Cross-border business compliance prediction method and device, computer device and medium

Abstract

The application relates to the technical field of artificial intelligence, and discloses a cross-border business compliance prediction method and device, computer equipment and a medium, wherein the method comprises the following steps: acquiring state information of all distributed nodes in a cross-border e-commerce platform to determine idle nodes and combine the idle nodes into a virtual computing resource pool; performing distributed training on a sub-model based on the virtual computing resource pool; collecting gradient information of each distributed node and updating a gradient of a total compliance prediction model to obtain a target total compliance prediction model, which predicts the compliance of a cross-border business to be analyzed.The application has the beneficial effects of improving overall resource utilization efficiency, reducing response delay, improving the training speed and accuracy of the model, and enabling the cross-border e-commerce platform to quickly adapt to market changes and improve user experience.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of cross-border business compliance prediction, and in particular to a cross-border business compliance prediction method and device, computer equipment and medium. BACKGROUND

[0002] With the rapid development of global trade, cross-border e-commerce business is rapidly growing in various regions. However, due to time zone differences and sudden traffic, the computing power demand of cross-border business fluctuates significantly. In peak periods, the average utilization rate of GPU resources is less than 40% due to the traditional fixed computing power allocation mode. At the same time, the response delay of the business is significantly increased, especially during peak periods, the response delay often exceeds 500 milliseconds, affecting user experience and transaction efficiency. Therefore, how to effectively improve the dynamic utilization rate of GPU resources to cope with the increasing computing power demand has become an important challenge for the development of cross-border business. The existing centralized training mode faces many challenges. SUMMARY

[0003] Therefore, it is necessary to propose a cross-border business compliance prediction method, device, computer equipment and medium for the existing cross-border business compliance prediction problem.

[0004] A cross-border business compliance prediction method, the method comprising:

[0005] Obtain the state information of all distributed nodes in the cross-border e-commerce platform;

[0006] Detect whether the state information of each distributed node reaches a first preset state, and mark the distributed node that reaches the first preset state as an idle node;

[0007] Obtain a training task and a compliance prediction total model corresponding to the training task, and combine multiple idle nodes into a virtual computing resource pool based on the training task through a load balancing algorithm;

[0008] Distributed training of sub-models based on the virtual computing resource pool;

[0009] Collect the gradient information of each distributed node in the virtual computing resource pool, aggregate and update the gradient of the compliance prediction total model through an aggregation algorithm, and obtain a target compliance prediction total model;

[0010] Obtain a cross-border business to be analyzed, and predict the compliance of the cross-border business to be analyzed based on the target compliance prediction total model.

[0011] Furthermore, after the steps of obtaining a training task and a compliance prediction overall model corresponding to the training task, and combining multiple idle nodes into a virtual computing resource pool through a load balancing algorithm based on the training task, the method further includes:

[0012] Real-time monitoring of the real-time status information of all distributed nodes;

[0013] Determining whether the real-time status information of the idle node has reached a second preset state, and determining whether the real-time status information of the non-idle nodes in all distributed nodes has reached a third preset state;

[0014] The idle nodes whose real-time status information reaches the second preset state are deleted from the virtual computing pool, and the non-idle nodes whose real-time status information reaches the third preset state are added to the virtual computing pool.

[0015] Furthermore, before the step of collecting the gradient information of each idle node in the virtual computing resource pool, aggregating the information through an aggregation algorithm and updating the gradient of the compliance prediction model to obtain the target compliance prediction model, the step further includes:

[0016] When a model update instruction of each of the idle nodes is triggered, each of the idle nodes obtains a plurality of cross-border business training samples in the region where the idle node is located;

[0017] Training the respective pre-stored compliance prediction sub-models based on the cross-border business training samples to obtain sub-model gradients;

[0018] Corresponding sub-gradient information is generated based on the sub-model gradient and uploaded to the distributed node where the compliance prediction overall model is located.

[0019] Furthermore, the step of generating corresponding sub-gradient information based on the sub-model gradient and uploading it to the distributed node where the compliance prediction overall model is located further includes:

[0020] Applying Laplace noise to the sub-model gradient to obtain a temporary sub-model gradient;

[0021] Filter each temporary sub-model gradient according to the preset gradient parameter filtering mechanism to obtain the target temporary sub-model gradient;

[0022] Based on the target temporary sub-model gradient, corresponding sub-gradient information is generated and uploaded to the distributed node where the compliance prediction overall model is located.

[0023] Furthermore, before the step of training the respective pre-stored compliance prediction sub-models based on the cross-border business training samples to obtain the sub-model gradients, the step further includes:

[0024] Each of the idle nodes downloads the teacher model from the cloud respectively;

[0025] Extract the attention matrix of the multi-layer Transformer layer in the teacher model;

[0026] A preset number of attention matrices are selected from the multi-layer Transformer layer, and a preset loss function is added to obtain a student model to serve as the compliance prediction sub-model.

[0027] Furthermore, the step of collecting the gradient information of each distributed node in the virtual computing resource pool, aggregating it through an aggregation algorithm and updating the gradient of the compliance prediction overall model to obtain the target compliance prediction overall model includes:

[0028] Obtain the total gradient information corresponding to the total compliance prediction model;

[0029] Calculating a differential parameter between the total gradient information and the sub-gradient information;

[0030] Based on the differential parameters, the compliance prediction overall model is updated by a preset data synchronization algorithm to obtain a temporary compliance prediction overall model;

[0031] Testing the performance of the overall model for predicting temporary compliance;

[0032] If the performance meets the preset requirements, the temporary compliance prediction model is recorded as the target compliance prediction model.

[0033] Furthermore, the step of combining multiple idle nodes into a virtual computing resource pool through a load balancing algorithm based on the training task includes:

[0034] Mapping the training task into a multidimensional feature vector;

[0035] Based on the multi-dimensional feature vector, the decision path is generated by the CART algorithm;

[0036] A plurality of idle nodes are obtained based on the decision path to form a virtual computing resource pool.

