Cross-border payment fraud order detection and prompt method based on machine learning and large language model

Abstract

The invention provides a cross-border payment fraud order detection and prompt method based on machine learning and a large language model, and relates to the technical field of artificial intelligence, and the method comprises the steps: obtaining the order information of an online cross-border transaction platform, and analyzing the order information through a large language model, so as to obtain the order analysis information; according to the order analysis information, vectorization is carried out by using a Word2Vec algorithm; according to the vectorized data, performing feature extraction by using a full-connection neural network to obtain order information features; according to the order information features, a reverse reconstruction auto-encoder is used to carry out dimension reduction; and performing classification by using a support vector machine according to the dimensionality-reduced features, performing risk detection on order information of cross-border payment, and performing prompting according to a threshold value. The invention provides a cross-border payment fraud order detection and prompt method based on machine learning and a large language model. The technical problem that in the prior art, limitation exists in the aspect of recognizing subtle abnormal behaviors and hidden risks in complex transactions is solved.

Description

Technical Field

[0001] The present invention relates to the field of artificial intelligence technology, and in particular to a method for detecting and prompting cross-border payment fraud orders based on machine learning and a large language model. Background Art

[0002] With the rapid development of global e-commerce and online transactions, the number of cross-border payment transactions has increased dramatically. Transactions often involve multiple currencies, regulations and payment systems, which increases the complexity of transactions and the risk of fraud. Traditional fraud detection systems mainly rely on simple rules or pattern matching technology, which is often insufficient in dealing with highly complex and dynamically changing cross-border payment environments. Existing technologies are limited in identifying subtle abnormal behaviors and hidden risks in complex transactions, and are difficult to adapt to changing fraud strategies and rapidly changing market conditions.

[0003] Existing technologies may lack effective pattern recognition and risk indicator extraction capabilities when processing complex structured data, limiting their scope of application in complex trading environments; traditional neural networks are susceptible to gradient vanishing or exploding during feature extraction, affecting the stability of training and the generalization ability of the model; in the process of feature dimensionality reduction, existing technologies cannot effectively minimize the dependencies between different features, resulting in the inability of the reduced data to accurately reflect the key characteristics of the original data, thereby affecting the subsequent classification accuracy and the practicality of the model. Summary of the invention

[0004] The present invention provides a method for detecting and prompting cross-border payment fraud orders based on machine learning and a large language model, which solves the technical problems that the prior art has limitations in identifying subtle abnormal behaviors and hidden risks in complex transactions and is difficult to adapt to changing fraud strategies and rapidly changing market conditions.

[0005] In order to solve the above technical problems, the technical solution of the present invention is as follows:

[0006] In a first aspect, a method for detecting and prompting cross-border payment fraud orders based on machine learning and a large language model, the method comprising:

[0007] Obtain order information from online cross-border trading platforms and use large language models to analyze the order information to obtain order analysis information;

[0008] According to the order analysis information, use the Word2Vec algorithm to vectorize to obtain vectorized data;

[0009] Based on the vectorized data, a fully connected neural network is used to extract features to obtain order information features;

[0010] According to the order information features, the inverse reconstruction autoencoder is used to reduce the dimension to obtain the features after dimensionality reduction;

[0011] Based on the features after dimensionality reduction, support vector machines are used to perform classification to obtain classification categories, wherein the classification categories include low risk, medium risk, and high risk;

[0012] Based on the classification categories, the order information of cross-border payments is checked for risks and prompts are given based on the thresholds.

[0013] Furthermore, order information of an online cross-border trading platform is obtained, and the order information is analyzed using a large language model to obtain order analysis information, wherein the large language model adopts a Transformer network model, and the Transformer network includes an encoder and a decoder;

[0014] The encoder consists of N trans The same encoding layers are composed of each encoding layer, where each encoding layer includes a multi-head attention mechanism sublayer and a feedforward network sublayer;

[0015] The feed-forward network sublayer includes two fully connected layers and a ReLU activation function;

[0016] The feed-forward network sublayer is trained using a method based on seismic wave propagation optimization;

[0017] The decoder consists of N trans The same decoding layers are composed of two multi-head attention mechanism sub-layers and one feed-forward network mechanism sub-layer;

[0018] In the encoder and decoder of the Transformer network, the output of each sublayer is layer normalized and residually connected, with the formula y = LN c (x trans +Sublayer(x trans )), where LN c () is the layer normalization function, Sublayer(x trans ) represents the output of the residual connected sublayer, x trans Represents the input of the sublayer.

[0019] Furthermore, the order information of the online cross-border trading platform is obtained, and the order information is analyzed using a large language model to obtain order analysis information, including:

[0020] According to the order information of the online cross-border trading platform, the first multi-head attention mechanism sublayer of the encoder uses linear transformation to obtain the query, key, and value vectors, and then divides the query, key, and value vectors into multiple heads, and calculates the dot product of the query and key vectors of each head to obtain the attention score;

[0021] Use the attention score Process and obtain the attention weight;

[0022] Use the attention weights to perform weighted summation on the value vector to get the output vector;

[0023] The output vectors of all heads are concatenated and linearly transformed to obtain the final output of the first multi-head self-attention layer;

[0024] According to the output of the first multi-head self-attention layer, the first fully connected layer is used to perform linear transformation in the first feedforward network sublayer of the encoder, and then the ReLU activation function is used for nonlinear transformation, and finally the second fully connected layer is used for linear transformation to obtain the output of the first feedforward network sublayer;

[0025] According to the output of the first feedforward network sublayer, in the second multi-head attention mechanism sublayer of the decoder, the second multi-head attention layer output is calculated;

[0026] According to the output of the second multi-head attention layer, the third multi-head attention mechanism sublayer uses the attention mechanism to perform cross-attention calculation on the output of the second multi-head attention layer to obtain a set of weighted decoder output features, namely, order analysis information.

