterogeneous families. The results reveal a clear family-level pattern: methods designed for heterogeneous graphs are strongest overall, multi-relation and fraud-oriented methods remain competitive, and target-only non-GNN baselines are limited by their inability to use relational financial context encoded by the benchmark graph structure during practical financial fraud detection tasks. The main contributions of this work are summarized as follows: • Heterogeneous financial fraud graph benchmarks. We construct two heterogeneous graph datasets from public transaction-level fraud sources. The datasets preserve multiple financial entity types, multiple relation types, and type-specific attributes, while also supporting deterministic simplified views for methods that require homogeneous or multi-relational graph inputs. • Reproducible limited-label evaluation protocol. We provide unified graph schemas, labeled and unlabeled training transactions, validation splits, balanced mini-test subsets, and standardized metrics to support controlled benchmarking under label scarcity and extreme class imbalance in a unified evaluation setting. • Empirical insights for heterogeneous fraud detection. Our experiments show that heterogeneous graph structures preserve rich signals encoded in typed financial entities and relations, which are essential for fraud modeling. These results point to a promising direction: extending fraud-oriented designs to heterogeneous architectures, so that typed structural information and domain-specific fraud cues can be jointly exploited. 2 Preliminaries and Related Work 2.1 Preliminaries Homogeneous graphs. We define the three graph formulations used in this paper. A homogeneous attributed graph is written as G=(,ℰ,)G=(V,E,X), where V is the node set, ℰE is the edge set, and ∈ℝ||×dX ^|V|× d is a shared node feature matrix. It assumes one node type and one edge type. Multi-relational graphs. A multi-relational graph is a single-node-type graph with multiple relation types, written as G=(,ℰ,ℛ,,ϕ)G=(V,E,R,X,φ), where ℛR is the relation-type set and ϕ:ℰ→ℛφ:E assigns each edge to a relation type. Equivalently, ℰ=⋃r∈ℛℰrE= _r E_r with ℰr=e∈ℰ:ϕ(e)=rE_r=\e :φ(e)=r\. This formulation captures relation heterogeneity while retaining a single node type. Heterogeneous graphs. A heterogeneous graph is written as G=(,ℰ,,ℛ,aa∈,τ,ϕ)G=(V,E,A,R,\X^a\_a ,τ,φ), where A is the node-type set, ℛR is the relation-type set, τ:→τ:V assigns each node to a node type, and ϕ:ℰ→ℛφ:E assigns each edge to a relation type. For each node type a∈a , nodes of that type are collected as a=v∈:τ(v)=aV_a=\v :τ(v)=a\ and have their own feature matrix a∈ℝ|a|×daX^a ^|V_a|× d_a. In our constructed benchmark, this stricter heterogeneous graph setting captures both relation-level and entity-level heterogeneity. Fraud detection task. Given a graph G in one of the above formulations, we formulate financial fraud detection as binary node classification over a target node type. Let a⋆a be the target type, a⋆V_a be its nodes, and L⊆a⋆V_L _a be the labeled training nodes. Unlabeled target nodes may remain in the graph as relational context but do not contribute to the supervised loss. Given labels yv∈0,1y_v∈\0,1\, a model assigns each target node a fraud score sv=fθ(G,v),y^v=[sv≥γ],s_v=f_θ(G,v), y_v=1[s_v≥γ], (1) where γ is a decision threshold. The supervised objective is θ⋆=argmin∑v∈Lθℓ(sv,yv).θ = _θ _v _L (s_v,y_v ). (2) In FinFraudBench, the target type is transaction, while customers, cards, merchants, categories, and locations provide relational context. 2.2 Related Work Graph-based fraud detection. Fraud detection has been studied using tabular transaction features, anomaly detection, and graph-based learning 2; 33; 5; 3; 25; 1; 30. General GNN backbones provide message-passing, convolutional, inductive, attention-based, and scalable neighborhood aggregation mechanisms 36; 16; 21; 17; 40; 7; 47; 43. In fraud detection, graph-based methods represent suspicious behaviors as relational structures among transactions, accounts, users, reviews, devices, and payment flows, while fraud-oriented GNNs further introduce assumptions about noisy neighborhoods, camouflage, label scarcity, partitioned propagation, or attribute association 28; 27; 11; 6; 49; 12; 23. These works motivate relational modeling for fraud detection, but their public evaluation settings are usually built on homogeneous graphs or single-node-type multi-relational graphs. Heterogeneous graph learning. Heterogeneous graph learning broadly studies graphs with typed nodes, typed edges, or both. To avoid ambiguity, we use multi-relational graphs for the single-node-type setting with multiple relation types, and heterogeneous graphs for the setting with multiple node types and multiple relation types. Heterogeneous information network research formalizes entity types, relation types, and meta-path semantics 39; 10, while neural models learn relation-specific transformations, meta-path-based attention, learned composite relations, and node- and edge-type-dependent attention 37; 41; 15; 46; 19; 45. These methods are relevant to financial fraud detection because financial risk data naturally involve different entity roles and relation semantics, yet public fraud benchmarks rarely expose these elements together with type-specific attributes. Fraud datasets and benchmarks. Public benchmarks strongly influence how fraud detection models are designed and evaluated. YelpChi and Amazon are widely used in graph-based fraud detection; although not financial datasets, they test whether models handle noisy and suspicious relational patterns 34; 31; 48. In common graph-fraud settings, however, they are usually treated as single-node-type multi-relational graphs, where all nodes share the same prediction target type and edge types encode alternative relations between target nodes 27; 11. Financial fraud