A Financial Fraud Detection Method and System Based on Isomatch Enhancement and Weakly Supervised Learning

By constructing a dual-channel grouping aggregation and global dependency modeling of the financial activity network, the problems of insufficient local neighborhood matching and weak supervision signals in financial fraud detection are solved, thereby improving the ability to identify disguised and gang-related fraud.

CN123070955APending Publication Date: 2026-11-10BEIHANG UNIV
View PDF 3 Cites 0 Cited by

Patent Information

Application Number
CN202611035379.9
Authority / Receiving Office
CN · China
Patent Type
Applications(China)
Current Assignee / Owner
Filing Date
2026-07-13
Publication Date
2026-11-10

AI Technical Summary

Technical Problem

Existing financial fraud detection methods suffer from insufficient local neighborhood matching and weak supervision signals when facing fraudsters disguised as legitimate entities, resulting in limited identification capabilities. In particular, they struggle to effectively capture the correlation between similar fraud nodes at long distances when labels are scarce.

Method used

By constructing normal category prototypes and fraud category prototypes, enhancing category consistency information through local neighborhood grouping and aggregation, and supplementing long-distance matching signals through global fraud dependency modeling, a dual-channel grouping and aggregation and attention propagation mechanism are adopted to generate final node representations to improve fraud detection accuracy.

Benefits of technology

In the local neighborhood, the dilution of the representation of fraudulent nodes by normal nodes is suppressed, which alleviates the lack of supervision signals under the condition of scarce labels, captures the association of fraudulent nodes in the global scope, and improves the identification ability of disguised and gang-related fraud.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN123070955A_ABST
    Figure CN123070955A_ABST
Patent Text Reader

Abstract

This invention discloses a financial fraud detection method and system based on isomatch enhancement and weakly supervised learning, belonging to the fields of financial risk control and graph neural network technology. Addressing the problems of fraudster masquerading, scarce fraud labels, and insufficient weakly supervised signals due to the sparse distribution of similar fraud nodes in financial fraud, this method constructs financial activity graph data, uses labeled nodes to build category prototypes and generates soft attribution weights for unlabeled nodes, allocates neighborhood information to normal and fraud channels for grouping and aggregation to enhance local isomatch, further filters high-confidence fraud candidate nodes to form an extended set to model global fraud dependency, and finally fuses global information with local representations to output the fraud probability. This invention can improve the accuracy and robustness of financial fraud detection under conditions of limited labeling and can be used for financial transaction risk control, anti-money laundering identification, and abnormal account screening.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] This invention relates to the fields of financial risk control and graph neural network technology. More specifically, this invention relates to a financial fraud detection method and system based on isomatch enhancement and weakly supervised learning. Background Technology

[0002] With the rapid development of digital financial infrastructure and the continuous deepening of online payment scenarios, financial activities have gradually formed a large-scale relationship network involving multiple entities such as users, accounts, merchants, payment institutions, transaction flows, and financial instruments. At the same time, financial fraud, telecommunications fraud, money laundering, and related black and gray market activities are showing a trend of organization, chaining, and disguise. Such fraudulent activities are no longer manifested as abnormal single accounts or transactions, but are hidden in the normal financial activity network through identity disguise, embedding of normal relationships, small and dispersed transactions, account switching, and splitting and transferring funds. This makes the fraudulent entities more similar to normal entities in terms of local structure and behavior, significantly increasing the difficulty of risk identification.

[0003] Existing methods for detecting financial fraud mainly include manual rule screening, expert experience models, traditional machine learning classification models, and detection models based on graph neural networks. Manual rules and expert models rely on preset thresholds and human experience, making it difficult to adapt to rapidly evolving new fraud patterns. Traditional machine learning methods mainly rely on the attribute characteristics of individual accounts, users, or transactions, making it difficult to fully utilize the complex interaction relationships between financial entities. Although graph neural networks can utilize adjacency relationships in financial activity networks for message passing and node representation learning, they usually implicitly rely on the assumption that adjacent nodes have similar categories or similar semantics. However, in financial fraud scenarios, fraudsters often actively connect with normal users, merchants, or trading partners to disguise themselves, making normal nodes the majority in their local neighborhood. This causes fraud-related signals to be diluted or even submerged by a large amount of normal information during message aggregation.

[0004] To address the limitations of graph neural networks in heterogeneous graph scenarios, researchers have proposed several improvement schemes. For example, Chinese patent application CN119202806A, "A Graph Fraud Risk Identification Method and System Based on Link Pattern Aggregation," attempts to compensate for the shortcomings of mainstream graph neural network algorithms in handling long-range dependent graph structures by using long-distance link sampling encoding and link information interaction aggregation. However, it still relies on relatively sufficient labeled samples, and the uncertainty of node category attribution is not fully considered during the link sampling process. Its performance is still greatly limited in scenarios where labeled samples are extremely scarce.

[0005] In financial fraud detection, genuine fraud labels typically rely on manual review, judicial determination, or risk control measures, which are costly and have long feedback cycles. The limited number of labeled fraud samples makes it difficult to provide sufficient supervision for the model. To address the node classification problem under scarce label conditions, US patent application US2023 / 0117980A1 proposes a graph prototype network system and method for few-shot learning on attribute networks. This system uses a meta-learning framework to learn category prototypes and predict the categories of unlabeled nodes under limited labeled sample conditions. However, it primarily addresses general few-shot node classification problems and does not specifically address the distortion of local neighborhood matching caused by fraudulent entities actively disguising themselves in financial fraud scenarios. Furthermore, it does not consider the difficulty of modeling long-distance dependencies caused by the sparse distribution of similar fraudulent nodes globally. Therefore, its applicability and detection accuracy in financial fraud detection scenarios still need improvement.

[0006] Chinese patent application CN121561730A, "Real-time Fraud Detection Method and System Based on Multimodal Fusion," attempts to improve the accuracy of fraud detection by fusing multimodal information such as payment behavior data, transaction text data, and account behavior data. However, it focuses on the temporal correlation analysis of multimodal data and does not involve the representation learning of topological relationships between nodes in graph structure data, nor does it solve the problem of effective aggregation of neighborhood information under conditions of insufficient homogeneity.

[0007] The above factors together lead to the coupled problems of insufficient matching and insufficient supervision signals in financial fraud detection: on the one hand, the scarcity of fraud labels makes it difficult for the model to effectively learn the category features of fraud nodes from a small number of labeled samples, and the category information cannot be fully propagated to unlabeled nodes, further exacerbating the lack of matching caused by the majority of normal nodes in the local neighborhood; on the other hand, the scarcity of effective fraud signals in the local neighborhood makes the model more dependent on the potential association between similar fraud nodes at a distance, and the lack of labeled samples precisely limits the model's ability to discover and utilize such long-distance matching associations on a global scale. Existing methods for financial fraud detection still have the following shortcomings: Fraudsters disguise themselves by actively connecting with legitimate entities, resulting in an excessively high proportion of legitimate nodes in local neighborhoods. Fraud-related signals are diluted during message aggregation, and existing graph neural network methods lack effective mechanisms for handling such heterogeneous neighborhoods. Similar fraud nodes often exhibit sparse distribution or weak connectivity in the global network, making it difficult to effectively capture long-distance class consistency associations through local message passing. Existing global modeling methods either introduce a large amount of interference by performing indiscriminate interactions on all nodes or fail to specifically optimize for the sparse distribution characteristics of fraud nodes. Obtaining genuine fraud labels is costly. Existing methods, with only a very small number of labeled samples, struggle to simultaneously enhance homogeneity in local neighborhoods and mine long-distance homogeneous associations globally, thus limiting their ability to identify disguised fraud, organized fraud, and cross-account fund chain fraud. Summary of the Invention

[0008] One object of the present invention is to solve at least the above-mentioned problems and to provide at least the advantages that will be described later.

[0009] To achieve these objectives and other advantages according to the present invention, a financial fraud detection method based on isomatch enhancement and weakly supervised learning is provided, comprising: S1. Obtain the raw data of financial activities, extract entity objects as nodes, construct the financial activity relationship between entity objects as edges, extract node attribute features to generate node features, and obtain graph data containing node features, adjacency relationships and weak supervision labels. Weak supervision labels include normal labels for labeled normal nodes and fraud labels for labeled fraud nodes. Unlabeled nodes remain unlabeled. S2. In the current propagation layer, construct normal category prototypes and fraud category prototypes based on the representations of labeled normal nodes and fraud nodes in the current propagation layer, calculate the similarity between the current representation of unlabeled nodes and the normal category prototypes and fraud category prototypes, generate soft classification weights for unlabeled nodes to belong to the normal category and fraud category, and set deterministic classification vectors for labeled nodes to obtain the classification weights of each node. S3. Obtain the local neighborhood node set of each node, with the node itself as the target node. Based on the category weight of the neighborhood nodes, assign the representations of the neighborhood nodes to the normal channel and the fraud channel respectively. Based on the current representation of the target node, generate adaptive aggregation weights for the normal channel and the fraud channel respectively. Then, weight and aggregate the representations of the neighborhood nodes in the normal channel and the fraud channel to generate the normal channel aggregated representation and the fraud channel aggregated representation respectively. Combine them with the current representation of the target node to generate the local enhancement representation of the next propagation layer. S4. Repeat steps S2 to S3 for multi-layer iterative propagation. In the last propagation layer, input the local enhanced representation of the node into the classifier to obtain the predicted probability of the unlabeled node belonging to the normal category and the fraud category. Filter high-confidence fraud candidate nodes according to the preset fraud probability threshold and merge them with the labeled fraud nodes to form an extended fraud node set. Calculate the long-distance dependency between the representations of each node on the extended fraud node set based on the attention propagation mechanism to generate a global fraud dependency representation. S5. Calculate the similarity between the global fraud dependency representation and the normal category prototype and the fraud category prototype to obtain the category belonging weight of the global fraud dependency representation, and map it to the normal channel and the fraud channel respectively. Obtain the global information of the mapped normal channel and the fraud channel respectively and fuse it into the normal channel aggregate representation and the fraud channel aggregate representation with the preset fusion coefficient. Combine it with the current representation of the target node to generate the final node representation, input it into the classifier and output the probability that each node is a fraud node.