[0027] Furthermore, the feedforward network sublayer is trained using a method based on seismic wave propagation optimization, including:

[0028] use Initializing the parameters of the feedforward network to obtain initialized parameters, wherein the parameters of the feedforward network include the weights and biases of the feedforward network, wherein: is the weight matrix of the lth layer of the feedforward network, is the bias vector of the lth layer of the feedforward network, is the number of nodes in the previous layer of the feedforward network, represents a normal distribution with a mean of 0 and a standard deviation of 1;

[0029] According to the initialized parameters, forward propagation is performed through the seismic wave propagation simulation layer to obtain the output of each layer, and the output of each layer is expressed as in, is the linear combination output of the lth layer, is the activation output of the l-1th layer, is the activation output of the lth layer, Sigs() is the activation function of the seismic wave propagation characteristics, and the formula of the seismic wave propagation characteristics activation function is: σ p is the amplitude adjustment parameter that simulates the diffusion characteristics of seismic waves, μ p and ω pare the amplitude and frequency of the sine wave part respectively;

[0030] According to the output of each layer, use The loss function of the feedforward network is calculated, where is the loss function of the feedforward network, Ns is the total number of samples input in the current batch, is the cross entropy loss for a single sample, is the output predicted by the model, y i is the true label, λ p is the regularization coefficient, ∥∥ F represents the Frobenius norm, express The Frobenius norm of x i is the i-th input sample, ∥∥ represents the L2 norm, α p is the weight controlling the intra-class distance penalty, S(y i ) indicates that x i A set of samples of the same category, Re() represents the ReLU activation function;

[0031] According to the loss function of the feedforward network, use Update the initialization parameters of the feedforward network to obtain the parameters after dimension reduction, where: is the weight update amount of the l-th layer feedforward network, is the bias update of the l-th layer feedforward network, η p is the feedforward network learning rate, γ p is the adjustment parameter of the seismic wave model; is a regularization term designed based on the propagation characteristics of seismic waves. The learning rate of the feedforward network is dynamically adjusted based on the gradient correction strategy. The gradient correction coefficient is set to The formula is in, and are the gradients of the t-th and t-1-th iterations respectively. The adaptive adjustment formula of the learning rate of the feedforward network learning rate is Where η0 is the initial feedforward network learning rate, η(t) is the feedforward network learning rate at the tth iteration, and α p It is a hyperparameter that controls the speed at which the learning rate decays;

[0032] Repeat the above steps until the preset maximum number of iterations is reached, which means the model training is completed.

[0033] Furthermore, according to the order analysis information, the Word2Vec algorithm is used for vectorization to obtain vectorized data, including:

[0034] According to the order analysis information, the Word2Vec algorithm is used to scan the order analysis information according to the preset corpus to obtain the scanning result;

[0035] According to the scanning results, each word is converted into a one-hot encoding vector to obtain vectorized data, and the dimension of the one-hot encoding vector is equal to the size of the preset vocabulary.

[0036] Furthermore, based on the vectorized data, a fully connected neural network is used to extract features to obtain order information features, including:

[0037] According to the vectorized data, use Initialize the network parameters to obtain the initialized parameters, wherein the network parameters include weights and biases, wherein ~ indicates that it obeys a specific distribution, W NN and b NN are the weights and biases of the neural network, σ sn is the initialization standard deviation, and are the initial weights and initial biases of the neural network, is a normal distribution;

[0038] According to the initialization parameters, use Calculate the loss L NN (W NN ,b NN ), and use E NN =exp(-β NN ·(W NN ,b NN )) Estimate the environmental fitness to obtain the loss L NN (W NN ,b NN ) and environmental adaptability, among which E NN is the environmental fitness, β NN is the sensitivity parameter, L NN () is the loss function of the neural network, C represents the number of categories, y c is the one-hot encoding of the true label, is the probability predicted by the model, which is calculated by the softmax function, and the formula is Among them, z c is the logistic regression probability output by the network;

[0039] According to the loss L NN (W NN ,b NN ),use Calculate the resources allocated to each parameter to obtain the result of resource competition, Rij The resources allocated for the parameters of the jth neuron in the i-th neural network layer, is the weight parameter of the jth neuron in the i-th neural network layer, is the weight parameter of the lth neuron in the kth neural network layer, k and l are both indices, γ NN is a hyperparameter that controls the sensitivity of resource allocation, is the loss function L NN About parameters The partial derivative of is the loss function L NN About parameters The partial derivative of

[0040] Use according to the result of resource competition Update weight W NN and bias b NN , to obtain the parameters after dimension reduction, where R i The resources allocated for the parameters of the i-th neural network layer, is the bias of the neurons in the i-th layer, is the weight parameter update of the lth neuron in the kth neural network layer, α NN is the learning rate, is the update amount of the bias parameter of the i-th layer of neurons;

[0041] According to the environmental adaptability, use The sensitivity parameter and the learning rate parameter are adjusted to obtain the adjusted sensitivity parameter and the learning rate parameter, where ΔE NN is the change in environmental fitness, that is, the difference between the environmental fitness of this iteration and the previous iteration; η NN and κ NN is the adjustment factor, is the updated learning rate parameter, is the updated sensitivity parameter;

[0042] According to the sensitivity parameter and the learning rate parameter, use Dynamically adjust the hyperparameters that control resource allocation sensitivity, where || is the absolute value symbol, is the hyperparameter of the sensitivity of the adjusted control resource allocation, ρ is the hyperparameter of the balance, and ρ NN Set to 0.1;

[0043] Repeat the above steps until the preset maximum number of iterations is reached, which means that the model training is completed and a fully connected neural network model is obtained;

[0044] A fully connected neural network model is used for feature extraction to obtain order information features.

[0045] Furthermore, according to the order information features, the inverse reconstruction autoencoder is used to reduce the dimension to obtain the features after dimensionality reduction, including:

[0046] use b r,init =0 to initialize the parameters of the autoencoder algorithm to obtain the initialized parameters, wherein the parameters of the autoencoder algorithm include the weight and bias parameters of the autoencoder, wherein W r,init is the initial weight matrix of the autoencoder, b r,init is the initial bias of the autoencoder, n r,in and n r,out Represent the number of neurons in the input and output layers respectively, and U(-1,1) represents a random number uniformly distributed in the interval [-1,1];

[0047] According to the order information characteristics and initialization parameters, the encoder uses z r =Sig(W r,enc ·x r +b r,enc ) is converted into a low-dimensional representation, which is then used by the decoder Reconstruct the output to the same dimension as the original data, and obtain the reconstructed output, where z r is the low-dimensional feature representation of the encoding, Sig() is the Sigmoid activation function, and W r,enc and W r,dec are the weights of the encoder and decoder, respectively, and b r,enc and b r,dec are the biases of the encoder and decoder respectively, x r is the input data, is the output of the reconstruction;

[0048] Based on the refactored output, use The inverse reconstruction loss of the autoencoder is calculated, where L r is the inverse reconstruction loss function, Ntr is the number of samples input in the current batch, ∥∥ represents the L2 norm, and λ r is the regularization parameter, D KL (P r ∥Q r ) is the KL divergence, which is calculated using Measure the original data distribution P r and reconstruct the data distribution Q r The difference between r,i and q r,iare the probability distributions of the original data and the reconstructed data in the i-th dimension, respectively, and Mtr is the feature dimension before feature dimensionality reduction;

[0049] According to the inverse reconstruction loss of the autoencoder, the weights and biases of the encoder and decoder are updated using the gradient descent algorithm to obtain the parameters after dimensionality reduction. The formula of the gradient descent algorithm is:

[0050] Among them, η r is the learning rate, is the updated encoder weight, is the updated encoder bias, is the updated decoder weight, is the updated decoder bias;

[0051] Repeat the above steps until the preset maximum number of iterations is reached, which means that the model training is completed and the reverse reconstruction autoencoder model is obtained;

[0052] Use the inverse reconstruction autoencoder model to reduce the dimension to obtain the reduced dimension features.

