A fraud detection method in the consumer finance field based on big data

By constructing financial transaction diagrams and user communication diagrams, using graph attention networks and sparse cross-attention mechanisms for feature extraction and integration, the problem that traditional fraud detection methods cannot adapt to the diversity of fraud behaviors is solved, and fraud detection in the consumer finance field is improved Accuracy and flexibility.

CN117408697BActive Publication Date: 2024-11-29BEIJING WIMI HUITONG INFORMATION TECH CO LTD
View PDF 1 Cites 0 Cited by

Patent Information

Application Number
CN202311355535.6
Authority / Receiving Office
CN · China
Patent Type
Patents(China)
Current Assignee / Owner
Filing Date
2023-10-19
Publication Date
2024-11-29
Estimated Expiration
2043-10-19

AI Technical Summary

Technical Problem

Traditional consumer finance fraud detection methods rely on prior knowledge and cannot effectively adapt to the diversity and complexity of fraudulent behavior, resulting in missed or false positives, making it difficult to keep up with the evolution of fraud methods, affecting financial losses and legitimate user experience.

Method used

By obtaining financial transaction data and user personal data, building financial transaction diagrams and user communication diagrams, using the graph attention network to extract features, and using sparse cross attention mechanisms to fusion, inputting the feedforward neural network for binary classification, predicting fraudulent transactions .

Benefits of technology

It realizes efficient identification of fraud gangs, improves the accuracy of fraud detection, deeply explores information in user transaction-related networks, and enhances the flexibility and accuracy of fraud detection.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN117408697B_ABST
    Figure CN117408697B_ABST
Patent Text Reader

Abstract

The present invention belongs to the technical field of big data processing, and specifically relates to a fraud detection method in the field of consumer finance based on big data, comprising: obtaining financial transaction data and user personal data and constructing them into a graph structure, using a graph attention network to extract financial transaction features and user features respectively, and outputting the result of financial transaction fraud detection after fusing them with a sparse cross attention mechanism; the present invention introduces the relationship between users during fraud detection, extracts features of the connections between transaction users in the form of a graph, and fuses them with the financial transaction features, so that the model has the ability to detect gang fraud, and at the same time uses sparse cross attention to fuse transaction features and user features, thereby achieving high-accuracy fraud detection in the financial field.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] The present invention belongs to the technical field of big data processing, and specifically relates to a fraud detection method in the field of consumer finance based on big data. Background Art

[0002] In today's digital age, with the popularization of the Internet and the rapid development of big data technology, the field of consumer finance faces huge opportunities and challenges. Consumer finance refers to financial services provided to individuals and families through financial technology, including loans, credit cards, borrowing, etc. However, with the continuous expansion of consumer finance business, fraud is also evolving and increasing, causing serious economic losses and trust crises to financial institutions and consumers.

[0003] Traditional fraud detection methods are usually based on rules and models. These methods often rely on prior knowledge and cannot adapt well to the diversity and complexity of fraudulent behaviors. And with the improvement of anti-fraud means and technologies, fraud is becoming increasingly difficult to be carried out by individuals, but relies on the organized conduct of fraud gangs. This traditional fraud detection method that relies on prior knowledge is powerless in the face of evolving fraud methods. Fraudsters constantly change their strategies and adopt new technical means and covert methods, making it difficult for traditional rules and models to keep up with the pace of fraud evolution. Traditional methods lack flexibility and cannot effectively adapt to the diversity and complexity of fraudulent behaviors. Therefore, they are prone to underreporting or false reporting, which not only causes financial losses but also affects the experience of legitimate users. Summary of the invention

[0004] In order to solve the above technical problems, the present invention proposes a fraud detection method in the field of consumer finance based on big data, comprising the following steps:

[0005] S1. Obtain financial transaction data and user personal data, and construct a financial transaction graph G F and transaction user communication graph G U ;

[0006] The financial transaction data includes: transaction time, transaction method, transaction amount, trader, account balance before the transaction, account balance after the transaction, recipient of the transaction, account balance of the recipient of the transaction before the transaction, and account balance of the recipient of the transaction after the transaction;

[0007] The user's personal data includes: gender, age, and friend list;

[0008] S2. For each financial transaction, from Figure G F and Figure G U Split out the financial transaction subgraph G where the transaction users are located f And the user communication subgraph G corresponding to the user u, using the graph attention network GAT to analyze the financial transaction subgraph G f and user interaction subgraph G u Perform feature extraction to obtain the financial transaction feature vector f and user vector u.

