An e-commerce platform account splitting management system based on cloud-edge collaboration

Abstract

The present application provides an e-commerce platform account management system based on cloud-edge collaboration, including: based on the normalized sales data set, selecting a support vector machine or a neural network to construct a prediction model that can capture the inherent laws of sales data; optimizing the prediction model through model compression technology to obtain a lightweight prediction model to reduce the number of parameters and calculations of the model; sending the lightweight prediction model to the edge node to perform real-time prediction at a location close to the data source; on the edge node, using an online learning algorithm to update and tune the lightweight prediction model in real time to adapt to changes in sales data; establishing a cloud-edge collaborative data transmission and synchronization mechanism to reduce the amount of data transmission between the cloud and the edge, and ensure that the prediction models on the cloud and edge nodes remain consistent.

Description

Technical Field

[0001] The present invention relates to the field of information technology, and in particular to an e-commerce platform account splitting management system based on cloud-edge collaboration. Background Art

[0002] In the account management system of e-commerce platforms, sales forecasting is a key link. Since the sales data of e-commerce platforms is huge and is affected by many factors, such as user behavior, market trends, promotional activities, etc., the sales data presents complex nonlinear characteristics and uncertainty. Traditional time series analysis methods, such as moving average method and exponential smoothing method, are difficult to capture the complex patterns and trend changes in sales data. Although machine learning algorithms, such as support vector machines and random forests, can establish nonlinear prediction models, they have high computational complexity and slow model training and prediction speed when processing large-scale data.

[0003] Therefore, how to improve prediction efficiency while ensuring prediction accuracy has become a technical challenge faced by the e-commerce platform account management system. This requires collaborative computing between the cloud and edge nodes, using the powerful computing and storage capabilities of the cloud to pre-process and feature-engineer historical sales data and build complex prediction models; while on the edge nodes, using lightweight online learning algorithms to update and optimize the prediction models sent from the cloud in real time to quickly respond to changes in sales data. At the same time, it is also necessary to establish an efficient data transmission and synchronization mechanism between the cloud and the edge to ensure the consistency and real-time performance of the prediction model. Summary of the invention

[0004] The present invention provides an e-commerce platform account splitting management system based on cloud-edge collaboration, which mainly includes:

[0005] Perform data cleaning and feature extraction on the historical sales data of the e-commerce platform to obtain a standardized sales data set with multi-dimensional features including user behavior, market trends, and promotional activities;

[0006] Based on the normalized sales data set, a support vector machine or a neural network is selected to construct a prediction model capable of capturing the inherent laws of the sales data;

[0007] The prediction model is optimized by using model compression technology to obtain a lightweight prediction model to reduce the number of model parameters and the amount of calculation;

[0008] Sending the lightweight prediction model to edge nodes to perform real-time prediction at locations close to data sources;

[0009] On the edge node, an online learning algorithm is used to update and tune the lightweight prediction model in real time to adapt to changes in sales data;

[0010] Establish a cloud-edge collaborative data transmission and synchronization mechanism to reduce the amount of data transmission between the cloud and the edge, and ensure that the prediction models on the cloud and edge nodes are consistent.

[0011] The technical solution provided by the embodiment of the present invention may have the following beneficial effects:

[0012] The present invention discloses an e-commerce sales forecasting method based on cloud-edge collaboration. The method first cleans and extracts features from historical sales data to construct a normalized data set containing multi-dimensional features such as user behavior and market trends. On this basis, a support vector machine or a neural network is selected to build a forecasting model, and a lightweight model is obtained by model compression. The lightweight model is sent to the edge node for real-time forecasting, and the model parameters are updated in real time using an online learning algorithm to adapt to the dynamic changes in sales data. At the same time, a cloud-edge collaboration mechanism is established to ensure the consistency of the cloud and edge node models. Through cloud-edge collaboration and model compression technology, the present invention realizes real-time and efficient forecasting of e-commerce sales data, improves forecasting accuracy and system response speed, and provides strong support for the operational decisions of the e-commerce platform. BRIEF DESCRIPTION OF THE DRAWINGS

[0013] Figure 1 This is a flow chart of an e-commerce platform account splitting management system based on cloud-edge collaboration of the present invention.

[0014] Figure 2 It is a schematic diagram of an e-commerce platform account management system based on cloud-edge collaboration of the present invention.

[0015] Figure 3 This is another schematic diagram of an e-commerce platform account management system based on cloud-edge collaboration according to the present invention. DETAILED DESCRIPTION

[0016] In order to further understand the content of the present invention, the present invention is described in detail in conjunction with the accompanying drawings and embodiments. The present application is further described in detail below in conjunction with the accompanying drawings and embodiments. It is understood that the specific embodiments described herein are only used to explain the relevant inventions, rather than to limit the invention. It is also necessary to explain that, for ease of description, only the parts related to the invention are shown in the accompanying drawings.

