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].”

“Wang et al. designed a semi-supervised graph attentive network to detect fraud from a multi-view (i.e., users’ relationships, attributes, and devices) graph [27].”

“distinct impacts of different entities, [37] devises a hierarchical attention model to better bridge different neighbors for online fraud detection.”

“Graphs serve as a ubiquitous form of data structure [1] in many real-world applications such as social networks [2], recommendation systems [3], and financial applications [4], where nodes represent entities and edges capture their relationships.”

“Graph-structured data are prevalent in numerous real-world applications, such as social networks [22], financial systems [52], and molecular graphs [53].”

“In view of insufficient labeled data, Wang et al. [369] propose a semi-supervised learning-based attention graph neural network method for fraud detection.”

“…(SVM), as well as seventeen state-of-the-art graph learning baselines for continual learning tasks, including two static GNN methods (SemiGNN [35] and CARE [3]), seven temporal GNN methods (TGAT [41], TADDY [18], MADG [47], GTAN [40], SAD [27], AdaDyGNN [16], and RGTAN [39]), four static GCL…”

“Representative models include GCN-style attribute propagation (Luo et al., 2024), inductive sampling-based methods such as Graph-SAGE (Hamilton et al., 2017), and attention-driven neighbor aggregation (Wang et al., 2019).”

“Attention based GNNs highlight important neighbours in large transaction networks [14], and risk diffusion style GNNs propagate risk scores along transaction paths to improve detection of collusive structures and long money trails [15].”

“• Semi-GNN [15]: This model combines semi-supervised loss function and label propagation mechanism to alleviate the problem of scarce labeled data in fraud detection.”

“Graph neural networks (GNNs) have shown powerful capability in graph representation learning and achieved remarkable performance for various applications on graph-structured data [39, 47, 52].”

“Early works (Wang et al. 2019) were motivated by GNN’s ability to consider local structure through neighborhood aggregation while succeeding ones focused on designing models based on certain characteristics associated with fraud.”

“…(GNNs) [10, 11, 23, 30, 58, 79] are a powerful approach for processing graph-structured data, making them particularly effective in applications such as recommendation systems [17, 25, 57, 66, 71, 77], social network analysis [2, 8, 40], and financial fraud detection [26, 37, 61, 67, 78, 82].”

“Graph neural networks have emerged as a complementary technique by modeling relationships between entities such as users, mer-chants, and devices [10], excelling at detecting collusive fraud patterns but remaining limited in their ability to process heterogeneous feature types and provide…”

“To address the issue, few-shot learning approaches [7, 12, 51] have been explored to leverage a limited number of labeled anomalies.”

“To tackle the aforementioned problems, numerous fraud detection models based on GNNs utilize techniques such as adjusting the importance of neighbours from different classes using predetermined indicators [23] or employing attention mechanisms that can be learned [19,28–30].”

“Real-world applications may operate under constraints such as lightweight computation, limited storage, and strict privacy requirements, which make reprocessing of source data generally infeasible (Wang et al., 2021c; Wu et al., 2021).”

“For example, a bunch of graph-based algorithms is proposed to detect credit card frauds either in real-world scenarios [10] or in cryptocurrency transactions [11].”

“Such methods based on attention mechanisms are frequent on NLAD tasks [39, 42].”

“Extensiveempiricalexperimentalresults onbothsyntheticandreal-worlddatasetsvalidateitssuperiority.”

“The database provides indexing, which facilitates the quick and efficient retrieval of data and improves query performance by reducing the number of nodes and edges that need to be traversed for retrieving data [20].”

“It is used in different fields such as insurance [23], [41], or finance [34].”

“…: LOF [3], Autoencoder [68], Deep-SAD [40] and MI-Net [22] where only the node attributes are used, and (2) graph-based : Radar [29], DOMINANT [8], SemiGNN [50], GDN [13], CARE-GNN [14], MI-GNN [48], CoLA [32], SL-GAD [67], and BWGNN [45] where both graph topological information and node…”

“…areas, including genomics (e.g., Eisen et al. 1998), image processing (Jain and Flynn 1996; Cheplygina, de Bruijne, and Pluim 2019), natural language processing (Liang 2005; Turian, Ratinov, and Bengio 2010) and anomaly detection (Akcay, Atapour-Abarghouei, and Breckon 2019; Wang et al. 2019).”

“Graph is a ubiquitous form in a variety of applications such as social networks [9,25], bioscience [4,20], finance [19], knowledge graph [21], traffic networks [1], etc.”

“Learning on Graphs (LoG) in multi-client systems has extensive applications such as Graph Neural Networks (GNNs) for recommendation [1–4], finance [5, 6], and traffic [7, 8].”

