Fraud detection method and apparatus based on correlation fraud awareness
By constructing a multi-relational social network graph structure and a graph neural network model, fraud behavior features in social networks are characterized and aggregated, solving the problem that existing methods fail to fully utilize the multi-relational correlation and achieving more efficient fraud behavior detection.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- WUHAN UNIV
- Filing Date
- 2023-03-09
- Publication Date
- 2026-06-19
Smart Images
Figure CN116662982B_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of behavior detection technology, and in particular to a method and apparatus for detecting fraudulent behavior based on association fraud perception. Background Technology
[0002] Fraud detection is a technique for identifying anomalous users engaging in fraudulent behavior within social networks. The anonymity of online networks makes combating fraud increasingly costly. Due to the advantages of graph representation learning, graph network-based fraud detection has made significant progress in recent years. Existing methods only study the differences in individual fraudulent behaviors across different relationship types, neglecting the correlations between fraudulent behaviors across multiple relationships.
[0003] Several research methods exist for fraud detection. Liu et al.'s team, in their paper "Alleviating the inconsistency problem of applying graph neural networks to fraud detection," and Dou et al.'s team, in their paper "Enhancing graph neural network-based fraud detectors against camouflaged fraudsters," proposed reconstructing the graph structure to address the graph inconsistency problem caused by fraudsters' masking behavior. Liu Yang et al.'s team, in their paper "Pick and choose: a GNN-based imbalanced learning approach for fraud detection," sampled neighbor information to eliminate the problem of fraudsters having fewer labels in the social network. Zhang et al.'s team, in their paper "Fraudre: fraud detection dual-resistant to graph inconsistency and imbalance," focused on solving the inconsistency and imbalance problems of graph structure in fraudulent social networks. Tang et al.'s team, in their paper "Rethinking graph neural networks for anomaly detection," explored the relationship between fraudulent behavior and spectral energy for the first time.
[0004] The above methods primarily utilize graph neural network (GNN) models as training models to achieve the task of classifying fraudster nodes. GNNs are deep learning models designed to solve the problem of extracting structural features from non-Euclidean space data (such as social networks). Thomas et al., in their paper "Semi-supervised classification with graph convolutional networks," proposed the Graph Convolutional Neural Network (GCN) model, extending traditional convolution operations to graph data. Petar et al., in their paper "Graph attention networks," introduced attention mechanisms into spatially based graph convolutional networks, proposing a more optimized GNN model. Hamiton et al., in their paper "Inductive Representation Learning on Large Graphs," proposed the GraphSage model, which uses a direct inductive approach to represent unknown graph nodes.
[0005] The above analysis shows that although various fraud detection methods have achieved good detection results, they still lack in-depth correlation analysis of fraudulent behaviors, and there is still much room for improvement in the effectiveness of fraud detection. Summary of the Invention
[0006] The main objective of this invention is to provide a method and apparatus for detecting fraudulent behavior based on association fraud perception, which aims to improve the effectiveness of fraud detection.
[0007] In a first aspect, the present invention provides a method for detecting fraudulent behavior based on association-based fraud perception, characterized in that the method for detecting fraudulent behavior based on association-based fraud perception includes:
[0008] Based on each user's original social data, a multi-relationship social network graph structure is constructed, consisting of multiple types of social behaviors among users. The multi-relationship social network graph structure includes a user set V, a user feature vector set X, a set of multiple relationship types R, and a set of user interaction behaviors under different relationship types E.
[0009] Based on the aforementioned multi-relationship social network graph structure, a graph neural network model is used to characterize the relation-specific node feature vectors and the feature vectors of each relation type.
[0010] Based on the relationship-specific node feature vectors, information is exchanged between the relationship-specific node feature vectors of different types of relationships for the same user, so as to obtain the node representation of fraud association perception.
[0011] Based on the feature vectors of each type of relationship and the node representation of fraud association perception, a graph attention model is used to aggregate features for each user to obtain the feature representation of each user.
[0012] Based on the feature representation of each user, the graph neural network learning model is optimized by calculating the cross-entropy loss.
