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

Abstract

The invention relates to the technical field of artificial intelligence, and discloses a cross-border business compliance prediction method and device, computer equipment and a medium, and the method comprises the steps: obtaining the state information of all distributed nodes in a cross-border e-commerce platform, so as to determine idle nodes, and combining the idle nodes into a virtual computing resource pool; performing distributed training of sub-models based on the virtual computing resource pool; and collecting sub-gradient information of each distributed node, and updating the gradient of the compliance prediction total model to obtain a target compliance prediction total model to predict the compliance of the cross-border service to be analyzed. The method has the beneficial effects that the overall resource utilization efficiency is improved, the response delay is reduced, and on the other hand, the training speed and precision of the model are also improved, so that a cross-border e-commerce platform can quickly adapt to market changes, and the user experience is improved.

Description

Technical Field

[0001] The present invention relates to the technical field of predicting cross-border business compliance, and particularly to a method, device, computer device and medium for predicting cross-border business compliance. Background Art

[0002] With the rapid development of global trade, cross-border e-commerce business has grown rapidly in various regions. However, due to time zone differences and sudden traffic, the computing power demand of cross-border business fluctuates significantly. During peak hours, the average utilization rate of GPU resources in the traditional fixed computing power allocation mode is less than 40%. At the same time, the latency of business response increases significantly, especially during peak periods, and the response latency 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 changing computing power demand has become an important challenge for the development of cross-border business.

[0003] In addition, data sovereignty regulations restrict the cross-border circulation of model training data, which poses challenges to the existing centralized training mode. Summary of the Invention

[0004] Based on this, it is necessary to propose a method, device, computer device and medium for predicting cross-border business compliance for the existing problem of predicting cross-border business compliance.

[0005] A method for predicting cross-border business compliance, the method comprising: Obtain the status information of all distributed nodes in a cross-border e-commerce platform; Detect whether the status information of each of the distributed nodes reaches a first preset state, and mark the distributed nodes that reach the first preset state as idle nodes; Obtain a training task and a total compliance prediction model corresponding to the training task, and combine a plurality of idle nodes into a virtual computing resource pool based on the training task through a load balancing algorithm; Perform distributed training of sub-models based on the virtual computing resource pool; Collect the partial gradient information of each distributed node in the virtual computing resource pool, aggregate and update the gradient of the total compliance prediction model through an aggregation algorithm to obtain a target total compliance prediction model; Obtain a cross-border business to be analyzed, and predict the compliance of the cross-border business to be analyzed based on the target total compliance prediction model.

[0006] Further, after the step of obtaining a training task and a total compliance prediction model corresponding to the training task, and combining a plurality of idle nodes into a virtual computing resource pool based on the training task through a load balancing algorithm, the method further comprises: Monitor the real-time status information of all distributed nodes in real time; Judge whether the real-time status information of the idle nodes reaches a second preset state, and judge whether the real-time status information of the non-idle nodes among all distributed nodes reaches a third preset state; Delete the idle nodes whose real-time status information reaches the second preset state from the virtual computing resource pool, and add the non-idle nodes whose real-time status information reaches the third preset state to the virtual computing resource pool.

[0007] Further, before the step of collecting the gradient information of each of the idle nodes in the virtual computing resource pool, summarizing and updating the gradient of the overall compliance prediction model through an aggregation algorithm to obtain a target overall compliance prediction model, further includes: When the 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 it is located; Train the pre-stored compliance prediction sub-model of each based on the cross-border business training samples to obtain a sub-model gradient; Generate corresponding gradient information based on the sub-model gradient and upload it to the distributed node where the overall compliance prediction model is located.

[0008] Further, the step of generating corresponding gradient information based on the sub-model gradient and uploading it to the distributed node where the overall compliance prediction model is located further includes: Apply Laplace noise to the sub-model gradient to obtain a temporary sub-model gradient; Filter each temporary sub-model gradient according to a preset gradient parameter filtering mechanism to obtain a target temporary sub-model gradient; Generate corresponding gradient information based on the target temporary sub-model gradient and upload it to the distributed node where the overall compliance prediction model is located.

[0009] Further, before the step of training the pre-stored compliance prediction sub-model of each based on the cross-border business training samples to obtain a sub-model gradient, further includes: Each of the idle nodes downloads a teacher model from the cloud respectively; Extract the attention matrix of multiple Transformer layers in the teacher model; Select a preset number of attention matrices from the multiple Transformer layers and add a preset loss function to obtain a student model as the compliance prediction sub-model.

[0010] Further, the step of collecting the sub-gradient information of each distributed node in the virtual computing resource pool, summarizing it through an aggregation algorithm, and updating the gradient of the overall compliance prediction model to obtain the target overall compliance prediction model includes: Obtain the overall gradient information corresponding to the overall compliance prediction model; Calculate the difference parameter between the overall gradient information and the sub-gradient information; Based on the difference parameter, update the overall compliance prediction model through a preset data synchronization algorithm to obtain a temporary overall compliance prediction model; Detect the performance of the temporary overall compliance prediction model; If the performance meets the preset requirements, record the temporary overall compliance prediction model as the target overall compliance prediction model.

[0011] Further, the step of combining multiple idle nodes into a virtual computing resource pool through a load balancing algorithm based on the training task includes: Map the training task to a multi-dimensional feature vector; Based on the multi-dimensional feature vector, generate a decision path through the CART algorithm; Based on the decision path, obtain multiple idle nodes to form a virtual computing resource pool.

