A sunshine exchange settlement model training method, device and transaction data classification method

Abstract

The present invention discloses a sunshine exchange settlement model training method, device and transaction data classification method, which specifically include the following steps: S1: acquiring transaction data and dividing it into a training set and a test set; S2: acquiring the weight of the sunshine exchange settlement model and reshaping it into a second-order matrix; S3: performing full-rank decomposition on the sunshine exchange settlement model to obtain two layers of convolution or reconstruct a fully connected layer; S4: applying orthogonal regularization and funnel regularization on the left singular matrix and the right singular matrix after the full-rank decomposition; S5: adopting an energy-based threshold method to reduce model parameters, thereby obtaining a low-rank model; S6: adjusting the hyperparameters of the low-rank model to obtain a target network model; the target network model is used to classify the transaction data.

Description

Technical Field

[0001] The present invention relates to the field of big data technology, and in particular to a sunshine exchange settlement model training method, device and transaction data classification method. Background Art

[0002] With the rapid development of the economy and society, international trade has become more frequent. The Sunshine Foreign Exchange Settlement Model can be used in many fields such as international trade, cross-border e-commerce, and personal cross-border remittances. However, Sunshine Foreign Exchange Settlement requires a lot of calculations every day, which undoubtedly puts tremendous pressure on the system. Faced with massive amounts of transaction data, traditional computing models are far from meeting actual needs. How to improve computing efficiency, reduce resource consumption, and ensure accuracy and low complexity of calculations are realistic problems that must be faced. At present, although the model compression technology of the Sunshine Foreign Exchange Settlement System has solved the problem of large-scale model deployment to a certain extent, how to balance the model compression rate and model performance is an urgent problem to be solved. Summary of the invention

[0003] In response to the technical problem of low accuracy in transaction data classification, the present invention proposes a sunshine settlement model training method, device and transaction data classification method. The present invention combines the advantages of orthogonal regularization and funnel regularization to achieve more efficient implementation of low-rank tensor decomposition, and combines energy-based thresholds to prune singular values, ensuring that the compressed model is not only small in size, but also reduces the loss of model accuracy. In addition, the method and device of the present invention have good versatility and can adapt to various types of deep learning models, including but not limited to CNN, recurrent neural network (RNN) and transformer (Transformer).

[0004] In order to achieve the above object, the present invention provides the following technical solutions:

[0005] A sunshine exchange model training method specifically comprises the following steps:

[0006] S1: Obtain transaction data and divide it into training set and test set;

[0007] S2: Get the weights of the sunshine exchange model and reshape it into a second-order matrix;

[0008] S3: Perform full rank decomposition on the sunshine exchange model. The decomposition results include left singular matrix, right singular matrix and diagonal matrix.

[0009] S4: Apply orthogonal regularization and funnel regularization to the left singular matrix and the right singular matrix after full rank decomposition;

[0010] S5: Use energy-based threshold method to reduce model parameters and obtain a low-rank model;

[0011] S6: Adjust the hyperparameters of the low-rank model to obtain a target network model; the target network model is used to classify the transaction data.

[0012] Preferably, in S1, the ratio of the training set to the test set is 8:2.

[0013] Preferably, in S2, the method for obtaining the second-order matrix includes:

[0014] The weights of the Sunshine Convergence Model are reshaped to reshape the original fourth-order weight tensor into a second-order matrix, that is, channel decomposition: D*D*M*N→M*DDN or spatial decomposition: D*D*M*N→MD*DN, where D is the size of the convolution kernel, and M and N are the input channel and output channel respectively.

[0015] Preferably, S3 includes:

[0016] S3-1: Perform SVD decomposition on the weights of the convolutional layer. The decomposition results include a left singular matrix, a right singular matrix, and a diagonal matrix. Then, the diagonal matrix is ​​fused with the left singular matrix to obtain a first matrix, and the diagonal matrix is ​​fused with the right singular matrix to obtain a second matrix. Finally, the first matrix and the second matrix are used as the update weights of the convolutional layer.

[0017] S3-2: The decomposition and construction process of the fully connected layer weights are as follows:

[0018] y T =x T W (1)

[0019]

[0020] In formulas (1), (2), and (3), x and y represent input and output respectively, W represents the weight matrix of the fully connected layer; z represents the output of the intermediate layer; T represents transpose; U and V represent the left singular matrix and the right singular matrix respectively; and s represents a diagonal matrix.