[0013] Fraudulent behavior is identified by using an optimized graph neural network learning model based on the social data of the user to be identified.
[0014] Optionally, based on the multi-relationship social network graph structure, the step of using a graph neural network model to characterize the relation-specific node feature vectors in each type of relation and the feature vectors of each type of relation includes:
[0015] Based on the user's feature vector set X, obtain relation-specific user behavior feature vectors.
[0016] The relation feature vector h is obtained by hot encoding the relation types contained in R. r ;
[0017] For eigenvectors and relational feature vector h r Preprocessing:
[0018]
[0019]
[0020] z r =W r h r ,
[0021] Among them, W v and W r It is a linear transition weight matrix;
[0022] Based on the processed feature vector and relational feature vector h r The graph neural network model is used to represent relation-specific node feature vectors in each type of relation.
[0023]
[0024]
[0025] Where tanh(·) denotes the hyperbolic tangent function, W align Let d be a linear transition weight matrix. v,rN represents the degree of node v in the subgraph structure of relation type r, i.e., the number of neighbors of node v in the subgraph structure. V Let λ be the set of neighboring nodes of node v, and let λ be a preset parameter.
[0026] The feature vectors representing each type of relationship are represented using an MLP model:
[0027]
[0028] in, and These are the learnable parameters for a typical MLP model.
[0029] Optionally, the step of exchanging information between relation-specific node feature vectors for different types of relationships of the same user to obtain a node representation for fraud association perception based on the relation-specific node feature vectors includes:
[0030] Based on the relation-specific node feature vectors Calculate the first attention coefficient between the feature vectors of nodes of different relation types:
[0031]
[0032] Where q r For the trainable attention vector coefficients, LeakyReLU(·) is an activation function;
[0033] Based on the first attention coefficient, information is exchanged between relation-specific node feature vectors of different relation types for the same user, thereby obtaining node representations for fraud association perception:
[0034]
[0035] Optionally, the step of aggregating features for each user using a graph attention model based on the feature vectors of each type of relationship and the node representations for fraud association perception to obtain the feature representation of each user includes:
[0036] Based on the feature vector h' of each type of relationship r And the node representation h'v,r of the fraud association perception, calculate the second attention coefficient:
[0037]
[0038] Among them, T r and U r Both are linear transition weight matrices;
[0039] Based on the second attention coefficient, the node feature vectors under different relationship types are aggregated to obtain the feature representation of each user:
[0040]
[0041] Optionally, calculating the cross-entropy loss includes:
[0042] The cross-entropy loss is calculated using the following formula:
[0043]
[0044] in, Y represents the set of user nodes belonging to the training dataset. v,m This represents the m-th dimension label value for each user v. Let m represent the value of the m-th dimension in the feature representation of user v.
[0045] Secondly, the present invention also provides a fraud behavior detection device based on association fraud perception, the fraud behavior detection device based on association fraud perception comprising:
[0046] The construction module is used to build a multi-relationship social network graph structure composed of multiple types of social behaviors between users based on each user's original social data. The multi-relationship social network graph structure includes a user set V, a user feature vector set X, a set of multiple relationship types R, and a set of user interaction behaviors under different relationship types E.
[0047] The first representation module is used to represent the relation-specific node feature vectors and the feature vectors of each type of relation respectively, based on the multi-relationship social network graph structure and using a graph neural network model.
[0048] The second representation module is used to exchange information between the relationship-specific node feature vectors for different types of relationships of the same user, based on the relationship-specific node feature vectors, so as to obtain the node representation of fraud association perception.
[0049] The third representation module is used to aggregate features for each user based on the feature vectors of each type of relationship and the node representations of fraud association perception, using a graph attention model to obtain the feature representation of each user.
[0050] The optimization module is used to optimize the graph neural network learning model by calculating cross-entropy loss based on the feature expression of each user.
[0051] The identification module is used to identify fraudulent behavior based on the social data of the user to be identified by learning an optimized graph neural network model.
[0052] Optional, the first characterization module is used for:
[0053] Based on the user's feature vector set X, obtain relation-specific user behavior feature vectors.