benchmarks such as Elliptic and DGraph are closer to our domain, but they are representative homogeneous graphs: Elliptic connects Bitcoin transaction nodes through payment flows, while DGraph models financial users with temporal interaction edges, strong class imbalance, and many unlabeled nodes 42; 20. Recent efforts such as H2GB construct large-scale heterogeneous graphs across domains, highlighting the value of realistic structural properties, but they are not specifically designed for fraud detection 24. Benchmark studies further show that dataset structure shapes the modeling assumptions methods are expected to exploit 18; 13; 29; 24. The lack of public financial fraud benchmarks with heterogeneous graph structure therefore limits the development and evaluation of models that reason over multiple financial entity types, relation types, and type-specific attributes. FinFraudBench fills this gap by providing a unified benchmark with two heterogeneous graph datasets constructed from public transaction-level sources and standardized splits for node-level fraud detection. 3 Benchmark Construction 3.1 Source Data and Construction Principles We construct CreditCard-Fraud and BankTrans-Fraud from two public tabular fraud datasets 38; 8. Both raw sources are transaction-level tables: each row corresponds to one transaction and contains a fraud label together with non-label fields describing the transaction itself, such as amount and time, and the observable financial context around it, such as customer- or payment-instrument identifiers, merchant and category information, and geographic or billing attributes. These fields are provided as columns in transaction records rather than as separate entity or relation tables. Our goal is to convert such tabular records into heterogeneous graphs while avoiding label leakage in graph or feature construction. Figure 2 summarizes this construction pipeline for both datasets. The construction follows three principles. First, recurring financial objects are represented as typed nodes rather than being kept only as transaction attributes. Second, relation types follow the semantic roles between typed entities, so different entity associations are preserved as distinct edge types. Third, labels remain defined only on transaction nodes, while non-transaction nodes provide unlabeled context for limited-label fraud detection. Figure 2: Overview of the graph construction pipeline. Public transaction-level tables are converted into heterogeneous graphs by mapping transactions and recurring financial objects to typed nodes, deriving semantic relations between observed entities, constructing type-specific attributes, and assigning fraud labels only to transaction nodes. 3.2 Graph Schema, Attributes, and Labels Both datasets share a transaction-centered heterogeneous schema with six node types and seven semantic edge types. Transaction nodes are the prediction targets, while customer, card, merchant, category, and location nodes provide typed context, where location denotes state or location fields in the source records. The seven edge types include five transaction-context associations and two context-context associations, customer–card ownership and customer–location association. Adding reverse edges yields fourteen directed edge types. Each source-table row is mapped to one transaction node, with entity-valued fields deduplicated within their own node types. The induced typed edges preserve intermediate entities that link transactions, rather than projecting shared contexts into direct transaction–transaction edges. Node attributes are built from non-label source fields and leakage-free deterministic features, with type-specific dimensions ranging from 7 to 21. Fraud labels are attached only to transaction nodes; all other node types provide unlabeled relational context. A compact schema and feature summary is provided in Appendix A. 4 Dataset Statistics and Analysis 4.1 Graph Scale and Heterogeneity Table 1 summarizes the graph-scale and label statistics of the two datasets. CreditCard-Fraud contains 1.86M nodes and 18.53M directed typed edges, while the larger BankTrans-Fraud contains 8.99M nodes and 89.23M directed typed edges. Both datasets share the same schema of six node types and fourteen directed edge types, but differ substantially in scale and fraud prevalence. Although transaction nodes dominate graph size, the non-transaction context nodes are central to the relational structure: they represent customers, cards, merchants, categories, and locations, and link transactions through shared financial entities and contexts instead of collapsing these connections into direct transaction–transaction edges. 4.2 Label Imbalance and Splits Figure 3: Transaction split proportions of the constructed fraud graphs. Figure 3 visualizes the main transaction split proportions. Each dataset is organized into labeled training transactions, unlabeled training transactions, validation transactions, a full test split, and five balanced mini-test subsets. The labeled training portion is intentionally small overall, about 2% of transactions in each dataset: CreditCard-Fraud contains 37.0K labeled and 1.22M unlabeled training transactions, while BankTrans-Fraud contains 178.3K labeled and 6.78M unlabeled training transactions. Unlabeled training transactions remain in the graph as context, so semi-supervised and transductive methods can use their node attributes and local neighborhoods without observing their labels. The full test set preserves the natural class distribution for realistic evaluation, while balanced mini-test subsets contain equal numbers of fraudulent and non-fraudulent transactions for broader and more stable comparison under extreme class imbalance in deployment-like settings. Table 1: Graph-scale and label statistics of the constructed datasets. Dataset #Trans. nodes #Total nodes #Context nodes #Directed edges #Fraud trans. Fraud rate CreditCard-Fraud 1.85M 1.86M 2.76K 18.53M 9,651 0.5210% BankTrans-Fraud 8.91M 8.99M 72.14K 89.23M 13,332 0.1495% 5 Benchmark Tasks and Evaluation Protocol The benchmark evaluates transaction-level fraud detection under the splits defined in Section 4.2. The canonical input is the heterogeneous graph constructed in Section 3, and deterministic projections support methods that require feature-only, homogeneous, or multi-relational inputs. 