[0021] Preferably, the S4 includes:

[0022] S4-1: Apply orthogonal regularization to the left singular matrix and the right singular matrix, that is, satisfy row orthogonality and column orthogonality. The specific process is as follows:

[0023]

[0024] In formulas (4) and (5), L1(U) represents the loss term of the left singular matrix; L1(V) represents the loss term of the right singular matrix; U and V represent the left singular matrix and the right singular matrix respectively; T represents the transpose; I represents the loss term of U T U is the same as the identity matrix; denote the row-orthogonal and column-orthogonal approximations of a left singular matrix, respectively. They represent the row orthogonal and column orthogonal approximations of the right singular matrix, r represents the rank of SVD (Singular Value Decomposition); ρ represents the hyperparameter;

[0025] S4-2: Apply funnel regularization to the left singular matrix and the right singular matrix. The implementation process is as follows:

[0026]

[0027] In formula (6), L2 represents the funnel regularization; W represents the weight matrix of the fully connected layer; c represents a positive number;

[0028] S4-3: Combine the L1 and L2 losses with the original network’s loss function L ce Fusion:

[0029] L ce =L1(U)+L1(V)+L2 (7)

[0030] In formula (7), L ce Represents the loss function of the original network.

[0031] Preferably, the S5 includes:

[0032] S5-1: Select an energy threshold and apply it to each layer of the neural network:

[0033]

[0034] In formula (8), j represents the serial number of the element in the set K, i represents a positive number from 1 to r, K represents the set of element j, and S j represents the jth singular value, S i represents the i-th singular value; e∈(0,1), represents the energy threshold; r represents the rank of SVD;

[0035] S5-2: Prune the singular values ​​in the diagonal matrix s according to the energy threshold to obtain a low-rank model.

[0036] Preferably, the S6 includes:

[0037] S6-1: Set the batch size to 128, the number of iterations to 20, the learning rate to 0.001, and use the stochastic gradient descent strategy to randomly decay the low-rank model;

[0038] S6-2: Retrain the decomposed low-rank model to obtain the target network model.

[0039] Preferably, the method further includes S7: testing the target network model using a test set and evaluating the test result.

[0040]

[0041] In formulas (9), (10), and (11), CR represents the compression ratio, n top1 Indicates the number of correctly classified samples, n top5 Indicates the number of true categories in the top five labels; P original and P compressed They represent the original parameter W and the compressed parameters U, S, and V respectively; N represents the total number of samples; Top-1 represents the proportion of the Top-1 results predicted by the model that match the true label, that is, if the first choice predicted by the model is the correct category, then this prediction is considered accurate; Top-5 refers to the proportion of at least one of the top five predictions provided by the model that matches the true label. Even if the model does not rank the correct category first, as long as the correct category is included in the top five predictions, the prediction is considered accurate.

[0042] The present invention also provides a sunlight exchange model training device, comprising:

[0043] A data receiving unit, used for receiving transaction data input into the system;

[0044] The tensor reshaping unit is used to reshape the weight tensor of the sunlight junction model. The reshaped shape is a second-order matrix.

[0045] Full rank decomposition unit, used for full rank decomposition of the sunlight sink model to obtain two layers of convolution or reconstruct the fully connected layer;

[0046] A regularization unit, used to apply orthogonal regularization and funnel regularization on the left singular matrix and the right singular matrix after full rank decomposition;

[0047] Energy threshold unit, used to select energy threshold, reduce model parameters, and obtain low-rank model;

[0048] A fine-tuning unit is used to fine-tune the low-rank model to obtain a target network model;

[0049] The test unit is used to test the target network model and evaluate the test results.

[0050] The present invention also provides a transaction data classification method, comprising:

[0051] A1: Obtain transaction data to be classified;

[0052] A2: Processing the unclassified transaction data by using the Sunshine Foreign Exchange Settlement Model, wherein the Sunshine Foreign Exchange Settlement Model is trained according to the training method;

[0053] A3: Display the classification results.

[0054] In summary, due to the adoption of the above technical solution, compared with the prior art, the present invention has at least the following beneficial effects:

[0055] The present invention provides a method and device for training a sunshine settlement model and a method for classifying transaction data, which adopts a general neural network compression framework, wherein orthogonal regularization and funnel regularization reduce redundant information of the model, thereby improving the robustness and generalization ability of the model; and a threshold-based weight pruning strategy reduces the accuracy loss while reducing model parameters. The method effectively solves the problem of embedded device deployment of the model, achieves a balance between model parameters and accuracy, and can be widely used in fields such as unmanned driving, aerospace, and international trade. BRIEF DESCRIPTION OF THE DRAWINGS

[0056] Figure 1 Schematic diagram of a sunlight exchange model training method according to an exemplary embodiment of the present invention.