[0054] The relation feature vector h is obtained by hot encoding the relation types contained in R. r ;
[0055] For eigenvectors and relational feature vector h r Preprocessing:
[0056]
[0057]
[0058] z r =W r h r ,
[0059] Among them, W v and W r It is a linear transition weight matrix;
[0060] Based on the processed feature vector and relational feature vector h r The graph neural network model is used to represent relation-specific node feature vectors in each type of relation.
[0061]
[0062]
[0063] Where tanh(·) denotes the hyperbolic tangent function, W align Let d be a linear transition weight matrix. v,r N represents the degree of node v in the subgraph structure of relation type r, i.e., the number of neighbors of node v in the subgraph structure. V Let λ be the set of neighboring nodes of node v, and let λ be a preset parameter.
[0064] The feature vectors representing each type of relationship are represented using an MLP model:
[0065]
[0066] in, and These are the learnable parameters for a typical MLP model.
[0067] Optional, a second characterization module, used for:
[0068] Based on the relation-specific node feature vectors Calculate the first attention coefficient between the feature vectors of nodes of different relation types:
[0069]
[0070] Where q r For the trainable attention vector coefficients, LeakyReLU(·) is an activation function;
[0071] Based on the first attention coefficient, information is exchanged between relation-specific node feature vectors of different relation types for the same user, thereby obtaining node representations for fraud association perception:
[0072]
[0073] Optional, a third characterization module, used for:
[0074] Based on the feature vector h' of each type of relationship r and the node representation h' of the fraud association perception v,r Calculate the second attention coefficient:
[0075]
[0076] Among them, T r and U r Both are linear transition weight matrices;
[0077] Based on the second attention coefficient, the node feature vectors under different relationship types are aggregated to obtain the feature representation of each user:
[0078]
[0079] Optional, optimization modules are used for:
[0080] The cross-entropy loss is calculated using the following formula:
[0081]
[0082] in Y represents the set of user nodes belonging to the training dataset. v，m This represents the m-th dimension label value for each user v. Let m represent the value of the m-th dimension in the feature representation of user v.
[0083] In this invention, a multi-relationship social network graph structure is constructed based on each user's original social data, consisting of multiple types of social behaviors among users. This multi-relationship social network graph structure includes a user set V, a user feature vector set X, a set of multiple relationship types R, and a set of user interaction behaviors under different relationship types E. Based on this multi-relationship social network graph structure, a graph neural network model is used to represent the relationship-specific node feature vectors in each type of relationship and the feature vectors of each type of relationship. Based on the relationship-specific node feature vectors, information is exchanged between the relationship-specific node feature vectors of different types of relationships for the same user to obtain node representations for fraud association perception. Based on the feature vectors of each type of relationship and the node representations for fraud association perception, a graph attention model is used to aggregate features for each user, obtaining the feature expression for each user. Based on the feature expression of each user, the graph neural network learning model is optimized by calculating cross-entropy loss. The optimized graph neural network learning model is then used to identify fraudulent behavior based on the social data of the user to be identified. This invention fully utilizes the patterns of fraudulent behavior associated with multiple relationship types, greatly improving the accuracy of fraud detection and effectively reducing errors. Attached Figure Description
[0084] Figure 1 This is a flowchart illustrating an embodiment of the fraud behavior detection method based on association fraud perception according to the present invention.
[0085] Figure 2 This is a schematic diagram of the functional modules of an embodiment of the fraud behavior detection device based on association fraud perception of the present invention.
[0086] The realization of the objective, functional features and advantages of the present invention will be further explained in conjunction with the embodiments and with reference to the accompanying drawings. Detailed Implementation
[0087] It should be understood that the specific embodiments described herein are merely illustrative of the invention and are not intended to limit the invention.
[0088] In a first aspect, embodiments of the present invention provide a fraud behavior detection method based on association fraud perception.
