A credit card fraud detection method based on adaptive meta-aggregation strategy

Abstract

The present application relates to a credit card fraud detection method based on an adaptive meta-aggregation strategy, including: assigning a working node to each subset divided in the credit card transaction data set in the meta-model, and initializing the weights and local model parameters of the working nodes; each working node in the training data independently performs model training based on its corresponding subset, and applies the momentum method to accumulate historical gradient information, introduces the L2 regularization term and uses the adaptive learning rate algorithm to optimize the local model parameters; the central node preliminarily updates the global model parameters of the meta-model based on each local model parameter; the weighted median aggregation method is used to aggregate the updated local model parameters of all working nodes, and the global model parameters are further updated based on the weighted median aggregation parameters; the test data is input into the updated meta-model to obtain the credit card fraud prediction result. This method significantly improves the adaptability and detection accuracy of the model when facing new fraud patterns.

Description

Technical Field

[0001] The present application relates to the technical field of credit card fraud detection, and in particular to a credit card fraud detection method based on an adaptive meta-aggregation strategy. Background Art

[0002] In recent years, the development of deep learning technology has brought new possibilities for credit card fraud detection. Deep learning models can automatically extract complex features from large amounts of data, improving the accuracy of detection. However, the training of deep learning models usually requires a large amount of labeled data and computing resources, and the model update and deployment process is relatively cumbersome.

[0003] Meta-learning, as an emerging learning paradigm, aims to enable models to adapt to new tasks more quickly. In credit card fraud detection, meta-learning can help models quickly learn and adapt with a small amount of transaction data when faced with new fraud patterns. However, existing meta-learning methods still face many challenges in practical applications, such as the generalization ability of the model, training efficiency, and robustness to outliers. In credit card fraud detection, distributed learning systems can accelerate model training and updating. However, in a distributed environment, each working node may encounter different data distributions or even suffer malicious attacks, resulting in deviations in the model update process; and how to ensure the robustness of model updates in a distributed learning environment is also an urgent problem to be solved. Summary of the invention

[0004] Based on this, it is necessary to provide a credit card fraud detection method based on an adaptive meta-aggregation strategy, which includes:

[0005] S1: construct a meta-model, in which a working node is assigned to each subset divided in the credit card transaction data set, and the weight and local model parameters of the working node are initialized; each of the subsets includes a plurality of historical data, and the historical data includes historical transaction records and their corresponding labels;

[0006] S2: Sampling a batch of training data from the credit card transaction data set;

[0007] S3: Initialize distributed learning;

[0008] S4: Each working node in the training data independently performs model training based on its corresponding subset, and applies the momentum method to accumulate historical gradient information, introduces the L2 regularization term, and uses the adaptive learning rate algorithm to optimize the local model parameters;

[0009] S5: Each working node sends the updated local model parameters to the central node, and the central node preliminarily updates the global model parameters of the meta-model based on each of the local model parameters;

[0010] S6: Aggregate the updated local model parameters of all working nodes using a weighted median aggregation method to obtain a weighted median aggregation parameter; and further update the global model parameter based on the weighted median aggregation parameter;

[0011] S7: Sampling a batch of test data from the credit card transaction data set, inputting the test data into the updated meta-model to obtain a credit card fraud prediction result.

[0012] Beneficial effects: This method significantly improves the model's adaptability and detection accuracy when facing new fraud patterns, effectively reduces financial risks, and enhances transaction security; this method can not only improve the model's learning and adaptation speed on new tasks, but also provide a robust model update mechanism in a distributed learning environment. BRIEF DESCRIPTION OF THE DRAWINGS

[0013] In order to more clearly illustrate the embodiments of the present application or the technical solutions in the prior art, the drawings required for use in the embodiments or the description of the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present application. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying creative work.

[0014] Figure 1 This is a flowchart of a credit card fraud detection method based on an adaptive meta-aggregation strategy according to an embodiment of the present application. DETAILED DESCRIPTION

[0015] In order to make the above-mentioned purposes, features and advantages of the present application more obvious and easy to understand, the specific implementation methods of the present application are described in detail below in conjunction with the accompanying drawings. In the following description, many specific details are set forth to facilitate a full understanding of the present application. However, the present application can be implemented in many other ways different from those described herein, and those skilled in the art can make similar improvements without violating the connotation of the present application, so the present application is not limited by the specific embodiments disclosed below.

