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].”

“Many applications such as recommender systems [13], [14], knowledge graphs [15], fraud detection [16], and drug discovery [17]–[19] have used GNNs to help them accomplish their tasks.”

“A key to fraud detection is to classify whether a user is a fraud user or non-fraud user [143].”

“To mitigate this, semi-supervised learning techniques, which leverage small amounts of labelled data combined with a large volume of unlabeled data, have been integrated into modern fraud detection frameworks [12], [13].”

“Index Terms —Fraud detection, semi-supervised learning, graph neural network.”

“Graph structured data are ubiquitous and are widely used in social network analysis [29], financial fraud detection [46], molecular design [17], search engines [44] and recommender systems [36, 38].”

“Likewise, Semi-GNN [56], CARE-GNN [17], and PC-GNN [34] construct multiple homo-graphs based on node relations.”

“Examples include [11], [15], [16] for review fraud detection, [2], [3] for fake news detection and [5], [6], [17]– [19] for financial fraud detection.”

“In the semi-supervised branch, Graph Neural Networks (GNNs) (Gao et al., 2024b; Wang et al., 2019; Liu et al., 2021a; He et al., 2024; Zhang et al., 2024; Chen et al., 2024b;a) have achieved remarkable success in detecting graph anomalies across various domains for its excellent representation…”

“[38] proposed a spatial-temporal model to detect both the structural and temporal signal for default user detection.”

“Graph Convolutional Network (GCN) has become a prevalent learning paradigm in many real-world applications, including financial fraud detection [45], drug discovery [19] and traffic prediction [7].”

“As the feature dimensions of users and items are usually the same or can be projected to the same space via linear transformation, we use X ∈ R ( 𝑚 + 𝑛 )× 𝑘 to denote the vertically concatenated feature matrix of X and X .”

“…balance the number of samples by over-sampling the minority class or under-sampling the majority class [12], [13], [14]; and 2) reweighting methods assign different weights to different classes or even different samples by cost-sensitive adjustments or metalearning-based methods [15], [16], [17].”

“The superiority of GNNs stems from iterative aggregation and update processes.”

“Today, GNNs have been widely used in online web services, e.g., social network analysis [9, 23], short-video recommendation [26, 53], shopping recommendation [40, 50], and financial fraud detection [28, 41].”

“Graph Neural Networks (GNNs) [24] have gained rapid adoption across a wide range of application domains, including social network analysis [40], financial fraud detection [4, 30], and drug discovery [13, 27].”

“Previous studies are either conducted on homogeneous [6], [14] or heterogeneous graphs [7], [8], [15].”

“GNNs aggregate information from neighboring nodes to generate representations for central nodes, facilitating automatic feature learning that considers interactive connections between transactions [4–6].”

“It has gained increasing attention from both academia and industry in recent years for the sparsity of anomaly labels and the significant real-world applications [1], [2], such as financial fraud detection [3], social rumor detection [4], and computer network intrusion detection [5].”

“In addition, there are other applications of heterogeneous graph embedding based on specific tasks, such as fraud detection [40, 41] and recommendation tasks [42, 43].”

“Graph Neural Networks (GNNs) have shown promising results in modeling graph-structured data such as social network analysis [17], finance [42], and drug discovery [21].”

“Prevailing methods may be built upon rule-based systems [4] and supervised [36] or semi-supervised [38] learning techniques.”

“Financial fraud includes credit card fraud [17], [19], [29], phone fraud [30], online transaction fraud [31], instant payment fraud [32], etc.”

“Inspired by previous work [41, 52], we design an S-Encoder that learns a compact representation of u from Ds for each partner mask M p .”

“Recently, graph neural networks (GNNs) have been employed for credit card fraud detection [Wang et al. , 2019; Cheng et al. , 2020; Zhang et al. , 2024] and achieve remarkable success as GNNs can more accurately infer the fraud probability by capturing patterns within relational graphs.”

“Graph representation learning and community detection are closely intertwined in network analysis, as graphs are widely used to model relationships between entities in various domains, including social networks [5], recommender systems [6], and financial risk control [7].”