[0017] like Figure 1-3 In this embodiment, an e-commerce platform account splitting management system based on cloud-edge collaboration may specifically include:

[0018] Step S101, data cleaning and feature extraction are performed on the historical sales data of the e-commerce platform to obtain a normalized sales data set containing multi-dimensional features of user behavior, market trends, and promotional activities.

[0019] Obtain sales data for the past year from the database of the e-commerce platform, wherein the sales data includes product information, transaction amount, and user behavior;

[0020] Importing the sales data into at least one data processing module, processing missing values ​​using the fillna() function, and identifying and removing outliers using the IQR method to obtain preliminarily cleaned sales data;

[0021] For the sales data after preliminary cleaning, the OneHotEncoder method is used to encode the category features of the product category and the promotion activity type to obtain the encoded sales data;

[0022] Building a sales prediction model based on the encoded sales data, wherein the sales prediction model uses an XGBoost algorithm to process nonlinear relationships and feature interactions in the sales data;

[0023] The sales forecasting model is deployed to the forecasting server, and a real-time sales forecasting function is provided through a RESTful API interface. According to the requested real-time sales data, the sales forecasting model is used to obtain the corresponding sales forecasting result, and the sales forecasting result is returned to the requesting end.

[0024] For example, 1 million sales records from the MySQL database of the e-commerce platform are obtained through the SQL statement SELECT *FROM sales_dataWHERE order_date BETWEEN '2022-01-01' AND '2022-12-31' in the past year, including fields such as product ID, product category, transaction amount, user ID, and user behavior. The sales data is imported into the Pandas data processing module, and the fillna() function is used to fill the missing values, where the numerical features are filled with the mean and the categorical features are filled with the mode. The outliers in the transaction amount field are identified by the IQR method, and values ​​less than Q1-5IQR or greater than Q3+5IQR are regarded as outliers and removed, and a total of 2,000 outliers are removed. For the sales data after preliminary cleaning, the OneHotEncoder method is used to perform one-hot encoding on the categorical features of product category and promotion activity type, converting them into multiple binary features. According to the encoded sales data, an XGBoost sales prediction model is constructed, with the number of trees set to 100, the maximum depth set to 5, the learning rate set to 1, the transaction amount set to the target variable, and other fields set to feature variables. The model is trained and cross-validated with 10 folds, and the average RMSE is 1,000 yuan. The trained XGBoost model is deployed to the RESTful API server built by Flask to provide real-time sales prediction functions. When receiving JSON format request data containing features such as product category and promotion activity type, the prediction model is called to predict the request data, and the predicted transaction amount is returned to the request end. The entire prediction process takes less than 100 milliseconds. By deploying the online sales prediction service, e-commerce platforms can predict the possible purchase amount of users based on their real-time behavior data, thereby conducting precision marketing and intelligent recommendations to increase sales and user experience.

[0025] Step S102, based on the normalized sales data set, select a support vector machine or a neural network to build a prediction model that can capture the inherent laws of the sales data.

[0026] After obtaining the sales data set, the data set is preprocessed, and the preprocessing includes: processing missing values ​​by using mean filling or deletion methods, identifying and processing outliers through box plots, and normalizing the data by using standardization methods.

[0027] The trained model is used to predict the test set and the accuracy, precision, recall and F1 score of the model are evaluated.

[0028] If the performance of the model meets the preset requirements, it will be saved as the final model and applied to the actual sales forecasting task, a sales strategy will be formulated based on the forecast results, and the forecast effect will be continuously tracked.

[0029] For example, sales data for the past year are extracted from the database of an e-commerce platform, including product information, transaction amount, user behavior, etc. Valid sales records are filtered out using SQL query statements, and data with an order status of "cancelled" or "refunded" are removed. The extracted raw data is imported into a Pandas data frame, and missing values ​​are processed by using the mean filling method, that is, the missing values ​​of continuous features are replaced by the mean of the feature, and the missing values ​​of discrete features are replaced by the mode. Then, outliers are identified through a box plot, and data that is less than the lower quartile Q1-1.5IQR or greater than the upper quartile Q3+1.5IQR is considered an outlier and deleted, where IQR is the interquartile range. Then, the minimum-maximum normalization method is used to normalize the numerical features, and the value range of each feature is scaled to the interval [0,1]. Using the processed data set, the XGBoost algorithm is used to build a sales prediction model, and it is trained on the training set by 5-fold cross validation. After the model training is completed, the performance of the model is evaluated on the test set. By calculating the evaluation indicators such as accuracy, precision, recall and F1 score, it is found that the model's accuracy reaches 85%, precision reaches 82%, recall reaches 80%, and F1 score reaches 0.81, which meets the preset performance requirements. Therefore, the model is saved as the final prediction model and deployed to the production environment. The sales situation for the next week is predicted regularly every day. According to the prediction results, the sales team formulates corresponding sales strategies, such as preparing stocks in advance for products with higher predicted sales and adjusting the pricing of products with lower predicted sales in a timely manner. At the same time, the prediction effect of the model is continuously tracked, and the model is retrained every month using the latest sales data to adapt to market changes and improve the accuracy of the prediction.

