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].”

“Then some further works like Metapath2vec [22] , HNE [23] are proposed to consider both the network topology”

“Some further works [20], [21] use the matrix factorization to factorize high-order relation matrix.”

“Some further works [20], [21] use the matrix factorization to factorize high-order relation matrix.”

“Commonly, these methods first extract statistical features from different aspects like user’s profiles and historical behaviours and then use some classical classifiers like SVM [9], treebased methods [10] or neural networks [7], [8] to decide fraud or not.”

“[8] selects some personal variables regarding the financial distress, debt structure, need for continuing growth, accounts receivable and totally uses a 27-dimensional vector as the input to the decision tree, neural network and bayesian belief network.”

“Then some further works like Metapath2vec [22] , HNE [23] are proposed to consider both the network topology”

“[9] extracts some statistic features about assets, liabilities, incomes, debt and sales and uses a multi-layer feedforward network to do financial statement fraud detection.”

“Then some following methods start to use graph embedding to incorporate the user interactions [5], [6].”

“Specifically, financial scenario also concerns about the KS [15], which is a metric to measure the risk differentiation of the model.”

“Therefore, a great number of literatures have been studied on different types of the fraud, like financial statement fraud [7]– [9], credit card fraud [10], [11], insurance fraud [12], [13]”

“Our work is also related to the graph-based semi-supervised learning (GSSL) method [26].”

“Then some following methods start to use graph embedding to incorporate the user interactions [5], [6].”

“These methods make prediction mainly based on the statistical features of a certain user and use classical classifiers like logistic regression, neural networks to do classification [2]–[4].”

“Therefore, a great number of literatures have been studied on different types of the fraud, like financial statement fraud [7]– [9], credit card fraud [10], [11], insurance fraud [12], [13]”

“These methods make prediction mainly based on the statistical features of a certain user and use classical classifiers like logistic regression, neural networks to do classification [2]–[4].”

“Therefore, a great number of literatures have been studied on different types of the fraud, like financial statement fraud [7]– [9], credit card fraud [10], [11], insurance fraud [12], [13] and so on.”

“Furthermore, rule-based methods are easy to be attacked [1].”

“com/FraudDetection/FRAUDRE SemiGNN [111] Python Tensorflow Static Attributed Graph https://github.”

“Semi-GNN [39] is a semi-supervised GNN model with hierarchical attention mechanism for explainable fraud prediction.”

“That is, (1) applying attention mechanisms to correlate different neighbors through various views (Wang et al., 2019a; Cui et al., 2020; Liu et al., 2021a), (2) using the resampling strategy to aggregate neighborhood information selectively (Dou et al., 2020; Liu et al., 2020; 2021b), (3) designing…”

“Semi-GNN and IHGAT [22] employ hierarchical attention mechanism for interpretable prediction, while based on camouflage behaviors and imbalanced problems, CARE-GNN and PC-GNN prune edges adaptively according to neighbor distribution.”

“posed by [183] can help learn better representations for fraud detection models.”

“, 2019) User-nickname graph Data Bipartite Loan (Wang et al., 2019) User-address graph Data Bipartite Loan (Wang et al.”

“proposed a semi-supervised attentive GNN, named SemiGNN, which applies a hierarchical attention mechanism to correlate different neighbors and different views better.(75) The aforementioned heterogeneous GNNs reveal illegal acts through aggregating nodes’ neighborhood information across different relations.”

“Graph structured data is very pervasive in real-world, such as social networks [9], financial transaction networks [37] and traffic networks [40].”

“The existing self-attention mechanism[18, 29, 38] is essentially a sum of neighbor information and not suitable for heterophilic neighbor nodes.”

“Fraud detection on e-commerce platforms: On e-commerce platforms, many fraud detection algorithms aim to detect financial criminals such as credit card fraudsters and risky accounts [4, 6, 8, 26, 27, 30, 36, 45, 48].”

“Semi-GNN [38] is a semisupervised graph neural model which adopts hierarchical attention to model the multi-view graph for fraud detection.”

“Over last few years, Graph Convolutional Networks (GCNs) have benefited many real world applications across different domains, such as molecule design [37], financial fraud detection [29], traffic prediction [30, 38], and user behavior analysis [11, 18, 27].”

“For example, by exploiting graph structural information, we can predict the chemical property of a given molecular graph [46], detect fraud activities in a nancial transaction graph [35], or recommend new friends to users in a social network [8].”

“Other GNN-based anomaly detection models such as GraphConsis [29] and Semi-GNN [47] have been shown to be less useful than CARE-GNN, thus they are ignored in the experiments.”

“Graph attentive network model and random walks (Wang et al. 2019a) were deployed on social graphs to link the unlabeled and labeled data and pass messages among them.”

“Anomaly Detection aims to identify entities that deviate significantly from the norm, which has been used for a variety of applications, such as revealing fraudulent or spam activity in social networks [26, 63, 72, 76, 84, 85] and financial transactions networks [10, 18, 19, 55, 67, 69, 74].”

“Even though GNNs have demonstrated outstanding performance in many graph mining tasks [19,35,30], it remains unclear how to exploit their potentiality for GAD problem.”