[0010] Preferably, in step S1, the raw data comes from the business database or risk monitoring database of financial activities; The extracted entity objects include at least one of the following: account entity, user entity, merchant entity, transaction entity, device entity, and financial instrument entity; Financial activity relationships include at least one of the following: there is a transaction, transfer, shared login device, shared account binding, shared payment and receipt, shared access to merchant or fund link relationship between two entities; The process involves generating an adjacency matrix or adjacency list for graph data based on the edge construction results. The adjacency matrix or adjacency list is used to represent the relationships between nodes. For each node, at least one of the following features is extracted: transaction frequency, transaction amount, transaction time distribution, account activity, device switching frequency, number of payment and receipt objects, historical risk record features, or fund flow path features. The extracted features are combined to generate a node feature vector, and all node feature vectors are combined into a node feature matrix.

[0011] Preferably, in step S1, the step of forming weakly supervised labels includes: Nodes corresponding to entities that have been manually reviewed, confirmed by regulators, found in historical blacklists, or confirmed as fraudulent through case retrospective analysis will be marked as fraudulent nodes. Nodes corresponding to entities that have been manually reviewed, have no abnormal transaction records, no risk records, or have been confirmed as normal by whitelists will be marked as normal nodes. Nodes that have not been confirmed or lack clear risk conclusions will be marked as unmarked nodes.

[0012] Preferably, step S2, which involves constructing category prototypes and calculating node category affiliation weights, includes: Obtain the current representation of each node in the current propagation layer, and aggregate the current representations of labeled normal nodes and labeled fraudulent nodes respectively to generate normal category prototypes and fraudulent category prototypes. For category c, the category prototype is represented as: Among them, V L This represents the set of labeled nodes, which is the union of all labeled normal nodes and labeled fraudulent nodes. 'be' represents the normal category, and 'fr' represents the fraudulent category. δ(y) represents the representation of node v at layer (l-1). v c) indicates whether the node label belongs to category c, and Δc represents the number of labeled nodes that belong to category c; The current representation of the node to be estimated, along with its normal category prototype and fraud category prototype, are mapped to the same latent space. The similarity between the node's representation and each prototype is calculated, generating soft attribution weights for the node's classification as either normal or fraudulent. (Node v) j The category affiliation vector is represented as: in, This represents a category prototype matrix consisting of normal category prototypes and fraud category prototypes. and As a learnable mapping matrix, a deterministic category attribution vector is set for labeled normal nodes and labeled fraudulent nodes respectively, and a soft attribution vector calculated from similarity is used for unlabeled nodes, forming a unified node-category attribution matrix.

[0013] Preferably, in step S3, the specific steps for assigning the neighbor node representations to the two information channels are as follows: For the target node, obtain the one-hop or multi-hop neighbors of the target node according to the adjacency relationship to form a local neighborhood node set. Read the normal category assignment weight and fraud category assignment weight of each neighbor node in the local neighborhood node set. According to the normal category assignment weight of each neighbor node, input the weighted representation of the neighbor node into the normal channel; according to the fraud category assignment weight of each neighbor node, input the weighted representation of the neighbor node into the fraud channel. For labeled normal nodes and labeled fraudulent nodes, they are respectively used as deterministic category assignment vectors in the corresponding category channels; for unlabeled nodes, they are contributed to the normal channel and fraud channel according to their normal category soft assignment weight and fraudulent category soft assignment weight, respectively. target node v i The neighborhood influence is decomposed into normal category channels and fraud category channels: Wherein, N(v)i ) represents the target node v i A local domain node set, and Representing node v respectively j The weights assigned to the normal and fraud categories. Represents the neighborhood aggregation coefficient. and Let represent the learnable transformation matrices for the normal category channel and the fraud category channel, respectively. Represents the target node v i The influence of the neighborhood on the l-th layer.

[0014] Preferably, step S4 specifically includes the following steps: Repeat steps S2 to S3 in claim 1 to perform multi-level iterative propagation. In each propagation layer, update the normal category prototype and fraud category prototype based on the node representation of the current propagation layer and calculate the category assignment weight of each node. Based on the updated category assignment weight, perform local neighborhood grouping and aggregation to generate the local enhanced representation of the next propagation layer. In the last propagation layer, the local augmented representation of each node is input into the classifier to obtain the predicted probability of unlabeled nodes belonging to the normal category and the fraud category. Based on the preset fraud probability threshold, nodes with fraud prediction probabilities not lower than the fraud probability threshold are selected from the unlabeled nodes as high-confidence fraud candidate nodes. The high-confidence fraud candidate nodes are merged with the labeled fraud nodes to form an extended fraud node set, and the current representation of each node in the extended fraud node set is extracted to form a global fraud node representation matrix. On the global fraud node representation matrix, the long-distance dependency between the representations of each node in the extended fraud node set is calculated based on the linear attention mechanism, and the global fraud dependency representation corresponding to each node is generated.

[0015] Preferably, step S5 specifically includes the following steps: Calculate the similarity between the global fraud dependency representation and the normal category prototype and the fraud category prototype to obtain the category affiliation weight of the global fraud dependency representation to the normal category and the fraud category; Based on the category attribution weight, the global fraud dependency representation is mapped to the normal channel and the fraud channel respectively, to obtain the global information mapped to the normal channel and the global information mapped to the fraud channel. These are then summed with preset fusion coefficients and added to the normal channel aggregation representation and the fraud channel aggregation representation of the target node in step S3 respectively, to obtain the global-local fused normal channel aggregation representation and fraud channel aggregation representation. The normal channel aggregation representation and fraud channel aggregation representation after global-local fusion are combined with the current propagation layer representation of the target node to generate the final node representation; The final node representation is input into the classifier, and the Softmax function outputs the probability that each node is a normal node and a fraudulent node. Nodes with a fraud probability higher than a preset detection threshold are identified as fraud risk nodes.

[0016] Preferably, in step S4, the linear attention mechanism computes the global fraud dependency representation as follows: Among them, X gt N represents the global fraud node representation matrix. gt W represents the number of nodes in the expanded fraud node set. Q and W K Let be a learnable mapping matrix, where 1 represents a column vector with all elements equal to 1. This represents the global fraud dependency representation matrix.

[0017] Preferably, in step S5, the category channel aggregation result after global-local fusion is represented as follows: in, "be" indicates the normal category, and "fr" indicates the fraud category. This represents the category channel aggregation function of the (l-1)th propagation layer. Represents node v i In the global fraudulent dependency representation at the (l-1)th propagation layer, β represents a preset fusion coefficient, ranging from 0 to 1, used to adjust the contribution strength of the global fraudulent dependency representation to the local aggregated representation. N(v) represents the mapping function that maps the global representation to the category channel. i ) represents node v i The set of local neighborhood nodes, and the final node representation is: in, Represents node v i The final node representation in the l-th propagation layer, Represents the fusion function. The mapping function represents the target node itself.