[0009] S3. Use the sparse cross attention mechanism to fuse the financial transaction feature vector f and the user vector u to obtain the fused feature vector H. Input the fused feature vector H into the feedforward neural network for binary classification to predict whether the transaction is a fraudulent transaction.

[0010] Beneficial effects of the present invention:

[0011] Starting from the connections between users, the present invention utilizes users' personal information and financial transaction information to construct user transaction networks and user association networks, uses graph attention networks to extract features from each node in the graph, and deeply explores the connections between users, thereby achieving efficient identification of fraud gangs. At the same time, financial transaction features are fused with user features using sparse cross attention, deeply exploring the information in user transaction association networks, and improving the accuracy of fraud detection in the field of consumer finance. BRIEF DESCRIPTION OF THE DRAWINGS

[0012] Figure 1 The present invention is a flow chart of a fraud detection method in the field of consumer finance based on big data. DETAILED DESCRIPTION

[0013] The following will be combined with the drawings in the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0014] A fraud detection method in the consumer finance field based on big data, such as Figure 1 As shown, including:

[0015] S1. Obtain financial transaction data and user personal data, and construct a financial transaction graph G F and transaction user social graph G U ;

[0016] The financial transaction data includes: transaction time, transaction method, transaction amount, trader, account balance before the transaction, account balance after the transaction, recipient of the transaction, account balance of the recipient of the transaction before the transaction, and account balance of the recipient of the transaction after the transaction;

[0017] The user's personal data includes: gender, age, and friend list;

[0018] S2. For each financial transaction, from Figure G F and Figure G U Split out the financial transaction subgraph G where the transaction users are located f And the user communication subgraph G corresponding to the user u , using the graph attention network GAT to analyze the financial transaction subgraph G f and user interaction subgraph G u Perform feature extraction to obtain the financial transaction feature vector f and user vector u.

[0019] S3. Use the sparse cross attention mechanism to fuse the financial transaction feature vector f and the user vector u to obtain the fused feature vector H. Input the fused feature vector H into the feedforward neural network for binary classification to predict whether the transaction is a fraudulent transaction.

[0020] Obtain financial transaction data and user personal data, and construct a financial transaction graph G for each transaction F and transaction user communication graph G U ,include:

[0021] Step 1: Financial transaction data at least includes transaction time, transaction method, transaction amount, trader, account balance before the transaction, account balance after the transaction, transaction recipient, account balance of the transaction recipient before the transaction, and account balance of the transaction recipient after the transaction; user personal data at least includes gender, age, and friend list;

[0022] Step 2: Construct the financial transaction graph G based on the traders and transaction recipients in the financial transaction data F (V F ,E F ), each node V in the graph F Each represents a user participating in the transaction, and other data in the financial transaction data are used as the dimensional feature vector of the node. F Representation graph G F The edge connecting two nodes in .

[0023] Graph Node V F ,include:

[0024]

[0025] Where m is the number of feature columns of financial data, Represents the i-th financial data feature value of the user.

[0026] Step 3: Construct the transaction user social graph G based on the friend list in the user's personal data U (V U ,E U), each node V in the graph U Each represents a user, E U Representation graph G U The edge connecting the two nodes in the user's personal data is used as the node V U The dimension of the feature vector.

[0027] Graph Node V U ,include:

[0028]

[0029] Where n is the number of feature columns of the user’s personal data, Represents the i-th user data feature value of the user.

[0030] For each financial transaction, from the financial transaction graph G F Split out the financial transaction subgraph G where the transaction users are located f And the user communication subgraph G corresponding to the user u ,include:

[0031] For each financial transaction involving a financial user, find the user in Figure G F and Figure G U The corresponding node in the graph is extracted, and all nodes with path r around the node are extracted to form a subgraph G. f and G u , extract the subgraph where the user node is located, so that the model does not have to operate on the entire graph, significantly reducing the computing cost;

[0032] The determination of the path r includes:

[0033]

[0034] Where N is the number of nodes in the graph, V k Represents the kth node in the graph, Degree(V k ) represents the degree of the kth node.