5.1 Task Setup Given the full graph or one of its projections, the task is binary node classification on transaction nodes. Models may use the structure and attributes available in their input view, but supervised labels are provided only for labeled training transactions. Validation labels are used only for model selection and early stopping. 5.2 Metrics and Reporting Protocol Our main experiments use the balanced mini-test subsets defined in Section 4.2 to compare ranking and classification behavior under equal fraud/non-fraud weight. We report six complementary metrics for each method. AUROC and AUPRC. AUROC measures whether fraudulent transactions receive higher risk scores than legitimate transactions across thresholds 14. AUPRC reflects the precision–recall trade-off for the fraud class and is especially important under class imbalance 9; 35. Together, they evaluate score ranking before threshold selection in deployment settings. Accuracy and Macro-F1. Accuracy is usually misleading on naturally imbalanced fraud data, but is informative on our balanced mini-test subsets because both classes contribute equally. Macro-F1 averages the fraud and non-fraud F1 scores and is more sensitive to failures on either class. Fraud F1 and Fraud Recall. Fraud F1 summarizes positive-class detection while penalizing excessive false positives. Fraud Recall measures the fraction of detected fraud cases and captures missed-fraud risk, but should be interpreted with AUPRC and Fraud F1. Reporting protocol. In the main benchmark tables, we report mean performance over five balanced mini-test subsets. Unless otherwise stated, model selection uses validation AUPRC. AUROC and AUPRC are ranking metrics, while accuracy, Macro-F1, Fraud F1, and Fraud Recall are threshold-dependent classification metrics. 6 Experiments We evaluate representative baselines that cover different modeling assumptions on the same benchmark. We organize methods into five comparison families. MLP and the target-only LLM prompting probes are non-GNN baselines. GCN and GraphSAGE represent homogeneous graph learning on collapsed graph views 21; 17. R-GCN represents relation-aware message passing around transaction nodes 37; ConsisGAD is grouped with multi-relation methods because its evaluated implementation explicitly uses relation-specific graph channels on multi-relation fraud graphs 6. PMP 49 and GAAP 12 represent fraud-oriented graph learning methods with assumptions tailored to graph fraud or anomaly detection. HAN, SeHGNN, and HGT represent heterogeneous graph learning methods with typed relation, meta-path, semantic, or node- and edge-type-aware message passing 41; 45; 19. The two LLM probes use Qwen2-7B-Instruct 44: zero-shot prompting 4; 22 and 4-shot in-context prompting 4; 32. These probes serve as additional non-GNN references in our benchmark. 6.1 Overall Benchmark Results Table 2 reports method-level results averaged over five balanced mini-test subsets, while Figure 4 summarizes the best method in each family for every dataset and metric. We report AUROC, AUPRC, accuracy, Macro-F1, Fraud F1, and Fraud Recall, grouping methods into non-GNN, homogeneous GNN, multi-relation, fraud-oriented, and heterogeneous families. Table 2: Overall benchmark results on balanced mini-test subsets. Results are averaged over five mini-test subsets and reported as mean ± standard deviation. The best result for each dataset and metric is highlighted with boldface and gray shading for visual emphasis. Dataset Method Model family AUROC AUPRC Acc. Macro-F1 Fraud F1 Fraud Recall CreditCard-Fraud MLP Non-GNN 0.880±0.0060.880±0.006 0.905±0.0050.905±0.005 0.754±0.0050.754±0.005 0.739±0.0060.739±0.006 0.678±0.0090.678±0.009 0.517±0.0100.517±0.010 LLM-ZS Non-GNN 0.659±0.0050.659±0.005 0.688±0.0090.688±0.009 0.551±0.0030.551±0.003 0.452±0.0040.452±0.004 0.219±0.0060.219±0.006 0.126±0.0040.126±0.004 LLM-ICL-4 Non-GNN 0.807±0.0420.807±0.042 0.850±0.0360.850±0.036 0.684±0.0120.684±0.012 0.650±0.0160.650±0.016 0.541±0.0260.541±0.026 0.374±0.0250.374±0.025 GCN Homogeneous GNN 0.820±0.0110.820±0.011 0.863±0.0070.863±0.007 0.542±0.0030.542±0.003 0.420±0.0070.420±0.007 0.155±0.0120.155±0.012 0.084±0.0070.084±0.007 GraphSAGE Homogeneous GNN 0.905±0.0060.905±0.006 0.928±0.0040.928±0.004 0.768±0.0100.768±0.010 0.756±0.0120.756±0.012 0.700±0.0170.700±0.017 0.542±0.0200.542±0.020 R-GCN Multi-relation 0.887±0.0070.887±0.007 0.916±0.0040.916±0.004 0.688±0.0070.688±0.007 0.654±0.0090.654±0.009 0.547±0.0140.547±0.014 0.377±0.0130.377±0.013 ConsisGAD Multi-relation 0.911±0.0040.911±0.004 0.932±0.0030.932±0.003 0.653±0.0060.653±0.006 0.605±0.0080.605±0.008 0.468±0.0140.468±0.014 0.305±0.0120.305±0.012 PMP Fraud-oriented 0.868±0.0080.868±0.008 0.903±0.0050.903±0.005 0.699±0.0070.699±0.007 0.670±0.0090.670±0.009 0.571±0.0130.571±0.013 0.400±0.0120.400±0.012 GAAP Fraud-oriented 0.924±0.0040.924 0.004 0.934±0.0030.934 0.003 0.716±0.0080.716±0.008 0.691±0.0100.691±0.010 0.605±0.0140.605±0.014 0.437±0.0140.437±0.014 HAN Heterogeneous 0.903±0.0040.903±0.004 0.927±0.0020.927±0.002 0.686±0.0080.686±0.008 0.652±0.0110.652±0.011 0.542±0.0170.542±0.017 0.372±0.0160.372±0.016 SeHGNN Heterogeneous 0.903±0.0050.903±0.005 0.928±0.0030.928±0.003 0.758±0.0060.758±0.006 0.743±0.0070.743±0.007 0.681±0.0110.681±0.011 0.517±0.0140.517±0.014 HGT Heterogeneous 0.915±0.0040.915±0.004 0.934±0.0040.934 0.004 0.832±0.0080.832 