[0053] Further, according to the features after dimensionality reduction, support vector machines are used for classification to obtain classification categories, which include low risk, medium risk, and high risk, including:

[0054] Based on the features after dimensionality reduction, support vector machine is used to find the best decision boundary in the data points;

[0055] Based on the optimal decision boundary, determine the classification category for risk assessment of cross-border payment data.

[0056] The above solution of the present invention includes at least the following beneficial effects:

[0057] The above-mentioned scheme of the present invention adopts a large language model based on Transformer to conduct in-depth analysis of order information in structured JSON format. The optimization algorithm is a feedforward network algorithm based on seismic wave propagation optimization. Through the combination of encoder and decoder, it can effectively extract and parse complex patterns and potential risk indicators in transaction data; the Word2Vec algorithm is used to convert the text output of the language model into a numerical vector, so that it can be compatible with the subsequent machine learning process, providing a basis for feature extraction and risk assessment; a three-layer fully connected neural network is adopted, and the network parameters are adjusted through an adaptive biological impact optimization algorithm to simulate the adaptation process of biological resources, thereby improving the model's ability to extract features and reducing instability during training; an autoencoder based on inverse reconstruction is used to perform feature dimensionality reduction, and the quality of data expression is optimized by minimizing the dependency between the encoded data and the original data, thereby enhancing the model's ability to capture key features of the original data. BRIEF DESCRIPTION OF THE DRAWINGS

[0058] Figure 1 It is a flow chart of a method for detecting and prompting cross-border payment fraud orders based on machine learning and a large language model provided by an embodiment of the present invention.

[0059] Figure 2 It is a schematic diagram of the Transformer structure of the cross-border payment fraud order detection and prompting method based on machine learning and a large language model provided in an embodiment of the present invention.

[0060] Figure 3 It is a schematic diagram of the feature extraction model training process of the cross-border payment fraud order detection and prompting method based on machine learning and a large language model provided by an embodiment of the present invention.

[0061] Figure 4 It is a schematic diagram of the feature dimensionality reduction model training process of the cross-border payment fraud order detection and prompting method based on machine learning and a large language model provided in an embodiment of the present invention. DETAILED DESCRIPTION

[0062] The exemplary embodiments of the present disclosure will be described in more detail below with reference to the accompanying drawings. Although the exemplary embodiments of the present disclosure are shown in the accompanying drawings, it should be understood that the present disclosure can be implemented in various forms and should not be limited by the embodiments set forth herein. On the contrary, these embodiments are provided to enable a more thorough understanding of the present disclosure and to fully convey the scope of the present disclosure to those skilled in the art.

[0063] like Figure 1 As shown, an embodiment of the present invention proposes a method for detecting and prompting cross-border payment fraud orders based on machine learning and a large language model, the method comprising:

[0064] 11. Obtain order information from online cross-border trading platforms and use large language models to analyze the order information to obtain order analysis information;

[0065] 12. According to the order analysis information, use the Word2Vec algorithm to vectorize to obtain vectorized data;

[0066] 13. Based on the vectorized data, use a fully connected neural network to extract features to obtain order information features;

[0067] 14. According to the order information characteristics, use the inverse reconstruction autoencoder to reduce the dimension to obtain the features after dimensionality reduction;

[0068] 15. Based on the features after dimensionality reduction, support vector machines are used to perform classification to obtain classification categories, wherein the classification categories include low risk, medium risk, and high risk;

[0069] 16. According to the classification category, the risk of cross-border payment order information is checked and prompts are given according to the threshold.

[0070] In an embodiment of the present invention, by using a large language model based on Transformer to perform in-depth analysis of order information in structured JSON format, complex patterns and potential risk indicators in transaction data can be more effectively extracted and parsed, which helps to discover fraud behavior patterns that are difficult to capture with traditional methods, thereby improving the accuracy of fraud detection; the Word2Vec algorithm is used to convert the text output of the language model into numerical vectors, so that these vectors can be compatible with subsequent machine learning processes, which not only simplifies the data processing process, but also makes it easier to integrate the model into existing cross-border payment systems, thereby enhancing the compatibility and scalability of the model; a three-layer fully connected neural network is used for feature extraction, and the network parameters are adjusted through an adaptive biological impact optimization algorithm, which can simulate the adaptation process of biological resources, improve the model's ability to extract features, and help reduce instability during training. Qualitative, thereby improving the stability and performance of the model; using an autoencoder based on inverse reconstruction for feature dimensionality reduction, by minimizing the dependency between the encoded data and the original data, the quality of data expression can be optimized, while reducing the dimension of the data, reducing the computational complexity of the model, and improving the computational efficiency, so that the model can process a large amount of cross-border payment order information more quickly; performing risk detection on cross-border payment order information according to classification categories (low risk, medium risk, high risk), and providing prompts based on thresholds, which helps users or payment platforms quickly identify and deal with potential fraudulent behaviors, so as to take necessary preventive measures in a timely manner; by improving the accuracy and efficiency of fraud detection, this method helps to reduce false positives and false negatives, improve user experience, and by timely prompting and preventing potential fraudulent behaviors, this method can also effectively ensure the security of cross-border payments and enhance users' trust in payment platforms.