[0018] This invention provides a financial fraud detection system based on isomatch enhancement and weakly supervised learning, comprising: The data acquisition and graph construction module is used to acquire raw data of financial activities, extract entity objects as nodes, construct the financial activity relationship between entity objects as edges, extract node attribute features to generate node features, and obtain graph data containing node features, adjacency relationships and weak supervision labels. The weak supervision labels include normal labels for labeled normal nodes and fraud labels for labeled fraud nodes. Unlabeled nodes remain unlabeled. The category prototype construction and classification estimation module is used to construct normal category prototypes and fraud category prototypes respectively in the current propagation layer based on the representations of labeled normal nodes and fraud nodes in the current propagation layer, calculate the similarity between the current representation of unlabeled nodes and the normal category prototypes and fraud category prototypes, generate soft classification weights for unlabeled nodes to be classified into normal and fraud categories, and set deterministic classification vectors for labeled nodes to obtain the classification weights of each node. The local neighborhood grouping and aggregation module is used to obtain the local neighborhood node set of each node, with the node itself as the target node. Based on the category belonging weight of the neighborhood nodes, the representations of the neighborhood nodes are assigned to the normal channel and the fraud channel respectively. Based on the current representation of the target node, adaptive aggregation weights are generated for the normal channel and the fraud channel respectively. The representations of the neighborhood nodes in the normal channel and the fraud channel are weighted and aggregated to generate the normal channel aggregated representation and the fraud channel aggregated representation respectively. These are then combined with the current representation of the target node to generate the local enhancement representation of the next propagation layer. The global dependency modeling module is used to repeatedly call the category prototype construction and attribution estimation module and the local neighborhood grouping and aggregation module to perform multi-layer iterative propagation. In the last propagation layer, the local enhanced representation of the node is input into the classifier to obtain the predicted probability that the unlabeled node belongs to the normal category and the fraud category. Based on the preset fraud probability threshold, high-confidence fraud candidate nodes are selected and merged with the labeled fraud nodes to form an extended fraud node set. On the extended fraud node set, the long-distance dependency between the representations of each node is calculated based on the attention propagation mechanism to generate a global fraud dependency representation. The fusion classification and output module is used to calculate the similarity between the global fraud dependency representation and the normal category prototype and the fraud category prototype, obtain the category belonging weight of the global fraud dependency representation, and map it to the normal channel and the fraud channel respectively, obtain the global information of the mapped normal channel and the fraud channel respectively, fuse it into the normal channel aggregate representation and the fraud channel aggregate representation respectively with a preset fusion coefficient, and combine it with the current representation of the target node to generate the final node representation, input it into the classifier and output the probability that each node is a fraud node.

[0019] The present invention has at least the following beneficial effects: First, this invention addresses the problem of insufficient local neighborhood matching caused by fraudulent entities actively connecting with normal entities to disguise themselves in financial fraud detection. By introducing category-based weights during local neighborhood aggregation, neighborhood information is allocated to normal channels and fraud channels for separate aggregation, achieving channel-level separation of neighborhood information. This suppresses the dilution effect of a large number of normal neighbors on the representation of fraudulent nodes, allowing fraudulent nodes to retain and transmit similar fraudulent information through the fraud channel even when surrounded by normal nodes in the local neighborhood. This improves the ability to identify disguised fraudulent nodes. Compared to the traditional graph neural network's method of indiscriminately aggregating all neighbors, this invention changes the way neighborhood information is fused through dual-channel group aggregation, enabling the model to distinguish between heterogeneous interference information and similar enhancement information in the neighborhood. This provides a feasible technical path for node representation learning in heterogeneous neighborhood scenarios.

[0020] Secondly, this invention addresses the challenges of weakly supervised learning caused by the high cost of acquiring fraud labels and the limited number of labeled samples in financial fraud detection. By utilizing a small number of labeled nodes to construct normal category prototypes and fraud category prototypes, and assigning soft category assignment weights to unlabeled nodes, category information can be propagated from labeled nodes to unlabeled nodes. The soft label estimation mechanism does not rely on external prior knowledge; it only relies on the similarity between the node representations of the current propagation layer and the category prototypes to provide a reference for the category assignment of unlabeled nodes. This alleviates the problem of insufficient supervision signals under conditions of scarce labels to a certain extent. At the same time, by introducing auxiliary constraints on category assignment into the training objective, the category prototype estimation and classification prediction can mutually verify and optimize each other, which helps to mine the structured information contained in unlabeled nodes.

[0021] Third, this invention addresses the problem of sparse distribution of similar fraudulent nodes in the global network and the difficulty in forming effective matching signals within local neighborhoods. By selecting high-confidence fraud candidate nodes from unlabeled nodes and combining them with labeled fraudulent nodes to form an extended fraud node set, and performing global attention propagation only within the set, fraudulent nodes sparsely distributed throughout the network can directly interact across local neighborhoods, capturing long-distance similar fraudulent dependencies. The global dependency modeling method differs from the traditional graph Transformer approach of performing indiscriminate global interaction on all nodes. It can supplement the long-distance matching signals that are difficult to obtain through local neighborhood aggregation, and avoid the information interference and computational overhead caused by introducing a large number of irrelevant normal nodes. By fusing global fraud dependency information and local enhanced representations in the same semantic space, the final representation of each node combines local neighborhood information and global matching association information, providing a feasible technical solution for the detection of sparsely distributed fraudulent nodes in large-scale financial activity networks.

[0022] Other advantages, objectives and features of the present invention will become apparent in part from the following description, and in part from those skilled in the art through study and practice of the invention. Attached Figure Description

[0023] Figure 1 This is a flowchart illustrating one of the technical solutions of the present invention: a financial fraud detection method based on isomatch enhancement and weakly supervised learning. Detailed Implementation

[0024] The present invention will now be described in further detail with reference to examples, so that those skilled in the art can implement it based on the description.

[0025] It should be noted that, unless otherwise specified, the experimental methods described in the following implementation plan are all conventional methods, and the reagents and materials described are all commercially available unless otherwise specified.

[0026] like Figure 1 As shown, this invention provides a financial fraud detection method based on isomatch enhancement and weakly supervised learning, comprising: S1. Obtain the raw data of financial activities, extract entity objects as nodes, construct the financial activity relationship between entity objects as edges, extract node attribute features to generate node features, and obtain graph data containing node features, adjacency relationships and weak supervision labels. Weak supervision labels include normal labels for labeled normal nodes and fraud labels for labeled fraud nodes. Unlabeled nodes remain unlabeled. S2. In the current propagation layer, construct normal category prototypes and fraud category prototypes based on the representations of labeled normal nodes and fraud nodes in the current propagation layer, calculate the similarity between the current representation of unlabeled nodes and the normal category prototypes and fraud category prototypes, generate soft classification weights for unlabeled nodes to belong to the normal category and fraud category, and set deterministic classification vectors for labeled nodes to obtain the classification weights of each node. S3. Obtain the local neighborhood node set of each node, with the node itself as the target node. Based on the category weight of the neighborhood nodes, assign the representations of the neighborhood nodes to the normal channel and the fraud channel respectively. Based on the current representation of the target node, generate adaptive aggregation weights for the normal channel and the fraud channel respectively. Then, weight and aggregate the representations of the neighborhood nodes in the normal channel and the fraud channel to generate the normal channel aggregated representation and the fraud channel aggregated representation respectively. Combine them with the current representation of the target node to generate the local enhancement representation of the next propagation layer. S4. Repeat steps S2 to S3 for multi-layer iterative propagation. In the last propagation layer, input the local enhanced representation of the node into the classifier to obtain the predicted probability of the unlabeled node belonging to the normal category and the fraud category. Filter high-confidence fraud candidate nodes according to the preset fraud probability threshold and merge them with the labeled fraud nodes to form an extended fraud node set. Calculate the long-distance dependency between the representations of each node on the extended fraud node set based on the attention propagation mechanism to generate a global fraud dependency representation. S5. Calculate the similarity between the global fraud dependency representation and the normal category prototype and the fraud category prototype to obtain the category belonging weight of the global fraud dependency representation, and map it to the normal channel and the fraud channel respectively. Obtain the global information of the mapped normal channel and the fraud channel respectively and fuse it into the normal channel aggregate representation and the fraud channel aggregate representation with the preset fusion coefficient. Combine it with the current representation of the target node to generate the final node representation, input it into the classifier and output the probability that each node is a fraud node.