[0012] A prediction device for cross-border business compliance, the device includes: An acquisition module, configured to acquire 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 distributed node reaches a first preset status, and mark the distributed nodes that reach the first preset status as idle nodes; A combination module, configured to acquire a training task and the overall 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 sub-models based on the virtual computing resource pool; A collection module, configured to collect the sub-gradient information of each distributed node in the virtual computing resource pool, summarize it through an aggregation algorithm, and update the gradient of the overall compliance prediction model to obtain a target overall compliance prediction model; A prediction module, configured to acquire a cross-border business to be analyzed, and predict the compliance of the cross-border business to be analyzed based on the target overall compliance prediction model.

[0013] A computer device, comprising a memory and a processor, the memory stores a computer program, when the computer program is executed by the processor, the processor is caused to perform the following steps: Obtain the status information of all distributed nodes in the cross-border e-commerce platform; Detect whether the status information of each of the distributed nodes reaches a first preset state, and mark the distributed nodes that reach the first preset state as idle nodes; Obtain a training task and the corresponding overall compliance prediction model for the training task, and combine a plurality of idle nodes into a virtual computing resource pool based on the training task through a load balancing algorithm; Perform distributed training of sub-models based on the virtual computing resource pool; Collect the sub-gradient information of each distributed node in the virtual computing resource pool, aggregate and update the gradient of the overall compliance prediction model through an aggregation algorithm to obtain a target overall compliance prediction model; Obtain a cross-border business to be analyzed, and predict the compliance of the cross-border business to be analyzed based on the target overall compliance prediction model.

[0014] A computer-readable storage medium stores a computer program, when the computer program is executed by a processor, the processor is caused to perform the following steps: Obtain the status information of all distributed nodes in the cross-border e-commerce platform; Detect whether the status information of each of the distributed nodes reaches a first preset state, and mark the distributed nodes that reach the first preset state as idle nodes; Obtain a training task and the corresponding overall compliance prediction model for the training task, and combine a plurality of idle nodes into a virtual computing resource pool based on the training task through a load balancing algorithm; Perform distributed training of sub-models based on the virtual computing resource pool; Collect the sub-gradient information of each distributed node in the virtual computing resource pool, aggregate and update the gradient of the overall compliance prediction model through an aggregation algorithm to obtain a target overall compliance prediction model; Obtain a cross-border business to be analyzed, and predict the compliance of the cross-border business to be analyzed based on the target overall compliance prediction model.

[0015] Advantages of the present invention: Through a dynamic resource management method that identifies and marks idle nodes, it is possible to flexibly adjust the computing resource allocation in an environment with significant fluctuations in computing power requirements, avoiding the problem of low utilization rate of GPU resources in the traditional fixed computing power allocation mode. Thus, the overall resource utilization efficiency is improved, the response latency is reduced. On the other hand, the training speed and accuracy of the model are also enhanced, enabling the cross-border e-commerce platform to quickly adapt to market changes and improve the user experience. BRIEF DESCRIPTION OF THE DRAWINGS

[0016] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the following will briefly introduce the drawings required for use in the description of the embodiments or the prior art. Obviously, the drawings in the following description are only some embodiments of the present invention. For those of ordinary skill in the art, without creative efforts, other drawings can be obtained based on these drawings.

[0017] Among them: Figure 1 It is an application environment diagram of a prediction method for cross-border business compliance in an embodiment; Figure 2 It is a flowchart of a prediction method for cross-border business compliance in an embodiment; Figure 3 It is a structural block diagram of a prediction device for cross-border business compliance in an embodiment; Figure 4 It is a structural block diagram of a computer device in an embodiment. DETAILED DESCRIPTION OF THE EMBODIMENTS

[0018] The following will clearly and completely describe the technical solutions in the embodiments of the present invention with reference to the drawings in the embodiments of the present invention. Obviously, the described embodiments are only some embodiments of the present invention, rather than all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those of ordinary skill in the art without creative efforts belong to the scope of protection of the present invention.

[0019] Figure 1 It is an application environment diagram of a prediction for cross-border business compliance in an embodiment. Referring to Figure 1 , the prediction method for cross-border business compliance is applied to a prediction system for cross-border business compliance. The prediction system for cross-border business compliance includes a terminal 110 and a server 120. The terminal 110 and the server 120 are connected through a network. The terminal 110 can specifically be a desktop terminal or a mobile terminal, and the mobile terminal can specifically be at least one of a mobile phone, a tablet computer, a laptop computer, etc. The server 120 can be realized by an independent server or a server cluster composed of multiple servers. The terminal 110 is used to collect status information, and the server 120 is used for compliance prediction.

[0020] As Figure 2 shown, in one embodiment, a method for predicting cross-border business compliance is provided. This method can be applied to both terminals and servers. In this embodiment, an example of applying it to a server is given. The method for predicting cross-border business compliance specifically includes the following steps: S1: Obtain the status information of all distributed nodes in the cross-border e-commerce platform; S2: Detect whether the status information of each of the distributed nodes reaches a first preset state, and mark the distributed nodes that reach the first preset state as idle nodes; S3: Obtain a training task and the corresponding overall compliance prediction model, and combine multiple idle nodes into a virtual computing resource pool based on the training task through a load balancing algorithm; S4: Conduct distributed training of sub-models based on the virtual computing resource pool; S5: Collect the partial gradient information of each distributed node in the virtual computing resource pool, summarize and update the gradient of the overall compliance prediction model through an aggregation algorithm to obtain a target overall compliance prediction model; 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 overall compliance prediction model.