[0057] Figure 2 Schematic diagram of a transaction data classification method according to an exemplary embodiment of the present invention.

[0058] Figure 3 Schematic diagram of a sunlight junction model training device according to an exemplary embodiment of the present invention. DETAILED DESCRIPTION

[0059] The present invention is further described in detail below in conjunction with the examples and specific implementation methods. However, this should not be understood as the scope of the above subject matter of the present invention being limited to the following examples, and all technologies realized based on the content of the present invention belong to the scope of the present invention.

[0060] In the description of the present invention, it is necessary to understand that the terms "longitudinal", "lateral", "up", "down", "front", "back", "left", "right", "vertical", "horizontal", "top", "bottom", "inside", "outside", etc., indicating the orientation or position relationship, are based on the orientation or position relationship shown in the accompanying drawings, and are only for the convenience of describing the present invention and simplifying the description, rather than indicating or implying that the device or element referred to must have a specific orientation, be constructed and operated in a specific orientation, and therefore cannot be understood as a limitation on the present invention.

[0061] It will be understood by those skilled in the art that, unless otherwise defined, all terms (including technical and scientific terms) used herein have the same meaning as generally understood by those skilled in the art in the art to which the present invention belongs. It should also be understood that terms such as those defined in general dictionaries should be understood to have meanings consistent with the meanings in the context of the prior art, and will not be interpreted with idealized or overly formal meanings unless specifically defined as herein.

[0062] like Figure 1 As shown, the present invention provides a sunlight junction model training method based on tensor decomposition, which specifically includes the following steps:

[0063] S1: Obtain the private transaction data of the Sunshine Exchange Settlement Model, including abnormal transaction data, and divide it into a training set and a test set;

[0064] In this embodiment, the official data set can be obtained from the pytorch deep learning framework or a designated website, and the training set and the test set can be divided into a certain ratio, for example, 8:2.

[0065] In this embodiment, the experimental hardware configuration uses an Intel Core i7-11800H@2.30GHz processor, NVIDIA GeForce RTX 3060, 6G video memory GPU, 32G memory, the operating system is Windows 10, the development platform is Pycharm, the programming language is Python, the deep learning framework is Pytorch, and the GPU acceleration library is CUDA11.2.

[0066] S2: Get the weights of the sunshine settlement model and reshape them into a second-order matrix (if it is a fully connected layer, it remains unchanged);

[0067] In this embodiment, the method for obtaining the second-order matrix includes:

[0068] The weights are reshaped to reshape the original fourth-order weight tensor into a second-order matrix, that is, channel decomposition: D*D*M*N→M*DDN or spatial decomposition: D*D*M*N→MD*DN, where D is the size of the convolution kernel, M and N are the input channel and output channel respectively. For the fully connected layer, no operation is performed.

[0069] S3: Perform full-rank decomposition on the sunlight exchange model to obtain two layers of convolution or reconstruct the fully connected layer to approximate the original weights.

[0070] S3-1: Perform SVD (Singular Value Decomposition) decomposition on the weights of the convolution layer. The result contains three parts, namely the left singular matrix, the right singular matrix, and the diagonal matrix. Then, the diagonal matrix is ​​fused with the left singular matrix to obtain the first matrix, and the diagonal matrix is ​​fused with the right singular matrix to obtain the second matrix. Finally, the first matrix and the second matrix are used as the update weights of the convolution layer.

[0071] S3-2: The decomposition and construction process of the fully connected layer weights are as follows:

[0072] y T =x T W (1)

[0073]

[0074] In formulas (1), (2), and (3), x and y represent input and output respectively, W represents the weight matrix of the fully connected layer; z represents the output of the intermediate layer; T represents transpose; U and V represent the left singular matrix and the right singular matrix respectively; and s represents a diagonal matrix.

[0075] Formula (1) represents the original fully connected layer, and formula (3) represents the reconstructed fully connected layer formed by the decomposed matrix.

[0076] S4: Apply orthogonal regularization and funnel regularization to the left singular matrix and the right singular matrix after full rank decomposition to obtain low rank characteristics while eliminating redundant information.