[0089] In one embodiment, reference is made to Figure 1 , Figure 1 This is a flowchart illustrating an embodiment of the fraud behavior detection method based on association fraud perception according to the present invention. Figure 1 As shown, fraud behavior detection methods based on association fraud perception include:
[0090] Step S10: Based on each user's original social data, construct a multi-relationship social network graph structure composed of multiple types of social behaviors between users. The multi-relationship social network graph structure includes a user set V, a user feature vector set X, a set of multiple relationship types R, and a set of user interaction behaviors under different relationship types E.
[0091] In this embodiment, based on the original social data, a multi-relationship social graph structure G = (V, X, E, R) is extracted in the following form, where V is the set of user nodes, and V is the set of user nodes. i ,v j ∈V represents two user nodes in the user node set; X represents the user's feature vector set; R represents a set of various types of relations, where r∈R is one type of relation; E represents the set of interaction behaviors between users under different relation types, e ij,r =(v i ,v j ,r)∈E represents an interaction behavior between two users under a certain relation type.
[0092] Step S20: Based on the multi-relationship social network graph structure, a graph neural network model is used to characterize the relation-specific node feature vectors and the feature vectors of each type of relation, respectively.
[0093] In this embodiment, we first construct node feature vectors with specific relationships. and the eigenvector h of the relation r .in It is obtained directly using the original node features X, i.e. Perform hot encoding operations on different types of relations to obtain feature vectors h for different types of relations. r For example, if there are three types of relations, then three feature vectors are obtained: [1,0,0], [0,1,0] and [0,0,1].
[0094] Further, in one embodiment, step S20 includes:
[0095] Based on the user's feature vector set X, obtain relation-specific user behavior feature vectors.
[0096] The relation feature vector h is obtained by hot encoding the relation types contained in R. r ;
[0097] For eigenvectors and relational feature vector h r Preprocessing:
[0098]
[0099]
[0100] z r =W r h r ,
[0101] Among them, W v and W r It is a linear transition weight matrix;
[0102] Based on the processed feature vector and relational feature vector h r The graph neural network model is used to represent relation-specific node feature vectors in each type of relation.
[0103]
[0104]
[0105] Where tanh(·) denotes the hyperbolic tangent function, W align Let d be a linear transition weight matrix. v,r N represents the degree of node v in the subgraph structure of relation type r, i.e., the number of neighbors of node v in the subgraph structure. V Let λ be the set of neighboring nodes of node v, and let λ be a preset parameter.
[0106] The feature vectors representing each type of relationship are represented using an MLP model:
[0107]
[0108] in, and These are the learnable parameters for a typical MLP model.
[0109] Step S30: Based on the relationship-specific node feature vectors, information is exchanged between the relationship-specific node feature vectors of different types of relationships for the same user, so as to obtain the node representation of fraud association perception.
[0110] In this embodiment, based on the relationship-specific node feature vectors, the attention coefficient between node feature vectors of different relationship types is first calculated. This attention coefficient is only calculated between node feature vectors of different relationship types of the same node. Based on the calculated attention coefficient, information is transferred between node features of different relationship types to obtain node representations for fraud association perception.
[0111] Further, in one embodiment, step S30 includes:
[0112] Based on the relation-specific node feature vectors Calculate the first attention coefficient between the feature vectors of nodes of different relation types:
[0113]
[0114] Where q r For the trainable attention vector coefficients, LeakyReLU(·) is an activation function;
[0115] Based on the first attention coefficient, information is exchanged between relation-specific node feature vectors of different relation types for the same user, thereby obtaining node representations for fraud association perception:
[0116]
[0117] Step S40: Based on the feature vector of each type of relationship and the node representation of fraud association perception, use the graph attention model to aggregate features for each user and obtain the feature expression of each user.
[0118] In this embodiment, firstly, based on the feature vector of each type of relationship and the node representation of fraud association perception, the attention coefficient is calculated. Based on the calculated attention coefficient, the node feature vectors under different relationship types are aggregated, and the feature expression of each user is obtained.