[0071] like Figure 1 As shown, 11. Obtain order information from an online cross-border trading platform, and use a large language model to analyze the order information to obtain order analysis information, wherein the large language model adopts a Transformer network model, and the Transformer network includes an encoder and a decoder;

[0072] The encoder consists of N trans The same encoding layers are composed of each encoding layer, where each encoding layer includes a multi-head attention mechanism sublayer and a feedforward network sublayer;

[0073] The feed-forward network sublayer includes two fully connected layers and a ReLU activation function;

[0074] The feed-forward network sublayer is trained using a method based on seismic wave propagation optimization;

[0075] The decoder consists of N transThe same decoding layers are composed of two multi-head attention mechanism sub-layers and one feed-forward network mechanism sub-layer;

[0076] In the encoder and decoder of the Transformer network, the output of each sublayer is layer normalized and residually connected, with the formula y = LN c (x trans +Sublayer(x trans )), where LN c () is the layer normalization function, Sublayer(x trans ) represents the output of the residual connected sublayer, x trans Represents the input of the sublayer.

[0077] In an embodiment of the present invention, order information is obtained from an online cross-border transaction platform. This information usually includes key fields such as product details, buyer information, seller information, transaction amount, transaction time, etc., and is often stored in a structured JSON format. A large language model based on Transformer is used to perform in-depth analysis of these order information. The Transformer network model is known for its powerful sequence processing capabilities and is particularly suitable for processing long texts or complex structured data. The Transformer network includes two parts: an encoder and a decoder. The encoder is responsible for converting the input data (i.e., order information) into a series of high-dimensional vector representations, while the decoder is usually used to generate outputs. It is composed of multiple identical encoding layers stacked together. Each encoding layer contains a multi-head attention mechanism sublayer and a feedforward network sublayer. The multi-head attention mechanism allows the model to focus on different parts of the input data at the same time, thereby improving the ability to capture complex patterns; it is composed of two fully connected layers and a ReLU activation function, which are used to further process the vector representation of the output of the encoding layer. The feedforward network sublayer is trained using a method based on seismic wave propagation optimization to improve the training efficiency and performance of the model; it is also composed of multiple decoding layers stacked together, but Each decoding layer contains two multi-head attention mechanism sub-layers and one feedforward network mechanism sub-layer; in the encoder and decoder of the Transformer network, the output of each sub-layer is layer normalized and residually connected. Layer normalization helps to stabilize the training process of the model, while residual connection allows the model to retain input information and avoid information loss in the deep network; the order information processed by the Transformer network is converted into a series of high-dimensional vector representations. The model will perform risk detection on the order information of cross-border payments based on the classification results (such as low risk, medium risk, and high risk), and issue prompts to users or payment platforms based on the set thresholds.

[0078] like Figure 1As shown, 11. Obtain order information from the online cross-border transaction platform, and use the large language model to analyze the order information to obtain order analysis information, including:

[0079] According to the order information of the online cross-border trading platform, the first multi-head attention mechanism sublayer of the encoder uses linear transformation to obtain the query, key, and value vectors, and then divides the query, key, and value vectors into multiple heads, and calculates the dot product of the query and key vectors of each head to obtain the attention score;

[0080] Use the attention score Process and obtain the attention weight;

[0081] Use the attention weights to perform weighted summation on the value vector to get the output vector;

[0082] The output vectors of all heads are concatenated and linearly transformed to obtain the final output of the first multi-head self-attention layer;

[0083] According to the output of the first multi-head self-attention layer, the first fully connected layer is used to perform linear transformation in the first feedforward network sublayer of the encoder, and then the ReLU activation function is used for nonlinear transformation, and finally the second fully connected layer is used for linear transformation to obtain the output of the first feedforward network sublayer;

[0084] According to the output of the first feedforward network sublayer, in the second multi-head attention mechanism sublayer of the decoder, the second multi-head attention layer output is calculated;

[0085] According to the output of the second multi-head attention layer, the third multi-head attention mechanism sublayer uses the attention mechanism to perform cross-attention calculation on the output of the second multi-head attention layer to obtain a set of weighted decoder output features, namely, order analysis information.

[0086] In an embodiment of the present invention, deep encoding of order information and multi-head self-attention processing: the order information obtained from the online cross-border trading platform is first input into the encoder part of the Transformer model; in the first multi-head attention mechanism sublayer of the encoder, the order information is first converted into query (Query), key (Key), and value (Value) vectors through linear transformation, and the vectors are then divided into multiple "heads" (i.e., the parts processed in parallel by multiple attention mechanisms) to capture different features in the data; each head calculates the dot product of the query and key vectors to obtain an attention score, which reflects the strength of association between different parts of information; the attention scores are softmax processed to obtain attention weights, which are used to perform weighted summation on the value vectors to generate the output vector of each head; finally, the output vectors of all heads are concatenated, and the final value of the first multi-head self-attention layer is obtained through linear transformation. Output; The first feedforward network sublayer of the encoder receives the output of the first multi-head self-attention layer as input; In this sublayer, the first fully connected layer is first used for linear transformation, and then the ReLU activation function is applied for nonlinear transformation to enhance the expressiveness of the model; Finally, the second fully connected layer is used for linear transformation to obtain the output of the first feedforward network sublayer; Cross-attention calculation of the decoder and generation of order analysis information: The second multi-head attention mechanism sublayer of the decoder receives the output of the first feedforward network sublayer as input, and calculates the output of the second multi-head attention layer; In the third multi-head attention mechanism sublayer of the decoder, the attention mechanism is used to perform cross-attention calculation on the output of the second multi-head attention layer; Through the cross-attention calculation, a set of weighted decoder output features are obtained, which not only improves the accuracy of fraud detection, but also enhances the interpretability and robustness of the model, providing a strong guarantee for the security and user experience of cross-border payments.

[0087] like Figure 1 As shown, the feedforward network sublayer is trained using a method based on seismic wave propagation optimization, including:

[0088] use Initializing the parameters of the feedforward network to obtain initialized parameters, wherein the parameters of the feedforward network include the weights and biases of the feedforward network, wherein: is the weight matrix of the lth layer of the feedforward network, is the bias vector of the lth layer of the feedforward network, is the number of nodes in the previous layer of the feedforward network, represents a normal distribution with a mean of 0 and a standard deviation of 1;

[0089] According to the initialized parameters, forward propagation is performed through the seismic wave propagation simulation layer to obtain the output of each layer, and the output of each layer is expressed as in, is the linear combination output of the lth layer, is the activation output of the l-1th layer, is the activation output of the lth layer, Sigs() is the activation function of the seismic wave propagation characteristics, and the formula of the seismic wave propagation characteristics activation function is: σ p is the amplitude adjustment parameter that simulates the diffusion characteristics of seismic waves, μ p and ω p are the amplitude and frequency of the sine wave part respectively;