[0037] A device for predicting cross-border business compliance, comprising:

[0038] The acquisition module is used to obtain the status information of all distributed nodes in the cross-border e-commerce platform;

[0039] a detection module, configured to detect whether the status information of each of the distributed nodes has reached a first preset state, and mark the distributed nodes that have reached the first preset state as idle nodes;

[0040] A combination module is used to obtain a training task and a compliance prediction model corresponding to the training task, and combine multiple idle nodes into a virtual computing resource pool through a load balancing algorithm based on the training task;

[0041] A training module, configured to perform distributed training of sub-models based on the virtual computing resource pool;

[0042] A collection module, configured to collect gradient information of each distributed node in the virtual computing resource pool, aggregate it through an aggregation algorithm, and update the gradient of the compliance prediction model to obtain a target compliance prediction model;

[0043] The prediction module is used to obtain the cross-border business to be analyzed and predict the compliance of the cross-border business to be analyzed based on the target compliance prediction model.

[0044] A computer device comprises a memory and a processor, wherein the memory stores a computer program, and when the computer program is executed by the processor, the processor performs the following steps:

[0045] Obtain status information of all distributed nodes in the cross-border e-commerce platform;

[0046] detecting whether the status information of each of the distributed nodes has reached a first preset state, and marking the distributed nodes that have reached the first preset state as idle nodes;

[0047] Obtaining a training task and a compliance prediction model corresponding to the training task, and combining multiple idle nodes into a virtual computing resource pool through a load balancing algorithm based on the training task;

[0048] Performing distributed training of sub-models based on the virtual computing resource pool;

[0049] Collecting the gradient information of each distributed node in the virtual computing resource pool, aggregating it through an aggregation algorithm and updating the gradient of the compliance prediction model to obtain a target compliance prediction model;

[0050] Obtain the cross-border business to be analyzed, and predict the compliance of the cross-border business to be analyzed based on the target compliance prediction model.

[0051] A computer-readable storage medium stores a computer program, which, when executed by a processor, causes the processor to perform the following steps:

[0052] Obtain status information of all distributed nodes in the cross-border e-commerce platform;

[0053] Detect whether the state information of each distributed node reaches a first preset state, and mark the distributed node reaching the first preset state as an idle node;

[0054] Obtain a training task and a compliance prediction total model corresponding to the training task, and combine multiple idle nodes into a virtual computing resource pool based on the training task through a load balancing algorithm;

[0055] Distributed training of a sub-model is performed based on the virtual computing resource pool;

[0056] Gradient information of each distributed node in the virtual computing resource pool is collected, and the gradient information is aggregated and updated to obtain a target compliance prediction total model;

[0057] Obtain a cross-border business to be analyzed, and predict the compliance of the cross-border business to be analyzed based on the target compliance prediction total model.

[0058] The beneficial effects of the present application are: through the dynamic resource management mode of identifying and marking idle nodes, in the environment where the computing power demand fluctuates significantly, the computing resource configuration can be flexibly adjusted, the problem of low GPU resource utilization in the traditional fixed computing power allocation mode is avoided, the overall resource utilization efficiency is improved, the response delay is reduced, on the other hand, the training speed and accuracy of the model are improved, so that the cross-border e-commerce platform can quickly adapt to market changes and improve user experience. BRIEF DESCRIPTION OF DRAWINGS

[0059] In order to more clearly illustrate the technical solutions in the embodiments of the present application or the prior art, the drawings needed to be used in the embodiments or prior art description will be briefly introduced below. Obviously, the drawings in the following description are only some embodiments of the present application, and other drawings can be obtained by those skilled in the art without creative labor.

[0060] Among them:

[0061] Figure 1 It is an application environment diagram of the cross-border business compliance prediction method in one embodiment;

[0062] Figure 2 It is a flowchart of the cross-border business compliance prediction method in one embodiment;

[0063] Figure 3 It is a structure block diagram of the cross-border business compliance prediction device in one embodiment;

[0064] Figure 4 It is a structure block diagram of the computer device in one embodiment. DETAILED DESCRIPTION

[0065] The following will clearly and completely describe the technical solutions in the embodiments of the present invention in conjunction with the accompanying drawings. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. All other embodiments obtained by ordinary technicians in this field based on the embodiments of the present invention without making any creative efforts shall fall within the scope of protection of the present invention.

[0066] Figure 1 This is a diagram of the application environment for predicting cross-border business compliance in an embodiment. Figure 1 This cross-border business compliance prediction method is applied to a cross-border business compliance prediction system. This cross-border business compliance prediction system includes a terminal 110 and a server 120. Terminal 110 and server 120 are connected via a network. Terminal 110 can be a desktop terminal or a mobile terminal. The mobile terminal can be at least one of a mobile phone, tablet computer, and laptop computer. Server 120 can be implemented as a standalone server or a server cluster consisting of multiple servers. Terminal 110 is used to collect status information, and server 120 is used to predict compliance.

[0067] like Figure 2 As shown, in one embodiment, a method for predicting cross-border business compliance is provided. This method can be applied to both a terminal and a server. This embodiment uses the server as an example for illustration. The method for predicting cross-border business compliance specifically includes the following steps:

[0068] S1: Obtain the status information of all distributed nodes in the cross-border e-commerce platform;

[0069] S2: Detecting whether the status information of each of the distributed nodes has reached a first preset state, and marking the distributed nodes that have reached the first preset state as idle nodes;

[0070] S3: Obtain a training task and a compliance prediction model corresponding to the training task, and combine multiple idle nodes into a virtual computing resource pool through a load balancing algorithm based on the training task;

[0071] S4: Performing distributed training of sub-models based on the virtual computing resource pool;

[0072] S5: collecting the gradient information of each distributed node in the virtual computing resource pool, aggregating it through an aggregation algorithm, and updating the gradient of the compliance prediction model to obtain a target compliance prediction model;

[0073] S6: Obtain the cross-border business to be analyzed, and predict the compliance of the cross-border business to be analyzed based on the target compliance prediction model.

[0074] As described in step S1 above, the status information of all distributed nodes in the cross-border e-commerce platform is obtained. The status information of all distributed nodes of the cross-border e-commerce platform is collected and monitored in real time. Distributed nodes can be different servers, computing units, or edge devices. To obtain accurate status information, a monitoring component can be deployed, such as a custom Exporter component based on Prometheus, to collect key performance indicators of each node every 500 milliseconds. These indicators specifically include CPU and GPU utilization, memory usage, network bandwidth and latency, storage usage, etc. By collecting this information in real time, the platform can fully understand resource usage and identify performance bottlenecks or potential idle resources. This comprehensive status monitoring is the basis for subsequent steps, ensuring that subsequent decisions have a reliable basis in terms of accuracy and timeliness. For example, if the GPU utilization of a node remains at a low level for a long time, it means that the node is not fully utilized under the current load, providing data support for subsequently marking it as an idle node.