[0077] S4-1: Since the decomposed matrix will be continuously updated, it is difficult to ensure orthogonality during the training process. Therefore, it is necessary to apply orthogonal regularization to the decomposed left singular matrix and right singular matrix, that is, to satisfy row orthogonality and column orthogonality. The specific process is as follows:

[0078]

[0079] In formulas (4) and (5), L1(U) represents the loss term of the left singular matrix; L1(V) represents the loss term of the right singular matrix; U and V represent the left singular matrix and the right singular matrix respectively; T represents the transpose; I represents the loss term of U T U is the same as the identity matrix; denote the row-orthogonal and column-orthogonal approximations of a left singular matrix, respectively. They represent the row orthogonal and column orthogonal approximations of the right singular matrix, r represents the rank of SVD (Singular Value Decomposition); ρ represents a hyperparameter.

[0080] S4-2: Applying funnel regularization to the decomposed left singular matrix and right singular matrix helps prevent model overfitting and limit the complexity of the model. The implementation process is as follows:

[0081]

[0082] In formula (6), L2 represents the funnel regularization; W represents the weight matrix of the fully connected layer; c represents a very small positive number to prevent the denominator from being 0. The purpose of the funnel regularization is to reduce the impact of noise and outliers in the data, improve the generalization ability of the model, and obtain low-rank characteristics.

[0083] S4-3: Combine the L1 and L2 losses with the original network’s loss function L ce The back propagation process is completed by the automatic differentiation tool (such as autograd) in the PyTorch deep learning framework, and there is no need to write formulas by hand.

[0084] L ce =L1(U)+L1(V)+L2 (7)

[0085] In this embodiment, the batch size is set to 128, the number of iterations is 200, the initial learning rate is 0.1, the learning rate is multiplied by 0.1 every 50 epochs, the weight decay is 0.001, and the stochastic gradient descent strategy is used for random decay (back propagation).

[0086] S5: Use energy-based threshold method to reduce model parameters and obtain a low-rank model.

[0087] S5-1: Select an energy threshold and apply the energy threshold to each layer of the neural network;

[0088] In this embodiment, for each layer, a set with the maximum number of singular values ​​is found and satisfies:

[0089]

[0090] In formula (8), j represents the serial number of the element in set K, i represents a positive number from 1 to r, K represents the set of element j, and S j represents the jth singular value, S i represents the i-th singular value; e∈(0,1), represents the energy threshold; r represents the rank of SVD (Singular Value Decomposition);

[0091] S5-2: Since the diagonal matrix s is a diagonal matrix (only the main diagonal has values, and the rest of the positions are 0), the singular values ​​from the upper left corner to the lower right corner will be arranged from large to small. The larger the singular value, the higher the importance, and vice versa. Therefore, the parameter e in formula (8) can be controlled to delete redundant features, for example, singular values ​​smaller than e can be deleted.

[0092] In this embodiment, the pruning process significantly reduces the rank of the decomposition layer, while the performance loss is negligible, thereby obtaining a low-rank model.

[0093] For example, for the kernel For a convolutional layer, if the rank of the decomposition layer can be reduced to r, the number of convolution parameters will be reduced to (M+DDN)r / DDMN when the channel is decomposed, and the number of convolution parameters will be reduced to (DM+DN)r / DDMN when the space is decomposed.

[0094] S6: Adjust the hyperparameters of the low-rank model and restore some accuracy to obtain the target network model, that is, the trained Sunshine Exchange model.

[0095] S6-1: Set the batch size to 128, the number of iterations to 20, the learning rate to 0.001, and use the stochastic gradient descent strategy for random decay;

[0096] S6-2: Fine-tune the low-rank model, retrain the decomposed low-rank model for a small number of cycles, and then fine-tune the weight parameters U, V to restore some accuracy of the model, and then obtain a robust target network model.

[0097] S7: Use the test set to test the target network model and evaluate the test results.

[0098] In this embodiment, the test set is compressed through the neural network (compression model), and finally outputs important results such as compression rate, classification accuracy, and training time:

[0099]

[0100] In formulas (9), (10), and (11), CR represents the compression ratio, n top1 Indicates the number of correctly classified samples, n top5 Indicates the number of true categories in the top five labels; P original and P compressedThey represent the original parameter W and the compressed parameters U, S, and V respectively; N represents the total number of samples; Top-1 represents the proportion of the Top-1 results predicted by the model that match the true label, that is, if the first choice predicted by the model (that is, the prediction with the highest probability) is the correct category, then this prediction is considered accurate; Top-5 refers to the proportion of at least one of the top five predictions provided by the model that matches the true label. Even if the model does not rank the correct category first, as long as the correct category is included in the top five predictions, the prediction is considered accurate.