[0035] Use the graph attention network GAT to analyze the financial transaction subgraph G f and user interaction subgraph G u Perform feature extraction to obtain the financial transaction feature vector f and user vector u including:

[0036] Step 1: For graph G f Each node V F , calculate the attention coefficient between all its neighboring nodes, and update the node V by weighted summing up the feature vectors of the neighboring nodes F The feature vector of itself, finally getting the node V FThe eigenvector F is calculated as follows:

[0037]

[0038] Where F represents the node V F The feature vector of , σ() represents the Sigmoid activation function, h i 、h j 、h k Respectively represent graph G f The dimension feature vector of any three different nodes i, j, k in is |V F |, W1 and W2 represent the dimensions 1×|V F |The first and second feature transformation matrices, Representation graph G f The set of adjacent nodes of the i-th node in;

[0039] Step 2: Convert image G f The feature vectors of each node in are concatenated to obtain the transaction feature vector f, with a dimension of a×|V F |, a represents graph G f The number of nodes in the

[0040] Step 3: Use the same method to obtain the user vector u, with a dimension of b×|V F |, where b represents the graph G u The number of nodes in ;

[0041] The sparse cross attention mechanism is used to fuse the financial transaction feature vector f and the user vector u, including:

[0042] The feature fusion formula of the financial transaction feature vector f and the user vector u is as follows:

[0043]

[0044]

[0045] Among them, W Q , W K , W V There are three feature conversion matrices, forming the Query, Key, and Value required for attention calculation. Q The dimension is |V F |×128,W K , W V The dimension is |V U |×128, Sample indicates sampling operation, represents the sampled Query matrix. In order to reduce the complexity of calculation, the Query matrix is ​​sampled to remove the redundant information. H represents the fused feature vector.

[0046] In order to reduce the complexity of calculation, the Query matrix is ​​sampled to remove redundant information, including:

[0047] Step 1: According to fW Q ×(uW K ) T The matrix A is calculated, with a dimension of a×b, and each element in the matrix is ​​A ij , calculate the importance score for each row in the matrix:

[0048]

[0049] Among them, Score i Represents the importance score of the i-th row in matrix A;

[0050] Step 2: Sort the scores from large to small, and select the a / 2 rows with the highest scores in matrix A to form the sampled matrix The dimension is

[0051] The fused feature vector H is input into the feedforward neural network for binary classification to predict whether the transaction is a fraudulent transaction. The formula is as follows:

[0052] Output = argmax(W out H)

[0053] Among them, Output represents the binary classification prediction result output by the feedforward neural network. Output = 0 means that this transaction is not a fraudulent transaction, and Output = 1 means that this transaction is a fraudulent transaction. out Represents the fusion feature conversion matrix, with a dimension of 128×2, W out H maps the eigenvector H to a one-dimensional vector containing two eigenvalues. argmax() represents the argmax function and outputs the subscript of the maximum eigenvalue in the one-dimensional vector.

[0054] Although embodiments of the present invention have been shown and described, it will be appreciated by those skilled in the art that various changes, modifications, substitutions and variations may be made to the embodiments without departing from the principles and spirit of the present invention, and that the scope of the present invention is defined by the appended claims and their equivalents.