[0021] As described in step S1 above, obtain the status information of all distributed nodes in the cross-border e-commerce platform. Real-time collect and monitor the status information of all distributed nodes of the cross-border e-commerce platform. The 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 the key performance indicators of each node every 500 milliseconds. These indicators specifically include the utilization rates of CPU and GPU, memory usage, network bandwidth and latency, storage usage, etc. Through the real-time collection of this information, the platform can comprehensively understand the resource usage situation, 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 rate 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 marking it as an idle node in the subsequent process.

[0022] As described in step S2 above, detect whether the state 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 state information to determine whether it meets the set first preset state. The first preset state can include a series of conditions, such as the GPU utilization rate being lower than a certain threshold (e.g., 20%), the CPU utilization rate being lower than a certain threshold (e.g., 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 response speed.

[0023] As described in step S3 above, obtain the training task and the overall compliance prediction 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. Obtain the training task to be executed and its corresponding overall compliance prediction model. Specifically, the overall 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 the load balancing algorithm may depend on the performance of the resources (such as the computing power and memory of the GPU), as well as the characteristics of the task to be executed (such as the priority of the task and the expected amount of computation). The goal is to reasonably allocate computing resources to ensure efficient task execution. During this process, the creation of the virtual computing resource pool not only improves the 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 state to adapt to the uncertain and volatile data processing requirements.

[0024] As described in step S4 above, perform distributed training of the sub-model based on the virtual computing resource pool. Utilize the created virtual computing resource pool 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 the cloud model, and trains the cloud model to obtain training parameters. During the training process, these nodes will independently iterate and update the model, and each node calculates its own gradient information and model update. 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 according to specific compliance requirements and business scenarios to ensure that the generated compliance prediction model can adapt to various complex actual application situations.

[0025] As described in step S5 above, collect the sub-gradient information of each distributed node in the virtual computing resource pool, and summarize and update the gradient of the overall compliance prediction model through an aggregation algorithm to obtain the target overall compliance prediction model. Collect the sub-gradient information calculated at each node before collection. After each idle node completes the training of the sub-model, it will transmit the corresponding gradient calculated by it back to the central node. These sub-gradients need to be summarized through a certain aggregation algorithm (such as the average method, weighted average method, etc.) to generate new model weights and gradients to update the overall compliance prediction model. The aggregation method is determined according to the specific task and the nature of the model to ensure accuracy and effectiveness during the integration process. By integrating this information, the target overall compliance prediction model will be updated to reflect the learning results from each training node. This process enables the model to have higher accuracy and reliability when processing compliance prediction tasks. The speed and efficiency of model update directly affect the subsequent real-time prediction performance, ensuring that the platform can quickly adapt to the changing compliance environment.

[0026] As described in step S6 above, obtain the cross-border business to be analyzed and predict the compliance of the cross-border business to be analyzed based on the target overall compliance prediction model. Conduct a compliance assessment on the cross-border business to be analyzed using the updated target overall compliance prediction model. The business to be analyzed refers to operations such as data processing, storage, and transmission 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 model calculation, output whether the business meets the requirements of various data sovereignty and compliance. This process can not only effectively judge the compliance degree of the business and avoid potential legal risks, but also help enterprises conduct dynamic monitoring and evaluation of compliance, thereby optimizing the overall compliance strategy. Through the implementation of this series of steps, the cross-border e-commerce platform can quickly respond to compliance challenges in the complex international regulatory environment and maintain its brand image and market competitiveness.

[0027] In one embodiment, after step S3 of obtaining the training task and the corresponding overall compliance prediction model, and combining multiple idle nodes into a virtual computing resource pool based on the training task through a load balancing algorithm, the following steps are further included: S401: Real-time monitor the real-time status information of all distributed nodes; S402: Judge whether the real-time status information of the idle nodes reaches a second preset state, and judge whether the real-time status information of the non-idle nodes among all distributed nodes reaches a third preset state; S403: Delete the idle nodes whose real-time status information reaches the second preset state from the virtual computing resource pool, and add the non-idle nodes whose real-time status information reaches the third preset state to the virtual computing resource pool.

[0028] As described in the above steps S401 - S403, it is necessary to continuously monitor the real - time status information of all distributed nodes in order to quickly respond to different computing requirements and status changes in a dynamic environment. Implementing this monitoring effectively helps the platform understand the current performance of each node, which can be achieved by deploying monitoring tools such as Prometheus. Combining with custom Exporter components, key performance data can be collected every 500 milliseconds. The monitored information includes CPU and GPU utilization, memory usage, network bandwidth, latency, etc., to ensure that the status of the nodes is accurately captured. In this way, the system can timely identify which nodes have become idle, which nodes are in a high - load state, and can even monitor network latency, bandwidth occupancy, etc., to 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 ensuring the best utilization efficiency of computing resources under different time periods and load conditions.