[0016] In addition, the terms "first" and "second" are used for descriptive purposes only and should not be understood as indicating or implying relative importance or implicitly indicating the number of the indicated technical features. Therefore, the features defined as "first" and "second" may explicitly or implicitly include at least one of the features. In the description of this application, the meaning of "plurality" is at least two, such as two, three, etc., unless otherwise clearly and specifically defined.

[0017] like Figure 1 As shown, this embodiment provides a credit card fraud detection method based on an adaptive meta-aggregation strategy, the method comprising:

[0018] S1: Construct a meta-model, in which a working node is assigned to each subset divided in the credit card transaction data set, and the weight and local model parameters of the working node are initialized; each of the subsets includes multiple historical data, and the historical data includes historical transaction records and their corresponding labels.

[0019] Specifically, the meta-model performs parameter initialization according to the historical data, and the initialization formula is:

[0020] ;

[0021] in, Represents the initialization metamodel parameters; Represents the network parameters of the metamodel; Represents historical data; Representation metamodel; x Represents historical transaction records; y Indicates a label, where the label includes normal or fraudulent; represents the loss function applicable to the binary classification problem; represents the regularization term;

[0022] Each working node initializes the local model parameters according to the initialized meta-model parameters, and the update formula is:

[0023] ;

[0024] in, Indicates i The local model parameters of each worker node.

[0025] Furthermore, the weight of the initialization working node includes:

[0026] A weight is assigned to each of the working nodes, and the weight is determined based on the size of the corresponding subset and the diversity of the feature distribution; the calculation formula is:

[0027] ;

[0028] in, Indicates i The weight of the worker nodes; Indicates i Worker nodes The size of the corresponding subset; represents the balance coefficient; A diversity metric function that represents the distribution of features; Indicates j Worker nodes The size of the corresponding subset; m Indicates the number of working nodes.

[0029] S2: Sample a batch of training data from the credit card transaction dataset.

[0030] In the meta-model of this embodiment, the credit card transaction data set is divided into a training set and a test set according to a set ratio 𝑝, where the training set is denoted as , the test set is denoted as , the training set and the test set satisfy the relationship, which is:

[0031] ;

[0032] ;

[0033] ;

[0034] in, represents the credit card transaction dataset; Indicates i The feature vector of each transaction; Indicates i The real label of each transaction; represents the number of transactions in the training set; represents the number of transactions in the test set;

[0035] Based on the feature vectors of transactions, both the transactions in the training set and the transactions in the test set are standardized; the standardization formula is:

[0036] ;

[0037] in, After standardization, i transactions; N represents the number of transactions in the credit card transaction dataset; The mean of the feature vectors representing transactions;

[0038] Use the standardized transactions in the training set and the transactions in the test set to construct training data and test data respectively;

[0039] ;

[0040] ;

[0041] in, represents training data; After standardization, i transactions; n Indicates the number of transactions after sampling; Represents test data;

[0042] The metamodel is represented as:

[0043] ;

[0044] in, Represents the metamodel.

[0045] S3: Initialize distributed learning, including:

[0046] In each working node, the global model parameters are initialized according to the local model parameters, and the update formula is:

[0047] ;

[0048] in, represents the global model parameters; Indicates i The local model parameters of each worker node.

[0049] S4: Each working node in the training data independently performs model training based on its corresponding subset, applies the momentum method to accumulate historical gradient information, introduces the L2 regularization term, and uses the adaptive learning rate algorithm to optimize the local model parameters.