[0119] Further, in one embodiment, step S40 includes:
[0120] Based on the feature vector h' of each type of relationship r and the node representation h' of the fraud association perception v,r Calculate the second attention coefficient:
[0121]
[0122] Among them, T r and U r Both are linear transition weight matrices;
[0123] Based on the second attention coefficient, the node feature vectors under different relationship types are aggregated to obtain the feature representation of each user:
[0124]
[0125] Step S50: Based on the feature representation of each user, optimize the graph neural network learning model by calculating cross-entropy loss;
[0126] Step S60: Fraudulent behavior is identified based on the social data of the user to be identified using the optimized graph neural network learning model.
[0127] In this embodiment, based on the feature representation of each user, a semi-supervised learning method is used to calculate the loss using partially known social relationship labels, classifying users' social relationships and thus optimizing the graph neural network learning model. Subsequently, the optimized graph neural network learning model can be used to identify fraudulent behavior based on the social data of the user to be identified.
[0128] Furthermore, in one embodiment, calculating the cross-entropy loss includes:
[0129] The cross-entropy loss is calculated using the following formula:
[0130]
[0131] in, Y represents the set of user nodes belonging to the training dataset. v,m This represents the m-th dimension label value for each user v. Let m represent the value of the m-th dimension in the feature representation of user v.
[0132] Secondly, embodiments of the present invention also provide a fraud behavior detection device based on association fraud perception.
[0133] In one embodiment, reference is made to Figure 2 , Figure 2 This is a functional module diagram of an embodiment of the fraud behavior detection device based on association fraud perception according to the present invention. Figure 2 As shown, the fraud behavior detection device based on association fraud perception includes:
[0134] Module 10 is used to construct a multi-relationship social network graph structure composed of multiple types of social behaviors between users based on each user's original social data. The multi-relationship social network graph structure includes a user set V, a user feature vector set X, a set of multiple relationship types R, and a set of user interaction behaviors under different relationship types E.
[0135] The first representation module 20 is used to represent the relation-specific node feature vectors and the feature vectors of each type of relation respectively, based on the multi-relationship social network graph structure and using a graph neural network model.
[0136] The second representation module 30 is used to exchange information with each other for the relationship-specific node feature vectors of different types of relationships of the same user based on the relationship-specific node feature vectors, so as to obtain the node representation of fraud association perception.
[0137] The third representation module 40 is used to aggregate features for each user based on the feature vectors of each type of relationship and the node representation of fraud association perception, using a graph attention model to obtain the feature expression of each user.
[0138] Optimization module 50 is used to optimize the graph neural network learning model by calculating cross-entropy loss based on the feature expression of each user.
[0139] The identification module 60 is used to identify fraudulent behavior based on the social data of the user to be identified by learning an optimized graph neural network model.
[0140] Furthermore, in one embodiment, the first characterization module 20 is used for:
[0141] Based on the user's feature vector set X, obtain relation-specific user behavior feature vectors.
[0142] The relation feature vector h is obtained by hot encoding the relation types contained in R. r ;
[0143] For eigenvectors and relational feature vector h r Preprocessing:
[0144]
[0145]
[0146] z r =W r h r ,
[0147] Among them, W v and W r It is a linear transition weight matrix;
[0148] Based on the processed feature vector and relational feature vector h r The graph neural network model is used to represent relation-specific node feature vectors in each type of relation.
[0149]
[0150]
[0151] Where tanh(·) denotes the hyperbolic tangent function, W align Let d be a linear transition weight matrix. v，r N represents the degree of node v in the subgraph structure of relation type r, i.e., the number of neighbors of node v in the subgraph structure. V Let λ be the set of neighboring nodes of node v, and let λ be a preset parameter.
[0152] The feature vectors representing each type of relationship are represented using an MLP model:
[0153]
[0154] in, and These are the learnable parameters for a typical MLP model.