[0090] According to the output of each layer, use The loss function of the feedforward network is calculated, where is the loss function of the feedforward network, Ns is the total number of samples input in the current batch, is the cross entropy loss of a single sample, is the output predicted by the model, y i is the true label, λ p is the regularization coefficient, ∥∥ F represents the Frobenius norm, express The Frobenius norm of x i is the i-th input sample, ∥∥ represents the L2 norm, α p is the weight controlling the intra-class distance penalty, S(y i ) indicates that x i A set of samples of the same category, Re() represents the ReLU activation function;

[0091] According to the loss function of the feedforward network, use Update the initialization parameters of the feedforward network to obtain the parameters after dimension reduction, where: is the weight update amount of the l-th layer feedforward network, is the bias update of the l-th layer feedforward network, η p is the feedforward network learning rate, γ p is the adjustment parameter of the seismic wave model; is a regularization term designed based on the propagation characteristics of seismic waves. The learning rate of the feedforward network is dynamically adjusted based on the gradient correction strategy. The gradient correction coefficient is set to The formula is in, and are the gradients of the t-th and t-1-th iterations respectively. The adaptive adjustment formula of the learning rate of the feedforward network learning rate is Where η0 is the initial feedforward network learning rate, η(t) is the feedforward network learning rate at the tth iteration, and α p It is a hyperparameter that controls the speed at which the learning rate decays;

[0092] Repeat the above steps until the preset maximum number of iterations is reached, which means the model training is completed.

[0093] In the embodiment of the present invention, the weights and biases of the feedforward network are initialized using a normal distribution (mean is 0, standard deviation is 1); an activation function of seismic wave propagation characteristics is introduced, the sinusoidal wave characteristics are integrated, the amplitude is adjusted by the α parameter, and A and ω respectively control the amplitude and frequency of the sinusoidal wave part, aiming to simulate the nonlinear characteristics of seismic wave propagation in the medium and enhance the expressive power of the model; the loss function of the feedforward network is calculated using the cross entropy loss and the regularization term, the cross entropy loss measures the difference between the model prediction output and the true label, and the regularization term is used to control the model complexity and prevent overfitting; the regularization term Including Frobenius norm and L2 norm, which are used to constrain the weight matrix and intra-class distance penalty respectively; according to the loss function, the weight and bias of the feedforward network are updated using an optimization algorithm; the seismic wave model adjustment parameter β and the regularization term designed based on the propagation characteristics of seismic waves are introduced to further guide the parameter update process; the learning rate η of the feedforward network adopts a dynamic adjustment method based on the gradient correction strategy, and the learning rate is adaptively adjusted through the gradient correction coefficient γ and the hyperparameter λ that controls the learning rate decay speed to improve the training efficiency and model performance; repeat the above steps until the preset maximum number of iterations is reached. In each iteration, the model calculates the output through forward propagation, evaluates the performance through the loss function, and optimizes the model structure through parameter update; as the number of iterations increases, the model gradually converges to the optimal solution, which is manifested as a decrease in the loss function value and an improvement in the performance of the model on the validation set; after the training is completed, the model can more accurately analyze the order information of the online cross-border trading platform, providing strong support for subsequent fraud detection, risk assessment and other tasks.

[0094] like Figure 1 As shown, 12. According to the order analysis information, the Word2Vec algorithm is used for vectorization to obtain vectorized data, including:

[0095] According to the order analysis information, the Word2Vec algorithm is used to scan the order analysis information according to the preset corpus to obtain the scanning result;

[0096] According to the scanning results, each word is converted into a one-hot encoding vector to obtain vectorized data, and the dimension of the one-hot encoding vector is equal to the size of the preset vocabulary.

[0097] In an embodiment of the present invention, the Word2Vec algorithm is used according to the order analysis information. The core of this algorithm is to use a preset large corpus to train word vectors; in the scanning stage, the Word2Vec algorithm will carefully analyze each word in the order analysis information, and these words will be compared and matched with the vocabulary in the corpus one by one; then, the algorithm will convert each matched word into a one-hot encoding vector, which is a commonly used representation method. In this representation, each word is assigned a vector equal to the size of the preset vocabulary; the characteristic of this vector is that only one dimension has a value of 1, while the values ​​of the other dimensions are all 0; the size of the vocabulary is preset and contains all the words that the algorithm can recognize and process; each word in the order analysis information is converted into a one-hot encoding vector, which together constitutes the vectorized data; using the Word2Vec algorithm to vectorize the order analysis information can effectively convert the text information into a numerical form that the computer can understand and process.

[0098] like Figure 1 As shown in 13, based on the vectorized data, a fully connected neural network is used to extract features to obtain order information features, including:

[0099] According to the vectorized data, use Initialize the network parameters to obtain the initialized parameters, wherein the network parameters include weights and biases, wherein ~ indicates that it obeys a specific distribution, W NN and b NN are the weights and biases of the neural network, σ sn is the initialization standard deviation, and are the initial weights and initial biases of the neural network, is a normal distribution;

[0100] According to the initialization parameters, use Calculate the loss L NN (W NN ,b NN ), and use E NN =exp(-β NN ·(W NN ,b NN )) Estimate the environmental fitness to obtain the loss L NN (W NN ,b NN ) and environmental adaptability, among which E NN is the environmental fitness, β NN is the sensitivity parameter, L NN () is the loss function of the neural network, C represents the number of categories, y c is the one-hot encoding of the true label, is the probability predicted by the model, which is calculated by the softmax function, and the formula is Among them, z c is the logistic regression probability output by the network;

[0101] According to the loss L NN (W NN ,b NN ),use Calculate the resources allocated to each parameter to obtain the result of resource competition, R ij The resources allocated for the parameters of the jth neuron in the i-th neural network layer, is the weight parameter of the jth neuron in the i-th neural network layer, is the weight parameter of the lth neuron in the kth neural network layer, k and l are both indices, γ NN is a hyperparameter that controls the sensitivity of resource allocation, is the loss function L NN About parameters The partial derivative of is the loss function L NN About parameter W NNkl The partial derivative of

[0102] Use according to the result of resource competition Update weight W NN and bias b NN , to obtain the parameters after dimension reduction, where R i The resources allocated for the parameters of the i-th neural network layer, is the bias of the neurons in the i-th layer, is the weight parameter update of the lth neuron in the kth neural network layer, α NN is the learning rate, is the update amount of the bias parameter of the i-th layer of neurons;

[0103] According to the environmental adaptability, use The sensitivity parameter and the learning rate parameter are adjusted to obtain the adjusted sensitivity parameter and the learning rate parameter, where ΔE NN is the change in environmental fitness, that is, the difference between the environmental fitness of this iteration and the previous iteration; η NN and κ NN is the adjustment factor, is the updated learning rate parameter, is the updated sensitivity parameter;

[0104] According to the sensitivity parameter and the learning rate parameter, use Dynamically adjust the hyperparameters that control resource allocation sensitivity, where || is the absolute value symbol, is the hyperparameter of the sensitivity of the adjusted control resource allocation, ρ is the hyperparameter of the balance, and ρ NN Set to 0.1;

[0105] Repeat the above steps until the preset maximum number of iterations is reached, which means that the model training is completed and a fully connected neural network model is obtained;

[0106] A fully connected neural network model is used for feature extraction to obtain order information features.