“FdGars [30], DCI [32] and SemiGNN [28] tend to summarize the abnormal patterns automatically using GNNs to detect fraudsters.”

“Over last few years, Graph Convolutional Networks (GCNs) have benefited many real world applications across different domains, such as molecular designs [30], financial fraud detection [22] and traffic prediction [23, 31].”

“Some existing methods (Wang et al., 2019; Hu et al., 2023; Liu et al., 2021; Hu et al., 2022) use graph neural networks (GNNs) to identify suspicious scam-mers or abnormal attributes in mobile social networks.”

“Early approaches like FdGars [Wang et al., 2019b] and CARE-GNN [Dou et al., 2020] focused on user classification and neighbor aggregation respectively.”

“Such rationales are not simply uninformative; they are auditability hazards, because they create an appearance of justification that does not correspond to the underlying evidence [16, 17].”

“Representative methods include the semi-supervised graph attention network of [39], the camouflage-resistant CARE-GNN [15], and the imbalance-aware PC-GNN [25].”

“Many existing methods [3], [6], [16] directly apply graph neural networks (GNNs) to classify which accounts are at risk.”

“[26 ] proposed a semi-supervised graph attention network (SemiGNN) to detect financial fraud.”

“6) SemiGNN (GB) [5]: It is a semi-supervised GNN that combines supervised and unsupervised learning.”

“Existing methods that aim to elevate this problem include applying an attention mechanism [19, 33] or performing neighborhood resampling [21, 22].”

“Wang et al. [28] proposed a semi-supervised graph attention network for financial deception detection on multi-view, labeled, and unlabeled data.”

“GNNs, such as CARE-GNN and SemiGNN, have demonstrated notable success in modeling dependencies across nodes and incorporating both relational and temporal information for improved detection performance [15, 39].”

“Group fairness mainly appears in high-risk applications, such as recidivism prediction [27], loan evaluation [28], etc. Sensitive information, such as race, skin color, and gender, is protected by law and should not be used as the basis for decision-making.”

“Advanced Graph Neural Network models like Graph Attention Networks [9], Care-GNN [8], and Semi-GNN [10] improved decision-making but remained black-box and costly.”

“Many scholars [10][11][12][13][14] consider using graph neural network (GNN) to conduct graph abnormal nodes detection. owing to their impressive performance in learning representations.”

“Graph-based methods model the interactions in a heterogeneous account–device graph (Liu et al., 2018; Wang DX et al., 2019) or a social graph (Liang et al., 2021) to detect malicious accounts or predict credit risk, based on the device aggregation or social aggregation of fraudsters.”

“Graph Neural Networks (GNNs) have emerged as a promising learning paradigm in modeling graph data [35, 38, 39].”

“Traditional anomaly-detection methods in financial supervision detect intuitive node(card)- or edge(transaction)-level anomalies by pre-defined rules [6], [7] or learning-based strategies [20], [21].”

“However, most of the prior methods [13, 14, 37, 40] focused exclusively on either accuracy or robustness of the fraud detectors, ignoring the detection fairness.”

“This dataset can predict whether a user can afford to pay and how much credit should be offered to the user[8].”

“For heterogeneous graphs, SemiGNN [7] use user-relation, user-app, user-nickname to construct multi-view graphs.”

“Our B-tests and W-tests partially solve this problem, but there are other strong methods for fraud feature engineering, such as graph methods [15, 60, 61, 67], entropy changing methods [17], and variance anomaly detection methods (V-tests, similar to B-tests).”

“The multi-relation exists in the dataset can thus be transformed into either a single homograph as seen in GraphConsis [48], or multiple homographs as utilized in SemiGNN [49], CARE-GNN [5] and PC-GNN [6].”

“Examples include SemiGNN [11] at ICDM 2019; HACUD [12] at AAAI 2019; ST-GNN [13] at IJCAI 2020; AMG-DP [14] at CIKM 2020; TemGNN [15] at SDM 2021; PC-GNN [16] at WWW 2021. bankruptcy prediction, and boost the development of financial risk research, especially SMEs bankruptcy research.”

“The device preprocessing methods suffer from low performance due to the lack of approximate local symmetry features in the extracted features [5].”

“Recently, the advances of graph neural networks (GNNs) [13], [23], [47] have inspired and empowered various attempts to adopt GNNs for detecting anomalies [10], [18], [32], [42], [50].”

“Previous GNN-based methods typically transformed multi-relation heterogeneous graphs into homogeneous graphs [17, 28, 29], or developed various strategies for learning directly from heterogeneous graphs [10, 13, 18, 39, 43].”

“Examples where queries must follow many links between data items include the investigation of financial fraud [2], supply chain and logistics analysis [3], protein-protein interactions in the cell [4], and social networks [5].”