0.008 0.827±0.0080.827 0.008 0.799±0.0110.799 0.011 0.669±0.0140.669 0.014 BankTrans-Fraud MLP Non-GNN 0.938±0.0040.938±0.004 0.946±0.0040.946±0.004 0.672±0.0030.672±0.003 0.633±0.0040.633±0.004 0.513±0.0070.513±0.007 0.346±0.0060.346±0.006 LLM-ZS Non-GNN 0.809±0.0070.809±0.007 0.813±0.0070.813±0.007 0.517±0.0040.517±0.004 0.372±0.0080.372±0.008 0.070±0.0130.070±0.013 0.036±0.0070.036±0.007 LLM-ICL-4 Non-GNN 0.873±0.0190.873±0.019 0.860±0.0120.860±0.012 0.541±0.0320.541±0.032 0.418±0.0630.418±0.063 0.151±0.1110.151±0.111 0.086±0.0680.086±0.068 GCN Homogeneous GNN 0.948±0.0030.948±0.003 0.949±0.0030.949±0.003 0.726±0.0050.726±0.005 0.706±0.0070.706±0.007 0.629±0.0100.629±0.010 0.464±0.0110.464±0.011 GraphSAGE Homogeneous GNN 0.926±0.0020.926±0.002 0.930±0.0030.930±0.003 0.546±0.0040.546±0.004 0.429±0.0080.429±0.008 0.170±0.0140.170±0.014 0.093±0.0090.093±0.009 R-GCN Multi-relation 0.962±0.0040.962±0.004 0.972±0.0030.972±0.003 0.854±0.0080.854±0.008 0.851±0.0090.851±0.009 0.829±0.0110.829±0.011 0.708±0.0160.708±0.016 ConsisGAD Multi-relation 0.966±0.0030.966±0.003 0.975±0.0020.975±0.002 0.822±0.0060.822±0.006 0.816±0.0070.816±0.007 0.784±0.0090.784±0.009 0.644±0.0130.644±0.013 PMP Fraud-oriented 0.959±0.0040.959±0.004 0.968±0.0030.968±0.003 0.816±0.0090.816±0.009 0.810±0.0100.810±0.010 0.775±0.0130.775±0.013 0.634±0.0180.634±0.018 GAAP Fraud-oriented 0.949±0.0050.949±0.005 0.954±0.0040.954±0.004 0.643±0.0040.643±0.004 0.591±0.0060.591±0.006 0.445±0.0100.445±0.010 0.286±0.0080.286±0.008 HAN Heterogeneous 0.969±0.0030.969±0.003 0.977±0.0020.977±0.002 0.881±0.0040.881±0.004 0.880±0.0040.880±0.004 0.865±0.0060.865±0.006 0.763±0.0090.763±0.009 SeHGNN Heterogeneous 0.969±0.0020.969±0.002 0.978±0.0010.978 0.001 0.884±0.0030.884 0.003 0.882±0.0030.882 0.003 0.869±0.0040.869 0.004 0.768±0.0070.768 0.007 HGT Heterogeneous 0.973±0.0020.973 0.002 0.978±0.0010.978 0.001 0.779±0.0060.779±0.006 0.768±0.0070.768±0.007 0.717±0.0100.717±0.010 0.559±0.0120.559±0.012 Figure 4: Family-level best results on balanced mini-test subsets. For each metric and model family, bars report the best-performing method within that family in Table 2; C denotes CreditCard-Fraud and BT denotes BankTrans-Fraud. 6.2 Representation Analysis Figure 5: Family-level mean-score radar across benchmark metrics. Figure 5 complements Figure 4 with a compact family-level mean-score view across the two datasets. The radar profiles show that heterogeneous methods are strongest overall, while multi-relation and fraud-oriented methods remain competitive and non-GNN baselines are consistently weaker. Transaction features provide a strong but incomplete baseline. MLP reaches Macro-F1 scores of 0.739 and 0.633 on the two datasets, confirming that target-node attributes carry substantial fraud signals. However, it cannot use the customer, card, merchant, category, and location nodes exposed by the benchmark. As shown in Table 2, homogeneous propagation is also not consistently beneficial: GCN trails MLP on CreditCard-Fraud, and GraphSAGE improves Macro-F1 there but falls below MLP on the larger graph. This suggests that the main challenge is not merely adding edges, but learning from the heterogeneous relational structure preserved in FinFraudBench. Heterogeneous graph models are strongest overall. HGT achieves the best accuracy, Macro-F1, Fraud F1, and Fraud Recall on CreditCard-Fraud, while GAAP slightly leads AUROC and AUPRC. On BankTrans-Fraud, HGT obtains the best AUROC and AUPRC, while SeHGNN leads the classification metrics. R-GCN also improves substantially on the larger graph, indicating that relation semantics remain useful even without full node-type heterogeneity. Overall, the results support the benchmark motivation that relation semantics, node types, and type-specific attributes provide complementary fraud signals 37; 41; 45; 19. Relation-aware and fraud-oriented biases remain useful but are not sufficient alone. GAAP obtains the best AUROC and AUPRC on CreditCard-Fraud, while ConsisGAD and PMP remain competitive on BankTrans-Fraud 6; 49; 12. Their gap to the strongest heterogeneous models on threshold-dependent classification metrics suggests that relation channels or fraud-oriented assumptions should be combined with explicit entity-type modeling. Several methods obtain high AUPRC but differ substantially in Fraud Recall; Figure 6 illustrates this metric mismatch and supports reporting both ranking metrics and threshold-dependent classification metrics. The LLM prompting baselines are weaker than most graph learning baselines, although 4-shot in-context prompting consistently improves over zero-shot prompting on both datasets, providing a target-only non-GNN reference point for the benchmark. Figure 6: AUPRC–Fraud Recall frontier for trained non-LLM baselines, showing that stronger ranking performance does not always lead to higher fraud recall at the selected decision threshold. Higher values indicate better performance on both axes. 6.3 Scalability and Stability Analysis Scalability is part of the benchmark rather than a separate engineering detail. CreditCard-Fraud contains 18.53M directed typed edges, and BankTrans-Fraud contains 89.23M, so practical methods must combine predictive quality with sampling or batching strategies. This is especially relevant for graph-based baselines, which can use unlabeled training transactions as context but must do so without full-batch propagation. The balanced mini-test protocol yields stable estimates for most trained baselines across the five subsets, while the two LLM prompting probes show larger standard deviations and are therefore less stable than the graph-learning methods. 