[0050] Specifically, the step includes:

[0051] S4.1: Each working node updates its local model parameters based on its corresponding subset using the gradient descent method;

[0052] S4.2: Calculate the gradient of the loss function on the working node relative to the local model parameters. The calculation formula is:

[0053] ;

[0054] in, Indicates i The gradient of the loss function on each worker node with respect to the local model parameters; Indicates i Local model parameters of each worker node; x Represents historical transaction records; y Indicates a label, where the label includes normal or fraudulent; represents the loss function applicable to the binary classification problem; Indicates i The subset corresponding to the working nodes; Expressing the i Local model parameters of worker nodes The gradient of

[0055] S4.3: Apply the momentum method in the working node to accumulate historical gradient information, reduce the updated noise based on the historical gradient information, and optimize the local model parameters for the first time. The calculation formula is:

[0056] ;

[0057] ;

[0058] in, Indicates i The working node is t + Historical gradient information accumulated at 1 iteration; represents the momentum coefficient; Indicates i The working node is t Historical gradient information accumulated at the iteration; Indicates i The working node is t + Local model parameters at 1 iteration; Indicates i The working node is t Local model parameters at iteration ; represents the second learning rate;

[0059] S4.4: Introduce L2 regularization term for local model parameters. The calculation formula of L2 regularization term is:

[0060] ;

[0061] in, Indicates i The L2 regularization term introduced in the local model parameters of each worker node; represents the regularization coefficient; d Represents the dimension of local model parameters; Indicates i The local model parameters of the worker node g elements;

[0062] S4.5: The working node further optimizes the local model parameters using the adaptive learning rate algorithm to obtain updated local model parameters; the update formula is:

[0063] ;

[0064] ;

[0065] ;

[0066] in, Indicatesi The working node is t +1 updated local model parameters; Indicates i The working node is t The local model parameters after the iteration update; Indicates i The working node is t +First momentum estimate at iteration 1; Indicates i The working node is t The first momentum estimate at iteration ; represents the first decay rate; Indicates i The working node is t +2nd momentum estimate at iteration 1; Indicates i The working node is t The second momentum estimate at iteration ; represents the second decay rate; Represents a constant.

[0067] S5: Each working node sends the updated local model parameters to the central node, and the central node preliminarily updates the global model parameters of the meta-model based on each of the local model parameters.

[0068] Specifically, the central node preliminarily updates the global model parameters by weighted average based on the local model parameters, and the update formula is:

[0069] ;

[0070] in, Indicates t +Global model parameters at 1 iteration; Indicates t Global model parameters at iteration ; m Indicates the number of working nodes; Indicates i The weight of the worker nodes; Indicates i The working node is t +1 updated local model parameters; Indicates i The working node is t The local model parameters after the iteration update; Represents the learning rate.

[0071] This process ensures that the system can effectively update the global model to improve the accuracy and robustness of credit card fraud detection even in the face of uneven data distribution or differences in node computing power.

[0072] S6: Adopting a weighted median aggregation method to aggregate the updated local model parameters of all working nodes to obtain weighted median aggregation parameters; and further updating the global model parameters based on the weighted median aggregation parameters.

[0073] The goal of aggregation is to effectively merge the local model parameters of multiple worker nodes when facing Byzantine nodes (worker nodes that provide erroneous updates). In this embodiment, a weighted median aggregation method is used to improve the robustness to outliers. Specifically, this step includes:

[0074] S6.1: Assign a second weight to the local model parameters of each working node, where the second weight is calculated based on the local loss and uncertainty of the working node, and the calculation formula is:

[0075] ;

[0076] in, Indicates k In the iteration i A second weight for assigning local model parameters to each working node; Indicates k Normalization factor in iterations; Indicates i The working node is k The mean of all losses in iterations; represents the uncertainty coefficient; represents a constant;

[0077] S6.2: Sort the local model parameters of all working nodes based on the second weight, and use the weighted median aggregation method to calculate the weighted median aggregation parameter, and the calculation formula is:

[0078] ;

[0079] in, Indicates k The weighted median aggregation parameter in the iterations; Represents the network parameters of the metamodel; m Indicates the number of working nodes; Indicates k In the iteration i Local model parameters of each worker node;

[0080] S6.3: Using the weighted median aggregation parameter to further update the global model parameters of the meta-model.