[0029] Based on the previously collected real - time status information, a detailed status judgment is made on idle nodes and non - idle nodes. The set second preset status includes some specific performance metrics, such as memory utilization and processing capacity, etc. Only when the status information of the idle nodes meets these thresholds can they be considered to be restored to an available state. Specifically, when the CPU load of the idle node < 30% for 5 minutes, it is considered that it does not meet the idle node criteria and is removed from the virtual computing resource pool. At the same time, the system also needs to check whether the status of all non - idle nodes reaches the third preset status, such as being lower than a certain load threshold, having too low a response time, etc., so that non - idle nodes can be considered for addition to the virtual computing resource pool. Specifically, it can be set that when the CPU load > 75% for 1 minute or P99 latency > 300ms, the non - idle node is identified as an idle node to be added to the virtual computing resource pool. Thus, it can timely reflect the loading capacity of the nodes and ensure the effective dynamic adjustment of the virtual resource pool. By accurately judging the status, the platform can find a balance between resource utilization and task requirements, ensuring the efficient configuration and timely response of resources.

[0030] In one embodiment, before step S5 of collecting the gradient information of each of the idle nodes in the virtual computing resource pool, aggregating it through an aggregation algorithm, and updating the gradient of the overall compliance prediction model to obtain the target overall compliance prediction model, it further includes: S411: When the model update instructions of each of the idle nodes are triggered, each of the idle nodes obtains a plurality of cross - border business training samples in the region where it is located; S412: Based on the cross - border business training samples, train the pre - stored compliance prediction sub - models of each of them to obtain sub - model gradients; S413: Generate corresponding sub-gradient information based on the sub-model gradients and upload it to the distributed node where the total compliance prediction model is located.

[0031] As described in the above steps S411 - S413, when a model update instruction for an idle node is detected, the node will automatically obtain multiple cross-border business training samples in its region. Training samples for cross-border e-commerce business usually contain various information related to legal compliance, user behavior, transaction data, etc. Due to data sovereignty regulations restricting the cross-border flow of model training data, each distributed node obtains cross-border business training samples in its region for training. The acquisition method can be obtaining from a database or communicating with external data sources to ensure that the collected samples are representative and diverse, so that the model can better adapt to the changing business environment. Each idle node uses the obtained cross-border business training samples to train its pre-stored compliance prediction sub-model. Specifically, using the supervised learning strategy in machine learning, the training process not only depends on the input features but also involves label data (such as compliance classification) to adjust the model weights. After each training, the node will calculate the corresponding sub-model gradients, and these gradients are the directions and magnitudes of the changes in model parameters in the current iteration. The obtained sub-model gradients can reflect the current learning efficiency and performance of the model, guiding the model to continuously optimize its prediction ability through feedback. Each idle node will generate 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 gradients data to ensure its effective utilization by subsequent aggregation algorithms but may also include appropriate compression and encoding to reduce the amount and time of data transmission and improve network efficiency. The uploaded sub-gradient information will be aggregated into the total model to form a complete update basis, providing a raw material foundation for the subsequent gradient aggregation process. By collecting sub-gradients, the learning results from different nodes can be continuously obtained, thereby achieving effective optimization and update of the total compliance prediction model, and thus realizing more efficient management and compliance detection of cross-border e-commerce.

[0032] In one embodiment, step S413 of generating corresponding sub-gradient information based on the sub-model gradients and uploading it to the distributed node where the total compliance prediction model is located further includes: S4131: Apply Laplace noise to the sub-model gradients to obtain temporary sub-model gradients; S4132: Filter each temporary sub-model gradient according to a preset gradient parameter filtering mechanism to obtain target temporary sub-model gradients; S4133: Generate corresponding sub-gradient information based on the target temporary sub-model gradients and upload it to the distributed node where the total compliance prediction model is located.

[0033] As described in the above steps S4131 - S4133, each idle node will impose Laplace noise on the sub - model gradients obtained during its training process. This process is to ensure the privacy and security of data, especially in application scenarios such as cross - border e - commerce that need to comply with data compliance requirements. The introduction of Laplace noise aims to protect the sensitive data information involved in the model training process and prevent potential information leakage. The specific operation is to add an uncertainty of Laplace distribution to each gradient value according to the set privacy budget and noise scale. In this way, the original precise gradient update is modified by the noise, 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, ΔW is the sub - model gradient, Lap represents the Laplace distribution, Lap(0,σ = 0.8 / ε) means the noise is symmetrically distributed around 0, σ is the scale parameter. The way of imposing Laplace noise is usually to apply random perturbation on each parameter, so that the final gradient can still reflect the learning direction of the model, but loses the sensitivity to individual data points to a certain extent. This can effectively avoid too much personal information in the model output and protect user privacy. Process each temporary sub - model gradient according to the preset gradient parameter filtering mechanism. The design of this filtering mechanism aims to eliminate data points with relatively large noise interference, so as to obtain a more accurate and reliable target temporary sub - model gradient. The preset gradient parameter filtering mechanism can be set according to the results of historical data analysis, the distribution characteristics of the gradients, and the expected model performance. For example, retain parameters with |ΔW|>0.05. In terms of direction filtering, local gradients with an angle >60° with the global average direction can be discarded. It is also possible to perform Top - K screening on the retained gradients (K = 20% of the total number of parameters). Specifically, insert a custom Hook function in the PyTorch framework and perform filtering during the backward() stage.

[0034] 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 through a TLS1.3 encrypted channel using the national secret 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.

[0035] 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: S4111: Each of the idle nodes downloads the teacher model from the cloud respectively; S4112: Extracting the attention matrix of the multi-layer Transformer layer in the teacher model; 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.