[0027] In the above technical solution, by constructing a financial activity network, and using local neighborhood grouping and aggregation to enhance category consistency information under the condition of only a small number of labeled samples, while supplementing long-distance matching signals by using global fraud dependency modeling, the accuracy and robustness of financial fraud detection are improved. Specifically, the following steps are included: Raw data of financial activities is obtained from financial business systems, transaction flow systems, account management systems, equipment risk control systems, or third-party risk monitoring systems. Various entity objects in the financial activity scenario are extracted from the raw data. Entity objects include, but are not limited to, accounts, users, merchants, transactions, equipment, or financial instruments. Each entity object is represented as a node in the graph data, and all nodes constitute the node set of the graph data. Edges of the graph data are constructed based on the financial activity relationships between entity objects. Specifically, if there is any one or more of the following relationships between two entity objects: transaction, transfer, shared device login, shared account binding, shared payment, shared merchant access, or fund link relationship, then an edge is established between the corresponding two nodes. All edges constitute the edge set of the graph data. An adjacency matrix or adjacency list of the graph data is generated based on the edge construction results to indicate whether there is a financial activity relationship between any two nodes. For each node, its attribute features are extracted, including information on transaction frequency, transaction amount, transaction time distribution, account activity, device switching frequency, number of payment and receipt parties, historical risk records, or fund flow paths. The extracted features from each node are combined to generate a feature vector for that node. All node feature vectors together constitute the node feature matrix of the graph data. Finally, weakly supervised labels are formed on the graph data. Specifically, nodes corresponding to entities confirmed as having fraud risk through manual review, regulatory confirmation, historical blacklisting, or case retrospective analysis are marked as fraudulent nodes. Nodes corresponding to entities confirmed as normal through manual review, long-term stable transaction records, risk-free records, or whitelisting are marked as normal nodes. Nodes that have not yet been confirmed or lack clear risk conclusions remain unlabeled nodes, i.e., they are not assigned a true label. The set of labeled normal nodes and the set of labeled fraudulent nodes together constitute the set of labeled nodes, and the set of unlabeled nodes constitutes the set of unlabeled nodes. This yields financial activity graph data containing node features, adjacency relationships, and weakly supervised labels, which serves as input for subsequent models. In the current propagation layer, initially, the current representation of each node is its initial feature vector. First, normal category prototypes and fraud category prototypes are constructed based on the labeled nodes. Specifically, the current representations of all labeled normal nodes are averaged and aggregated to obtain the normal category prototype; the current representations of all labeled fraud nodes are averaged and aggregated to obtain the fraud category prototype. The normal category prototype reflects the average representation direction of labeled normal nodes in the current layer, and the fraud category prototype reflects the average representation direction of labeled fraud nodes in the current layer. For each node to be estimated, including unlabeled nodes and optional labeled nodes, its current representation is mapped to the same comparable latent space along with the normal category prototype and the fraud category prototype, respectively. By calculating the similarity between the mapped representation and the two category prototypes, the soft weights for classifying the node as normal or fraud are estimated. The higher the similarity, the greater the weights. A higher similarity value indicates that the node is closer to the prototype of the category, and the greater the weight it has to be assigned to that category. The two calculated similarity values ​​are normalized so that the sum of the weights of the node belonging to the normal category and the fraud category is 1. These are used as the soft assignment weights for the normal category and the fraud category, respectively. For labeled normal and fraudulent nodes, the above soft assignment weights are not used. Instead, their category assignment vectors are set to deterministic values: the category assignment vector for labeled normal nodes is 1 for the normal category and 0 for the fraud category; the category assignment vector for labeled fraudulent nodes is 0 for the normal category and 1 for the fraud category. For unlabeled nodes, the soft assignment weights calculated through similarity are used as their category assignment weights. In this way, a unified node-category assignment weight matrix is ​​formed for all nodes, providing a basis for subsequent local neighborhood grouping and aggregation. For each target node in the graph data, firstly, the one-hop neighbor nodes of the target node are obtained based on adjacency relationships. This can be extended to multi-hop neighbors as needed, forming a local neighborhood node set for the target node. Then, the normal category assignment weight and fraud category assignment weight of each neighbor node in this local neighborhood node set are read. Based on the normal category assignment weight of each neighbor node, the representation of the neighbor node is assigned to the normal channel according to its weight. Based on the fraud category assignment weight of each neighbor node, the representation of the neighbor node is assigned to the fraud channel according to its weight. For labeled normal neighbor nodes, since their weight for the normal category is essentially 1, their representations all enter the normal channel. For labeled fraud neighbor nodes, since their weight for the fraud category is essentially 1, their representations all enter the fraud channel. For unlabeled neighbor nodes, their representations are contributed to the normal channel and fraud channel respectively according to their normal soft assignment weight and fraud soft assignment weight. This grouping method avoids forcibly classifying unlabeled neighbors into a single category, effectively reducing the negative impact of mixing different category information and oversmoothing on node representation learning. Furthermore, to more accurately aggregate information from different channels, adaptive aggregation weights are generated for the target node based on its current representation, specifically for the normal channel and the fraud channel. These adaptive aggregation weights can dynamically adjust the aggregation intensity according to the target node's own feature state, making the aggregation process more targeted and flexible. Using these adaptive aggregation weights, the representations of neighboring nodes in the normal and fraud channels are weighted and aggregated to obtain normal channel aggregated representations and fraud channel aggregated representations. These two channel aggregated representations reflect the combined influence of normal semantic information and fraudulent semantic information in the target node's neighborhood, respectively. Finally, the target node's current representation, normal channel aggregated representation, and fraud channel aggregated representation are input into a fusion function. This fusion function integrates the three parts of information to generate a locally enhanced representation of the target node in the next propagation layer. This locally enhanced representation retains the target node's own information while strengthening the category consistency information that may be related to the target node in the neighborhood and suppressing interference from heterogeneous neighbors, thereby improving the representation quality of the fraud node in the spoofing scenario. Repeat the above steps to perform multi-level iterative propagation. In each propagation layer, the category prototype and the weight of each category are recalculated based on the node representation of the current layer. Local neighborhood grouping and aggregation are performed based on the updated weights to generate the local enhanced representation of the next layer. Through multi-level propagation, the node representation gradually integrates the information of higher-order neighborhoods while maintaining the separation and enhancement of the channel information of each category. After completing the propagation of the preset number of layers, the local augmented representations of each node output from the last layer are input into a classifier. The classifier calculates the predicted probability of each unlabeled node belonging to the normal category and the fraud category. Based on the preset fraud probability threshold, nodes with a fraud prediction probability not lower than the threshold are selected from all unlabeled nodes and designated as high-confidence fraud candidate nodes. The high-confidence fraud candidate nodes are then merged with the labeled fraud nodes to form an extended fraud node set. Based on the attention propagation mechanism, the long-distance dependencies between the representations of each node in the set are calculated, enabling any two nodes in the set to directly interact across the original local neighborhood. This generates a global fraud dependency representation for each node. This global fraud dependency representation can capture the potential matching associations of similar fraud nodes in the global scope, effectively compensating for the difficulty in obtaining long-distance supervision signals in the local neighborhood due to the sparse distribution of fraud nodes. The obtained global fraud dependency representation is compared with the constructed normal category prototype and fraud category prototype to calculate the category affiliation weight of the global representation to the normal category and fraud category. Based on these weights, the global fraud dependency representation is mapped to the normal channel and fraud channel respectively, forming global normal channel information and global fraud channel information. Then, with a preset fusion coefficient, the above global normal channel information and global fraud channel information are fused into the normal channel aggregation representation and fraud channel aggregation representation of the target node in step S3, respectively, to obtain the fused normal channel aggregation representation and fused fraud channel aggregation representation. The fusion coefficient is a preset hyperparameter used to control the contribution strength of global information to the local aggregation result. The target node's current representation and the fused normal channel... The three aggregated representations of channels and the fused fraud channel aggregated representations are input into a fusion function to generate the final node representation for each node. This final node representation simultaneously includes channel information enhanced by matching in the local neighborhood and long-distance fraud matching association information in the global scope. The two complement each other, significantly enhancing the ability to identify fraud nodes under weak supervision. Finally, the final node representation is input into a classifier, which outputs the probability of each node belonging to the normal category and the fraud category. The output fraud probability is compared with a preset detection threshold: if the fraud probability is higher than the threshold, the node is identified as a fraud risk node; if the fraud probability is not higher than the threshold, the node is identified as a normal or low-risk node, thus completing the fraud risk detection of all nodes in the financial activity network. During the model training phase, a supervised classification loss is constructed based on the true labels of labeled nodes and the model's predicted probabilities. Simultaneously, a category-based auxiliary constraint loss is constructed based on the true category attribution of labeled nodes and the predicted category attribution of unlabeled nodes. The two types of losses are weighted and summed to obtain the total training loss. All learnable parameters in the model are updated by minimizing this total training loss until the model converges. After training, the model can be used to quickly infer and output the fraud risk probability of any new input financial activity node.

[0028] Specifically, step S2, which involves constructing category prototypes and calculating node category affiliation weights, includes: Obtain the current representation of each node in the current propagation layer, and aggregate the current representations of labeled normal nodes and labeled fraudulent nodes respectively to generate normal category prototypes and fraudulent category prototypes. For category c, the category prototype is represented as: Where "be" represents the normal category and "fr" represents the fraud category. δ(y) represents the representation of node v at layer (l-1). v c) indicates whether the node label belongs to category c, and Δc represents the number of labeled nodes that belong to category c; The current representation of the node to be estimated, along with its normal category prototype and fraud category prototype, are mapped to the same latent space. The similarity between the node's representation and each prototype is calculated, generating soft attribution weights for the node's classification as either normal or fraudulent. (Node v) j The category affiliation vector is represented as: in, This represents a category prototype matrix consisting of normal category prototypes and fraud category prototypes. and As a learnable mapping matrix, a deterministic category attribution vector is set for labeled normal nodes and labeled fraudulent nodes respectively, and a soft attribution vector calculated from similarity is used for unlabeled nodes, forming a unified node-category attribution matrix.

[0029] By constructing normal category prototypes and fraud category prototypes, the real label information of a small number of labeled nodes is condensed into representative representations of the two categories. By calculating the similarity between the representations of unlabeled nodes and the two prototypes, a reasonable soft attribution weight is assigned to each unlabeled node, so that category information can be effectively propagated from a small number of labeled nodes to a large number of unlabeled nodes. This provides a reliable weight basis for subsequent local neighborhood grouping and aggregation, and solves the problem of insufficient labeling signals under weak supervision.