[0101] In this embodiment, the above process is a classification problem, and the quality of the model can be judged based on Top-1, Top-5 and CR.

[0102] Based on the above-mentioned sunshine exchange settlement model training method based on tensor decomposition, the present invention also provides a transaction data classification method, such as Figure 2 As shown, the specific steps include:

[0103] A1: Obtain the transaction data to be classified.

[0104] In this embodiment, there are normal behaviors and abnormal behaviors in the transaction data, so it is necessary to identify and classify normal behaviors and abnormal behaviors, so as to discover potential abnormal behaviors or abnormal transactions, which helps to improve the accuracy of decision-making.

[0105] A2: The trained Sunshine Exchange Settlement Model (target network model) is used to process the classified transaction data.

[0106] In this embodiment, after the terminal device obtains the transaction data to be classified, the transaction data to be classified is input into the trained Sunshine Foreign Exchange Settlement Model, which is trained according to the neural network model training method provided in the above embodiment.

[0107] A3: Display the classification results.

[0108] In this embodiment, the trained lightweight model can classify the transaction data of Sunshine Exchange Settlement (classify normal transactions or abnormal transactions), and the final output result of the model is the classification accuracy. Potential abnormal behaviors or abnormal transactions can be discovered based on the classification accuracy, which helps to improve the accuracy of decision-making.

[0109] Based on the above training method, such as Figure 3 As shown, the present invention also provides a sunlight junction model training device based on tensor decomposition, comprising:

[0110] A data receiving unit is used to receive data input into the system, divide the data set into certain proportions, and then store the data set into the data storage unit;

[0111] The tensor reshaping unit is used to reshape the weight tensor of the convolutional neural network into a second-order matrix.

[0112] The full-rank decomposition unit is used to perform full-rank decomposition on the sunlight settlement model to obtain two layers of convolution or reconstruct the fully connected layer to approximate the original weights;

[0113] The regularization unit is used to apply orthogonal regularization and funnel regularization to the factor matrix after full rank decomposition; while eliminating redundant information, low rank characteristics are obtained; orthogonal regularization is applied to the decomposed full rank matrix, that is, row orthogonality and column orthogonality are satisfied;

[0114] Energy threshold unit, used to select energy threshold, reduce model parameters, and obtain low-rank model;

[0115] A fine-tuning unit is used to fine-tune the low-rank model to obtain a target network model;

[0116] The test unit is used to test the target network model and evaluate the test results.

[0117] Those skilled in the art will appreciate that the above-mentioned embodiments are specific examples for implementing the present invention, and in actual applications, various changes may be made thereto in form and detail without departing from the spirit and scope of the present invention.

Claims

1. A method for training a sunshine settlement model, characterized in that: The specific steps include: S1: Obtain transaction data and divide it into training set and test set; S2: Get the weights of the sunshine exchange model and reshape it into a second-order matrix; S3: Perform full rank decomposition on the Sunshine Remittance Model, and the decomposition results include a left singular matrix, a right singular matrix, and a diagonal matrix; S4: Apply orthogonal regularization and funnel regularization to the left singular matrix and the right singular matrix after the full rank decomposition; S4-1: Apply orthogonal regularization to the left singular matrix and the right singular matrix, that is, satisfy row orthogonality and column orthogonality. The specific process is as follows: In formulas (4) and (5), L1(U) represents the loss term of the left singular matrix; L1(V) represents the loss term of the right singular matrix; U and V represent the left singular matrix and the right singular matrix respectively; T represents the transpose; I represents the loss term of U T U is the same as the identity matrix; denote the row-orthogonal and column-orthogonal approximations of a left singular matrix, respectively. They represent the row and column orthogonal approximation of the right singular matrix, r represents the rank of SVD (Singular Value Decomposition); ρ represents the hyperparameter; S4-2: Apply funnel regularization to the left singular matrix and the right singular matrix. The implementation process is as follows: In formula (6), L2 represents the funnel regularization; W represents the weight matrix of the fully connected layer in the model; c represents a positive number; S4-3: Combine the L1 and L2 losses with the original network’s loss function L ce Fusion: L ce =L1(U)+L1(V)+L2 (7) In formula (7), L ce Represents the loss function of the original network; S5: Use energy-based threshold method to reduce model parameters and obtain a low-rank model; S6: Adjust the hyperparameters of the low-rank model to obtain a target network model; the target network model is used to classify the transaction data into normal transactions or abnormal transactions.