[0155] Furthermore, in one embodiment, the second characterization module 30 is used for:
[0156] Based on the relation-specific node feature vectors Calculate the first attention coefficient between the feature vectors of nodes of different relation types:
[0157]
[0158] Where q r For the trainable attention vector coefficients, LeakyReLU(·) is an activation function;
[0159] Based on the first attention coefficient, information is exchanged between relation-specific node feature vectors of different relation types for the same user, thereby obtaining node representations for fraud association perception:
[0160]
[0161] Furthermore, in one embodiment, the third characterization module 40 is used for:
[0162] Based on the feature vector h' of each type of relationship r And the node representation h'v,r of the fraud association perception, calculate the second attention coefficient:
[0163]
[0164] Among them, T r and U r Both are linear transition weight matrices;
[0165] Based on the second attention coefficient, the node feature vectors under different relationship types are aggregated to obtain the feature representation of each user:
[0166]
[0167] Furthermore, in one embodiment, the optimization module 50 is used for:
[0168] The cross-entropy loss is calculated using the following formula:
[0169]
[0170] in, Y represents the set of user nodes belonging to the training dataset. v,m This represents the m-th dimension label value for each user v. Let m represent the value of the m-th dimension in the feature representation of user v.
[0171] The functions of each module in the above-mentioned fraud behavior detection device based on association fraud perception correspond to the steps in the above-mentioned fraud behavior detection method based on association fraud perception. Their functions and implementation processes will not be described in detail here.
[0172] It should be noted that, in this document, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or system that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or system. Unless otherwise specified, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, method, article, or system that includes that element.
[0173] The sequence numbers of the above embodiments of the present invention are for descriptive purposes only and do not represent the superiority or inferiority of the embodiments.
[0174] Through the above description of the embodiments, those skilled in the art can clearly understand that the methods of the above embodiments can be implemented by means of software plus necessary general-purpose hardware platforms. Of course, they can also be implemented by hardware, but in many cases the former is a better implementation method. Based on this understanding, the technical solution of the present invention, in essence, or the part that contributes to the prior art, can be embodied in the form of a software product. This computer software product is stored in a storage medium (such as ROM / RAM, magnetic disk, optical disk) as described above, and includes several instructions to cause a terminal device to execute the methods described in the various embodiments of the present invention.
[0175] The above are merely preferred embodiments of the present invention and do not limit the scope of the patent. Any equivalent structural or procedural transformations made based on the description and drawings of the present invention, or direct or indirect applications in other related technical fields, are similarly included within the scope of patent protection of the present invention.
Claims
1. A fraud behavior detection method based on association fraud perception, characterized in that, The fraud behavior detection method based on association fraud perception includes: Based on each user's raw social data, a multi-relationship social network graph structure is constructed, consisting of various types of social behaviors among users. This multi-relationship social network graph structure includes user sets. User feature vector set Multiple relation types set and the set of user interaction behaviors under different relationship types ; Based on the aforementioned multi-relationship social network graph structure, a graph neural network model is used to characterize the relation-specific node feature vectors and the feature vectors of each relation type. Based on the relationship-specific node feature vectors, information is exchanged between the relationship-specific node feature vectors of different types of relationships for the same user, so as to obtain the node representation of fraud association perception. Based on the feature vectors of each type of relationship and the node representation of fraud association perception, a graph attention model is used to aggregate features for each user to obtain the feature representation of each user. Based on the feature representation of each user, the graph neural network learning model is optimized by calculating the cross-entropy loss. Fraudulent behavior is identified by using an optimized graph neural network learning model based on the social data of the user to be identified. The step of exchanging information between relation-specific node feature vectors for different types of relationships of the same user to obtain a node representation for fraud association perception includes: Based on the relation-specific node feature vectors Calculate the first attention coefficient between the feature vectors of nodes of different relation types: in These are the trainable attention vector coefficients. It is an activation function; Based on the first attention coefficient, information is exchanged between relation-specific node feature vectors of different relation types for the same user, thereby obtaining node representations for fraud association perception: ； The step of aggregating features for each user based on the feature vectors of each type of relationship and the node representation of fraud association perception, using a graph attention model, to obtain the feature representation of each user includes: Based on the feature vector of each type of relationship and the node representation of the fraud association perception. Calculate the second attention coefficient: in, and Both are linear transition weight matrices; Based on the second attention coefficient, the node feature vectors under different relationship types are aggregated to obtain the feature representation of each user: 。 2. The fraud behavior detection method based on association fraud perception as described in claim 1, characterized in that, Based on the multi-relationship social network graph structure, the steps of characterizing the relation-specific node feature vectors and the feature vectors of each relation class using a graph neural network model include: Based on the user's feature vector set Obtain relation-specific user behavior feature vectors ; right The relationship feature vector is obtained by hot encoding the included relationship types. ; For eigenvectors and relational feature vectors Preprocessing: ， in, and It is a linear transition weight matrix; Based on the processed feature vector and relational feature vectors The graph neural network model is used to represent relation-specific node feature vectors in each type of relation. : in Represents the hyperbolic tangent function. It is a linear transition weight matrix. For nodes In relation types The degree of a node in a subgraph structure The number of neighbors of in the subgraph structure. For nodes The set of neighboring nodes, These are preset parameters; The feature vectors representing each type of relationship are represented using an MLP model: in, and These are the learnable parameters of the MLP model.