[0075] As described in step S2 above, detect whether the status information of each of the distributed nodes has reached a first preset state, and mark the distributed nodes that have reached the first preset state as idle nodes. Analyze the collected status information to determine whether the information meets the set first preset state. The first preset state may include a series of conditions, such as GPU utilization is lower than a certain threshold (such as 20%), CPU utilization is lower than a certain threshold (such as 15%), etc. If some distributed nodes meet these conditions, they will be marked as idle nodes. During high-pressure workloads, quickly identifying and reallocating idle nodes can ensure stability and responsiveness.

[0076] As described in step S3 above, a training task and a compliance prediction model corresponding to the training task are obtained, and based on the training task, multiple idle nodes are combined into a virtual computing resource pool through a load balancing algorithm. The training task to be executed and its corresponding compliance prediction model are obtained. Specifically, the compliance prediction model can be stored in the cloud. By applying the load balancing algorithm, the idle nodes marked in the previous step are combined to form a virtual computing resource pool. The choice of load balancing algorithm may depend on the performance of the resources (such as the computing power and memory of the GPU) and the characteristics of the task to be executed (such as the priority of the task, the expected computing amount, etc.). The goal is to reasonably allocate computing resources to ensure efficient execution of tasks. In this process, the creation of a virtual computing resource pool not only improves resource utilization efficiency, but also provides the possibility for concurrent task processing. The formed virtual pool is elastic and flexible, and can be dynamically adjusted according to the current system status to adapt to uncertain and volatile data processing needs.

[0077] As described in step S4 above, distributed training of the sub-model is performed based on the virtual computing resource pool. The created virtual computing resource pool is used to perform distributed training of the sub-model. This process allows multiple idle nodes to work in parallel to jointly train the sub-model of the task. The key to distributed training is that each idle node obtains the corresponding training data and cloud model, and trains the cloud model to obtain training parameters. During the training process, these nodes will iteratively update the model independently, and each node calculates its own gradient information and model updates. In this mode, the training time can be significantly shortened and the efficiency of parallel computing can be improved. Due to the characteristics of the cross-border e-commerce platform, this process can provide accurate model updates based on specific compliance requirements and business scenarios to ensure that the generated compliance prediction model can adapt to various complex actual application situations.

[0078] As described in step S5 above, the sub-gradient information from each distributed node in the virtual computing resource pool is collected and aggregated using an aggregation algorithm to update the gradient of the overall compliance prediction model, resulting in the target overall compliance prediction model. The sub-gradient information previously calculated at each node is collected. After completing sub-model training, each idle node transmits the corresponding gradient calculated back to the central node. These sub-gradients are aggregated using an aggregation algorithm (such as averaging or weighted averaging) to generate new model weights and gradients to update the overall compliance prediction model. The aggregation method is determined based on the specific task and model characteristics to ensure accuracy and effectiveness during the integration process. By integrating this information, the target overall compliance prediction model is updated to reflect the learning results from each training node. This process enables the model to achieve higher accuracy and reliability when handling compliance prediction tasks. The speed and efficiency of model updates directly impact the subsequent real-time prediction performance, ensuring the platform can quickly adapt to the ever-changing compliance environment.

[0079] As described in step S6 above, the cross-border business to be analyzed is obtained, and the compliance of the cross-border business to be analyzed is predicted based on the target compliance prediction model. The compliance of the cross-border business to be analyzed is evaluated using the updated target compliance prediction model. The business to be analyzed refers to the data processing, storage, transmission and other operations involved in cross-border e-commerce transactions. When using the target model for prediction, the system will input the relevant features of the business to be analyzed into the account model, and after the model calculation, it will output whether the business meets the compliance requirements. This process can not only effectively judge the compliance level of the business and avoid potential legal risks, but also help enterprises to dynamically monitor and evaluate compliance, thereby optimizing the overall compliance strategy. Through the implementation of this series of steps, cross-border e-commerce platforms can quickly respond to compliance challenges and maintain brand image and market competitiveness.

[0080] In one embodiment, after step S3 of obtaining a training task and a compliance prediction overall model corresponding to the training task, and combining multiple idle nodes into a virtual computing resource pool through a load balancing algorithm based on the training task, the process further includes:

[0081] S401: Real-time monitoring of the real-time status information of all distributed nodes;

[0082] S402: Determine whether the real-time status information of the idle node has reached a second preset state, and determine whether the real-time status information of the non-idle nodes in all distributed nodes has reached a third preset state;

[0083] S403: Deleting the idle nodes whose real-time status information reaches the second preset state from the virtual computing pool, and adding the non-idle nodes whose real-time status information reaches the third preset state to the virtual computing pool.

[0084] As described in steps S401-S403 above, the real-time status information of all distributed nodes needs to be continuously monitored in order to quickly respond to different computing needs and status changes in a dynamic environment. Implementing this monitoring effectively helps the platform understand the current performance of each node. This can be achieved by deploying monitoring tools such as Prometheus, combined with a custom Exporter component, which can collect key performance data every 500 milliseconds. The monitored information includes CPU and GPU utilization, memory usage, network bandwidth, latency, etc. to ensure that the status of the node is accurately captured. In this way, the system can promptly identify which nodes have become idle and which nodes are under high load, and can even monitor network latency and bandwidth, occupancy, and ensure the effective allocation of resources. The results of real-time monitoring provide basic support for subsequent analysis, enabling the system to make more scientific scheduling decisions and ensure the optimal use efficiency of computing resources under different time periods and load conditions.

[0085] Based on previously collected real-time status information, a detailed status assessment is performed on idle and non-idle nodes. A second, pre-set status is defined, encompassing specific performance metrics such as memory utilization and processing capacity. Idle nodes are considered available only when their status meets these thresholds. Specifically, if an idle node's CPU load remains below 30% for five minutes, it is considered non-idle and removed from the virtual computing pool. Simultaneously, the system checks whether all non-idle nodes meet a third, pre-set status, such as below a load threshold or experiencing low response time. Only then can these nodes be considered for inclusion in the virtual computing resource pool. Specifically, a CPU load exceeding 75% for one minute or a P99 latency exceeding 300ms can be set to identify them as idle and allow them to join the virtual computing pool. This ensures timely reflection of node load capacity and efficient dynamic adjustment of the virtual resource pool. By accurately assessing status, the platform can balance resource utilization with task demands, ensuring efficient resource allocation and timely response.