[0030] Specifically, in step S3, the specific steps for assigning the neighbor node representations to the two information channels are as follows: For the target node, obtain the one-hop or multi-hop neighbors of the target node according to the adjacency relationship to form a local neighborhood node set. Read the normal category assignment weight and fraud category assignment weight of each neighbor node in the local neighborhood node set. According to the normal category assignment weight of each neighbor node, input the weighted representation of the neighbor node into the normal channel; according to the fraud category assignment weight of each neighbor node, input the weighted representation of the neighbor node into the fraud channel. For labeled normal nodes and labeled fraudulent nodes, they are respectively used as deterministic category assignment vectors in the corresponding category channels; for unlabeled nodes, they are contributed to the normal channel and fraud channel according to their normal category soft assignment weight and fraudulent category soft assignment weight, respectively. target node v i The neighborhood influence is decomposed into normal category channel and fraud category channel: in, and Representing node v respectively j The weights assigned to the normal and fraud categories. Represents the neighborhood aggregation coefficient. and Let represent the learnable transformation matrices for the normal category channel and the fraud category channel, respectively. Represents the target node v i The influence of the neighborhood on the l-th layer.

[0031] By assigning the representations of neighboring nodes to the normal channel and the fraud channel according to the category affiliation weight, channel-level separation of neighborhood information is achieved, enabling the target node to obtain normal semantic information and fraud semantic information from the neighborhood respectively. This avoids the information mixing and oversmoothing problems caused by forcibly classifying unlabeled nodes into a single category, effectively suppresses the dilution and over-drowning of fraud node representations by normal neighbors, and ensures the accurate transmission of category consistency information during the local aggregation process. Specifically, step S4 includes the following steps: Repeat steps S2 to S3 in claim 1 to perform multi-level iterative propagation. In each propagation layer, update the normal category prototype and fraud category prototype based on the node representation of the current propagation layer and calculate the category assignment weight of each node. Based on the updated category assignment weight, perform local neighborhood grouping and aggregation to generate the local enhanced representation of the next propagation layer. In the last propagation layer, the local augmented representation of each node is input into the classifier to obtain the predicted probability of unlabeled nodes belonging to the normal category and the fraud category. Based on the preset fraud probability threshold, nodes with fraud prediction probabilities not lower than the fraud probability threshold are selected from the unlabeled nodes as high-confidence fraud candidate nodes. The high-confidence fraud candidate nodes are merged with the labeled fraud nodes to form an extended fraud node set, and the current representation of each node in the extended fraud node set is extracted to form a global fraud node representation matrix. On the global fraud node representation matrix, the long-distance dependency between the representations of each node in the extended fraud node set is calculated based on the linear attention mechanism to generate the global fraud dependency representation corresponding to each node. Through multi-layer iterative propagation, node representations are gradually integrated with higher-order neighborhood information. At the same time, each layer refines the category prototype and attribution weight, forming a virtuous cycle. On this basis, by screening high-confidence fraud candidate nodes and merging them with labeled fraud nodes to form an extended set, the sample size of fraud nodes is effectively expanded. More importantly, global attention propagation is only performed within the extended fraud node set, enabling similar fraud nodes sparsely distributed in the global network to directly interact across local neighborhoods, capturing long-distance fraud matching dependencies, while avoiding the computational overhead and interference from irrelevant normal nodes caused by global attention of all nodes. Specifically, step S5 includes the following steps: Calculate the similarity between the global fraud dependency representation and the normal category prototype and the fraud category prototype to obtain the category affiliation weight of the global fraud dependency representation to the normal category and the fraud category; Based on the category attribution weight, the global fraud dependency representation is mapped to the normal channel and the fraud channel respectively, to obtain the global information mapped to the normal channel and the global information mapped to the fraud channel. These are then summed with preset fusion coefficients and added to the normal channel aggregation representation and the fraud channel aggregation representation of the target node in step S3 respectively, to obtain the global-local fused normal channel aggregation representation and fraud channel aggregation representation. The normal channel aggregation representation and fraud channel aggregation representation after global-local fusion are combined with the current propagation layer representation of the target node to generate the final node representation; The final node representation is input into the classifier, and the probability of each node being a normal node and a fraud node is output through the Softmax function. Nodes with a fraud probability higher than the preset detection threshold are identified as fraud risk nodes. By mapping the global fraud dependency representation to the category channel and then summing it into the local channel aggregation result with a fusion coefficient, the organic fusion of global and local information under the same channel semantics is achieved. The final node representation carries both the channel information enhanced by isomatch in the local neighborhood and the long-distance isomatch association information between similar fraud nodes in the global scope. The two types of information complement each other. The introduction of the fusion coefficient allows the model to flexibly control the contribution intensity of global information, avoiding excessive dominance or insufficient contribution of global information, and significantly improving the ability to identify fake fraud nodes under weak supervision.

[0032] In another technical solution, step S4, where the linear attention mechanism computes the global fraud dependency representation, is represented as follows: Among them, X gt N represents the global fraud node representation matrix. gt W represents the number of nodes in the expanded fraud node set. Q and W K Let be a learnable mapping matrix, where 1 represents a column vector with all elements equal to 1. This represents the global fraud dependency representation matrix.

[0033] In the above technical solution, after screening high-confidence fraud candidate nodes, constructing an expanded fraud node set, and extracting the features of all nodes in the set to generate a global fraud node feature matrix X, the process is as follows: gt Then, a linear attention mechanism is used to model long-distance dependencies between nodes within the expanded set of fraudulent nodes. The specific calculation steps are as follows: The global fraud node feature matrix X gt Through two sets of learnable linear projection matrices W respectively Q W K After completing the feature mapping, the query matrix Q and key matrix K are output. Each row of the query matrix Q corresponds to the query vector of a single node, representing the current node's interest in associating with other nodes. Each row of the key matrix K corresponds to the key vector of a single node, serving as the feature identifier for that node that can be matched and recognized by other nodes. Q W K The parameters of both projection matrices are updated iteratively through backpropagation during the model training phase; The normalized diagonal matrix D required for linear attention is calculated using the query matrix Q and the key matrix K. The calculation logic is as follows: First, the key matrix K is multiplied by a column vector 1, where all elements are 1, to obtain an intermediate vector. The values ​​of each dimension of this vector are the sum of the key vectors of all nodes in the corresponding feature dimension. Then, the query matrix Q is multiplied by this intermediate vector to generate a column vector with the same dimensions as the total number of expanded fraudulent nodes. All elements in this column vector are incremented by 1. The elements of the processed vector are used as diagonal elements to construct the diagonal matrix D. The incrementing operation ensures the stability of the matrix operation values, and the accompanying scaling constraint can avoid excessively sharp attention distribution that could cause gradient oscillations. The i-th diagonal element of the diagonal matrix D is the normalization coefficient specific to the i-th node, which is used to perform row-by-row normalization processing on subsequent global dependent features. Directly use the global fraud node feature matrix X gt The result is reused as a value matrix, and the global fraud dependency feature matrix is ​​solved by combining the results of the interaction operation of Q and K. The final global dependency feature of a single node consists of two superimposed terms: the first term is the node's original feature X. gt The first term is an equivalent residual connection structure, ensuring that the inherent features of nodes are not diluted or lost during global information interaction. The second term is the global context aggregation information obtained by multiplying the query matrix, the transpose of the key matrix, and the value matrix. After scaling, each node can receive long-distance similar correlation feature signals transmitted by all other fraudulent nodes in the extended set. After summing the two features, row-wise normalization is performed using the inverse of the diagonal matrix D. Finally, the global fraud dependency feature matrix corresponding to each node in the extended fraud node set is output. The collaborative design of residual connectivity and scaling can alleviate the graph oversmoothing phenomenon that easily occurs during multi-layer feature propagation and maintain the distinguishability of feature representations of different nodes. Through the above linear attention operation, each node in the expanded fraud node set can obtain global features of all fraud-related node information in the fusion set, and complete the same-match dependency modeling of similar fraud nodes at the global level. This can effectively make up for the shortcomings of the local neighborhood aggregation strategy, which is difficult to capture long-distance supervision signals due to the sparse distribution of similar fraud nodes. The obtained global fraud dependency features can be directly input into the subsequent feature fusion and classification module to complete downstream tasks.

[0034] In another technical solution, the category channel aggregation result after global-local fusion in step S5 is represented as follows: in, "be" indicates the normal category, and "fr" indicates the fraud category. This represents the category channel aggregation function of the (l-1)th propagation layer. Represents node v iIn the global fraud dependency representation of the (l−1)th propagation layer, β represents the preset fusion coefficient. N(v) represents the mapping function that maps the global representation to the category channel. i ) represents node v i The set of local neighborhood nodes, and the final node representation is: in, Represents node v i The final node representation in the l-th propagation layer, Represents the fusion function. The mapping function represents the target node itself.

