A Semi-supervised Graph Attentive Network for Financial Fraud Detection

“In our model SemiGNN, we randomly initialize parameters and optimize the model with Adam [14].”

“• Xgboost [28]: It is a tree-based model, which is very”

“• GCN [24]: It is a deep graph neural network and each node’s embedding is ensembled by its own embedding”

“DeepWalk [16] and Node2vec [17] propose to use the random walk and skip-gram to learn the node representations.”

“LINE [18] and SDNE [19] propose explicit objective functions for preserving first-and second-order proximity.”

“Through several layers of non-linear functions to map the raw data into the high-level semantic space, it is easier to correlate multiview data [27].”

“LINE [18] and SDNE [19] propose explicit objective functions for preserving first-and second-order proximity.”

“Recently, graph convolution network based methods are very popular [24], [25].”

“Operational environments further demand reliability, inter-pretability, and maintainability [68].”

“Graph Anomaly Detection (GAD) has emerged as a pivotal task in transaction monitoring and financial risk control systems, dedicated to identifying atypical graph entities (nodes and edges) that significantly deviate from global structures or local affinities [13, 20].”

“The preprocessing and transaction behavior feature extraction stages aim to improve the quality of data representation before anomaly detection and graph learning processes [10], [15], [16].”

“Closer aligned with the methodology used in this article is the literature on detecting irregularities in economic or financial networks, where, in order to incorporate structural information, node representations are used for downstream classification tasks [45–47].”

“Money laundering schemes typically involve coordinated interactions across multiple accounts and temporal sequences [34].”

“…PC-GNN [Liu et al. , 2021], CARE-GNN [Dou et al. , 2020], GAGA [Wang et al. , 2023], FRAUDRE [Zhang et al. , 2021], HOGRL [Zou and Cheng, 2024], DGA-GNN [Duan et al. , 2024]; and three semi-supervised ones: SemiGNN [Wang et al. , 2019], GTAN [Xiang et al. , 2023], RGTAN [Xiang et al. , 2025].”

“G RAPH neural networks (GNNs) have been widely used across various graph-structured data domains, including recommendation systems [1], financial systems [2], [3], and bioinformatics [4], [5].”

“…of internet-based systems and online services that encompass e-commerce platforms, social networks, and financial transaction systems has brought unprecedented convenience while also giving rise to sophisticated fraudulent activities (He et al. 2022; Kudugunta and Fer-rara 2018; Wang et al. 2019).”

“For example, a triggered fraudulent account could be aligned with a benign class and evade fraud screening [8], while attacker-chosen items could be surfaced to users in a recommender [6].”

“Prior work [35] assumed feature similarity between neighboring nodes and utilized random negative sampling to create an unsupervised loss.”

“As a result, they struggle to capture collusion, shared devices, synthetic identity rings, or coordinated merchant usage that manifest as structured patterns across accounts and infrastructure [8], [12].”

“This absence of a relational awareness presents a structural blind spot for fraud detection systems [6], [7],[18].”

“Wang et al. [10] constructed user-merchant bipartite graphs and achieved 97% recall on e-commerce fraud by leveraging merchant reputation propagation.”

“Representative examples include heterogeneous graph neural networks (GNN)-based fraud detection [57], graph attention-based modeling [55, 61], and dual-layer graph modeling for suspicious interaction patterns [45].”

“This is consistent with studies showing that GCNs can capture local graph topology for fraud detection [18, 19].”

“Anomaly detection in attributed graphs is critical for real-world applications such as financial fraud detection [10,11], cybersecurity monitoring, and social network analysis [12,13].”

“Research has demonstrated that GNNs outperform traditional machine learning methods by capturing dependencies among linked transactions [15].”

“Modern online information systems, such as e-commerce platforms, review communities, social networks, and financial services, generate rich relational data among users, items, content, and transactions [2, 20, 28, 32].”

“Wang et al. [7] developed SemiGNN, a semi-supervised graph attentive network deployed on Alipay with 4 million labeled and 100 million unlabeled users, demonstrating scalability through hierarchical attention mechanisms.”

“Wang et al. [17] applied GNNs to anti-money laundering with state-of-the-art results on the Elliptic Bitcoin dataset.”

“Wang et al. [22] built a transaction graph and used Graph Neural Networks, which obtained more detection performance on financial networks, especially against organised fraud rings.”

“GraphConsis [11] addresses the issue of label inconsistency among neighbours, and Wang et al. [10] apply graph attention to credit-card fraud in a semi-supervised manner.”