7 Discussion 7.1 Key Findings The experiments suggest that FinFraudBench is useful not only for ranking baselines, but also for diagnosing which parts of the released schema a method can exploit. Transaction features, relation channels, entity types, type-specific attributes, and fraud-oriented assumptions contribute complementary signals for multi-entity financial fraud detection. Heterogeneous graph design gives the strongest overall signal. Across experiments, heterogeneous methods provide the most consistent gains. This suggests that FinFraudBench exposes useful signals through entity types, relation semantics, and type-specific attributes. Homogeneous or single-node-type multi-relational projections simplify this structure and can discard part of the original fraud entity-relation information. Fraud-oriented designs can build on heterogeneous graphs. Multi-relation and fraud-oriented methods remain competitive on several metrics, showing that relation channels and fraud-oriented inductive biases are useful for financial fraud detection. However, their remaining gap to the strongest heterogeneous methods suggests that these cues are most promising when combined with entity types and type-specific attributes in heterogeneous financial graphs. 7.2 Limitations and Future Work FinFraudBench is constructed from public transaction-level tables, so its graph signals are limited by the fields exposed in those sources and may not fully reflect richer real-world fraud contexts. Future work can extend the benchmark with more complete operational signals and develop fraud-oriented models designed directly for heterogeneous fraud graphs under limited labels and extreme class imbalance in realistic deployment settings. 8 Conclusion We presented FinFraudBench, a benchmark with two heterogeneous graph datasets for financial fraud detection constructed from public transaction-level fraud sources. The proposed graphs preserve multiple financial entity types, relation types, type-specific attributes, limited labels, and realistic fraud imbalance. Our baseline reproduction provides family-level insights into how model designs use heterogeneous fraud signals. The results show that heterogeneous GNNs perform best. Multi-relation and fraud-oriented methods remain competitive but less consistent, highlighting the value of preserving heterogeneous financial structure. Data Availability The original tabular datasets used in this work are publicly available on Kaggle: https://w.kaggle.com/datasets/kartik2112/fraud-detection and https://w.kaggle.com/datasets/computingvictor/transactions-fraud-datasets. Ethics and Reproducibility Statement FinFraudBench is built from public transaction-level data and does not involve human-subject experiments or new data collection. The benchmark is intended for fraud detection research rather than operational decision making without additional validation. All preprocessing, split construction, and reported experiments are described in the paper and appendix. The code repository has been released openly, and the constructed datasets are available on Hugging Face with fixed random splits to support reproduction of the main results. AI Use Statement We used AI-assisted tools to support literature search and survey, manuscript writing, language polishing, and formatting during paper preparation. These tools were not used to design the proposed benchmark, construct the datasets, define the experimental protocol, or draw scientific conclusions. All scientific claims, dataset construction, experimental analysis, and final wording were carefully reviewed and approved by the authors. References Akoglu et al. (2015) L. Akoglu, H. Tong, and D. Koutra Graph based anomaly detection and description: a survey. Data mining and knowledge discovery 29 (3), p. 626–688. Cited by: §1, §2.2. Bolton and Hand (2002) R. J. Bolton and D. J. Hand Statistical fraud detection: a review. Statistical science 17 (3), p. 235–255. Cited by: §1, §2.2. Breunig et al. (2000) M. M. Breunig, H. Kriegel, R. T. Ng, and J. Sander LOF: identifying density-based local outliers. In Proceedings of the 2000 ACM SIGMOD international conference on Management of data, p. 93–104. Cited by: §2.2. Brown et al. (2020) T. Brown, B. Mann, N. Ryder, M. Subbiah, J. D. Kaplan, P. Dhariwal, A. Neelakantan, P. Shyam, G. Sastry, A. Askell, et al. Language models are few-shot learners. Advances in neural information processing systems 33, p. 1877–1901. Cited by: §6. Chandola et al. (2009) V. Chandola, A. Banerjee, and V. Kumar Anomaly detection: a survey. ACM computing surveys (CSUR) 41 (3), p. 1–58. Cited by: §2.2. Chen et al. (2024) N. Chen, Z. Liu, B. Hooi, B. He, R. Fathony, J. Hu, and J. Chen Consistency training with learnable data augmentation for graph anomaly detection with limited supervision. In The twelfth international conference on learning representations, Cited by: §2.2, §6.2, §6. Chiang et al. (2019) W. Chiang, X. Liu, S. Si, Y. Li, S. Bengio, and C. Hsieh Cluster-gcn: an efficient algorithm for training deep and large graph convolutional networks. In Proceedings of the 25th ACM SIGKDD International Conference on Knowledge Discovery & Data Mining, KDD ’19, New York, NY, USA, p. 257–266. External Links: ISBN 9781450362016, Link, Document Cited by: §2.2. ComputingVictor (2024) ComputingVictor Financial transactions dataset: analytics. Note: Kaggle datasetAccessed: 2026-07-18 External Links: Link Cited by: §1, §3.1. Davis and Goadrich (2006) J. Davis and M. Goadrich The relationship between precision-recall and roc curves. In Proceedings of the 23rd International Conference on Machine Learning, ICML ’06, New York, NY, USA, p. 233–240. External Links: ISBN 1595933832, Link, Document Cited by: §5.2. Dong et al. (2017) Y. Dong, N. V. Chawla, and A. Swami Metapath2vec: scalable representation learning for heterogeneous networks. In Proceedings of the 23rd ACM SIGKDD International Conference on Knowledge Discovery and Data Mining, KDD ’17, New York, NY, USA, p. 135–144. External Links: ISBN 9781450348874, Link, Document Cited by: §1, §2.2. Dou et