[0035] In the above technical solution, after completing the calculation through the aforementioned linear attention mechanism and obtaining the global fraud dependency representation of each node in the extended fraud node set, the global long-distance association information is deeply fused with the local neighborhood category grouping and aggregation result obtained in step S3 to make up for the feature loss problem caused by the sparse distribution of fraud nodes in the local neighborhood, and generate a high-precision node representation that simultaneously contains local matching enhancement information and global long-distance matching association information. This solution adopts a category channel alignment fusion strategy to unify global features and local features into the same category semantic space, avoiding feature misalignment and information distortion caused by cross-semantic fusion. The global fraud dependency representation of the target node is subjected to category adaptation processing. The feature similarity between the global representation and the prototype of the normal category and the prototype of the fraud category in the current layer is calculated to obtain the soft attribution weights of the global representation to the normal category and the fraud category respectively. The soft weights represent the degree of adaptation of the global features to the two semantic categories. Based on the learnable category channel mapping function, the global fraud dependency representation is projected to the feature space that is completely consistent with the local category channel aggregation result. The projected features are weighted and corrected in combination with the aforementioned soft attribution weights, so that the global features are reasonably allocated to the normal category channel and the fraud category channel according to their own correlation with different categories, thus completing the category channel adaptation of the global features. Using a preset fusion coefficient, the global normal channel information and global fraud channel information after channel adaptation are superimposed on the local normal channel aggregation representation and the local fraud channel aggregation representation corresponding to the target node, respectively, to realize the enhancement correction of the local aggregation features by the global features, and obtain the category channel aggregation results after global-local fusion. In this embodiment, the fusion coefficient is a hyperparameter with a value between 0 and 1, which is used to accurately control the contribution ratio of global information in the feature fusion process and adapt to the fraud node distribution characteristics of different financial networks. Specifically, when fraudulent nodes are sparsely distributed globally and the effective monitoring signals in local neighborhoods are weak, the value of β can be increased to strengthen the dominant role of global long-distance matching information and compensate for the defects of local feature failure. When the local neighborhood of a node has strong class consistency and high local feature recognition, the value of β can be decreased to prioritize local aggregated features and supplement global features, thus preserving local fine semantic information. In actual deployment, the optimal value of β can be determined through iterative optimization of the validation set performance, or it can be fixed in the range of 0.3 to 0.5 to stably balance the fusion weights of global and local features. This additive fusion method achieves feature superposition based on a unified category semantic channel, which can accurately introduce global long-distance association information while preserving the original local aggregated features, thus completely avoiding semantic mismatch and feature redundancy problems caused by cross-channel fusion. After obtaining the normal category channel aggregation representation and the fraud category channel aggregation representation after global-local fusion, the two sets of category channel fusion features, together with the target node's own representation after correction by the self-mapping function, are input into the fusion function to integrate and output the final node representation of the target node in the l-th propagation layer. The fusion function can be flexibly designed according to the actual task scenario and can be implemented in various ways such as feature concatenation + linear transformation, gated adaptive fusion, and multilayer perceptron nonlinear fusion. Its core function is to efficiently integrate local neighborhood category enhancement information and global long-distance matching association information while preserving the inherent semantic features of the node itself. Through the above hierarchical fusion design, the final output node representation has both local fine-grained category features and global coarse-grained correlation features, which solves the technical defects of traditional local aggregation models that cannot capture the long-distance dependence of sparse fraud nodes and the loss of local detail information by single global features. Subsequently, the enhanced final node representation can be input into a classifier to calculate the predicted probability of each node being a fraud node and complete the fraud node identification task. This implementation relies on a global-local fusion mechanism based on category channel alignment to achieve organic complementarity between global fraud-dependent representation and local enhancement representation. At the same time, by adjusting the fusion coefficient to adapt to different fraud distribution scenarios, it significantly improves the identification accuracy and model generalization ability of disguised fraud nodes in weak supervision and sparse node distribution scenarios, ensuring the stability and effectiveness of fraud node detection in various financial transaction networks and social connection networks.

[0036] In another technical solution, after obtaining the final node representation of each node through the aforementioned steps, the present invention further provides a preferred implementation method for model training and fraud risk output. Specifically, in each round of iterative training, the final node representation of each node is... Input the classifier, where L is the preset total number of propagation layers, and obtain the model's predicted probabilities. ,in and These represent the probabilities that node v is predicted as either fraudulent or normal, respectively. Construct the total training loss function L total It is composed of supervised classification loss L cls And category attribution auxiliary constraint loss L aux The weighted sum of the two parts yields: Wherein, λ is a preset tradeoff hyperparameter used to adjust the contribution intensity of the auxiliary constraint loss, and its value range is preferably 0.1 to 1.0; Supervised classification loss L cls It only applies to the set of labeled nodes V L The cross-entropy loss function is used: in, The indicator function takes a value of 1 when the true label of node v is equal to c, and 0 otherwise. This loss forces the model to produce accurate classification predictions on a small number of labeled nodes and is the main source of model supervision signals. Category attribution auxiliary constraint loss L aux The loss is used to constrain the consistency between the class assignment weights of each node, including labeled and unlabeled nodes, and the model's predictions, thereby enhancing the class discrimination ability under weak supervision. The loss is defined as: in, The class assignment weight of node v in the last propagation layer is a deterministic assignment vector for labeled nodes and a soft assignment weight after Softmax normalization for unlabeled nodes. The auxiliary loss makes the predicted probability of unlabeled nodes move closer to its estimated class assignment direction, so that the class prototype estimation and classification prediction can enhance each other and optimize each other. In this embodiment, the total loss function L total All learnable parameters in the model are updated end-to-end using the backpropagation algorithm, including the mapping matrix in the category prototype construction and attribution estimation modules. , The transformation matrix in the local neighborhood grouping and aggregation module , and adaptive aggregation weight generation function The parameters in the modeling module include the linear projection matrix W. Q W K The classifier parameters and fusion function in the classification and output module are fused. Learnable parameters in the propagation layer and neighborhood aggregation coefficients in each propagation layer Training employs a mini-batch stochastic gradient descent optimizer, such as Adam, with a learning rate set to 0.001–0.01. The number of iterations is determined based on early stopping performance on the validation set to ensure the model's generalization ability. After the model converges, for any target node v, represent its final node. Input a trained classifier and output the fraud probability. The probability of fraud is compared with a preset detection threshold. Comparison: If If the node is identified as a fraud risk node, it is considered a normal or low-risk node, and the detection threshold is set accordingly. The precision / recall ratio can be flexibly set according to different requirements of business scenarios. The value range is usually 0.3 to 0.7, with 0.5 being preferred. In risk control practice, this threshold can be adjusted to balance the risks of false positives and false negatives: if it is necessary to reduce false negatives, it is better to accept false positives and the threshold can be appropriately lowered; if it is necessary to improve precision and reduce false positives, the threshold can be appropriately increased. The above training and prediction process enables the model to make full use of the soft supervision information of a large number of unlabeled nodes with only a small number of labeled samples and the auxiliary constraint of category classification. At the same time, it supplements the long-distance matching signal through global fraud dependency representation, and finally achieves high-precision fraud node identification. It can be flexibly deployed in real-time risk control systems for financial transactions, anti-money laundering monitoring platforms and abnormal account screening tools to quickly assess risks and output early warnings for newly connected accounts, transactions or merchant nodes.

[0037] This invention provides a financial fraud detection system based on isomatch enhancement and weakly supervised learning, comprising: The data acquisition and graph construction module is used to acquire raw data of financial activities, extract entity objects as nodes, construct the financial activity relationship between entity objects as edges, extract node attribute features to generate node features, and obtain graph data containing node features, adjacency relationships and weak supervision labels. The weak supervision labels include normal labels for labeled normal nodes and fraud labels for labeled fraud nodes. Unlabeled nodes remain unlabeled. The category prototype construction and classification estimation module is used to construct normal category prototypes and fraud category prototypes respectively in the current propagation layer based on the representations of labeled normal nodes and fraud nodes in the current propagation layer, calculate the similarity between the current representation of unlabeled nodes and the normal category prototypes and fraud category prototypes, generate soft classification weights for unlabeled nodes to be classified into normal and fraud categories, and set deterministic classification vectors for labeled nodes to obtain the classification weights of each node. The local neighborhood grouping and aggregation module is used to obtain the local neighborhood node set of each node, with the node itself as the target node. Based on the category belonging weight of the neighborhood nodes, the representations of the neighborhood nodes are assigned to the normal channel and the fraud channel respectively. Based on the current representation of the target node, adaptive aggregation weights are generated for the normal channel and the fraud channel respectively. The representations of the neighborhood nodes in the normal channel and the fraud channel are weighted and aggregated to generate the normal channel aggregated representation and the fraud channel aggregated representation respectively. These are then combined with the current representation of the target node to generate the local enhancement representation of the next propagation layer. The global dependency modeling module is used to repeatedly call the category prototype construction and attribution estimation module and the local neighborhood grouping and aggregation module to perform multi-layer iterative propagation. In the last propagation layer, the local enhanced representation of the node is input into the classifier to obtain the predicted probability that the unlabeled node belongs to the normal category and the fraud category. Based on the preset fraud probability threshold, high-confidence fraud candidate nodes are selected and merged with the labeled fraud nodes to form an extended fraud node set. On the extended fraud node set, the long-distance dependency between the representations of each node is calculated based on the attention propagation mechanism to generate a global fraud dependency representation. The fusion classification and output module is used to calculate the similarity between the global fraud dependency representation and the normal category prototype and the fraud category prototype, obtain the category belonging weight of the global fraud dependency representation, and map it to the normal channel and the fraud channel respectively, obtain the global information of the mapped normal channel and the fraud channel respectively, fuse it into the normal channel aggregate representation and the fraud channel aggregate representation respectively with a preset fusion coefficient, and combine it with the current representation of the target node to generate the final node representation, input it into the classifier and output the probability that each node is a fraud node.