“Outside Elliptic, GNN fraud detection has been applied to e-commerce at scale [Wang et al., 2019], financial networks [Liu et al., 2019], and insurance claims [Pourhabibi et al., 2020].”

“Although several proposed GNN-based fraud detection frameworks [5, 24, 28, 31, 35, 50, 55, 56, 67] have shown promising results in detecting non-camouflage malicious actors, they have struggled to tackle those malicious actors who employ camouflage [23].”

“Standard imbalance-learning techniques such as re-sampling and re-weighting (Wang et al., 2019; Cui et al., 2020; Dou et al., 2020; Liu et al., 2020) are often effective for static graphs, but they rely on independent and identically distributed (i.i.d.) assumptions that are are not generally…”

“…as they often correspond to detrimental events and abnormal behaviors, such as financial fraud rings characterized by dense clique formations [4], [5] or misinformation campaigns exhibiting star-like spreading patterns [6], [7], [8], which can ultimately lead to severe financial losses and…”

“By witnessing the success of GNNs, many cloud applications have incorporated GNNs in their service, such as e-commerce [45, 64], recommendation [62, 66], financial fraud detection [39, 57], and social network analysis [32].”

“The Semi-supervised Attention-based Graph Neural Network(SAGNN) proposed by Wang et al. achieved 0.8 accuracy on credit card transaction datasets [14] .”

“[17] (175 citations) and [18] (150) attracted interdisciplinary attention through early use of graph-based models and machine learning.”

“[3] Unlike anomaly detection in time series or images, graphs incorporate both node features and topology, making it challenging to detect anomalies using only one type Identify applicable funding agency here.”

“…cycle of a financial loan service, which consists of five phases: (1) the service provider admits a non-fraudulent user by performing the preemptive fraud user detection task [3], [4]; (2) when the admitted user, i.e., borrower, initiates a loan, the service Guofang Ma is the corresponding author.”

“No-tably, SemiGNN [13] proposed a GNN model for semi-supervised classification, combining the strengths of GCN and GAT and optimizing information propagation via an adaptive depth function.”

“For example, SemiGNN [11] utilizes node-level and view-level attention mechanisms to aggregate neighbor information under different relationships to obtain node representation.”

“Consequently, graph databases integrated with GNN frameworks can support web services such as recommendation systems [5], [6], traffic prediction [7], and fraud detection [8], [9].”

“SemiGNN [25]: Employs hierarchical attention to fuse multi-view graph information (user relationships, behavior sequences) for improved fraud detection accuracy.”

“…transaction networks, such as capturing stock temporal features via the Weisfeiler-Lehman graph kernel [14], hierarchical graph neural networks integrating multi-view market state information [15], and semi-supervised graph attention networks leveraging multi-source data for fraud detection [16].”

“Typically, techniques like re-sampling or re-weighting during the GNN aggregation process are widely investigated [18]–[20].”

“They have been used across a diverse range of real-world graph-based applications, such as molecular discovery [1], estimated time of arrival (ETA) prediction [2], financial fraud detection [3], recommendation systems [4], and fake news detection [5].”

“And the diagnostic process for these fraudulent behaviors is highly intricate, resulting in a considerable degree of uncertainty in behavioral decision-making when influenced by a multitude of complex factors[10].”

“The general paradigm of GNN-based anomaly detection is to input the attribute and structural information of the graph into the GNN and distinguish anomalies by learning the representation differences between normal and abnormal nodes [2].”

“They have been used for a variety of tasks ranging from node classification [1], graph classification [2] to edge predictions [1], with applications to various fields like financial fraud detection [3], recommender systems [4], etc.”

“Many proposed GNN models are applied to fake news and rumor detection [Bian et al. 2020; Wang et al. 2019b], fraud detection [Liu et al. 2021; Wang et al. 2019a], and other financial applications [Dastidar et al. 2022; Liu et al. 2020; Yang et al. 2020; Zhang et al. 2022].”

“• SemiGNN [21]: SemiGNN is a semi-supervised graph attention network that treats unlabeled and labeled data as separate views for anomaly detection.”

“Some other methods perform user identification by doing fraud detection to avoid fraudulent verification [3, 37, 38, 42].”

“Wang et al. [55] developed a semi-supervised graph attentive network specifically designed for fraud detection using a multi-view graph (i.e., users’ relationships, attributes, and devices).”

“Likewise, it has been proven that classifier assembly techniques help improve robustness in contexts with highly unbalanced data [12].”

“Graphs serve as a common language for modeling a wide spectrum of structured and relational systems, ranging from social networks [43] to knowledge graphs [36], from financial transaction networks [35] to e-commerce product graphs [22].”