[0107] In an embodiment of the present invention, the network parameters are initialized using normal distribution according to the vectorized data; the initialized weights and biases are the starting point of neural network learning, and will be adjusted by optimization algorithms in the subsequent training process; a loss function is used to calculate the loss value under the current network state, and environmental fitness is introduced as another indicator to measure network performance, combining the loss value and sensitivity parameters to guide subsequent parameter adjustments; the loss function usually reflects the difference between the model prediction output and the true label, while the environmental fitness considers more the adaptability of the network in a specific environment; in each iteration, the amount of resources allocated to each parameter is calculated according to the loss value, which can be regarded as a resource competition mechanism; the result of resource competition determines which parameters will receive more attention, that is, more resources, in the subsequent training process; based on the result of resource competition, the weights and biases are updated using optimization algorithms to minimize the loss value and Improve network performance; dynamically adjust sensitivity parameters and learning rate parameters according to changes in environmental fitness. The sensitivity parameter controls the degree of influence of environmental fitness on parameter updates, while the learning rate determines the step size of parameter updates. By adjusting these parameters, the stability and learning ability of the network can be better balanced; introduce a hyperparameter to control the sensitivity of resource allocation. This hyperparameter is dynamically adjusted according to changes in sensitivity parameters and learning rate parameters to ensure that resource allocation is more reasonable and effective; repeat the above steps until the preset maximum number of iterations is reached or other stopping conditions are met. During the training process, the network will gradually converge to a stable state, at which time the performance of the model is optimal or near-optimal; use the trained fully connected neural network model for feature extraction. The model can convert high-dimensional features extracted from vectorized data into low-dimensional, easy-to-understand and process feature vectors. The feature vectors are order information features.

[0108] like Figure 1 As shown in 14, according to the order information characteristics, the inverse reconstruction autoencoder is used to reduce the dimension to obtain the features after dimensionality reduction, including:

[0109] use b r,init=0 to initialize the parameters of the autoencoder algorithm to obtain the initialized parameters, wherein the parameters of the autoencoder algorithm include the weight and bias parameters of the autoencoder, wherein W r,init is the initial weight matrix of the autoencoder, b r,init is the initial bias of the autoencoder, n r,in and n r,out Represent the number of neurons in the input and output layers respectively, and U(-1,1) represents a random number uniformly distributed in the interval [-1,1];

[0110] According to the order information characteristics and initialization parameters, the encoder uses z r =Sig(W r,enc ·x r +b r,enc ) is converted into a low-dimensional representation, which is then used by the decoder Reconstruct the output to the same dimension as the original data, and obtain the reconstructed output, where z r is the low-dimensional feature representation of the encoding, Sig() is the Sigmoid activation function, and W r,enc and W r,dec are the weights of the encoder and decoder, respectively, and b r,enc and b r,dec are the biases of the encoder and decoder respectively, x r is the input data, is the output of the reconstruction;

[0111] Based on the refactored output, use The inverse reconstruction loss of the autoencoder is calculated, where L r is the inverse reconstruction loss function, Ntr is the number of samples input in the current batch, ∥∥ represents the L2 norm, and λ r is the regularization parameter, D KL (P r ∥Q r ) is the KL divergence, which is calculated using Measure the original data distribution P r and reconstruct the data distribution Q r The difference between r,i and q r,i are the probability distributions of the original data and the reconstructed data in the i-th dimension, respectively, and Mtr is the feature dimension before feature dimensionality reduction;

[0112] According to the inverse reconstruction loss of the autoencoder, the weights and biases of the encoder and decoder are updated using the gradient descent algorithm to obtain the parameters after dimension reduction. The formula of the gradient descent algorithm is:

[0113] Among them, η r is the learning rate, is the updated encoder weight, is the updated encoder bias, is the updated decoder weight, is the updated decoder bias;

[0114] Repeat the above steps until the preset maximum number of iterations is reached, which means that the model training is completed and the reverse reconstruction autoencoder model is obtained;

[0115] Use the inverse reconstruction autoencoder model to reduce the dimension to obtain the reduced dimension features.

[0116] In an embodiment of the present invention, the parameters of the autoencoder algorithm are initialized using uniformly distributed random numbers, which provides a starting point for the training of the autoencoder and is adjusted through an optimization algorithm in the subsequent training process; order information features are used as input data and converted into a representation of a low-dimensional space through an encoder, thereby retaining the key information of the original data and removing redundancy and noise; the decoder uses this low-dimensional representation to reconstruct an output with the same dimension as the original data; the performance of the autoencoder can be evaluated by comparing the difference between the reconstructed output and the original input; in order to quantify the reconstruction error, an inverse reconstruction loss function is used to calculate the loss value, and the loss function combines the L2 norm and the KL divergence to measure the difference between the original data distribution and the reconstructed data distribution. The L2 norm is used The Euclidean distance between the reconstructed output and the original input is evaluated, while the KL divergence is used to measure the similarity between two probability distributions. According to the calculated inverse reconstruction loss, the gradient descent algorithm is used to update the weights and biases of the encoder and decoder. Through continuous iteration, the parameters of the autoencoder can be gradually optimized so that it can reconstruct the input data more accurately. When the preset maximum number of iterations is reached or other stopping conditions are met, the model has been trained. The trained inverse reconstruction autoencoder model is used to reduce the dimensionality of the order information features, converting the high-dimensional order information features into low-dimensional, easy-to-understand and process feature vectors, which retains the key information of the original data and reduces the dimensionality and complexity of the data, providing strong support for subsequent analysis and decision-making.

[0117] like Figure 1 As shown, 15. Based on the features after dimensionality reduction, support vector machines are used for classification to obtain classification categories, wherein the classification categories include low risk, medium risk, and high risk, including:

[0118] Based on the features after dimensionality reduction, support vector machine is used to find the best decision boundary in the data points;

[0119] Based on the optimal decision boundary, determine the classification category for risk assessment of cross-border payment data.