[0038] In the above technical solution, the learnable parameters such as mapping matrix, transformation matrix, aggregation weight, and fusion coefficient involved in the execution of each module are jointly optimized by the supervised classification loss of labeled nodes and the category attribution auxiliary constraint loss. During model training, these parameters are updated end-to-end until convergence. The signal flow and data interaction relationships between the modules of the system are as follows: The graph data output by the data acquisition and graph construction module is sequentially processed through multiple layers of iteration by the category prototype construction and attribution estimation module and the local neighborhood grouping and aggregation module. The output of each layer serves as the input of the next layer. After completing the propagation of all preset layers, the global dependency modeling module constructs an extended fraud node set based on the output of the last layer and generates a global fraud dependency representation. The fusion classification and output module merges the global fraud dependency representation with the local aggregation results and outputs the final fraud risk probability. The modules work together to achieve accurate detection of fraud nodes in the financial activity network under the condition of a small number of labeled samples. Example

[0039] For the experimental dataset, this embodiment uses the Elliptic dataset, which is used in financial transaction fraud scenarios, for validation. The Elliptic dataset is a transaction network dataset designed to identify illegal or fraudulent transaction nodes in financial transaction networks. The dataset contains 46,564 nodes and 73,248 edges, with fraudulent nodes accounting for 9.76% and an overall matching rate of 69.17%. The dataset is characterized by complex financial transaction relationships, scarce fraud labels, a low proportion of fraudulent nodes, and a sparse distribution of fraudulent nodes, which can effectively verify the effectiveness of this invention in weakly supervised financial fraud detection scenarios. To comprehensively evaluate the performance of the present invention, this embodiment compares the present invention with the following three types of methods: Traditional graph neural network methods, including GCN and GAT, mainly rely on neighborhood message aggregation for node representation learning, which is easily affected by normal neighbor information introduced by fraudulent node spoofing behavior. Heterogeneous mapping neural network methods include BWGNN, GHRN, GDN, GAGA, PMP, and ConsisGAD, which alleviate the problems of heterogeneous neighborhood, class imbalance, or neighborhood noise in fraud detection in different ways. Graph Transformer methods include SGFormer, which can introduce global information, but if global interaction is performed directly on all nodes, it may introduce irrelevant normal node information and weaken fraudulent matching signals. This invention enhances the effective representation of financial fraud nodes under weak supervision by learning soft attribution of unlabeled node categories, local grouping aggregation of normal / fraudulent nodes, and acquisition of global matching information on an expanded fraud node set. Considering the imbalance of positive and negative samples in the financial fraud detection dataset, the evaluation metrics should take into account both the overall ranking ability and the ability to identify minority fraud nodes. In this embodiment, ROC-AUC and PR-AUC are selected as two evaluation metrics to evaluate the performance of each method. The performance comparison results of each method on the Elliptic dataset are shown in Table 1.

[0040] Table 1. Performance comparison of each method on the Elliptic dataset (%) As shown in Table 1, the ROC-AUC of this invention reaches 98.93% and the PR-AUC reaches 95.38% on the Elliptic financial transaction fraud dataset, both significantly outperforming all comparative methods. These experimental results fully verify the effectiveness and superiority of this invention. Compared with traditional graph neural network methods (GCN, GAT), this invention, through a normal / fraud dual-channel grouping aggregation mechanism, effectively avoids the dilution of fraud node representations by a large amount of normal neighbor information during local neighborhood aggregation, significantly improving the identification ability of fraud nodes. Compared with existing heterogeneous graph neural network methods (BWGNN, GHRN, GDN, GAGA, PMP, ConsisGAD), this invention further utilizes the soft classification of unlabeled nodes to enhance class consistency information under weak supervision conditions, making the classification more consistent with the labeled nodes. Even with a very small sample size, this invention can still obtain high-quality node representations. Compared with the Graph Transformer method (SGFormer), this invention does not perform indiscriminate global interaction on all nodes. Instead, it selects high-confidence fraud candidate nodes and selectively obtains global fraud association information only within the expanded fraud node set. This captures the dependencies between distant, similar fraud nodes and effectively reduces interference from irrelevant normal nodes, while also reducing computational overhead. This invention can effectively alleviate the problems of insufficient weak supervision signals caused by fraud node masquerading, scarcity of fraud labels, and sparse distribution of similar fraud nodes in financial transaction fraud detection. It significantly improves the accuracy, robustness, and scalability of financial fraud detection and can be widely applied to practical scenarios such as financial transaction risk control, anti-money laundering identification, and abnormal account screening.

[0041] Although embodiments of the present invention have been disclosed above, they are not limited to the applications listed in the specification and embodiments. They can be applied to various fields suitable for the present invention. Other modifications can be easily made by those skilled in the art. Therefore, without departing from the general concept defined by the claims and their equivalents, the present invention is not limited to the specific details and examples shown and described herein.

Claims

1. A financial fraud detection method based on isomatch enhancement and weakly supervised learning, characterized in that, include: S1. Obtain the raw data of financial activities, extract entity objects as nodes, construct the financial activity relationship between entity objects as edges, extract node attribute features to generate node features, and obtain graph data containing node features, adjacency relationships and weak supervision labels. Weak supervision labels include normal labels for labeled normal nodes and fraud labels for labeled fraud nodes. Unlabeled nodes remain unlabeled. S2. In the current propagation layer, construct normal category prototypes and fraud category prototypes based on the representations of labeled normal nodes and fraud nodes in the current propagation layer, calculate the similarity between the current representation of unlabeled nodes and the normal category prototypes and fraud category prototypes, generate soft classification weights for unlabeled nodes to belong to the normal category and fraud category, and set deterministic classification vectors for labeled nodes to obtain the classification weights of each node. S3. Obtain the local neighborhood node set of each node, with the node itself as the target node. Based on the category weight of the neighborhood nodes, assign the representations of the neighborhood nodes to the normal channel and the fraud channel respectively. Based on the current representation of the target node, generate adaptive aggregation weights for the normal channel and the fraud channel respectively. Then, weight and aggregate the representations of the neighborhood nodes in the normal channel and the fraud channel to generate the normal channel aggregated representation and the fraud channel aggregated representation respectively. Combine them with the current representation of the target node to generate the local enhancement representation of the next propagation layer. S4. Repeat steps S2 to S3 for multi-layer iterative propagation. In the last propagation layer, input the local enhanced representation of the node into the classifier to obtain the predicted probability of the unlabeled node belonging to the normal category and the fraud category. Filter high-confidence fraud candidate nodes according to the preset fraud probability threshold and merge them with the labeled fraud nodes to form an extended fraud node set. Calculate the long-distance dependency between the representations of each node on the extended fraud node set based on the attention propagation mechanism to generate a global fraud dependency representation. S5. Calculate the similarity between the global fraud dependency representation and the normal category prototype and the fraud category prototype to obtain the category belonging weight of the global fraud dependency representation, and map it to the normal channel and the fraud channel respectively. Obtain the global information of the mapped normal channel and the fraud channel respectively and fuse it into the normal channel aggregate representation and the fraud channel aggregate representation with the preset fusion coefficient. Combine it with the current representation of the target node to generate the final node representation, input it into the classifier and output the probability that each node is a fraud node.

2. The financial fraud detection method based on isomatch enhancement and weakly supervised learning as described in claim 1, characterized in that, In step S1, the raw data comes from the business database or risk monitoring database of financial activities; The extracted entity objects include at least one of the following: account entity, user entity, merchant entity, transaction entity, device entity, and financial instrument entity; Financial activity relationships include at least one of the following: there is a transaction, transfer, shared login device, shared account binding, shared payment and receipt, shared access to merchant or fund link relationship between two entities; The process involves generating an adjacency matrix or adjacency list for graph data based on the edge construction results. The adjacency matrix or adjacency list is used to represent the relationships between nodes. For each node, at least one of the following features is extracted: transaction frequency, transaction amount, transaction time distribution, account activity, device switching frequency, number of payment and receipt objects, historical risk record features, or fund flow path features. The extracted features are combined to generate a node feature vector, and all node feature vectors are combined into a node feature matrix.

3. The financial fraud detection method based on isomatch enhancement and weakly supervised learning as described in claim 1, characterized in that, In step S1, the steps for forming weak supervision labels include: Nodes corresponding to entities that have been manually reviewed, confirmed by regulators, found in historical blacklists, or confirmed as fraudulent through case retrospective analysis will be marked as fraudulent nodes. Nodes corresponding to entities that have been manually reviewed, have no abnormal transaction records, no risk records, or have been confirmed as normal by whitelists will be marked as normal nodes. Nodes that have not been confirmed or lack clear risk conclusions will be marked as unmarked nodes.