[0036] As described in the above steps S4111 - S4113, each idle node will download the teacher model from the cloud. The teacher model is usually a pre - trained deep - learning model with high complexity and accuracy, capable of handling rich features and complex data patterns. In the context of cross - border e - commerce, such a teacher model can be a compliance prediction model trained with a large amount of data, which can effectively capture key features related to compliance. In one embodiment, the teacher model can be the above - mentioned overall compliance prediction model. During the download process, it is necessary to ensure the security of communication and the integrity of data to avoid data being tampered with or leaked during transmission. Usually, data transmission can be carried out through the HTTPS protocol, and means such as checksums are used to ensure the integrity of the file. In addition, each idle node may select the most suitable version of the teacher model according to its own computing resources, storage capacity, etc., to achieve optimal resource allocation. The purpose of downloading the teacher model is to prepare a basic framework for subsequent model training, helping the nodes inherit the knowledge and feature - learning ability of the teacher model when training their respective compliance prediction sub - models, thereby improving the training effect and accuracy of the final student model. The attention matrices of multiple Transformer layers will be extracted from the downloaded teacher model. The Transformer model uses the self - attention mechanism, enabling each input feature to establish dynamic weight relationships with other features, and the attention matrix is a direct manifestation of these relationships. It records how different input features interact with each other within the model and their importance in the decision - making process. The knowledge distillation process can help the student model still perform well with relatively fewer parameters. When extracting the attention matrix, the system needs to ensure that the extraction level and parameter configuration meet the requirements of the subsequent model. In addition, the multiple extracted attention matrices will provide a basis for constructing the student model, enabling the subsequent generated compliance prediction sub - models to better understand the feature distribution and its internal relationships during their learning process. This process ultimately enables the student model to have higher initial performance and efficiency compared to training from scratch. A preset number of matrices will be selected from the extracted attention matrices of multiple Transformer layers, and these matrices will be combined to construct the student model. The selected attention matrices are usually chosen based on their relevance or importance in the output results to ensure that the constructed student model can effectively learn the key features and knowledge in the teacher model. At the same time, when constructing the student model, the system will design a preset loss function to provide feedback during the model training process, enabling the model to continuously adjust parameters to reduce the prediction error. Common loss functions include cross - entropy loss (for classification tasks) or mean - squared error (for regression tasks). Through these loss functions, the learning ability and effect of the model during training can be evaluated.The obtained student model will serve as a compliance prediction sub-model, inheriting the characteristics and knowledge of the teacher model. At the same time, due to the optimization of parameters and structure, it has higher computational efficiency and adaptability. This distillation process enables the student model to maintain high accuracy and performance when processing cross-border business compliance predictions within the region, while reducing the demand for computing resources, which is an efficient model construction strategy.

[0037] In one embodiment, step S5 of collecting the sub-gradient information of each distributed node in the virtual computing resource pool, aggregating and updating the gradient of the total compliance prediction model through an aggregation algorithm to obtain the target total compliance prediction model includes: S501: Obtain the total gradient information corresponding to the total compliance prediction model; S502: Calculate the difference parameter between the total gradient information and the sub-gradient information; S503: Based on the difference parameter, update the total compliance prediction model through a preset data synchronization algorithm to obtain a temporary total compliance prediction model; S504: Detect the performance of the temporary total compliance prediction model; S505: If the performance meets the preset requirements, record the temporary total compliance prediction model as the target total compliance prediction model.

[0038] As described in the above steps S501 - S505, first, extract the corresponding total gradient information from the overall compliance prediction model. The total gradient information is the gradient required for parameter update calculated through the backpropagation algorithm during the model training process, reflecting the sensitivity of the model to each parameter in the current state. The total gradient information indicates the direction and magnitude of adjustment for each parameter (such as weights and biases) to improve the model's prediction performance. The process of extracting the total gradient information usually depends on the performance of the current overall compliance prediction model on the training dataset. The total gradient that the system needs to process may contain information of multiple parameters, involving all connections in a complex model structure such as a multi - layer neural network. By effectively extracting this information, the current training state of the model can be accurately evaluated, providing the necessary basic data for subsequent updates. Calculate the differential parameters between the total gradient information and the partial gradient information collected in the previous step. Compare these two parts of gradient information to find out the specific parameters that need to be adjusted during the model update process. The total gradient information represents the parameter update that the current overall compliance prediction model expects to perform, while the partial gradient information comes from the local training results of each idle node. By calculating the differential parameters, the specific update effects generated during the distributed training of each node can be identified and compared with the requirements of the overall model. This comparison can effectively reveal the learning information differences at different levels and regions, and thus provide a more accurate basis for parameter update. In addition, the calculation of differential parameters can also be regarded as the aggregation of learning achievements between nodes, which helps to adjust the parameters that show strong learning ability in some nodes. By effectively calculating the differential parameters, it can be ensured that the update of the overall compliance prediction model integrates the learning experiences of different nodes, and thus better adapts to the complex cross - border business environment. According to the calculated differential parameters, update the overall compliance prediction model through a preset data synchronization algorithm. This step is the key to realizing model adjustment and optimization, that is, applying the obtained differential parameters 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 various algorithm strategies, including but not limited to the traditional SGD (Stochastic Gradient Descent), Adam optimization algorithm, etc. This algorithm is responsible for determining how to incorporate the differential parameters into the existing model parameters at what ratio to achieve effective learning and performance improvement. Through this parameter update, the overall compliance prediction model can integrate the newly obtained learning achievements while maintaining the original knowledge, thus improving the overall performance. It should be noted that the update process must consider the model stability. If the parameter update is too large, it may cause the model not to converge or oscillate.