[0086] In one embodiment, before step S5 of collecting the gradient information of each idle node in the virtual computing resource pool, aggregating it through an aggregation algorithm and updating the gradient of the compliance prediction model to obtain the target compliance prediction model, the process further includes:

[0087] S411: When a model update instruction of each idle node is triggered, each idle node obtains a plurality of cross-border business training samples in the region where the idle node is located;

[0088] S412: Training the respective pre-stored compliance prediction sub-models based on the cross-border business training samples to obtain sub-model gradients;

[0089] S413: Generate corresponding sub-gradient information based on the sub-model gradient, and upload it to the distributed node where the compliance prediction overall model is located.

[0090] As described in steps S411-S413 above, when an idle node detects a model update instruction, the node automatically retrieves multiple cross-border business training samples from its region. Cross-border e-commerce business training samples typically contain information related to legal compliance, user behavior, transaction data, and so on. Due to restrictions on the cross-border circulation of model training data, each distributed node retrieves cross-border business training samples from its region for training. This can be done from a database or through communication with external data sources to ensure that the collected samples are representative and diverse, allowing the model to better adapt to the ever-changing business environment. Each idle node uses the obtained cross-border business training samples to train its pre-stored compliance prediction sub-model. Specifically, utilizing supervised learning strategies in machine learning, the training process relies not only on input features but also on labeled data (such as compliance classification) to adjust model weights. After each training session, the node calculates the corresponding sub-model gradients. These gradients represent the direction and magnitude of the change in model parameters during the current iteration. The resulting sub-model gradients reflect the model's current learning efficiency and performance, guiding the model to continuously optimize its predictive capabilities through feedback. Each idle node generates corresponding sub-gradient information based on the sub-model gradients calculated by its trained compliance prediction sub-model. This process not only involves organizing and formatting the gradient data to ensure its effective utilization by the subsequent aggregation algorithm, but also includes appropriate compression and encoding to reduce the amount and time of data transmission and improve network efficiency. The uploaded sub-gradient information is aggregated into the overall model, forming a complete update basis and providing the raw materials for the subsequent gradient aggregation process. By collecting sub-gradients, learning results from different nodes can be continuously obtained, thereby effectively optimizing and updating the overall compliance prediction model, thereby achieving more efficient management and compliance testing for cross-border e-commerce.

[0091] In one embodiment, the step S413 of generating corresponding sub-gradient information based on the sub-model gradient and uploading it to the distributed node where the compliance prediction overall model is located further includes:

[0092] S4131: applying Laplace noise to the sub-model gradient to obtain a temporary sub-model gradient;

[0093] S4132: Filter each temporary sub-model gradient according to a preset gradient parameter filtering mechanism to obtain a target temporary sub-model gradient;

[0094] S4133: Generate corresponding sub-gradient information based on the target temporary sub-model gradient, and upload it to the distributed node where the compliance prediction overall model is located.

[0095] As described in steps S4131-S4133 above, each idle node applies Laplace noise to the sub-model gradients obtained during its training process. This process ensures data privacy and security, particularly in scenarios such as cross-border e-commerce that require compliance with data compliance requirements. The introduction of Laplace noise aims to protect sensitive data involved in model training and prevent potential information leakage. Specifically, based on a set privacy budget and noise scale, a Laplace-distributed uncertainty is added to each gradient value. This noise modifies the original precise gradient update, forming a "temporary sub-model gradient." Specifically, ΔW' = ΔW + Lap(0, σ = 0.8 / ε), where ε is the privacy budget (default ε = 1.2), ΔW' is the temporary sub-model gradient, and ΔW is the sub-model gradient. Laplace noise is typically applied by applying random perturbations to each parameter, ensuring that the final gradient still reflects the model's learning direction, but with a certain degree of loss of sensitivity to individual data points. This effectively prevents excessive personal information from being included in the model output, protecting user privacy. Each temporary sub-model gradient is processed according to a preset gradient parameter filtering mechanism. This filtering mechanism is designed to remove data points with relatively high noise interference, thereby obtaining more accurate and reliable target temporary sub-model gradients. The preset gradient parameter filtering mechanism can be set based on the results of historical data analysis, the distribution characteristics of gradients, and the expected model performance. For example, parameters with |ΔW| greater than 0.05 are retained. For directional filtering, local gradients with an angle greater than 60° with the global average direction can be discarded. Top-K filtering (K = 20% of the total number of parameters) can also be performed on the retained gradients. Specifically, a custom Hook function is inserted into the PyTorch framework to perform filtering during the backward() phase.

[0096] Based on the filtered target temporary sub-model gradients, corresponding sub-gradient information is generated. This process involves formatting and encoding the target temporary sub-model gradients as necessary to enable effective integration with the overall compliance prediction model. During the upload process, the system uses encryption and compression to improve data transmission efficiency and security. The uploaded sub-gradient information not only contains the gradients after Laplace noise processing and filtering, but may also include other metadata, such as node identifiers and update timestamps, to facilitate management and tracking of this information. This upload step is crucial, as it not only transfers the learning results of each node to the overall compliance prediction model but also provides a foundation for integrating learning results across different nodes and further aggregating and optimizing the model. Specifically, secure transmission can be achieved by establishing a TLS1.3 encrypted channel using the SM2 algorithm. Access control can be achieved through tenant isolation using JWT tokens containing the tenant ID (SHA256 hash value) and data permission tags (e.g., {region: EU, type: financial}). Ultimately, the processed sub-gradient information provides an accurate basis for model updates and revisions, improving the accuracy and reliability of the overall compliance prediction model and effectively supporting compliance risk management in cross-border e-commerce businesses.

[0097] In one embodiment, before the step S412 of training the respective pre-stored compliance prediction sub-models based on the cross-border business training samples to obtain the sub-model gradients, the following steps are further included:

[0098] S4111: Each of the idle nodes downloads the teacher model from the cloud respectively;

[0099] S4112: Extracting the attention matrix of the multi-layer Transformer layer in the teacher model;

[0100] S4113: Select a preset number of attention matrices from the multi-layer Transformer layer and add a preset loss function to obtain a student model as the compliance prediction sub-model.