4. The financial fraud detection method based on isomatch enhancement and weakly supervised learning as described in claim 1, characterized in that, Step S2, which involves constructing category prototypes and calculating node category affiliation weights, includes: Obtain the current representation of each node in the current propagation layer, and aggregate the current representations of labeled normal nodes and labeled fraudulent nodes respectively to generate normal category prototypes and fraudulent category prototypes. For category c, the category prototype is represented as: Among them, V L This represents the set of labeled nodes, which is the union of all labeled normal nodes and labeled fraudulent nodes. 'be' represents the normal category, and 'fr' represents the fraudulent category. δ(y) represents the representation of node v at layer (l-1). v c) indicates whether the node label belongs to category c, and Δc represents the number of labeled nodes that belong to category c; The current representation of the node to be estimated, along with its normal category prototype and fraud category prototype, are mapped to the same latent space. The similarity between the node's representation and each prototype is calculated, generating soft attribution weights for the node's classification as either normal or fraudulent. (Node v) j The category affiliation vector is represented as: in, This represents a category prototype matrix consisting of normal category prototypes and fraud category prototypes. and As a learnable mapping matrix, a deterministic category attribution vector is set for labeled normal nodes and labeled fraudulent nodes respectively, and a soft attribution vector calculated from similarity is used for unlabeled nodes, forming a unified node-category attribution matrix.

5. The financial fraud detection method based on isomatch enhancement and weakly supervised learning as described in claim 1, characterized in that, In step S3, the specific steps for assigning the neighbor node representations to the two information channels are as follows: For the target node, obtain the one-hop or multi-hop neighbors of the target node according to the adjacency relationship to form a local neighborhood node set. Read the normal category assignment weight and fraud category assignment weight of each neighbor node in the local neighborhood node set. According to the normal category assignment weight of each neighbor node, input the weighted representation of the neighbor node into the normal channel. Based on the fraud category weight of each neighboring node, the representations of the neighboring nodes are weighted and input into the fraud channel. For labeled normal nodes and labeled fraudulent nodes, they are respectively used as deterministic category assignment vectors in the corresponding category channels; for unlabeled nodes, they are contributed to the normal channel and fraud channel according to their normal category soft assignment weight and fraudulent category soft assignment weight, respectively. target node v i The neighborhood influence is decomposed into normal category channel and fraud category channel: Wherein, N(v) i ) represents the target node v i A local domain node set, and Representing node v respectively j The weights assigned to the normal and fraud categories. Represents the neighborhood aggregation coefficient. and Let represent the learnable transformation matrices for the normal category channel and the fraud category channel, respectively. Represents the target node v i The influence of the neighborhood on the l-th layer.

6. The financial fraud detection method based on isomatch enhancement and weakly supervised learning as described in claim 1, characterized in that, Step S4 specifically includes the following steps: Repeat steps S2 to S3 in claim 1 to perform multi-level iterative propagation. In each propagation layer, update the normal category prototype and fraud category prototype based on the node representation of the current propagation layer and calculate the category assignment weight of each node. Based on the updated category assignment weight, perform local neighborhood grouping and aggregation to generate the local enhanced representation of the next propagation layer. In the last propagation layer, the local augmented representation of each node is input into the classifier to obtain the predicted probability of unlabeled nodes belonging to the normal category and the fraud category. Based on the preset fraud probability threshold, nodes with fraud prediction probabilities not lower than the fraud probability threshold are selected from the unlabeled nodes as high-confidence fraud candidate nodes. The high-confidence fraud candidate nodes are merged with the labeled fraud nodes to form an extended fraud node set, and the current representation of each node in the extended fraud node set is extracted to form a global fraud node representation matrix. On the global fraud node representation matrix, the long-distance dependency between the representations of each node in the extended fraud node set is calculated based on the linear attention mechanism, and the global fraud dependency representation corresponding to each node is generated.

7. The financial fraud detection method based on isomatch enhancement and weakly supervised learning as described in claim 1, characterized in that, Step S5 specifically includes the following steps: Calculate the similarity between the global fraud dependency representation and the normal category prototype and the fraud category prototype to obtain the category affiliation weight of the global fraud dependency representation to the normal category and the fraud category; Based on the category attribution weight, the global fraud dependency representation is mapped to the normal channel and the fraud channel respectively, to obtain the global information mapped to the normal channel and the global information mapped to the fraud channel. These are then summed with preset fusion coefficients and added to the normal channel aggregation representation and the fraud channel aggregation representation of the target node in step S3 respectively, to obtain the global-local fused normal channel aggregation representation and fraud channel aggregation representation. The normal channel aggregation representation and fraud channel aggregation representation after global-local fusion are combined with the current propagation layer representation of the target node to generate the final node representation; The final node representation is input into the classifier, and the Softmax function outputs the probability that each node is a normal node and a fraudulent node. Nodes with a fraud probability higher than a preset detection threshold are identified as fraud risk nodes.

8. The financial fraud detection method based on isomatch enhancement and weakly supervised learning as described in claim 1, characterized in that, In step S4, the linear attention mechanism computes the global fraud dependency representation as follows: Among them, X gt N represents the global fraud node representation matrix. gt W represents the number of nodes in the expanded fraud node set. Q and W K Let be a learnable mapping matrix, where 1 represents a column vector with all elements equal to 1. This represents the global fraud dependency representation matrix.

9. The financial fraud detection method based on isomatch enhancement and weakly supervised learning as described in claim 1, characterized in that, In step S5, the category channel aggregation result after global-local fusion is represented as follows: in, "be" indicates the normal category, and "fr" indicates the fraud category. This represents the category channel aggregation function of the (l-1)th propagation layer. Represents node v i In the global fraud dependency representation of the (l−1)th propagation layer, β represents the preset fusion coefficient. N(v) represents the mapping function that maps the global representation to the category channel. i ) represents node v i The set of local neighborhood nodes, and the final node representation is: in, Represents node v i The final node representation in the l-th propagation layer, Represents the fusion function. The mapping function represents the target node itself.

10. A financial fraud detection system based on isomatch enhancement and weakly supervised learning, characterized in that, include: The data acquisition and graph construction module is used to acquire raw data of financial activities, extract entity objects as nodes, construct the financial activity relationship between entity objects as edges, extract node attribute features to generate node features, and obtain graph data containing node features, adjacency relationships and weak supervision labels. The weak supervision labels include normal labels for labeled normal nodes and fraud labels for labeled fraud nodes. Unlabeled nodes remain unlabeled. The category prototype construction and classification estimation module is used to construct normal category prototypes and fraud category prototypes respectively in the current propagation layer based on the representations of labeled normal nodes and fraud nodes in the current propagation layer, calculate the similarity between the current representation of unlabeled nodes and the normal category prototypes and fraud category prototypes, generate soft classification weights for unlabeled nodes to be classified into normal and fraud categories, and set deterministic classification vectors for labeled nodes to obtain the classification weights of each node. The local neighborhood grouping and aggregation module is used to obtain the local neighborhood node set of each node, with the node itself as the target node. Based on the category belonging weight of the neighborhood nodes, the representations of the neighborhood nodes are assigned to the normal channel and the fraud channel respectively. Based on the current representation of the target node, adaptive aggregation weights are generated for the normal channel and the fraud channel respectively. The representations of the neighborhood nodes in the normal channel and the fraud channel are weighted and aggregated to generate the normal channel aggregated representation and the fraud channel aggregated representation respectively. These are then combined with the current representation of the target node to generate the local enhancement representation of the next propagation layer. The global dependency modeling module is used to repeatedly call the category prototype construction and attribution estimation module and the local neighborhood grouping and aggregation module to perform multi-layer iterative propagation. In the last propagation layer, the local enhanced representation of the node is input into the classifier to obtain the predicted probability that the unlabeled node belongs to the normal category and the fraud category. Based on the preset fraud probability threshold, high-confidence fraud candidate nodes are selected and merged with the labeled fraud nodes to form an extended fraud node set. On the extended fraud node set, the long-distance dependency between the representations of each node is calculated based on the attention propagation mechanism to generate a global fraud dependency representation. The fusion classification and output module is used to calculate the similarity between the global fraud dependency representation and the normal category prototype and the fraud category prototype, obtain the category belonging weight of the global fraud dependency representation, and map it to the normal channel and the fraud channel respectively, obtain the global information of the mapped normal channel and the fraud channel respectively, fuse it into the normal channel aggregate representation and the fraud channel aggregate representation respectively with a preset fusion coefficient, and combine it with the current representation of the target node to generate the final node representation, input it into the classifier and output the probability that each node is a fraud node.

Citation Information

Patent Citations

  • Link mode aggregated graph fraud risk identification method and system

    CN119202806A

  • Real-time fraud detection method and system based on multi-modal fusion

    CN121561730A

  • Systems and methods for graph prototypical networks for few-shot learning on attributed networks

    US20230117980A1