al. (2020) Y. Dou, Z. Liu, L. Sun, Y. Deng, H. Peng, and P. S. Yu Enhancing graph neural network-based fraud detectors against camouflaged fraudsters. In Proceedings of the 29th ACM international conference on information & knowledge management, p. 315–324. Cited by: §1, §2.2, §2.2. Duan et al. (2025) M. Duan, D. He, T. Zheng, L. Jia, M. Song, X. Wang, and Z. Feng Global attribute-association pattern aggregation for graph fraud detection. In Proceedings of the AAAI Conference on Artificial Intelligence, Vol. 39, p. 11616–11624. Cited by: §2.2, §6.2, §6. Dwivedi et al. (2023) V. P. Dwivedi, C. K. Joshi, A. T. Luu, T. Laurent, Y. Bengio, and X. Bresson Benchmarking graph neural networks. Journal of Machine Learning Research 24 (43), p. 1–48. Cited by: §2.2. Fawcett (2006) T. Fawcett An introduction to roc analysis. Pattern Recognition Letters 27 (8), p. 861–874. Note: ROC Analysis in Pattern Recognition External Links: ISSN 0167-8655, Document, Link Cited by: §5.2. Fu et al. (2020) X. Fu, J. Zhang, Z. Meng, and I. King Magnn: metapath aggregated graph neural network for heterogeneous graph embedding. In Proceedings of the web conference 2020, p. 2331–2341. Cited by: §2.2. Gilmer et al. (2017) J. Gilmer, S. S. Schoenholz, P. F. Riley, O. Vinyals, and G. E. Dahl Neural message passing for quantum chemistry. In Proceedings of the 34th International Conference on Machine Learning, D. Precup and Y. W. Teh (Eds.), Proceedings of Machine Learning Research, Vol. 70, p. 1263–1272. External Links: Link Cited by: §2.2. Hamilton et al. (2017) W. Hamilton, Z. Ying, and J. Leskovec Inductive representation learning on large graphs. Advances in neural information processing systems 30. Cited by: §2.2, §6. Hu et al. (2020a) W. Hu, M. Fey, M. Zitnik, Y. Dong, H. Ren, B. Liu, M. Catasta, and J. Leskovec Open graph benchmark: datasets for machine learning on graphs. Advances in neural information processing systems 33, p. 22118–22133. Cited by: §2.2. Hu et al. (2020b) Z. Hu, Y. Dong, K. Wang, and Y. Sun Heterogeneous graph transformer. In Proceedings of the web conference 2020, p. 2704–2710. Cited by: §1, §2.2, §6.2, §6. Huang et al. (2022) X. Huang, Y. Yang, Y. Wang, C. Wang, Z. Zhang, J. Xu, L. Chen, and M. Vazirgiannis Dgraph: a large-scale financial dataset for graph anomaly detection. Advances in Neural Information Processing Systems 35, p. 22765–22777. Cited by: §1, §2.2. Kipf and Welling (2016) T. N. Kipf and M. Welling Semi-supervised classification with graph convolutional networks. arXiv preprint arXiv:1609.02907. Cited by: §2.2, §6. Kojima et al. (2022) T. Kojima, S. S. Gu, M. Reid, Y. Matsuo, and Y. Iwasawa Large language models are zero-shot reasoners. Advances in neural information processing systems 35, p. 22199–22213. Cited by: §6. Li et al. (2026) Y. Li, J. Hu, B. Hooi, B. He, and C. Chen Dgp: a dual-granularity prompting framework for fraud detection with graph-enhanced llms. In Proceedings of the AAAI Conference on Artificial Intelligence, Vol. 40, p. 15171–15179. Cited by: §2.2. Lin et al. (2025) J. Lin, X. Guo, S. Zhang, Y. Zhu, and J. Shun When heterophily meets heterogeneity: challenges and a new large-scale graph benchmark. In Proceedings of the 31st ACM SIGKDD Conference on Knowledge Discovery and Data Mining V. 2, p. 5607–5618. Cited by: §2.2. Liu et al. (2008) F. T. Liu, K. M. Ting, and Z. Zhou Isolation forest. In 2008 eighth ieee international conference on data mining, p. 413–422. Cited by: §2.2. Liu et al. (2022) K. Liu, Y. Dou, Y. Zhao, X. Ding, X. Hu, R. Zhang, K. Ding, C. Chen, H. Peng, K. Shu, et al. Bond: benchmarking unsupervised outlier node detection on static attributed graphs. Advances in Neural Information Processing Systems 35, p. 27021–27035. Cited by: §1. Liu et al. (2020) Z. Liu, Y. Dou, P. S. Yu, Y. Deng, and H. Peng Alleviating the inconsistency problem of applying graph neural network to fraud detection. In Proceedings of the 43rd international ACM SIGIR conference on research and development in information retrieval, p. 1569–1572. Cited by: §1, §2.2, §2.2. Liu et al. (2018) Z. Liu, C. Chen, X. Yang, J. Zhou, X. Li, and L. Song Heterogeneous graph neural networks for malicious account detection. In Proceedings of the 27th ACM international conference on information and knowledge management, p. 2077–2085. Cited by: §2.2. Lv et al. (2021) Q. Lv, M. Ding, Q. Liu, Y. Chen, W. Feng, S. He, C. Zhou, J. Jiang, Y. Dong, and J. Tang Are we really making much progress? revisiting, benchmarking and refining heterogeneous graph neural networks. In Proceedings of the 27th ACM SIGKDD conference on knowledge discovery & data mining, p. 1150–1160. Cited by: §2.2. Ma et al. (2023) X. Ma, J. Wu, S. Xue, J. Yang, C. Zhou, Q. Z. Sheng, H. Xiong, and L. Akoglu A comprehensive survey on graph anomaly detection with deep learning. IEEE Transactions on Knowledge and Data Engineering 35 (12), p. 12012–12038. External Links: Document Cited by: §1, §2.2. McAuley and Leskovec (2013) J. J. McAuley and J. Leskovec From amateurs to connoisseurs: modeling the evolution of user expertise through online reviews. In Proceedings of the 22nd International Conference on World Wide Web, W ’13, New York, NY, USA, p. 897–908. External Links: ISBN 9781450320351, Link, Document Cited by: §1, §2.2. Min et al. (2022) S. Min, X. Lyu, A. Holtzman, M. Artetxe, M. Lewis, H. Hajishirzi, and L. Zettlemoyer Rethinking the role of demonstrations: what makes in-context learning work?. In Proceedings of the 2022 Conference on Empirical Methods in Natural Language Processing, Y. Goldberg, Z. Kozareva, and Y. Zhang (Eds.), Abu Dhabi, United Arab Emirates, p. 11048–11064. External Links: Link, Document Cited by: §6. Phua et al. (2010) C. Phua, V. Lee, K. Smith, and R. Gayler A comprehensive survey of data mining-based fraud detection research. arXiv preprint arXiv:1009.6119. Cited by: §1, §2.2. Rayana and Akoglu (2015) S. Rayana and L. Akoglu Collective opinion