“However, from the results of some of the baselines in the table, although GCN, GAT, GraphSAGE, and GeniePath models run on a single relation graph, they are better than RGCN, Player2Vec, and SemiGNN in terms of accuracy of the Yelp dataset, which are run on the multi-relational graphs.”

“Graph Neural Networks (GNNs) [15, 22] have made remarkable achievements in modeling graphs from various domains such as social networks [15], financial system [35], and recommendation”

“To assess the performance of our proposed framework REDRESS, we choose thewidely adopted semi-supervised node classification task [19, 36, 39] and link prediction task [20, 26] as the downstream learning tasks.”

“A few of recent works start to utilize the graph-based methods to model complex interactive patterns of users [13, 18, 19, 32].”

“To tackle the unique challenges of graph anomalies, such as imbalance, heterophily, and camouflage, several adaptations of standard GNNs have been proposed [20, 96, 24, 53, 52, 21, 79, 54, 90].”

“Based on classic GNN architectures, various models are invented by reconstructing graph for fraud detection tasks, such as Fdgars [39], GraphConsis [22] and SemiGNN [38].”

“Graph neural networks (GNNs) [3, 12, 18] have made remarkable achievements in modeling graph structured data from various domains such as social networks [8, 12], financial system [41], and recommendation system [42].”

“Recent methods realize the importance of incorporating multiple relationships into graph learning [12, 24–26, 42, 44].”

“SemiGNN [40], a semisupervised attention GNN, is proposed to address the lack of labeled data in the network and uses a hierarchical attention mechanism to perform fraudulent user detection by embedding learning from multiple views.”

“SemiGNN (Wang et al. 2019) proposes to expand the labeled data through their social relations to get the unlabeled data, utilizing the multi-view labeled and unlabeled data for fraud detection.”

“Graph-structured data are very pervasive in the real-world such as social networks [14], fnance system [29], and molecular graphs [17].”

“Graph neural networks are used on financial networks, typically in a semi-supervised setting [77, 79, 80].”

“By delving deeper into the robustness of GCL, GCL techniques can be better applied to risk-sensitive critical scenarios like financial fraud detection [157].”

“…aforementioned problems, many existing GNN-based fraud detection models adopt strategies such as re-weighting neighbors from different classes using predefined indicators (Liu et al., 2021c) or learnable attention values (Wang et al., 2019; Liu et al., 2021a; Shi et al., 2022; Cui et al., 2020).”

“GNNs have shown remarkable success across diverse knowledge discovery and data mining scenarios, including drug discovery [17], fraud detection [29, 47], and knowledge reasoning [34].”

“[22,36,43,44,55,68,71, 72,80,88,89,92,93, 95,96,97,98, 101,102,103,104,105, 106,107] 24 Yes Analyze Relationships Between Subjects This task is associated with the analysis of a transaction destination.”

“For example, Semi-GNN [39] decomposes heterogeneous graphs into multi-views and then then fuses features from different perspectives through an attentionbased hierarchical fusion method.”

“Ding et al. (2019) generates anomaly scores of nodes in an AutoEncoder paradigm based on GNNs; Li et al. (2019) and Wang et al. (2019a) introduce the advanced GNNs techniques for spam detection and fraud detection, respectively; Liu et al. (2019) designs a specific multi-hop aggregation mechanism…”

“Zhou et al. [17] examined the application of graph-based methods in fraud detection, introducing a novel perspective that considers transaction relationships for improved accuracy.”

“Even though there are a few works under semi-supervised learning settings [14], [15], it is difficult for them to effectively confront the challenges of the scarcity of labeled anomalies and class imbalance associated with anomaly detection.”

“These methods fail to automatically detect fraud patterns and express important behavior information [6] since there are many interactions of transactions [7] and the transaction behaviors of users are dynamic [8].”

“1) Neighbor Information Aggregation: Previous works tend to use attention mechanisms [39], [40] or design weighting parameters [41] for message aggregation.”

“The authors in [27] expand the labeled data through their social relations to get the unlabeled data and propose a semisupervised attentive Graph Neural Network, named SemiGNN, to utilize the multiview labeled and unlabeled data for fraud detection.”

“Our method is implemented in PyTorch 1.13.1 (Paszke et al. 2019) and PyG (Fey and Lenssen 2019) with Python 3.9, and DGL (Wang et al. 2019c) is used for preprocessing the data.”

“The performance improvements of DeepSAD and SemiGNN are steadily increased with more labeled data.”

“Recent advances in graph neural networks (GNNs) have demonstrated remarkable potential in fraud detection by naturally modeling the relational structure inherent in financial networks [9], [10].”

“[95] designed a hierarchical attention structure in graph neural network from node-level attention to view-level attention when generating graph embeddings.”

“Compared to previous work [2, 8, 9, 17, 22, 23], our defined AMG is a more general framework to model real-world unstructured network data, which consists of hundreds of millions of nodes and edges of multiple types, and each node and edge is associated with plentiful attributes.”