[0101] As described in steps S4111-S4113 above, each idle node downloads a teacher model from the cloud. The teacher model is typically a pre-trained deep learning model with high complexity and accuracy, capable of processing rich features and complex data patterns. In the cross-border e-commerce scenario, this teacher model can be a compliance prediction model trained on a large amount of data, effectively capturing key compliance-related characteristics. In one embodiment, the teacher model can be the aforementioned overall compliance prediction model. During the download process, communication security and data integrity must be ensured to prevent data tampering or leakage during transmission. Data transmission is typically performed via the HTTPS protocol, with checksums and other methods used to ensure file integrity. In addition, each idle node may select the teacher model version that best suits its workload based on its own computing resources, storage capacity, and other factors, to achieve optimal resource allocation. The purpose of downloading the teacher model is to prepare a basic framework for subsequent model training, enabling nodes to inherit the teacher model's knowledge and feature learning capabilities when training their respective compliance prediction sub-models, thereby improving the training effectiveness and accuracy of the final student model. The attention matrix of the multi-layer Transformer layer is extracted from the downloaded teacher model. The Transformer model utilizes a self-attention mechanism, enabling each input feature to establish dynamic weighted relationships with other features. The attention matrix directly reflects these relationships. It records how different input features influence each other within the model and their importance in the decision-making process. The knowledge distillation process helps the student model achieve good performance while using a relatively small number of parameters. When extracting the attention matrix, the system ensures that the extracted layers and parameter configuration meet the requirements of the subsequent model. Furthermore, the extracted multiple attention matrices provide a foundation for building the student model, enabling the subsequent compliance prediction model to better understand the feature distribution and its internal relationships during its learning process. This process ultimately results in a student model with higher initial performance and efficiency compared to training from scratch. A preset number of matrices are selected from the extracted attention matrices of multiple Transformer layers and combined to construct the student model. The selected attention matrices are typically chosen based on their relevance or importance in the output, ensuring that the constructed student model effectively learns the key features and knowledge of the teacher model. Furthermore, when constructing the student model, the system designs a preset loss function to provide feedback during model training, enabling the model to continuously adjust parameters to reduce prediction error. Common loss functions include cross entropy loss (for classification tasks) or mean squared error (for regression tasks). These loss functions can be used to evaluate the learning ability and effect of the model during training.The obtained student model will serve as a compliance prediction sub-model, inheriting the characteristics and knowledge of the teacher model, while having higher computational efficiency and adaptability due to optimization in parameters and structure. This distillation process enables the student model to maintain high accuracy and performance in predicting cross-border business compliance within the region while reducing the demand for computing resources, making it an efficient model construction strategy.

[0102] In one embodiment, the step S5 of collecting the gradient information of each distributed node in the virtual computing resource pool, aggregating and updating the gradient of the compliance prediction general model through an aggregation algorithm to obtain the target compliance prediction general model, includes:

[0103] S501: Obtain the total gradient information corresponding to the compliance prediction general model;

[0104] S502: Calculate the difference parameters of the total gradient information and the gradient information;

[0105] S503: Based on the difference parameters, update the compliance prediction general model through a preset data synchronization algorithm to obtain a temporary compliance prediction general model;

[0106] S504: Detect the performance of the temporary compliance prediction general model;

[0107] S505: If the performance meets the preset requirement, record the temporary compliance prediction general model as the target compliance prediction general model.

[0108] As described in steps S501-S505 above, the corresponding total gradient information is first extracted from the overall compliance prediction model. Total gradient information is the gradient required for parameter updates, calculated via the backpropagation algorithm during model training. It reflects the model's sensitivity to various parameters in its current state. This total gradient information indicates the direction and magnitude of adjustments to various parameters (such as weights and biases) to improve the model's predictive performance. The process of extracting this total gradient information typically depends on the performance of the current overall compliance prediction model on the training dataset. The total gradients the system needs to process may contain information on multiple parameters, involving complex model structures such as all connections in a multi-layer neural network. By effectively extracting this information, the current training status of the model can be accurately assessed, providing the necessary foundational data for subsequent updates. The difference parameter between the total gradient information and the partial gradient information collected in the previous step is calculated. These two gradients are compared to identify the specific parameters that need to be adjusted during the model update process. The total gradient information represents the desired parameter updates for the current overall compliance prediction model, while the partial gradient information represents the local training results from each idle node. By calculating differential parameters, we can identify the specific update effects generated during distributed training at each node and compare these effects with the requirements of the overall model. This comparison effectively reveals differences in learning information across different layers and regions, providing a more precise basis for parameter updates. Furthermore, the calculation of differential parameters can be viewed as an aggregation of learning outcomes across nodes, helping to adjust parameters that demonstrate strong learning capabilities at certain nodes. Effectively calculating differential parameters ensures that updates to the overall compliance prediction model incorporate the learning experiences of different nodes, enabling better adaptation to the complex cross-border business environment. Based on the calculated differential parameters, the overall compliance prediction model is updated using a pre-set data synchronization algorithm. This step is key to model adjustment and optimization. The derived differential parameters are applied to the current overall model to generate a new, temporary overall compliance prediction model. It should be noted that the "data synchronization algorithm" here may involve a variety of algorithmic strategies, including but not limited to traditional SGD (stochastic gradient descent) and the Adam optimization algorithm. This algorithm determines how and in what proportion to incorporate the differential parameters into the existing model parameters to achieve effective learning and performance improvement. This parameter update allows the overall compliance prediction model to maintain existing knowledge while integrating newly acquired learning outcomes, thereby improving overall performance. It is important to note that the update process must consider model stability. Excessive parameter updates may cause the model to fail to converge or even oscillate.