[0081] In this embodiment, the loss function applicable to the binary classification problem includes a logarithmic loss function, and the expression of the logarithmic loss function is:

[0082] ;

[0083] in, represents the logarithmic loss function; Represents the network parameters of the metamodel; x Represents historical transaction records; y Indicates a label; N Indicates the number of transactions; Indicates i Transactions The true label of is 1, which indicates fraud and 0, which indicates normal. Indicated in the parameter Make a trade probability.

[0084] S7: Sampling a batch of test data from the credit card transaction data set, inputting the test data into the updated meta-model to obtain a credit card fraud prediction result.

[0085] In this embodiment, the test data is also used to evaluate the model performance, and the calculation formula is:

[0086] ;

[0087] ;

[0088] ;

[0089] in, represents the performance evaluation function; represents the global model parameters for further update; Represents test data; n Indicates the number of transactions after sampling; Indicates the test data i transactions; Indicates transaction The corresponding true label; represents the first indicator function; express sign function; represents the prediction function; represents the first weight matrix; represents the second weight matrix; represents the first bias term; Represents the second bias term; ReLU(⋅) represents the linear rectification function; Represents the sigmoid activation function;

[0090] In the final calculation, the sign function converts the predicted probability into a class label, which is defined as:

[0091] .

[0092] Furthermore, the weight of the loss function is adjusted according to the evaluation results, and the update formula is:

[0093] ;

[0094] in, Represents the loss function after adjusting the weights; represents the logarithmic loss function; Represents the network parameters of the metamodel; x Represents historical transaction records; y Indicates a label; represents the third weight; represents the second indicator function, which is related to the quality of the model evaluation result; Represents the model judgment definition; Indicates the preset threshold value;

[0095] ;

[0096] when When , the performance of the model is judged to be "excellent", then the second indicator function The output of is 0, indicating that no additional adjustment is made to the loss function;

[0097] when When , the performance of the model is judged to be "poor", then the second indicator function The output of is 1, indicating that the loss function needs to be adjusted.

[0098] The credit card fraud detection method based on the adaptive meta-aggregation strategy provided in this embodiment is based on an innovative meta-learning method, and is specifically designed to improve the accuracy and robustness of credit card fraud detection. The method uses a meta-aggregator algorithm to optimize the model update process in a distributed learning environment. The meta-aggregator algorithm effectively integrates updates from different working nodes through a carefully designed weighting mechanism and pruning strategy, while excluding potential malicious updates and outliers, thereby significantly improving the model's adaptability and detection accuracy when facing new fraud patterns.

[0099] The meta-learning framework in this embodiment allows the model to quickly learn new tasks with only a small amount of annotated data, which is critical in the rapidly changing field of fraud detection. In addition, this embodiment also provides an anomaly detection mechanism that can monitor and identify abnormal patterns in data in real time, dynamically adjust the weights of working nodes, and ensure the robustness of model updates. It also provides a model performance evaluation method that enables users to quantitatively analyze the performance of the model on the validation data set and optimize the model accordingly.

[0100] This system architecture of the present embodiment not only ensures high scalability and adapts to application requirements of different scales, but also provides the function of user-defined parameters, allowing users to adjust key parameters such as learning rate and regularization coefficient according to specific circumstances to achieve optimal detection performance. Through these technical solutions, an innovative solution is provided for credit card fraud detection, which effectively reduces financial risks and enhances transaction security.

[0101] This innovation of combining the concept of meta-learning with the meta-aggregator algorithm in this embodiment not only improves the model's ability to learn quickly, but also enhances its robustness in distributed systems, bringing breakthrough progress to the field of credit card fraud detection.

[0102] The technical features of the above-described embodiments may be arbitrarily combined. To make the description concise, not all possible combinations of the technical features in the above-described embodiments are described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this specification.

[0103] The above-described embodiments only express several implementation methods of the present application, and the descriptions thereof are relatively specific and detailed, but they cannot be construed as limiting the scope of the patent application. It should be pointed out that, for a person of ordinary skill in the art, several variations and improvements can be made without departing from the concept of the present application, and these all belong to the protection scope of the present application. Therefore, the protection scope of the patent application shall be subject to the attached claims.