spam detection: bridging review networks and metadata. In Proceedings of the 21th ACM SIGKDD International Conference on Knowledge Discovery and Data Mining, KDD ’15, New York, NY, USA, p. 985–994. External Links: ISBN 9781450336642, Link, Document Cited by: §1, §2.2. Saito and Rehmsmeier (2015) T. Saito and M. Rehmsmeier The precision-recall plot is more informative than the roc plot when evaluating binary classifiers on imbalanced datasets. PloS one 10 (3), p. e0118432. Cited by: §5.2. Scarselli et al. (2009) F. Scarselli, M. Gori, A. C. Tsoi, M. Hagenbuchner, and G. Monfardini The graph neural network model. IEEE Transactions on Neural Networks 20 (1), p. 61–80. External Links: Document Cited by: §2.2. Schlichtkrull et al. (2018) M. Schlichtkrull, T. N. Kipf, P. Bloem, R. Van Den Berg, I. Titov, and M. Welling Modeling relational data with graph convolutional networks. In European semantic web conference, p. 593–607. Cited by: §2.2, §6.2, §6. Shenoy (2020) K. Shenoy Credit card transactions fraud detection dataset. Kaggle. Available online: https://w. kaggle. com/datasets/kartik2112/fraud-detection (accessed on 29 July 2023). Cited by: §1, §3.1. Shi et al. (2017) C. Shi, Y. Li, J. Zhang, Y. Sun, and P. S. Yu A survey of heterogeneous information network analysis. IEEE Transactions on Knowledge and Data Engineering 29 (1), p. 17–37. External Links: Document Cited by: §1, §2.2. Veličković et al. (2017) P. Veličković, G. Cucurull, A. Casanova, A. Romero, P. Lio, and Y. Bengio Graph attention networks. arXiv preprint arXiv:1710.10903. Cited by: §2.2. Wang et al. (2019) X. Wang, H. Ji, C. Shi, B. Wang, Y. Ye, P. Cui, and P. S. Yu Heterogeneous graph attention network. In The world wide web conference, p. 2022–2032. Cited by: §2.2, §6.2, §6. Weber et al. (2019) M. Weber, G. Domeniconi, J. Chen, D. K. I. Weidele, C. Bellei, T. Robinson, and C. E. Leiserson Anti-money laundering in bitcoin: experimenting with graph convolutional networks for financial forensics. arXiv preprint arXiv:1908.02591. Cited by: §1, §2.2. Wu et al. (2021) Z. Wu, S. Pan, F. Chen, G. Long, C. Zhang, and P. S. Yu A comprehensive survey on graph neural networks. IEEE Transactions on Neural Networks and Learning Systems 32 (1), p. 4–24. External Links: Document Cited by: §2.2. Yang et al. (2024) A. Yang, B. Yang, B. Hui, B. Zheng, B. Yu, C. Zhou, C. Li, C. Li, D. Liu, F. Huang, G. Dong, H. Wei, H. Lin, J. Tang, J. Wang, J. Yang, J. Tu, J. Zhang, J. Ma, J. Yang, J. Xu, J. Zhou, J. Bai, J. He, J. Lin, K. Dang, K. Lu, K. Chen, K. Yang, M. Li, M. Xue, N. Ni, P. Zhang, P. Wang, R. Peng, R. Men, R. Gao, R. Lin, S. Wang, S. Bai, S. Tan, T. Zhu, T. Li, T. Liu, W. Ge, X. Deng, X. Zhou, X. Ren, X. Zhang, X. Wei, X. Ren, X. Liu, Y. Fan, Y. Yao, Y. Zhang, Y. Wan, Y. Chu, Y. Liu, Z. Cui, Z. Zhang, Z. Guo, and Z. Fan Qwen2 technical report. External Links: 2407.10671, Link Cited by: §6. Yang et al. (2023) X. Yang, M. Yan, S. Pan, X. Ye, and D. Fan Simple and efficient heterogeneous graph neural network. In Proceedings of the AAAI conference on artificial intelligence, Vol. 37, p. 10816–10824. Cited by: §2.2, §6.2, §6. Yun et al. (2019) S. Yun, M. Jeong, R. Kim, J. Kang, and H. J. Kim Graph transformer networks. Advances in neural information processing systems 32. Cited by: §2.2. Zeng et al. (2019) H. Zeng, H. Zhou, A. Srivastava, R. Kannan, and V. Prasanna Graphsaint: graph sampling based inductive learning method. arXiv preprint arXiv:1907.04931. Cited by: §2.2. Zhang et al. (2020) S. Zhang, H. Yin, T. Chen, Q. V. N. Hung, Z. Huang, and L. Cui GCN-based user representation learning for unifying robust recommendation and fraudster detection. In Proceedings of the 43rd International ACM SIGIR Conference on Research and Development in Information Retrieval, SIGIR ’20, New York, NY, USA, p. 689–698. External Links: ISBN 9781450380164, Link, Document Cited by: §2.2. Zhuo et al. (2024) W. Zhuo, Z. Liu, B. Hooi, B. He, G. Tan, R. Fathony, and J. Chen Partitioning message passing for graph fraud detection. arXiv preprint arXiv:2412.00020. Cited by: §2.2, §6.2, §6. Appendix A Additional Dataset Details A.1 Type-Specific Feature Fields Table 3 lists representative non-label feature groups used to build node attributes. The two graphs share the same node and relation schema, but the feature dimension of the same node type can differ across datasets because the two source tables expose different raw fields. In all cases, both graphs avoid label leakage in graph and feature construction. Table 3: Representative non-label feature groups by node type. Node type Representative feature groups transaction amount, time fields, location fields, distance or transaction context customer profile encodings, location fields, transaction counts and amount statistics card billing/context encodings, transaction counts, amount statistics, unique entity counts merchant category encoding, transaction counts, amount statistics, unique customer/card/state counts category transaction counts, amount statistics, unique merchant/customer/card counts location transaction counts, amount statistics, unique entity counts, location/population summaries A.2 Detailed Transaction Splits Table 4 reports the split statistics behind the compact split summary in the main paper. The unlabeled training split keeps transaction nodes available as graph context while hiding their labels from supervised training. Table 4: Detailed transaction split statistics. Dataset Split #Transactions #Fraud #Benign Fraud rate CreditCard-Fraud train_labeled 37.05K 214 36.83K 0.5776% train_unlabeled 1.22M 7.08K 1.22M 0.5789% validation 37.05K 214 36.83K 0.5776% full_test 555.72K 2.15K 553.57K 0.3860% BankTrans-Fraud train_labeled 178.30K 268 178.03K 0.1503% train_unlabeled 6.78M 10.13K 6.77M 0.1495% validation 178.30K 268 178.03K 0.1503% full_test 1.78M 2.67K 1.78M 0.1495% A.3 Complete Balanced Mini-Test Results Tables 5 and 