[0109] The generated interim compliance prediction model undergoes performance testing. This process is a crucial step in ensuring that the new model can meet anticipated requirements in real-world applications. Performance testing typically involves considering multiple evaluation metrics, such as accuracy, precision, and recall. These metrics help the system understand the model's performance on a given test set. A pre-set validation dataset may be used during performance testing to ensure fairness and reliability. Evaluating the interim model provides insights into its potential performance in real-world applications and identifies potential issues and deficiencies. For example, if the model's prediction accuracy in a specific scenario falls below the target, the system needs to analyze the cause, potentially due to insufficient model learning on specific data types. The results of the performance testing directly influence subsequent decisions. If the interim compliance prediction model performs well, demonstrating high accuracy and good adaptability, it will proceed to the next stage of verification. Conversely, if performance falls short of expectations, further model adjustments or a review of the update process may be necessary to ensure that the final model is sufficient to meet cross-border business compliance requirements. This testing process is crucial to ensuring continuous model improvement and accurately capturing business changes. The test results are compared against pre-defined criteria, including thresholds for key performance indicators (such as precision, recall, and F1-score). If the performance meets the pre-set requirements, the system officially marks the provisional compliance prediction model as the target compliance prediction model. This transition signifies the success of the model optimization process and its readiness for implementation in real-world scenarios such as cross-border e-commerce.

[0110] In one embodiment, the step S3 of combining multiple idle nodes into a virtual computing resource pool based on the training task through a load balancing algorithm includes:

[0111] S301: Mapping the training task into a multi-dimensional feature vector;

[0112] S302: Generate a decision path using the CART algorithm based on the multidimensional feature vector;

[0113] S303: Acquire multiple idle nodes based on the decision path to combine them into a virtual computing resource pool.

[0114] As described in steps S301-S303 above, the training task is first mapped into a multidimensional feature vector. A training task typically has multiple attributes. In a specific embodiment, the multidimensional feature vector is a three-dimensional vector (priority P∈[1-3], latency requirement D∈[100-500ms], cost level C∈[1-5]). The resulting multidimensional feature vector is then used to generate a decision path using the CART (Classification and Regression Trees) algorithm. CART is a machine learning algorithm used to construct decision trees, which predict outcomes based on input feature vectors. In this step, the CART algorithm analyzes the feature vectors to construct a decision tree. The branching of the tree effectively identifies the appropriate computing resource allocation strategy. Each decision branch reflects the pros and cons of selecting a resource under specific conditions. This tree structure offers the advantage of visualizing complex decision-making processes, providing clear guidance for resource allocation. The resulting decision path is a tree structure consisting of nodes and edges, indicating how to select and utilize idle nodes under different feature conditions, thereby achieving efficient load balancing. Guided by the decision path, idle nodes that meet the training task requirements are selected. Selected nodes generally possess sufficient computing power, storage capacity, and other performance indicators to support the upcoming training task. By rationally combining these idle nodes, the system forms a flexible and dynamic computing resource pool. This virtual resource pool can be rapidly allocated based on task requirements, ensuring that required computing resources are immediately available when needed, thereby improving efficiency and maximizing resource utilization. This achieves effective load balancing, enhances overall system performance, and reduces resource waste.

[0115] Reference Figure 3 The present invention also provides a device for predicting cross-border business compliance, the device comprising:

[0116] Acquisition module 902, used to obtain status information of all distributed nodes in the cross-border e-commerce platform;

[0117] A detection module 904 is configured to detect whether the status information of each of the distributed nodes has reached a first preset state, and mark the distributed nodes that have reached the first preset state as idle nodes;

[0118] A combining module 906 is configured to obtain a training task and a compliance prediction model corresponding to the training task, and combine multiple idle nodes into a virtual computing resource pool based on the training task using a load balancing algorithm;

[0119] A training module 908 is configured to perform distributed training of sub-models based on the virtual computing resource pool;

[0120] The collection module 910 is used to collect the gradient information of each distributed node in the virtual computing resource pool, aggregate it through an aggregation algorithm, and update the gradient of the compliance prediction model to obtain a target compliance prediction model;

[0121] The prediction module 912 is used to obtain the cross-border business to be analyzed and predict the compliance of the cross-border business to be analyzed based on the target compliance prediction model.

[0122] In one embodiment, the cross-border business compliance prediction device further includes:

[0123] Real-time status information monitoring module, used to monitor the real-time status information of all distributed nodes in real time;

[0124] A real-time status information determination module, configured to determine whether the real-time status information of the idle node has reached a second preset state, and to determine whether the real-time status information of the non-idle nodes in all distributed nodes has reached a third preset state;

[0125] The virtual computing pool adjustment module is used to delete the idle nodes whose real-time status information reaches the second preset state from the virtual computing pool, and add non-idle nodes whose real-time status information reaches a third preset state to the virtual computing pool.

[0126] In one embodiment, the cross-border business compliance prediction device further includes:

[0127] A training sample acquisition module, configured to enable each idle node to acquire a plurality of cross-border business training samples in its region when a model update instruction of each idle node is triggered;

[0128] A compliance prediction sub-model training module, configured to train the respective pre-stored compliance prediction sub-models based on the cross-border business training samples to obtain sub-model gradients;

[0129] The sub-gradient information generation module is used to generate corresponding sub-gradient information based on the sub-model gradient and upload it to the distributed node where the compliance prediction overall model is located.

[0130] In one embodiment, the gradient information generation module further includes:

[0131] a noise adding submodule, configured to apply Laplace noise to the sub-model gradient to obtain a temporary sub-model gradient;

[0132] A filtering submodule is used to filter each temporary sub-model gradient according to a preset gradient parameter filtering mechanism to obtain a target temporary sub-model gradient;

[0133] The sub-gradient information generation sub-module is used to generate corresponding sub-gradient information based on the target temporary sub-model gradient and upload it to the distributed node where the compliance prediction overall model is located.

[0134] In one embodiment, the cross-border business compliance prediction device further includes:

[0135] A teacher model downloading module, used for each of the idle nodes to download the teacher model from the cloud;

[0136] An attention matrix extraction module, used to extract the attention matrix of the multi-layer Transformer layer in the teacher model;

[0137] The attention matrix selection module is used to select a preset number of attention matrices from the multi-layer Transformer layer and add a preset loss function to obtain a student model as the compliance prediction sub-model.

[0138] In one embodiment, the collection module 910 includes:

[0139] The total gradient information acquisition submodule is used to obtain the total gradient information corresponding to the compliance prediction total model;

[0140] A differential parameter calculation submodule, configured to calculate differential parameters between the total gradient information and the sub-gradient information;

[0141] An updating submodule, configured to update the compliance prediction overall model based on the differential parameters by using a preset data synchronization algorithm to obtain a temporary compliance prediction overall model;

[0142] A performance detection submodule, configured to detect the performance of the overall temporary compliance prediction model;

[0143] The marking submodule is used to record the temporary compliance prediction model as the target compliance prediction model if the performance reaches the preset requirement.