“The propagation and transformation process are integrated as a single GNN layer.”

“[2] Daixin Wang, Jianbin Lin, Peng Cui, Quanhui Jia, Zhen Wang, Yanming Fang, Quan Yu, Jun Zhou, Shuang Yang, and Yuan Qi.”

“However, many overdue, fraudulent, and cash-out behaviors are inevitable on the platform [2], [3], [7]–[9].”

“FdGars (Wang et al. 2019b) differentiates between scammers and regular users in Tencent’s App Store.”

“Many studies have explored for few-shot learning, using techniques like meta-learning or contrastive learning [5, 20, 36, 38, 42, 46, 48].”

“On the one hand, there are many interactions among transactions [7].”

“, malicious user detection in social media [7, 42], interest group discovery in recommender systems [14, 43, 45, 46], event prediction in service networks [28], to just name a few.”

“Among semisupervised methods, SemiGNN [20] is a representative method that leverages hierarchical attention to learn from multi-view graphs for fraud detection.”

“Graph Neural Networks (GNNs) have been widely applied in a variety of fields, such as social network analysis [5], financial risk control [26], and recommender systems [27], where both structural and attribute information are learned via message passing on the graphs [11].”

“These entities can encompass various subjects, such as reviews on the website [24], [25] or transactions on a trading system [26], [27].”

“They extend the convolution to non-Euclidean space and have become one of the hottest topics in numerous domains, including fraud detection [46, 59, 66], computer vision [31, 64] and other areas [12, 40, 56].”

“Firstly, we select popular GNN-based fraud detection methods including GraphSage [Hamilton et al. , 2017], GEM [Liu et al. , 2018], FdGars [Wang et al. , 2019b], Graph-Consis [Liu et al. , 2020], CARE-GNN [Dou et al. , 2020], PC-GNN [Liu et al. , 2021], FRAUDRE [Zhang et al. , 2021], H 2 -…”

“KS value balances the models’ ability to catch more frauds (high Recall) and reduce the efects on normal loan applications (low Disturbance), which is the de facto standard in the ield of loan anti-fraud [18, 54].”

“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 [18].”

“The main goal of fraud detection is to predict whether a given entity (anything from an object to a user to other things) will be involved in fraudulent activities or not [2].”

“Therefore, semi-supervised techniques [9, 47] are further utilized to conduct fraud detection, which can be viewed as a combination of the above two kinds of methods.”

“Recent advances leverage graph neural networks [10], [11] and temporal models [12], but cannot leverage cross-platform fraud intelligence without violating privacy regulations [13].”

“Semi-GNN [35] breaks down a heterogeneous graph into multiple views and combines features from various perspectives using an attention-based hierarchical fusion approach to identify credit card fraud.”

“Firstly, our method consistently outperforms the vanilla baselines trained with the original node features or the LM-finetuned node embeddings in most cases.”

“It has a higher detection accuracy and can provide a safer and faster financial fraud and fraud detection method in practical applications (Wang et al. 2019).”

“For instance [48] addressed the homogeneity and heterogeneity of issue of networks for fraudulent invitation detection, anomaly detection on dynamic graphs with sudden bouts of increased and decreased activities is proposed in [10], [41] proposed a semi-supervised approach and an attention-based approach for fraud det…”

“The graph serves as a fundamental data structure for modeling a diverse range of relational data, ranging from financial transaction networks [7, 42] to social science [9].”

“Various GNN models have been widely applied on a variety of industrial domains such as misinformation detection (Benamira et al., 2019), financial fraud detection (Wang et al., 2019), traffic prediction (Zhao et al., 2019), and so- 1Carnegie Mellon University 2Google Research.”

“[25] utilized the multi-view labeled and unlabeled data for fraud detection, by expanding the labeled data through their social relations to the unlabeled data and proposing a semi-supervised attentive GNN.”

“This innovation has made GATs particularly useful in applications in multivariate time series prediction [38], financial fraud detection [37] and in other financial modeling framework, where some market indices have a more significant influence on others.”