[0039] Perform performance testing on the generated temporary compliance prediction total model. This process is an important step to ensure that the new model can meet the expected requirements in actual applications. Performance testing usually includes considerations of multiple evaluation metrics such as the accuracy, precision, recall rate, etc. of the model, and these metrics will help the system understand the performance of the model on a given test set. When performing performance testing, a preset validation data set may be used to ensure the fairness and reliability of the testing. By evaluating the temporary model, its potential performance in actual use can be obtained, and possible problems and deficiencies can be discovered. For example, if the prediction accuracy of the model in a specific scenario is lower than the set expected target, the system needs to analyze the reasons, which may involve insufficient learning of the model on specific types of data. The results of performance testing will directly affect subsequent decisions. If the temporary compliance prediction total model performs well, has a high accuracy rate and good adaptability, it will enter the next confirmation stage; on the contrary, if the performance fails to meet the expectations, it may be necessary to further adjust the model or re-examine the update process to ensure that the finally selected model is sufficient to meet the requirements of cross-border business compliance. This testing process is an important link to ensure the continuous progress of the model and accurately capture business changes. Compare the test results with the preset standards, including the thresholds of key performance indicators (such as accuracy rate, recall rate, F1-score, etc.). If the performance meets the preset requirements, the system will officially mark the temporary compliance prediction total model as the target compliance prediction total model. This transformation marks the success of the model optimization process and prepares it for use in actual scenarios such as cross-border e-commerce.

[0040] In one embodiment, step S3 of combining multiple idle nodes into a virtual computing resource pool based on the training task through a load balancing algorithm includes: S301: Map the training task to a multi-dimensional feature vector; S302: Generate a decision path through the CART algorithm based on the multi-dimensional feature vector; S303: Obtain multiple idle nodes based on the decision path to form a virtual computing resource pool.

[0041] As described in the above steps S301 - S303, the training task is first mapped to a multi - dimensional feature vector. A training task usually has multiple attributes. In a specific embodiment, the multi - dimensional feature vector is specifically a three - dimensional vector (priority P ∈ [1 - 3], latency requirement D ∈ [100 - 500ms], cost level C ∈ [1 - 5]). Using the obtained multi - dimensional feature vector, a decision path is generated through the CART (Classification And Regression Trees) algorithm. CART is a machine learning algorithm for constructing decision trees, which can predict results through the input feature vector. In this step, by analyzing the feature vector, the CART algorithm will construct a decision tree. Through the path branches, a suitable computing resource allocation strategy can be effectively distinguished. Each branch decision reflects the advantages and disadvantages of selecting a certain resource under specific conditions. The advantage of this tree - like structure is that it can visualize the complex decision - making process, thus providing clear guidance for resource allocation. The finally generated decision path will be a tree - like structure composed of nodes and edges, indicating how to select and utilize idle nodes under different feature conditions, so as to achieve efficient load balancing. According to the guidance of the decision path, idle nodes that meet the requirements of the training task are screened out. The selected nodes generally have sufficient computing power, storage capacity, and other performance indicators to support the upcoming training task. By reasonably combining these idle nodes, the system will form a flexible and dynamic computing resource pool. This virtual resource pool can be quickly allocated according to task requirements, ensuring that the required computing resources are provided immediately when needed, thereby improving efficiency and promoting the maximum utilization of resources. Thus, effective load balancing is achieved, the overall system performance is improved, and resource waste is reduced.

[0042] Referring to Figure 3 , the present invention also provides a prediction device for cross - border business compliance. The device includes: An acquisition module 902, configured to acquire the status information of all distributed nodes in the cross - border e - commerce platform; A detection module 904, configured to detect whether the status information of each of the distributed nodes reaches a first preset state, and mark the distributed nodes that reach the first preset state as idle nodes; A combination module 906, configured to acquire a training task and the overall compliance prediction 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; A training module 908, configured to perform distributed training of sub - models based on the virtual computing resource pool; A collection module 910, configured to collect the sub-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 to obtain a target compliance prediction total model; A prediction module 912, configured to 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.

[0043] In one embodiment, the prediction device for cross-border business compliance further includes: A real-time status information monitoring module, configured to monitor the real-time status information of all distributed nodes in real time; A real-time status information judgment module, configured to judge whether the real-time status information of the idle nodes reaches a second preset state, and judge whether the real-time status information of the non-idle nodes among all distributed nodes reaches 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 the non-idle nodes whose real-time status information reaches the third preset state to the virtual computing resource pool.

[0044] In one embodiment, the prediction device for cross-border business compliance further includes: A training sample acquisition module, configured to, when a model update instruction of each of the idle nodes is triggered, each of the idle nodes acquires a plurality of cross-border business training samples in its area; A compliance prediction sub-model training module, configured to train the pre-stored compliance prediction sub-model based on the cross-border business training samples to obtain a sub-model gradient; A sub-gradient information generation module, configured to generate corresponding sub-gradient information based on the sub-model gradient and upload it to the distributed node where the compliance prediction total model is located.

[0045] In one embodiment, the sub-gradient information generation module further includes: A noise addition sub-module, configured to apply Laplace noise to the sub-model gradient to obtain a temporary sub-model gradient; A filtering sub-module, configured to filter each temporary sub-model gradient according to a preset gradient parameter filtering mechanism to obtain a target temporary sub-model gradient; A sub-gradient information generation sub-module, configured 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 total model is located.