[0144] In one embodiment, the combining module 906 includes:

[0145] A multidimensional feature vector mapping submodule, used to map the training task into a multidimensional feature vector;

[0146] The decision path generation submodule is used to generate a decision path based on the multi-dimensional feature vector using the CART algorithm;

[0147] The idle node acquisition submodule is used to acquire multiple idle nodes based on the decision path to combine them into a virtual computing resource pool.

[0148] Figure 4FIG1 shows an internal structure diagram of a computer device in an embodiment. The computer device can be a terminal or a server. Figure 4 As shown, the computer device includes a processor, a memory, and a network interface connected via a system bus. The memory includes a non-volatile storage medium and an internal memory. The non-volatile storage medium of the computer device stores an operating system and may also store a computer program. When the computer program is executed by the processor, the processor can implement a method for predicting cross-border business compliance. The internal memory may also store a computer program. When the computer program is executed by the processor, the processor can implement a method for predicting cross-border business compliance. Those skilled in the art will understand that Figure 4 The structure shown in the figure is only a block diagram of a part of the structure related to the solution of the present application, and does not constitute a limitation on the computer device to which the solution of the present application is applied. The specific computer device may include more or fewer components than shown in the figure, or combine certain components, or have a different component arrangement.

[0149] In one embodiment, a computer device is provided, comprising a memory and a processor, wherein the memory stores a computer program, and when the computer program is executed by the processor, the processor performs the following steps:

[0150] Obtain status information of all distributed nodes in the cross-border e-commerce platform;

[0151] detecting whether the status information of each of the distributed nodes has reached a first preset state, and marking the distributed nodes that have reached the first preset state as idle nodes;

[0152] Obtaining a training task and a compliance prediction model corresponding to the training task, and combining multiple idle nodes into a virtual computing resource pool through a load balancing algorithm based on the training task;

[0153] Performing distributed training of sub-models based on the virtual computing resource pool;

[0154] Collecting the sub-gradient information of each distributed node in the virtual computing resource pool, aggregating it through an aggregation algorithm and updating the gradient of the compliance prediction overall model to obtain a target compliance prediction overall model;

[0155] Obtain the cross-border business to be analyzed, and predict the compliance of the cross-border business to be analyzed based on the target compliance prediction model.

[0156] By identifying and marking idle nodes through a dynamic resource management method, computing resource allocation can be flexibly adjusted in an environment where computing power demand fluctuates significantly, avoiding the problem of low GPU resource utilization in the traditional fixed computing power allocation model, thereby improving overall resource utilization efficiency and reducing response latency. On the other hand, it also improves the training speed and accuracy of the model, allowing cross-border e-commerce platforms to quickly adapt to market changes and improve user experience.

[0157] In one embodiment, a computer-readable storage medium is provided, storing a computer program. When the computer program is executed by a processor, the processor performs the following steps:

[0158] Obtain status information of all distributed nodes in the cross-border e-commerce platform;

[0159] detecting whether the status information of each of the distributed nodes has reached a first preset state, and marking the distributed nodes that have reached the first preset state as idle nodes;

[0160] Obtaining a training task and a compliance prediction model corresponding to the training task, and combining multiple idle nodes into a virtual computing resource pool through a load balancing algorithm based on the training task;

[0161] Performing distributed training of sub-models based on the virtual computing resource pool;

[0162] Collecting the sub-gradient information of each distributed node in the virtual computing resource pool, aggregating it through an aggregation algorithm and updating the gradient of the compliance prediction overall model to obtain a target compliance prediction overall model;

[0163] Obtain the cross-border business to be analyzed, and predict the compliance of the cross-border business to be analyzed based on the target compliance prediction model.

[0164] By identifying and marking idle nodes through a dynamic resource management method, computing resource allocation can be flexibly adjusted in an environment where computing power demand fluctuates significantly, avoiding the problem of low GPU resource utilization in the traditional fixed computing power allocation model, thereby improving overall resource utilization efficiency and reducing response latency. On the other hand, it also improves the training speed and accuracy of the model, allowing cross-border e-commerce platforms to quickly adapt to market changes and improve user experience.

[0165] Those skilled in the art will appreciate that all or part of the processes in the above-mentioned embodiments can be implemented by instructing the relevant hardware through a computer program. The program can be stored in a non-volatile computer-readable storage medium. When the program is executed, it can include the processes of the embodiments of the above-mentioned methods. Among them, any reference to memory, storage, database or other media used in the embodiments provided in this application can include non-volatile and / or volatile memory. Non-volatile memory can include read-only memory (ROM), programmable ROM (PROM), electrically programmable ROM (EPROM), electrically erasable programmable ROM (EEPROM) or flash memory. Volatile memory can include random access memory (RAM) or external cache memory. By way of illustration and not limitation, RAM is available in various forms, such as static RAM (SRAM), dynamic RAM (DRAM), synchronous DRAM (SDRAM), double data rate SDRAM (DDRSDRAM), enhanced SDRAM (ESDRAM), synchronous link (Synchlink) DRAM (SLDRAM), memory bus (Rambus) direct RAM (RDRAM), direct memory bus dynamic RAM (DRDRAM), and memory bus dynamic RAM (RDRAM).

[0166] The technical features of the above embodiments can be combined arbitrarily. To make the description concise, not all possible combinations of the technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this specification.

[0167] The above-described embodiments merely represent several implementation methods of the present application. While the descriptions are relatively specific and detailed, they should not be construed as limiting the scope of the present invention. It should be noted that a person of ordinary skill in the art may make various modifications and improvements without departing from the spirit of the present invention, all of which fall within the scope of protection of the present invention. Therefore, the scope of protection of the present invention shall be determined by the appended claims.