6 report the complete balanced mini-test results, including Fraud Precision, which is omitted from the main table to keep the main comparison compact. Table 5: Complete balanced mini-test results on CreditCard-Fraud. Values are means and standard deviations across five mini-test subsets. Method AUROC AUPRC Acc. Macro-F1 Fraud F1 Fraud Prec. Fraud Recall MLP 0.8798±0.00570.8798± 0.0057 0.9046±0.00520.9046± 0.0052 0.7539±0.00500.7539± 0.0050 0.7393±0.00590.7393± 0.0059 0.6775±0.00870.6775± 0.0087 0.9822±0.00200.9822± 0.0020 0.5172±0.01030.5172± 0.0103 GCN 0.8199±0.01070.8199± 0.0107 0.8632±0.00670.8632± 0.0067 0.5416±0.00330.5416± 0.0033 0.4204±0.00650.4204± 0.0065 0.1554±0.01150.1554± 0.0115 0.9861±0.00990.9861± 0.0099 0.0844±0.00680.0844± 0.0068 GraphSAGE 0.9046±0.00590.9046± 0.0059 0.9279±0.00440.9279± 0.0044 0.7681±0.01020.7681± 0.0102 0.7556±0.01180.7556± 0.0118 0.7003±0.01670.7003± 0.0167 0.9890±0.00440.9890± 0.0044 0.5422±0.01950.5422± 0.0195 R-GCN 0.8870±0.00720.8870± 0.0072 0.9163±0.00440.9163± 0.0044 0.6876±0.00660.6876± 0.0066 0.6542±0.00880.6542± 0.0088 0.5466±0.01370.5466± 0.0137 0.9957±0.00300.9957± 0.0030 0.3768±0.01280.3768± 0.0128 HAN 0.9025±0.00380.9025± 0.0038 0.9269±0.00240.9269± 0.0024 0.6858±0.00800.6858± 0.0080 0.6515±0.01080.6515± 0.0108 0.5422±0.01700.5422± 0.0170 0.9978±0.00230.9978± 0.0023 0.3724±0.01600.3724± 0.0160 SeHGNN 0.9026±0.00480.9026± 0.0048 0.9279±0.00260.9279± 0.0026 0.7575±0.00590.7575± 0.0059 0.7426±0.00730.7426± 0.0073 0.6807±0.01110.6807± 0.0111 0.9959±0.00410.9959± 0.0041 0.5172±0.01370.5172± 0.0137 HGT 0.9150±0.00350.9150± 0.0035 0.9339±0.00360.9339± 0.0036 0.8317±0.00770.8317± 0.0077 0.8271±0.00820.8271± 0.0082 0.7990±0.01060.7990± 0.0106 0.9911±0.00420.9911± 0.0042 0.6694±0.01400.6694± 0.0140 ConsisGAD 0.9110±0.00420.9110± 0.0042 0.9315±0.00310.9315± 0.0031 0.6526±0.00580.6526± 0.0058 0.6049±0.00840.6049± 0.0084 0.4676±0.01350.4676± 0.0135 1.0000±0.00001.0000± 0.0000 0.3052±0.01150.3052± 0.0115 PMP 0.8683±0.00750.8683± 0.0075 0.9029±0.00530.9029± 0.0053 0.6992±0.00680.6992± 0.0068 0.6697±0.00860.6697± 0.0086 0.5709±0.01290.5709± 0.0129 0.9949±0.00480.9949± 0.0048 0.4004±0.01200.4004± 0.0120 GAAP 0.9237±0.00370.9237± 0.0037 0.9344±0.00330.9344± 0.0033 0.7155±0.00770.7155± 0.0077 0.6915±0.00960.6915± 0.0096 0.6054±0.01420.6054± 0.0142 0.9873±0.00620.9873± 0.0062 0.4366±0.01420.4366± 0.0142 LLM-ZS 0.6591±0.00460.6591± 0.0046 0.6877±0.00860.6877± 0.0086 0.5510±0.00320.5510± 0.0032 0.4519±0.00390.4519± 0.0039 0.2189±0.00590.2189± 0.0059 0.8412±0.02490.8412± 0.0249 0.1258±0.00360.1258± 0.0036 LLM-ICL-4 0.8067±0.04170.8067± 0.0417 0.8495±0.03600.8495± 0.0360 0.6839±0.01200.6839± 0.0120 0.6501±0.01640.6501± 0.0164 0.5414±0.02600.5414± 0.0260 0.9841±0.00510.9841± 0.0051 0.3739±0.02470.3739± 0.0247 Table 6: Complete balanced mini-test results on BankTrans-Fraud. Values are means and standard deviations across five mini-test subsets. Method AUROC AUPRC Acc. Macro-F1 Fraud F1 Fraud Prec. Fraud Recall MLP 0.9384±0.00430.9384± 0.0043 0.9456±0.00410.9456± 0.0041 0.6718±0.00300.6718± 0.0030 0.6329±0.00420.6329± 0.0042 0.5133±0.00680.5133± 0.0068 0.9926±0.00320.9926± 0.0032 0.3462±0.00630.3462± 0.0063 GCN 0.9476±0.00320.9476± 0.0032 0.9486±0.00300.9486± 0.0030 0.7257±0.00520.7257± 0.0052 0.7055±0.00650.7055± 0.0065 0.6285±0.00990.6285± 0.0099 0.9732±0.00510.9732± 0.0051 0.4642±0.01110.4642± 0.0111 GraphSAGE 0.9263±0.00190.9263± 0.0019 0.9302±0.00260.9302± 0.0026 0.5460±0.00430.5460± 0.0043 0.4286±0.00820.4286± 0.0082 0.1696±0.01440.1696± 0.0144 0.9915±0.00930.9915± 0.0093 0.0928±0.00860.0928± 0.0086 R-GCN 0.9615±0.00390.9615± 0.0039 0.9717±0.00280.9717± 0.0028 0.8542±0.00820.8542± 0.0082 0.8510±0.00870.8510± 0.0087 0.8292±0.01120.8292± 0.0112 1.0000±0.00001.0000± 0.0000 0.7084±0.01640.7084± 0.0164 HAN 0.9687±0.00280.9687± 0.0028 0.9766±0.00210.9766± 0.0021 0.8813±0.00430.8813± 0.0043 0.8796±0.00440.8796± 0.0044 0.8654±0.00550.8654± 0.0055 0.9995±0.00120.9995± 0.0012 0.7630±0.00880.7630± 0.0088 SeHGNN 0.9689±0.00190.9689± 0.0019 0.9784±0.00120.9784± 0.0012 0.8838±0.00320.8838± 0.0032 0.8822±0.00340.8822± 0.0034 0.8686±0.00410.8686± 0.0041 0.9995±0.00070.9995± 0.0007 0.7680±0.00650.7680± 0.0065 HGT 0.9729±0.00200.9729± 0.0020 0.9784±0.00090.9784± 0.0009 0.7793±0.00620.7793± 0.0062 0.7680±0.00710.7680± 0.0071 0.7170±0.01000.7170± 0.0100 0.9989±0.00160.9989± 0.0016 0.5592±0.01160.5592± 0.0116 ConsisGAD 0.9657±0.00310.9657± 0.0031 0.9751±0.00190.9751± 0.0019 0.8221±0.00640.8221± 0.0064 0.8163±0.00710.8163± 0.0071 0.7835±0.00950.7835± 0.0095 1.0000±0.00001.0000± 0.0000 0.6442±0.01290.6442± 0.0129 PMP 0.9592±0.00360.9592± 0.0036 0.9685±0.00270.9685± 0.0027 0.8163±0.00900.8163± 0.0090 0.8099±0.00990.8099± 0.0099 0.7752±0.01340.7752± 0.0134 0.9978±0.00090.9978± 0.0009 0.6340±0.01790.6340± 0.0179 GAAP 0.9493±0.00490.9493± 0.0049 0.9544±0.00420.9544± 0.0042 0.6427±0.00390.6427± 0.0039 0.5907±0.00600.5907± 0.0060 0.4447±0.00990.4447± 0.0099 0.9973±0.00370.9973± 0.0037 0.2862±0.00830.2862± 0.0083 LLM-ZS 0.8087±0.00720.8087± 0.0072 0.8127±0.00740.8127± 0.0074 0.5174±0.00410.5174± 0.0041 0.3720±0.00760.3720± 0.0076 0.0697±0.01310.0697± 0.0131 0.9587±0.05040.9587± 0.0504 0.0362±0.00700.0362± 0.0070 LLM-ICL-4 0.8725±0.01860.8725± 0.0186 0.8598±0.01200.8598± 0.0120 0.5406±0.03210.5406± 0.0321 0.4179±0.06280.4179± 0.0628 0.1513±0.11110.1513± 0.1111 0.9561±0.01440.9561± 0.0144 0.0856±0.06780.0856± 0.0678 A.4 Mini-Test Stability Figure 7 reports the coefficient of variation across the five balanced mini-test subsets for each metric and trained non-LLM baseline. The figure complements the averaged results in the main paper by showing that the balanced mini-test protocol gives stable estimates for most graph-learning baselines. Figure 7: Stability of balanced mini-test evaluation. Heatmaps report the coefficient of variation across five mini-test subsets; lighter cells indicate more stable estimates.