“Graph Convolutional Networks (GNNs) [1], [2], [7] have achieved great success in many real world applications across different domains, such as recommender system [11], molecule design [12], financial fraud detection [13], traffic prediction [14], and user behavior analysis [15].”

“While other methods aggregate information intra-view and inter-∗ zhengya.sun@ia.ac.cn (Z. Sun) ORCID (s): view hierarchically, which convolute each node over its within-view neighborhood and then integrate information across views [22, 32, 46, 11].”

“The continuation of Financial Statement Fraud (FSF) poses significant concerns to global capital market stability [12].”

“In real-world networked systems, such as social networks [3] and payment transaction graphs [4], even a few anomalies, e.”

“…models in a variety of tasks, for example, intrusion detection [29], [30], [31], [32], load forecasting [33], [34], [35], [36], health monitoring [37], [38], [39], [40], privacy preserving [41], [42], [43], [44], traffic prediction [45], [46], [47], [48], and fraud detection [49], [50], [51], [52].”

“It is essential for decision-makers to be aware of the abnormal patterns in data in the social system, such as financial fraud detection [2], spam detection [3], and network intrusion detection [4], [5].”

“com/stories/us machine learning based researches for CRF [14, 31, 35] in the AI research field, CLS has often been overlooked by the same research community.”

“…towards graph-based approaches as the use of graph convolutional networks on datasets with partial labels has been effective for predicting node attributes within citation networks so that many GNN-based fraud detectors have been proposed to detect fraud [ Wang et al. , 2019; Liu et al. , 2018 ] .”

“With the increasing societal concerns on the fairness of GNNs [31], explorations have been made to alleviate the bias exhibited in GNNs.”

“, medical diagnosis, defect detection, animal behavior understanding, and fraud detection [31,2,29].”

“For example, it was designed based on the attention mechanism in Heterogeneous Graph Attention Network (HAN) [193], which is also a prevalent choice in many cybersecurity solutions [69, 181, 235].”

“The authors in [20], expand the labelled data through their social relations to get the unlabelled data and propose a semi‐supervised attentive graph neural network, named SemiGNN to utilize the multi‐view labelled and unlabelled data for fraud detection.”

“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.”

“Graph mining has become the cornerstone in a wealth of real-world applications, such as social media mining [44], brain connectivity analysis [1], computational epidemiology [19] and financial fraud detection [37].”

“As a result, spatial-based GNN was gradually proposed in [Niepert et al. , 2016; Hamilton et al. , 2017; Hu et al. , 2019; Wang et al. , 2019].”

“…divided into three main categories, namely, GNN-based telecom fraud detection [4], [25], [26], [27], [28], [29], [30], [31], [32], [33], GNN-based social network fraud detection [5], [34], [35], [36], [37], [38], and GNN-based financial fraud detection [39], [40], [41], [42], [43], [44], [45], [46].”

“Conventional graph-based detectors [14, 15] reveal fraudsters by aggregating their neighbors’ features based on the assumption of homophily (i.”

“, drug design [8], fake news detection in social media [15], recommendation system [9], and even financial sector [7].”

“They are increasingly adopted in various tasks such as anomaly detection [26], social network recommendation [28], and graph classification [27].”

“In the last few years, GNNs have become a hot topic in machine learning and benefited many graph-related tasks [39, 40] and real world applications across a verity of areas, including molecule design [14], financial fraud detection [8], traffic prediction [8, 13], user behavior analysis [15, 41, 42], and recommender s…”

“One perspective to obtain explanations for fraud detection is applying attention mechanisms that generate soft masks on the input graph [12, 29].”

“Many previous works conduct semi-supervised node classification tasks on single-relation graphs [18] or define multi-relation graph [4, 17, 20, 32] to tackle fraud detection tasks.”

“Graph data is pervasive among a plethora of realms, e.g., financial fraud detection (Wang et al. 2019; Pourhabibi et al. 2020; Cheng et al. 2020), social recommendation (Fan et al. 2019; Song et al. 2019; Guo and Wang 2020), and chemical reaction prediction (Do, Tran, and Venkatesh 2019; Shi et al.…”

“GEM[26], SemiGNN[41], ASA[44], Player2Vec[50], and CARE-GNN[6] all construct multiple homogeneous graphs based on node relations in corresponding datasets.”

“In the context of cross-silo FL, where the clients pool is typically made up from 2 to 100 [12] organizations, there is a growing need to deploy a federated XGBoost system on specific tasks such as survival analysis [1] and financial fraud detection [6, 28].”

“For instance, in fraud detection within financial networks, label acquisition may be constrained to specific network regions due to varying international legal frameworks and diverse data collection periods (Wang et al., 2019; Dou et al., 2020).”

“As a natural result, AUC is more appropriate than accuracy for long-tail classification problems such as disease prediction [24], [80] and rare event detection [37], [41], [69], [71], due to this appealing property [19], [22].”

“To address the specific challenges of graph anomalies such as label imbalance [34, 31], relation camouflage [12, 38], and feature heterophily [50, 15], numerous adaptations of standard GNNs have been proposed [9, 67, 13, 40, 39, 10, 53, 41, 62].”