Claims

1. A method for predicting cross-border business compliance, characterized in that: The method comprises: Obtain status information of all distributed nodes in the cross-border e-commerce platform; detecting whether the status information of each of the distributed nodes has reached a first preset state, and marking the distributed nodes that have reached the first preset state as idle nodes; Obtaining a training task and a compliance prediction model corresponding to the training task, and combining multiple idle nodes into a virtual computing resource pool through a load balancing algorithm based on the training task; Distributed training of the sub-model is performed based on the virtual computing resource pool; wherein each distributed node independently performs iterative updates, gradient information calculations, and model updates on the sub-model; Collect the sub-gradient information of each distributed node in the virtual computing resource pool, summarize the sub-gradient information of each distributed node through an aggregation algorithm and update the gradient of the compliance prediction overall model to obtain a target compliance prediction overall model; specifically: obtain the total gradient information corresponding to the compliance prediction overall model; calculate the differential parameter between the total gradient information and the sub-gradient information; based on the differential parameter, update the compliance prediction overall model through a preset data synchronization algorithm to obtain a temporary compliance prediction overall model; test the performance of the temporary compliance prediction overall model; if the performance meets the preset requirements, record the temporary compliance prediction overall model as the target compliance prediction overall model; Obtaining cross-border business to be analyzed, and predicting the compliance of the cross-border business to be analyzed based on the target compliance prediction model; After the step of combining multiple idle nodes into a virtual computing resource pool through a load balancing algorithm based on the training task, the method further includes: Real-time monitoring of the real-time status information of all distributed nodes; Determining whether the real-time status information of the idle node has reached a second preset state, and determining whether the real-time status information of the non-idle nodes in all distributed nodes has reached a third preset state; Deleting the idle nodes whose real-time status information reaches the second preset state from the virtual computing resource pool, and adding the non-idle nodes whose real-time status information reaches the third preset state to the virtual computing resource pool; The step of combining multiple idle nodes into a virtual computing resource pool through a load balancing algorithm based on the training task includes: Mapping the training task into a multidimensional feature vector; wherein the multidimensional feature vector is specifically a three-dimensional vector including three dimensional features of priority, latency requirement, and cost level; Based on the multi-dimensional feature vector, the decision path is generated by the CART algorithm; A plurality of idle nodes are obtained based on the decision path to form a virtual computing resource pool.

2. The method for predicting cross-border business compliance according to claim 1, characterized in that: Before the step of collecting the sub-gradient information of each distributed node in the virtual computing resource pool, aggregating the sub-gradient information of each distributed node through an aggregation algorithm and updating the gradient of the compliance prediction overall model to obtain the target compliance prediction overall model, the method further includes: When a model update instruction of each of the idle nodes is triggered, each of the idle nodes obtains a plurality of cross-border business training samples in the region where the idle node is located; Training the respective pre-stored compliance prediction sub-models based on the cross-border business training samples to obtain sub-model gradients; Corresponding sub-gradient information is generated based on the sub-model gradient and uploaded to the distributed node where the compliance prediction overall model is located.

3. The method for predicting cross-border business compliance according to claim 2, characterized in that: The step of generating corresponding sub-gradient information based on the sub-model gradient and uploading it to the distributed node where the compliance prediction overall model is located further includes: Applying Laplace noise to the sub-model gradient to obtain a temporary sub-model gradient; Filter each temporary sub-model gradient according to the preset gradient parameter filtering mechanism to obtain the target temporary sub-model gradient; Based on the target temporary sub-model gradient, corresponding sub-gradient information is generated and uploaded to the distributed node where the compliance prediction overall model is located.

4. The method for predicting cross-border business compliance according to claim 2, characterized in that: Before the step of training the respective pre-stored compliance prediction sub-models based on the cross-border business training samples to obtain the sub-model gradients, the step further includes: Each of the idle nodes downloads the teacher model from the cloud respectively; Extract the attention matrix of the multi-layer Transformer layer in the teacher model; A preset number of attention matrices are selected from the multi-layer Transformer layer, and a preset loss function is added to obtain a student model to serve as the compliance prediction sub-model.

5. A device for predicting compliance of cross-border business, characterized in that: The device comprises: The acquisition module is used to obtain the status information of all distributed nodes in the cross-border e-commerce platform; a detection module, configured to detect whether the status information of each of the distributed nodes has reached a first preset state, and mark the distributed nodes that have reached the first preset state as idle nodes; A combination module is used to obtain a training task and a compliance prediction model corresponding to the training task, and combine multiple idle nodes into a virtual computing resource pool through a load balancing algorithm based on the training task; A training module, configured to perform distributed training of the sub-model based on the virtual computing resource pool; wherein each distributed node independently performs iterative updates, gradient information calculations, and model updates on the sub-model; A collection module is used to collect the sub-gradient information of each distributed node in the virtual computing resource pool, aggregate the sub-gradient information of each distributed node through an aggregation algorithm, and update the gradient of the compliance prediction overall model to obtain a target compliance prediction overall model; specifically: obtain the total gradient information corresponding to the compliance prediction overall model; calculate the differential parameter between the total gradient information and the sub-gradient information; based on the differential parameter, update the compliance prediction overall model through a preset data synchronization algorithm to obtain a temporary compliance prediction overall model; test the performance of the temporary compliance prediction overall model; if the performance meets the preset requirements, record the temporary compliance prediction overall model as the target compliance prediction overall model; A prediction module, configured to obtain cross-border business to be analyzed and predict the compliance of the cross-border business to be analyzed based on the target compliance prediction model; The device further comprises: Real-time status information monitoring module, used to monitor the real-time status information of all distributed nodes in real time; A real-time status information determination module, configured to determine whether the real-time status information of the idle node has reached a second preset state, and to determine whether the real-time status information of the non-idle nodes in all distributed nodes has reached a third preset state; a virtual computing resource pool adjustment module, configured to delete the idle nodes whose real-time status information reaches the second preset state from the virtual computing resource pool, and add non-idle nodes whose real-time status information reaches a third preset state to the virtual computing resource pool; The combined module comprises: A multidimensional feature vector mapping submodule is used to map the training task into a multidimensional feature vector; wherein the multidimensional feature vector is specifically a three-dimensional vector including three dimensional features of priority, latency requirement, and cost level; The decision path generation submodule is used to generate a decision path based on the multi-dimensional feature vector using the CART algorithm; The idle node acquisition submodule is used to acquire multiple idle nodes based on the decision path to combine them into a virtual computing resource pool.

6. A computer-readable storage medium, characterized in that A computer program is stored, and when the computer program is executed by a processor, the processor is caused to perform the steps of the method for predicting cross-border business compliance as claimed in any one of claims 1 to 4.

7. A computer device, characterized in that: The device includes a memory and a processor, wherein the memory stores a computer program, and when the computer program is executed by the processor, the processor executes the steps of the method for predicting cross-border business compliance as claimed in any one of claims 1 to 4.

Images