[0046] In one embodiment, the prediction device for cross-border business compliance further includes: A teacher model download module for each of the idle nodes to download the teacher model from the cloud respectively; An attention matrix extraction module for extracting the attention matrices of multiple Transformer layers in the teacher model; An attention matrix selection module for selecting a preset number of attention matrices from the multiple Transformer layers and adding a preset loss function to obtain a student model as the compliance prediction sub-model.

[0047] In one embodiment, the collection module 910 includes: A total gradient information acquisition sub-module for acquiring the total gradient information corresponding to the compliance prediction total model; A differential parameter calculation sub-module for calculating the differential parameters between the total gradient information and the sub-gradient information; An update sub-module for updating the compliance prediction total model based on the differential parameters through a preset data synchronization algorithm to obtain a temporary compliance prediction total model; A performance detection sub-module for detecting the performance of the temporary compliance prediction total model; A marking sub-module for marking the temporary compliance prediction total model as the target compliance prediction total model if the performance meets the preset requirements.

[0048] In one embodiment, the combination module 906 includes: A multi-dimensional feature vector mapping sub-module for mapping the training task into a multi-dimensional feature vector; A decision path generation sub-module for generating a decision path based on the multi-dimensional feature vector through the CART algorithm; An idle node acquisition sub-module for acquiring multiple idle nodes based on the decision path to form a virtual computing resource pool.

[0049] Figure 4 The internal structure diagram of a computer device in one embodiment is shown. The computer device can specifically be a terminal or a server. As Figure 4 shown, the computer device includes a processor, a memory, and a network interface connected through a system bus. Among them, 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 can also store a computer program. When the computer program is executed by the processor, the processor can implement the prediction method for cross-border business compliance. The internal memory can also store a computer program. When the computer program is executed by the processor, the processor can execute the prediction method for cross-border business compliance. Those skilled in the art can understand, Figure 4The structure shown is only a block diagram of some structures related to the solution of this application, and does not constitute a limitation on the computer device to which the solution of this application is applied. The specific computer device may include more or fewer components than those shown in the figure, or combine some components, or have different component arrangements.

[0050] In one embodiment, a computer device is proposed, including a memory and a processor. The memory stores a computer program. When the computer program is executed by the processor, the processor is caused to perform the following steps: Obtain the status information of all distributed nodes in the cross-border e-commerce platform; 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; Obtain a training task and the corresponding overall compliance prediction model for the training task, and combine multiple idle nodes into a virtual computing resource pool based on the training task through a load balancing algorithm; Perform distributed training of sub-models based on the virtual computing resource pool; Collect the partial gradient information of each distributed node in the virtual computing resource pool, aggregate and update the gradient of the overall compliance prediction model through an aggregation algorithm to obtain a target overall compliance prediction model; Obtain a cross-border business to be analyzed, and predict the compliance of the cross-border business to be analyzed based on the target overall compliance prediction model.

[0051] Through a dynamic resource management method of identifying and marking idle nodes, in an environment with significant fluctuations in computing power requirements, the computing resource configuration can be flexibly adjusted, avoiding the problem of low utilization rate of GPU resources in the traditional fixed computing power allocation mode, thereby improving the overall resource utilization efficiency, reducing the response latency, and on the other hand, also improving the training speed and accuracy of the model, enabling the cross-border e-commerce platform to quickly adapt to market changes and enhancing the user experience.

[0052] In one embodiment, a computer-readable storage medium is proposed, storing a computer program. When the computer program is executed by a processor, the processor is caused to perform the following steps: Obtain the status information of all distributed nodes in the cross-border e-commerce platform; 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; Obtain a training task and the corresponding overall compliance prediction model for the training task, and combine multiple idle nodes into a virtual computing resource pool based on the training task through a load balancing algorithm; Perform distributed training of sub-models based on the virtual computing resource pool; Collect the sub-gradient information of each distributed node in the virtual computing resource pool, aggregate and update the gradient of the total compliance prediction model through an aggregation algorithm to obtain the target total compliance prediction model; Obtain the cross-border business to be analyzed, and predict the compliance of the cross-border business to be analyzed based on the target total compliance prediction model.

[0053] Through a dynamic resource management method that identifies and marks idle nodes, in an environment with significant fluctuations in computing power requirements, the computing resource configuration can be flexibly adjusted, avoiding the problem of low GPU resource utilization in the traditional fixed computing power allocation mode, thereby improving the overall resource utilization efficiency, reducing the response latency, and on the other hand, also improving the training speed and accuracy of the model, enabling the cross-border e-commerce platform to quickly adapt to market changes and enhance the user experience.

[0054] Those of ordinary skill in the art can understand that all or part of the processes of implementing the methods in the above embodiments can be completed by instructing 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 methods. Among them, any reference to a memory, storage, database, or other medium used in the various embodiments provided in this application can include non-volatile and / or volatile memories. 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 many forms, such as static RAM (SRAM), dynamic RAM (DRAM), synchronous DRAM (SDRAM), double data rate SDRAM (DDR SDRAM), 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), etc.

[0055] The technical features of the above embodiments can be combined arbitrarily. For the sake of brevity of description, 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, it should be considered as the scope recorded in this specification.