Claims

1. A fraud detection method in the field of consumer finance based on big data, It is characterized in that include: S1. Obtain financial transaction data and user personal data, and construct a financial transaction graph G F and transaction user communication graph G U ; The financial transaction data includes: transaction time, transaction method, transaction amount, trader, account balance before the transaction, account balance after the transaction, recipient of the transaction, account balance of the recipient of the transaction before the transaction, and account balance of the recipient of the transaction after the transaction; The user's personal data includes: gender, age, and friend list; S2. For each financial transaction, from Figure G F and Figure G U Split out the financial transaction subgraph G where the transaction users are located f And the user communication subgraph G corresponding to the user u , using the graph attention network GAT to analyze the financial transaction subgraph G f and user interaction subgraph G u Perform feature extraction to obtain the financial transaction feature vector f and user vector u; S21: For Figure G f Each node V F , calculate the attention coefficient between all its neighboring nodes, and update the node V by weighted summing up the feature vectors of the neighboring nodes F The feature vector of itself, finally getting the node V F The eigenvector F of For Figure G f Each node V F , calculate the attention coefficient between all its neighboring nodes, and update the node V by weighted summing up the feature vectors of the neighboring nodes F The feature vector of the node VF is finally obtained, including: Where F represents the node V F The feature vector of , σ( ) represents the Sigmoid activation function, h i 、h j 、h k Respectively represent graph G f The dimension feature vector of any three different nodes i, j, k in is |V F |, W 1 , W 2 Respectively represent the dimensions of 1×|V F |The first and second feature transformation matrices, Representation graph G f The set of adjacent nodes of the i-th node in; S22: Map G f The feature vectors of each node in are concatenated to obtain the transaction feature vector f, with a dimension of a×|V F |, where a represents the graph G f The number of nodes in the S23: Repeat steps S21-S22 to extract the user vector u, with a dimension of b×|V F |, where b represents the graph G u The number of nodes in ; S3, using the sparse cross attention mechanism to fuse the financial transaction feature vector f and the user vector u to obtain the fused feature vector H, and input the fused feature vector H into the feedforward neural network for binary classification to predict whether the transaction is a fraudulent transaction; The sparse cross attention mechanism is used to fuse the financial transaction feature vector f and the user vector u to obtain the fused feature vector H, which includes: H=Softmax(Sample(fW Q ×(uW K ) T )×(uW K ) T )your V Among them, H represents the fused feature vector, W Q , W K , W V They represent the dimensions respectively |V F |×128,|V U |×128,|V U |×128 query vector matrix, key vector matrix, value vector matrix, Sample( ) represents the query vector matrix of fW Q ×(uW K ) T The obtained matrix is ​​sampled, f represents the financial transaction feature vector, u represents the user vector, and T represents the matrix transpose.

2. According to the big data-based fraud detection method in the consumer finance field according to claim 1, It is characterized in that Constructing the financial transaction graph G F and transaction user communication graph G U ,include: S11: The traders and transaction recipients in the financial transaction data are used as graph nodes V F , taking other data in the financial transaction data as the dimension feature vector of the node, the graph node V F The transaction behaviors between them are the edges E of the graph F , we get the financial transaction graph G F (V F , E F ); S12: Construct the transaction user social graph G based on the friend list in the user's personal data U (V U , E U ), each node V in the graph U Each represents a user, E U Representation graph G U The edge connecting the two nodes in the user's personal data is used as the node V U The dimension of the feature vector.

3. According to claim 2, a fraud detection method in the field of consumer finance based on big data, It is characterized in that The graph node V F ,include: Among them, m is the number of characteristic columns of financial data, Represents the i-th financial data feature value of the user.

4. According to claim 2, a fraud detection method in the field of consumer finance based on big data, It is characterized in that The graph node V U ,include: Where n is the number of feature columns of the user’s personal data, Represents the i-th user data feature value of the user.

5. According to claim 1, a fraud detection method in the field of consumer finance based on big data, It is characterized in that From Figure G F and Figure G U Split out the financial transaction subgraph G where the transaction users are located f And the user communication subgraph G corresponding to the user u ,include: For each financial transaction involving a financial user, find the user in Figure G. F and Figure G U The corresponding node in the graph is extracted, and all nodes with path r around the node are extracted to form a subgraph G. f and G u .

6. A fraud detection method in the field of consumer finance based on big data according to claim 5, It is characterized in that The determination of path r includes: Where N is the number of nodes in the graph, V k Represents the kth node in the graph, Degree(V k ) represents the degree of the kth node.

7. A fraud detection method in the field of consumer finance based on big data according to claim 1, It is characterized in that The fused feature vector H is input into the feedforward neural network for binary classification to predict whether the transaction is a fraudulent transaction, including: Output=argmax(W out H) Among them, Output represents the binary classification prediction result output by the feedforward neural network. Output = 0 means that this transaction is not a fraudulent transaction, and Output = 1 means that this transaction is a fraudulent transaction. out Represents the fusion feature transformation matrix with a dimension of 128×2, W out H represents mapping the feature vector H to a one-dimensional vector, and argmax() represents the argmax function.

Citation Information

Patent Citations

  • Financial transaction anomaly detection method and cross-regional sustainable training method thereof

    CN115660688A