“…on graph-structured data (Wu et al. 2020; Zhou et al. 2020), achieving remarkable success in domains such as social network analysis (Fan et al. 2019), recommendation systems (Wu et al. 2019), drug discovery (Bongini, Bianchini, and Scarselli 2021), and financial fraud detection (Wang et al. 2019).”

“See Hand (2005), Anderson and Hardin (2009), Abdou and Pointon (2011), Roesch et al. (2016), Wang et al. (2019), Fang and Chen (2019), Alaraj et al. (2021) and many references therein, by way of illustration.”

“These architectural innovations have collectively established graph neural networks as state-of-the-art approaches for network-based fraud detection tasks [13].”

“A semi-supervised GNN model [6] employing node-level and view-level attention by integrating labeled and unlabeled data significantly enhances the detection performance compared to the baselines.”

“Subsequent studies, such as Wang et al. [3] in 2019 and Yang et al. [4] in 2019 and Hu et al. [5] in 2020, expanded the labeled data through their social relations to get the unlabeled data and proposed a semi-supervised model respectively to get a more accurate classification result.”

“Traditional machine learning models are widely used but struggle to leverage the structural and relational information present in these networks [4], [5].”

“…in the field of image processing with numerous citations [1], Graph Neural Networks (GNNs) have recently emerged as a powerful framework for addressing graph-based learning tasks across diverse domains such as social networks [2, 3], bioinformatics [4], wireless communication [5] and finance [6].”

“Graph structures provide an improved representation of user behavior within their network context [6].”

“These applications include financial risk control, social network analytics [2], and e-commerce recommendation systems [3].”

“Additionally, SS settings are increasingly relevant in many fields like image processing (Cheplygina et al., 2019), anomaly detection (Wang et al., 2019), empirical risk (Yuval and Rosset, 2022).”

“For supervised GAD studies, some methods [15], [29] use attention mechanisms to differentiate between neighbors during message aggregation, while others [4], [17] employ sampling or resampling strategies or design specialized loss functions [35], [36].”

“Wang et al. [10] proposed a semi-supervised attention graph neural network, which incorporated a hierarchical attention mechanism to enhance model interpretability and fraud detection performance.”

“The model integrated node-level and view-level attention mechanisms, and optimized the use of labeled and unlabeled data through specific loss functions [6].”

“Complex network data is prevalent in various real-world domains, including social media networks [1], financial transaction networks [2], and biochemical structures [3].”

“Some methods, such as Fdgars [17] and GraphConsis [18], construct a single homo-graph contains multiple relations, while some others, e.g, GEM [19] and SemiGNN [20], convert the heter-graphs to multiple homo-graphs based on the node relationships.”

“In recent years, we have witnessed the appearance of a significant number of works on graph anomaly detection using deep learning techniques, such as [26–28].”

“Recently, many studies introduce graph neural networks to capture the relation among the data samples in a variety of prediction applications[10-13].”

“Graph Neural Networks (GNNs) have been widely applied in a variety of fields, such as social network analysis Hamilton et al. (2017), drug discovery Li et al. (2021a), financial risk control Wang et al. (2019), and recommender systems Wu et al. (2022).”

“It is also noteworthy that various studies have investigated the more broad category of auditing tasks such as financial statement fraud detection [17], financial distress prediction [18], tax evasion risk management [19], and so on.”

“Although Graph Neural Network (GNN) has demonstrated its success in many graph learning applications such as link predictions [1], [2], fraud detection [3], [4], personalized recommendations [5]–[7], and pattern recognition [8], [9], most existing GNNs are only applicable to homogeneous graphs with single type of node…”

“As an illustrational example, in financial social networks, users and their social connections can be formed as a graph, and the credit of users in the graph can be predicted by GCNs [40], [41].”

“Supervised/semi-supervised methods[48],[49],[50],[52],[54]focusonleveraginglabeled normalandanomaloussamplestolearnpatternboundariesof bothcategoriestodetectanomalies.”

“Other than the Leiden algorithm and the spectral clustering approach, deep-learning-based fraud detection is the dominant paradigm under intense study [52], [53], [54].”

“Extensiveempiricalexperimentalresults onbothsyntheticandreal-worlddatasetsvalidateitssuperiority.”

“As the development of GNNs in recent years, many GNN-based fraud detectors have been proposed [46, 45, 73, 74, 90, 91, 77], where FdGars [74], GAS [45], and ASA [77] adopt graph convolutional networks (GCN), a variant of GNN, while SemiGNN [73] and Player2Vec [90] adopt graph attention network (GAT), another GNN varia…”