Claims

1. A credit card fraud detection method based on adaptive meta-aggregation strategy, characterized in that: include: S1: construct a meta-model, in which a working node is assigned to each subset divided in the credit card transaction data set, and the weight and local model parameters of the working node are initialized; each of the subsets includes a plurality of historical data, and the historical data includes historical transaction records and their corresponding labels; S2: Sampling a batch of training data from the credit card transaction data set; S3: Initialize distributed learning, including: In each working node, the global model parameters are initialized according to the local model parameters, and the update formula is: ； in, represents the global model parameters; Indicates i Local model parameters of each worker node; S4: Each working node in the training data independently performs model training based on its corresponding subset, and applies the momentum method to accumulate historical gradient information, introduces the L2 regularization term, and uses the adaptive learning rate algorithm to optimize the local model parameters; S5: Each working node sends the updated local model parameters to the central node, and the central node preliminarily updates the global model parameters of the meta-model based on each of the local model parameters, including: The central node preliminarily updates the global model parameters by weighted average based on the local model parameters. The update formula is: ； in, Indicates t +Global model parameters at 1 iteration; Indicates t Global model parameters at iteration ; m Indicates the number of working nodes; Indicates i The weight of the worker nodes; Indicates i The working node is t +1 updated local model parameters; Indicates i The working node is t The local model parameters after the iteration update; represents the learning rate; S6: Aggregate the updated local model parameters of all working nodes using a weighted median aggregation method to obtain a weighted median aggregation parameter; and further update the global model parameter based on the weighted median aggregation parameter; S7: Sampling a batch of test data from the credit card transaction data set, inputting the test data into the updated meta-model to obtain a credit card fraud prediction result.

2. The credit card fraud detection method based on adaptive meta-aggregation strategy according to claim 1, characterized in that: S1 also includes: the meta-model performs parameter initialization according to the historical data, and the initialization formula is: ； in, Represents the initialization metamodel parameters; Represents the network parameters of the metamodel; Represents historical data; Representation metamodel; x Represents historical transaction records; y Indicates a label, where the label includes normal or fraudulent; represents the loss function applicable to the binary classification problem; represents the regularization term; Each working node initializes the local model parameters according to the initialized meta-model parameters, and the update formula is: ； in, Indicates i The local model parameters of each worker node.

3. The credit card fraud detection method based on adaptive meta-aggregation strategy according to claim 1, characterized in that: In S1, the weights of the initialized working nodes include: A weight is assigned to each of the working nodes, and the weight is determined based on the size of the corresponding subset and the diversity of the feature distribution; the calculation formula is: ； in, Indicates i The weight of the worker nodes; Indicates i Worker nodes The size of the corresponding subset; represents the balance coefficient; A diversity metric function that represents the distribution of features; Indicates j Worker nodes The size of the corresponding subset; m Indicates the number of working nodes.

4. The credit card fraud detection method based on the adaptive meta-aggregation strategy according to claim 1, characterized in that S4 include: S4.1: Each working node updates its local model parameters based on its corresponding subset using the gradient descent method; S4.2: Calculate the gradient of the loss function on the working node relative to the local model parameters. The calculation formula is: ； in, Indicates i The gradient of the loss function on each worker node with respect to the local model parameters; Indicates i Local model parameters of each worker node; x Represents historical transaction records; y Indicates a label, where the label includes normal or fraudulent; represents the loss function applicable to the binary classification problem; Indicates i The subset corresponding to the working nodes; Indicates i Local model parameters of worker nodes The gradient of S4.3: Apply the momentum method in the working node to accumulate historical gradient information, reduce the updated noise based on the historical gradient information, and optimize the local model parameters for the first time. The calculation formula is: ； ； in, Indicates i The working node is t + Historical gradient information accumulated at 1 iteration; represents the momentum coefficient; Indicates i The working node is t Historical gradient information accumulated at the iteration; Indicates i The working node is t + Local model parameters at 1 iteration; Indicates i The working node is t Local model parameters at iteration ; represents the learning rate; S4.4: Introduce L2 regularization term for local model parameters. The calculation formula of L2 regularization term is: ； in, Indicates i The L2 regularization term introduced in the local model parameters of each worker node; represents the regularization coefficient; d Represents the dimension of local model parameters; Indicates i The local model parameters of the worker node g elements; S4.5: The working node further optimizes the local model parameters using the adaptive learning rate algorithm to obtain updated local model parameters; the update formula is: ； ； ； in, Indicates i The working node is t +1 updated local model parameters; Indicates i The working node is t The local model parameters after the iteration update; Indicates i The working node is t +First momentum estimate at iteration 1; Indicates i The working node is t The first momentum estimate at iteration ; represents the first decay rate; Indicates i The working node is t +2nd momentum estimate at iteration 1; Indicates i The working node is t The second momentum estimate at iteration ; represents the second decay rate; Represents a constant.