[0120] In an embodiment of the present invention, the features after dimensionality reduction are used as input data, and the support vector machine algorithm is used to find the best decision boundary in the data points. The decision boundary is the core of the support vector machine algorithm and can effectively separate data points of different categories. In the support vector machine, the decision boundary is determined by maximizing the interval (i.e., the edge) between two categories, which helps to improve the accuracy and robustness of classification. According to the best decision boundary found, the risk of cross-border payment data is evaluated and its classification category is determined. The classification categories include three levels: low risk, medium risk and high risk. The support vector machine algorithm will assign the data points to the corresponding risk categories according to their positions on the decision boundary. Through support vector machine classification, the risk category label of each cross-border payment data point can be obtained for subsequent decision-making, risk management and prevention measures. It not only provides effective risk warning and management tools for financial institutions and payment platforms, but also provides strong support for risk research and strategy formulation in the field of cross-border payments. In practical applications, this method can be widely used in various risk management and decision-making tasks related to cross-border payments, providing solid protection for the stable operation of enterprises and the financial security of users.

[0121] The above is a preferred embodiment of the present invention. It should be pointed out that for ordinary technicians in this technical field, several improvements and modifications can be made without departing from the principles of the present invention. These improvements and modifications should also be regarded as the scope of protection of the present invention.

Claims

1. A method for detecting and prompting cross-border payment fraud orders based on machine learning and large language models, characterized in that: The method comprises: Obtain order information from online cross-border trading platforms and use large language models to analyze the order information to obtain order analysis information; According to the order analysis information, use the Word2Vec algorithm to vectorize to obtain vectorized data; Based on the vectorized data, a fully connected neural network is used to extract features to obtain order information features; According to the order information features, the inverse reconstruction autoencoder is used to reduce the dimension to obtain the features after dimensionality reduction; Based on the features after dimensionality reduction, support vector machines are used to perform classification to obtain classification categories, wherein the classification categories include low risk, medium risk, and high risk; Based on the classification categories, the order information of cross-border payments is checked for risks and prompts are given based on the thresholds.

2. The cross-border payment fraud order detection and prompting method based on machine learning and large language model according to claim 1 is characterized in that: Obtain order information from an online cross-border trading platform, and analyze the order information using a large language model to obtain order analysis information, wherein the large language model adopts a Transformer network model, and the Transformer network includes an encoder and a decoder; The encoder consists of N trans The same encoding layers are composed of each encoding layer, where each encoding layer includes a multi-head attention mechanism sublayer and a feedforward network sublayer; The feed-forward network sublayer includes two fully connected layers and a ReLU activation function; The feed-forward network sublayer is trained using a method based on seismic wave propagation optimization; The decoder consists of N trans The same decoding layers are composed of two multi-head attention mechanism sub-layers and one feed-forward network mechanism sub-layer; In the encoder and decoder of the Transformer network, the output of each sublayer is layer normalized and residually connected, with the formula y = LN c (x trans +Sublayer(x trans )), where LN c () is the layer normalization function, Sublayer(x trans ) represents the output of the residual connected sublayer, x trans Represents the input of the sublayer.

3. The cross-border payment fraud order detection and prompting method based on machine learning and large language model according to claim 2 is characterized in that: Obtain order information from online cross-border trading platforms and use a large language model to analyze the order information to obtain order analysis information, including: According to the order information of the online cross-border trading platform, the first multi-head attention mechanism sublayer of the encoder uses linear transformation to obtain the query, key, and value vectors, and then divides the query, key, and value vectors into multiple heads, and calculates the dot product of the query and key vectors of each head to obtain the attention score; Use the attention score Process and obtain the attention weight; Use the attention weights to perform weighted summation on the value vector to get the output vector; The output vectors of all heads are concatenated and linearly transformed to obtain the final output of the first multi-head self-attention layer; According to the output of the first multi-head self-attention layer, the first fully connected layer is used to perform linear transformation in the first feedforward network sublayer of the encoder, and then the ReLU activation function is used for nonlinear transformation, and finally the second fully connected layer is used for linear transformation to obtain the output of the first feedforward network sublayer; According to the output of the first feedforward network sublayer, in the second multi-head attention mechanism sublayer of the decoder, the second multi-head attention layer output is calculated; According to the output of the second multi-head attention layer, the third multi-head attention mechanism sublayer uses the attention mechanism to perform cross-attention calculation on the output of the second multi-head attention layer to obtain a set of weighted decoder output features, namely, order analysis information.

4. The cross-border payment fraud order detection and prompting method based on machine learning and large language model according to claim 3 is characterized in that: The feedforward network sublayer is trained using a method based on seismic wave propagation optimization, including: use Initializing the parameters of the feedforward network to obtain initialized parameters, wherein the parameters of the feedforward network include the weights and biases of the feedforward network, wherein: is the weight matrix of the lth layer of the feedforward network, is the bias vector of the lth layer of the feedforward network, is the number of nodes in the previous layer of the feedforward network, represents a normal distribution with a mean of 0 and a standard deviation of 1; According to the initialized parameters, forward propagation is performed through the seismic wave propagation simulation layer to obtain the output of each layer, and the output of each layer is expressed as in, is the linear combination output of the lth layer, is the activation output of the l-1th layer, is the activation output of the lth layer, Sigs() is the activation function of the seismic wave propagation characteristics, and the formula of the seismic wave propagation characteristics activation function is: σ p is the amplitude adjustment parameter that simulates the diffusion characteristics of seismic waves, μ p and ω p are the amplitude and frequency of the sine wave part respectively; According to the output of each layer, use The loss function of the feedforward network is calculated, where is the loss function of the feedforward network, Ns is the total number of samples input in the current batch, is the cross entropy loss for a single sample, is the output predicted by the model, y i is the true label, λ p is the regularization coefficient, ∥∥ F represents the Frobenius norm, express The Frobenius norm of x i is the i-th input sample, ∥∥ represents the L2 norm, α p is the weight controlling the intra-class distance penalty, S(y i ) indicates that x i A set of samples of the same category, Re() represents the ReLU activation function; According to the loss function of the feedforward network, use Update the initialization parameters of the feedforward network to obtain the parameters after dimension reduction, where: is the weight update amount of the l-th layer feedforward network, is the bias update of the l-th layer feedforward network, η p is the feedforward network learning rate, γ p is the adjustment parameter of the seismic wave model; is a regularization term designed based on the propagation characteristics of seismic waves. The learning rate of the feedforward network is dynamically adjusted based on the gradient correction strategy. The gradient correction coefficient is set to The formula is in, and are the gradients of the t-th and t-1-th iterations respectively. The adaptive adjustment formula of the learning rate of the feedforward network learning rate is Where η0 is the initial feedforward network learning rate, η(t) is the feedforward network learning rate at the tth iteration, and α p It is a hyperparameter that controls the speed at which the learning rate decays; Repeat the above steps until the preset maximum number of iterations is reached, which means the model training is completed.