2. A method for training a sunlight exchange model according to claim 1, characterized in that: In S1, the ratio of the training set to the test set is 8:

2.

3. A method for training a sunlight junction model according to claim 1, characterized in that: In S2, the method for obtaining the second-order matrix includes: The weights of the Sunshine Convergence Model are reshaped to reshape the original fourth-order weight tensor into a second-order matrix, that is, channel decomposition: D*D*M*N→M*DDN or spatial decomposition: D*D*M*N→MD*DN, where D is the size of the convolution kernel, and M and N are the input channel and output channel respectively.

4. A method for training a sunlight exchange model according to claim 1, characterized in that: The S3 includes: S3-1: Perform SVD decomposition on the weights of the convolutional layer. The decomposition results include a left singular matrix, a right singular matrix, and a diagonal matrix. Then, the diagonal matrix is ​​fused with the left singular matrix to obtain a first matrix, and the diagonal matrix is ​​fused with the right singular matrix to obtain a second matrix. Finally, the first matrix and the second matrix are used as the update weights of the convolutional layer. S3-2: The decomposition and construction process of the fully connected layer weights are as follows: y T =x T W (1) In formulas (1), (2), and (3), x and y represent input and output respectively, W represents the weight matrix of the fully connected layer; z represents the output of the intermediate layer; T represents transpose; U and V represent the left singular matrix and the right singular matrix respectively; and s represents a diagonal matrix.

5. A method for training a sunlight junction model according to claim 1, characterized in that: The S5 includes: S5-1: Select an energy threshold and apply it to each layer of the neural network: In formula (8), j represents the serial number of the element in the set K, i represents a positive number from 1 to r, K represents the set of element j, S j represents the jth singular value, S i represents the i-th singular value; e∈(0,1), represents the energy threshold; r represents the rank of SVD; S5-2: Prune the singular values ​​in the diagonal matrix s according to the energy threshold to obtain a low-rank model.

6. A method for training a sunlight junction model according to claim 1, characterized in that: The S6 includes: S6-1: Set the batch size to 128, the number of iterations to 20, the learning rate to 0.001, and use the stochastic gradient descent strategy to randomly decay the low-rank model; S6-2: Retrain the decomposed low-rank model to obtain the target network model.

7. A method for training a sunlight exchange model according to claim 1, characterized in that: It also includes S7: using the test set to test the target network model and evaluating the test results: In formulas (9), (10), and (11), CR represents the compression ratio, n top1 Indicates the number of correctly classified samples, n top5 Indicates the number of true categories in the top five labels; P original and P compressed Represent the original parameters and the parameters after compression respectively; N represents the total number of samples; Top-1 represents the proportion of Top-1 results predicted by the model that match the true label, that is, if the first choice predicted by the model is the correct category, then this prediction is considered accurate; Top-5 refers to the proportion of at least one of the top five predictions provided by the model that matches the true label. Even if the model does not rank the correct category first, as long as the correct category is included in the top five predictions, the prediction is considered accurate.

8. A sunlight junction model training device based on the method according to any one of claims 1 to 7, characterized in that: include: A data receiving unit, used for receiving transaction data input into the system; The tensor reshaping unit is used to reshape the weight tensor of the sunlight junction model. The reshaped shape is a second-order matrix. Full rank decomposition unit, used for full rank decomposition of the sunlight sink model to obtain two layers of convolution or reconstruct the fully connected layer; A regularization unit, used to apply orthogonal regularization and funnel regularization on the left singular matrix and the right singular matrix after full rank decomposition; Energy threshold unit, used to select energy threshold, reduce model parameters, and obtain low-rank model; A fine-tuning unit is used to fine-tune the low-rank model to obtain a target network model, which is used to classify transaction data into normal transactions or abnormal transactions; The test unit is used to test the target network model and evaluate the test results.

9. A transaction data classification method, characterized in that: include: A1: Obtain transaction data to be classified; A2: Processing the unclassified transaction data by means of a Sunshine Foreign Exchange Settlement Model, wherein the Sunshine Foreign Exchange Settlement Model is obtained by training according to the training method according to any one of claims 1 to 7; A3: Display the classification results.

Images