5. The credit card fraud detection method based on the adaptive meta-aggregation strategy according to claim 4, characterized in that S6 include: S6.1: Assign a second weight to the local model parameters of each working node, where the second weight is calculated based on the local loss and uncertainty of the working node, and the calculation formula is: ； in, Indicates k In the iteration i A second weight for assigning local model parameters to each working node; Indicates k Normalization factor in iterations; Indicates i The working node is k The mean of all losses in iterations; represents the uncertainty coefficient; represents a constant; S6.2: Sort the local model parameters of all working nodes based on the second weight, and use the weighted median aggregation method to calculate the weighted median aggregation parameter, and the calculation formula is: ； in, Indicates k The weighted median aggregation parameter in the iterations; Represents the network parameters of the metamodel; m Indicates the number of working nodes; Indicates k In the iteration i Local model parameters of each worker node; S6.3: Using the weighted median aggregation parameter to further update the global model parameters of the meta-model.

6. The credit card fraud detection method based on adaptive meta-aggregation strategy according to claim 5, characterized in that: The loss function suitable for binary classification problems includes the logarithmic loss function, and the expression of the logarithmic loss function is: ； in, represents the logarithmic loss function; Represents the network parameters of the metamodel; x Represents historical transaction records; y Indicates a label; N Indicates the number of transactions; Indicates i Transactions The true label of is 1, which indicates fraud and 0, which indicates normal. Indicated in the parameter Make a trade probability.

7. The credit card fraud detection method based on adaptive meta-aggregation strategy according to claim 1, characterized in that: In the meta-model, the credit card transaction dataset is divided into a training set and a test set according to the set ratio 𝑝, where the training set is denoted as , the test set is denoted as , the training set and the test set satisfy the relationship, which is: ； ； ； in, represents the credit card transaction dataset; Indicates i The feature vector of each transaction; Indicates i The real label of each transaction; represents the number of transactions in the training set; represents the number of transactions in the test set; The transactions in the training set and the transactions in the test set are standardized based on the feature vectors of the transactions; Use the standardized transactions in the training set and the transactions in the test set to construct training data and test data respectively; ； ； in, represents training data; After standardization, i transactions; n Indicates the number of transactions after sampling; Represents test data; The metamodel is represented as: ； in, Represents the metamodel.

8. The credit card fraud detection method based on adaptive meta-aggregation strategy according to claim 1, characterized in that: It also includes evaluating the model performance using the test data, and the calculation formula is: ； ； ； in, represents the performance evaluation function; represents the global model parameters for further update; Represents test data; n Indicates the number of transactions after sampling; Indicates the test data i transactions; Indicates transaction The corresponding true label; represents the first indicator function; express sign function; represents the prediction function; represents the first weight matrix; represents the second weight matrix; represents the first bias term; Represents the second bias term; ReLU(⋅) represents the linear rectification function; Represents the sigmoid activation function; Adjust the weight of the loss function according to the evaluation results, and the update formula is: ； ； in, Represents the loss function after adjusting the weights; represents the logarithmic loss function; Represents the network parameters of the metamodel; x Represents historical transaction records; y Indicates a label; represents the third weight; represents the second indicator function; Represents the model judgment definition; Indicates the preset threshold value; when When the performance of the model is judged to be "excellent", the second indicator function The output of is 0, indicating that no additional adjustment is made to the loss function; when When the performance of the model is judged to be "poor", the second indicator function The output of is 1, indicating that the loss function needs to be adjusted.

Images