5. The cross-border payment fraud order detection and prompting method based on machine learning and large language model according to claim 4 is characterized in that: According to the order analysis information, the Word2Vec algorithm is used for vectorization to obtain vectorized data, including: According to the order analysis information, the Word2Vec algorithm is used to scan the order analysis information according to the preset corpus to obtain the scanning result; According to the scanning results, each word is converted into a one-hot encoding vector to obtain vectorized data, and the dimension of the one-hot encoding vector is equal to the size of the preset vocabulary.

6. The cross-border payment fraud order detection and prompting method based on machine learning and large language model according to claim 5 is characterized in that: Based on the vectorized data, a fully connected neural network is used to extract features to obtain order information features, including: According to the vectorized data, use Initialize the network parameters to obtain the initialized parameters, wherein the network parameters include weights and biases, wherein ~ indicates that it obeys a specific distribution, W NN and b NN are the weights and biases of the neural network, σ sn is the initialization standard deviation, and are the initial weights and initial biases of the neural network, is a normal distribution; According to the initialization parameters, use Calculate the loss L NN (W NN ,b NN ), and use E NN =exp(-β NN ·(W NN ,b NN )) Estimate the environmental fitness to obtain the loss L NN (W NN ,b NN ) and environmental adaptability, among which E NN is the environmental fitness, β NN is the sensitivity parameter, L NN () is the loss function of the neural network, C represents the number of categories, y c is the one-hot encoding of the true label, is the probability predicted by the model, which is calculated by the softmax function, and the formula is Among them, z c is the logistic regression probability output by the network; According to the loss L NN (W NN ,b NN ),use Calculate the resources allocated to each parameter to obtain the result of resource competition, R ij The resources allocated for the parameters of the jth neuron in the i-th neural network layer, is the weight parameter of the jth neuron in the i-th neural network layer, is the weight parameter of the lth neuron in the kth neural network layer, k and l are both indices, γ NN is a hyperparameter that controls the sensitivity of resource allocation, is the loss function L NN About parameters The partial derivative of is the loss function L NN About parameters The partial derivative of Use according to the result of resource competition Update weight W NN and bias b NN , to obtain the parameters after dimension reduction, where R i The resources allocated for the parameters of the i-th neural network layer, is the bias of the neurons in the i-th layer, is the weight parameter update of the lth neuron in the kth neural network layer, α NN is the learning rate, is the update amount of the bias parameter of the i-th layer of neurons; According to the environmental adaptability, use The sensitivity parameter and the learning rate parameter are adjusted to obtain the adjusted sensitivity parameter and the learning rate parameter, where ΔE NN is the change in environmental fitness, that is, the difference between the environmental fitness of this iteration and the previous iteration; η NN and κ NN is the adjustment factor, is the updated learning rate parameter, is the updated sensitivity parameter; According to the sensitivity parameter and the learning rate parameter, use Dynamically adjust the hyperparameters that control resource allocation sensitivity, where || is the absolute value symbol, is the hyperparameter of the sensitivity of the adjusted control resource allocation, ρ is the hyperparameter of the balance, and ρ NN Set to 0.1; Repeat the above steps until the preset maximum number of iterations is reached, which means that the model training is completed and a fully connected neural network model is obtained; A fully connected neural network model is used for feature extraction to obtain order information features.

7. The cross-border payment fraud order detection and prompting method based on machine learning and large language model according to claim 6 is characterized in that: According to the order information features, the inverse reconstruction autoencoder is used to reduce the dimension to obtain the features after dimensionality reduction, including: use b r,init =0 to initialize the parameters of the autoencoder algorithm to obtain the initialized parameters, wherein the parameters of the autoencoder algorithm include the weight and bias parameters of the autoencoder, wherein W r,init is the initial weight matrix of the autoencoder, b r,init is the initial bias of the autoencoder, n r,in and n r,out Represent the number of neurons in the input and output layers respectively, and U(-1,1) represents a random number uniformly distributed in the interval [-1,1]; According to the order information characteristics and initialization parameters, the encoder uses z r =Sig(W r,enc ·x r +b r,enc ) is converted into a low-dimensional representation, which is then used by the decoder Reconstruct the output to the same dimension as the original data, and obtain the reconstructed output, where z r is the low-dimensional feature representation of the encoding, Sig() is the Sigmoid activation function, and W r,enc and W r,dec are the weights of the encoder and decoder, respectively, and b r,enc and b r,dec are the biases of the encoder and decoder respectively, x r is the input data, is the output of the reconstruction; Based on the refactored output, use The inverse reconstruction loss of the autoencoder is calculated, where L r is the inverse reconstruction loss function, Ntr is the number of samples input in the current batch, ∥∥ represents the L2 norm, and λ r is the regularization parameter, D KL (P r ∥Q r ) is the KL divergence, which is calculated using Measure the original data distribution P r and reconstruct the data distribution Q r The difference between r,i and q r,i are the probability distributions of the original data and the reconstructed data in the i-th dimension, respectively, and Mtr is the feature dimension before feature dimensionality reduction; According to the inverse reconstruction loss of the autoencoder, the weights and biases of the encoder and decoder are updated using the gradient descent algorithm to obtain the parameters after dimensionality reduction. The formula of the gradient descent algorithm is: Among them, η r is the learning rate, is the updated encoder weight, is the updated encoder bias, is the updated decoder weight, is the updated decoder bias; Repeat the above steps until the preset maximum number of iterations is reached, which means that the model training is completed and the reverse reconstruction autoencoder model is obtained; Use the inverse reconstruction autoencoder model to reduce the dimension to obtain the reduced dimension features.

8. The cross-border payment fraud order detection and prompting method based on machine learning and large language model according to claim 7 is characterized in that: According to the features after dimensionality reduction, support vector machine is used for classification to obtain classification categories, which include low risk, medium risk, and high risk, including: Based on the features after dimensionality reduction, support vector machine is used to find the best decision boundary in the data points; Based on the optimal decision boundary, determine the classification category for risk assessment of cross-border payment data.

Images