“Graph Neural Networks (GNNs) [22] have emerged as a powerful paradigm for representation learning and predictive modeling on graph-structured data across various problem domains [2, 8, 25, 27, 35].”

“SemiGNN [14] utilizes a hierarchical attention mechanism to better associate different neighbors and different views.”

“[38] proposed a semi-supervised attentive graph neural network to utilize the social relations in the multi-view labeled and unlabeled data for fraud detection.”

“For example, Wang et al. [28] linked the labeled and unlabeled data via their social relations and proposed a semi-supvervised graph attentive network to utilize the multi-view labeled and unlabeled data for fraud detection, moreover, the work designed a hierarchical attention mechanism to better…”

“We choose FdGars [26], GraphConsis [35], GEM [27], SemiGNN [31], and BTG [6] as state-of-the-art GNN-based fraud detectors.”

“Out of ∼ 310 initially screened articles, 29 were deemed relevant for synergy-based BFSI contexts (see [1], [2], [3], [4], [5], [6], [7], [8], [9], [10], [11], [12], [13], [14], [15], [16], [17], [18], [19], [20], [21], [22], [23], [24], [25], [26], [27], [28].”

“Many problems in data analysis are inherently imbalanced in areas such as finance (fraud detection) [36], marketing (churn prediction) [24], medicine (medical diagnosis) [18], industry (predictive maintenance) [34], image recognition [32, 33], etc.”

“SemiGNN [37] used a hierarchical attention mechanism to correlate different neighbors better.”

“In real-world scenarios, it is common for data to be perturbed during the training or testing phases (Dai et al. 2023; Wang et al. 2019), highlighting the importance of studying the robustness of GNNs.”

“Due to their strong power in learning representations on graph data, they have achieved great success in many real-world applications such as recommender systems [8, 41], computer vision [19], drug discovery [7] and financial fraud detection [34].”

“The fraud and spam detection tasks require millisecond-level inference on the million-scale graph to identify the malicious users and avoid the property loss of the victim users [18, 21, 28].”

“Although there are works dedicated to corrupted user detection [19], [20], [21], [22], [23], they primarily operate on the premise of pre-existing user information in an offline context, rendering them ineffective for online detection based on bandit feedback.”

“Graph-based methods [4–6, 9, 15, 17, 19, 21, 22, 29, 31] use predefined features like devices, IPs, and purchase patterns to connect users, but these features may not reflect behavioral similarity and are easy to manipulate by fraudsters [5, 22].”

“In the era of big data, many high-impact domains, such as e-commerce [1, 2], finance [3], biomedical science [4, 5], and cyber security [6, 7], naturally exhibit long-tailed data distributions, where a few head categories 1 are well-studied with abundant data, while massive tail categories are…”

“Machine learning-based methods usually resort to feature extraction [5].”

“Graph structured data are ubiquitous in real world, with prominent examples including financial networks [24], molecular finger-prints [34], and recommender systems [28].”

“It finds extensive applications in capturing high-risk entities and behaviors, including but not limited to spam detection [29], financial fraud detection [38], and fake news detection [11].”

“Graph Neural Networks (GNNs) have been widely applied in a variety of fields, such as social network analysis [5], financial risk control [26], and recommender systems [27], where both structural and attribute information are learned via message passing on the graphs [11].”

“Besides, some researches have also attempted to use genetic algorithm [45], evolutionary algorithm [46] and other methods [47, 6] to detect financial statements fraud.”

“However, the effectiveness of these measures is limited due to challenges such as class imbalance and model overfitting in detection algorithms (Wang et al., 2019).”

“In [40], the authors utilized both labeled data (users are labeled as fraud or not) and unlabeled data (social relations and attributes) and proposed SemiGNN to exploit the representations of users.”

“SemiGNN [24] proposes a hierarchical attention mechanism to detect fraudsters.”

“People bring in the graph methods to model complex relational data in fraud detection scenarios[3, 6, 12, 17, 18, 21, 22, 26, 33, 35, 47].”