[0056] The above-described embodiments merely represent several implementation manners of the present application. The description thereof is relatively specific and detailed, but it should not be construed as a limitation on the patent scope of the present application. It should be noted that for those of ordinary skill in the art, without departing from the concept of the present application, several modifications and improvements can still be made, and these all fall within the protection scope of the present application. Therefore, the protection scope of the patent of the present application shall be subject to the appended claims.

Claims

1. A prediction method for cross-border business compliance, characterized in that The method includes: Obtaining the status information of all distributed nodes within a cross-border e-commerce platform; Detecting whether the status information of each of the distributed nodes reaches a first preset status, and marking the distributed nodes that reach the first preset status as idle nodes; Obtaining a training task and a corresponding overall compliance prediction model, and combining multiple idle nodes into a virtual computing resource pool based on the training task through a load balancing algorithm; Performing distributed training of sub-models based on the virtual computing resource pool; Collecting the partial gradient information of each distributed node in the virtual computing resource pool, summarizing and updating the gradient of the overall compliance prediction model through an aggregation algorithm to obtain a target overall compliance prediction model; Obtaining a cross-border business to be analyzed, and predicting the compliance of the cross-border business to be analyzed based on the target overall compliance prediction model.

2. The predictive method for cross-border business compliance according to claim 1, wherein After the step of obtaining a training task and a corresponding overall compliance prediction model, and combining multiple idle nodes into a virtual computing resource pool based on the training task through a load balancing algorithm, it further includes: Real-time monitoring of the real-time status information of all distributed nodes; Judging whether the real-time status information of the idle nodes reaches a second preset status, and judging whether the real-time status information of the non-idle nodes among all distributed nodes reaches a third preset status; Deleting the idle nodes whose real-time status information reaches the second preset status from the virtual computing resource pool, and adding the non-idle nodes whose real-time status information reaches the third preset status to the virtual computing resource pool.

3. The predictive method for cross-border business compliance according to claim 1, wherein Before the step of collecting the partial gradient information of each of the idle nodes in the virtual computing resource pool, summarizing and updating the gradient of the overall compliance prediction model through an aggregation algorithm to obtain a target overall compliance prediction model, it further includes: When the 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 its region; Training the pre-stored compliance prediction sub-models of each based on the cross-border business training samples to obtain sub-model gradients; Generating corresponding partial gradient information based on the sub-model gradients and uploading it to the distributed node where the overall compliance prediction model is located.

4. The predictive method for cross-border business compliance according to claim 3, wherein The step of generating corresponding partial gradient information based on the sub-model gradients and uploading it to the distributed node where the overall compliance prediction model is located further includes: Applying Laplace noise to the sub-model gradients to obtain temporary sub-model gradients; Filtering each of the temporary sub-model gradients according to a preset gradient parameter filtering mechanism to obtain target temporary sub-model gradients; Generating corresponding partial gradient information based on the target temporary sub-model gradients and uploading it to the distributed node where the overall compliance prediction model is located.

5. The prediction method for cross-border business compliance according to claim 3, characterized in that Before the step of training the pre-stored compliance prediction sub-models of each based on the cross-border business training samples to obtain sub-model gradients, it further includes: Each of the idle nodes separately downloads a teacher model from the cloud; Extracting the attention matrix of multiple Transformer layers in the teacher model; Select a preset number of attention matrices from the multi-layer Transformer layers and add them to a preset loss function to obtain a student model, which is used as the compliance prediction sub-model.

6. The predictive method for cross-border business compliance according to claim 1, wherein The step of collecting the sub-gradient information of each distributed node in the virtual computing resource pool, summarizing it through an aggregation algorithm, and updating the gradient of the compliance prediction total model to obtain the target compliance prediction total model includes: Obtain the total gradient information corresponding to the compliance prediction total model; Calculate the difference parameter between the total gradient information and the sub-gradient information; Based on the difference parameter, update the compliance prediction total model through a preset data synchronization algorithm to obtain a temporary compliance prediction total model; Detect the performance of the temporary compliance prediction total model; If the performance meets the preset requirements, record the temporary compliance prediction total model as the target compliance prediction total model.

7. The predictive method for cross-border business compliance according to claim 1, wherein The step of combining multiple idle nodes into a virtual computing resource pool based on the training task through a load balancing algorithm includes: Map the training task to a multi-dimensional feature vector; Generate a decision path based on the multi-dimensional feature vector through the CART algorithm; Obtain multiple idle nodes based on the decision path to form a virtual computing resource pool.

8. A prediction device for cross-border business compliance, characterized in that, The device includes: An acquisition module, configured to acquire 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 distributed node reaches a first preset status, and mark the distributed nodes that reach the first preset status as idle nodes; A combination module, configured to acquire a training task and the 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; A training module, configured to perform distributed training of the sub-model based on the virtual computing resource pool; A collection module, configured to collect the sub-gradient information of each distributed node in the virtual computing resource pool, summarize it through an aggregation algorithm, and update the gradient of the compliance prediction total model to obtain the target compliance prediction total model; A prediction module, configured to 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.

9. A computer-readable storage medium, characterized in that, Stores a computer program, which when executed by a processor, causes the processor to execute the steps of the method for predicting cross-border business compliance according to any one of claims 1 to 7.

10. A computer device, characterized in that, The device includes a memory and a processor, the memory stores a computer program, which when executed by the processor, causes the processor to execute the steps of the method for predicting cross-border business compliance according to any one of claims 1 to 7.

Images