3. The fraud behavior detection method based on association fraud perception as described in claim 1, characterized in that, Calculating the cross-entropy loss includes: The cross-entropy loss is calculated using the following formula: in, This represents the set of user nodes belonging to the training dataset. Represent each user The Dimensional label value, Indicates user In the feature expression of the first The value of the dimension.
4. A fraud behavior detection device based on association fraud perception, characterized in that, The fraud behavior detection device based on association fraud perception includes: The building module is used to construct a multi-relationship social network graph structure composed of various types of social behaviors among users, based on each user's original social data. The multi-relationship social network graph structure includes user sets. User feature vector set Multiple relation types set and the set of user interaction behaviors under different relationship types ; The first representation module is used to represent the relation-specific node feature vectors and the feature vectors of each type of relation respectively, based on the multi-relationship social network graph structure and using a graph neural network model. The second representation module is used to exchange information between the relationship-specific node feature vectors for different types of relationships of the same user, based on the relationship-specific node feature vectors, so as to obtain the node representation of fraud association perception. The third representation module is used to aggregate features for each user based on the feature vectors of each type of relationship and the node representations of fraud association perception, using a graph attention model to obtain the feature representation of each user. The optimization module is used to optimize the graph neural network learning model by calculating cross-entropy loss based on the feature expression of each user. The identification module is used to identify fraudulent behavior based on the social data of the user to be identified by learning a model using an optimized graph neural network. The second characterization module is used for: Based on the relation-specific node feature vectors Calculate the first attention coefficient between the feature vectors of nodes of different relation types: in These are the trainable attention vector coefficients. It is an activation function; Based on the first attention coefficient, information is exchanged between relation-specific node feature vectors of different relation types for the same user, thereby obtaining node representations for fraud association perception: ； The third representation module is used for: Based on the feature vector of each type of relationship and the node representation of the fraud association perception. Calculate the second attention coefficient: in, and Both are linear transition weight matrices; Based on the second attention coefficient, the node feature vectors under different relationship types are aggregated to obtain the feature representation of each user: 。 5. The fraud behavior detection device based on association fraud perception as described in claim 4, characterized in that, The first representation module is used for: Based on the user's feature vector set Obtain relation-specific user behavior feature vectors ; right The relationship feature vector is obtained by hot encoding the included relationship types. ; For eigenvectors and relational feature vectors Preprocessing: ， in, and It is a linear transition weight matrix; Based on the processed feature vector and relational feature vectors The graph neural network model is used to represent relation-specific node feature vectors in each type of relation. : in Represents the hyperbolic tangent function. It is a linear transition weight matrix. For nodes In relation types The degree of a node in a subgraph structure The number of neighbors of in the subgraph structure. For nodes The set of neighboring nodes, These are preset parameters; The feature vectors representing each type of relationship are represented using an MLP model: in, and These are the learnable parameters of the MLP model.
6. The fraud behavior detection device based on association fraud perception as described in claim 4, characterized in that, The optimization module is used for: The cross-entropy loss is calculated using the following formula: in, This represents the set of user nodes belonging to the training dataset. Represent each user The Dimensional label value, Indicates user In the feature expression of the first The value of the dimension.