[0030] Step S103, optimizing the prediction model by model compression technology to obtain a lightweight prediction model to reduce the number of model parameters and the amount of calculation.

[0031] For the prediction model, the TensorFlow Pruning API is used to perform pruning operations, the weight threshold is set to 0.1, and the weights in the model whose absolute values ​​are less than the threshold are removed.

[0032] The pruned model is converted to TensorFlow Lite format through the TensorFlow Lite converter and dynamic range quantization is applied to it to quantize the floating point weights of the model to 8-bit integers.

[0033] The quantized model is used as the student model, and the original uncompressed model is used as the teacher model. The knowledge of the teacher model is transferred to the student model through the knowledge distillation API in the TensorFlow Model Optimization Toolkit.

[0034] The TensorFlow Profiler tool is used to analyze the computational graph of the compressed model to identify the computationally intensive operations and memory bottlenecks. Based on the analysis results, the TensorFlow Graph Transform Tool is used to optimize the computational graph of the model, reducing unnecessary storage and transmission of intermediate results through operator fusion and constant folding.

[0035] Use TensorFlow Lite Benchmark Tool to perform performance tests on the optimized model on the target device to obtain its latency and memory usage indicators. According to the test results, adjust the compression parameters of the model, and find the best balance between model size, inference speed, and accuracy through multiple rounds of iterative optimization.

[0036] Exemplarily, when compressing and optimizing the sales forecasting model, the model is first pruned using the TensorFlow Pruning API. By setting the weight threshold to 1, the weights in the model whose absolute value is less than the threshold are removed to achieve the purpose of slimming the model. After pruning, the number of parameters of the model is reduced from the original 10 million to 8 million, a 20% reduction in size. Then, the pruned model is converted to the TensorFlow Lite format through the TensorFlow Lite converter, and the dynamic range quantization technology is applied to it. During the quantization process, the KL divergence method is used to estimate the distribution of the activation data of the model, and the weights are quantized according to the estimation results, and the original 32-bit floating-point weights are compressed to 8-bit integers. After quantization, the model size is further reduced to 1 / 4 of the original, with only 2.5 million parameters. In order to ensure the performance of the compressed model, the knowledge distillation technology is used to use the original uncompressed model as the teacher model and the quantized model as the student model. The knowledge of the teacher model is transferred to the student model by minimizing the cross entropy loss between the prediction results of the student model and the teacher model. After 50 epochs of training, the student model's accuracy on the test set was improved from the original 85% to 87%, which is close to the performance of the teacher model. To further optimize the model, the TensorFlow Profiler tool was used to analyze the computational graph of the compressed model, and it was found that the bottleneck was mainly the matrix multiplication operation of the fully connected layer. The TensorFlow Graph Transform Tool was used to optimize the computational graph, merge repeated matrix multiplication operators, and reduce unnecessary storage and transmission of intermediate results. The computational efficiency of the optimized model was improved by 15%. Finally, the TensorFlow Lite Benchmark Tool was used on the target mobile device to perform performance testing on the optimized model. The test results showed that the average inference time of the model was reduced from the original 500ms to 200ms, and the memory usage was reduced from the original 100MB to 30MB, which met the requirements of actual deployment. After multiple rounds of adjusting the compression parameters, an optimized model that achieved the best balance between model size, inference speed and accuracy was finally obtained, laying the foundation for the online application of the sales forecasting system.

[0037] Step S104: Send the lightweight prediction model to the edge node to perform real-time prediction at a location close to the data source.

[0038] Obtain lightweight prediction model parameters and structural information of a pre-established machine learning model; optimize and compress the lightweight prediction model using model pruning technology based on the hardware configuration of the edge node to obtain a lightweight prediction model suitable for the edge node; transmit the lightweight prediction model to the edge node deployed near the data source through a secure network transmission protocol; the edge node receives real-time data from the data source, performs data cleaning and feature extraction on the real-time data; predicts the real-time data using a lightweight prediction model deployed on the edge node based on the extracted features; calculate the confidence of the prediction result, and if the confidence is lower than a preset threshold, the data is filtered out through a secure network transmission protocol. The data is transmitted to the cloud node through an encrypted channel; after receiving the low-confidence data, the cloud node uses a more complex and accurate prediction model for secondary prediction; the cloud prediction result is returned to the edge node through an encrypted channel; the edge node uses a weighted average method based on the local prediction result and the cloud prediction result to obtain the final prediction result; for the final prediction result, the predefined judgment conditions are used to generate corresponding control instructions; the edge node sends the control instructions to the execution device through a dedicated interface; the cloud uses new samples to regularly retrain the prediction model and update the model parameters; the updated model parameters are distributed to each edge node for updating the lightweight prediction model.

[0039] Exemplarily, when obtaining a pre-established machine learning model, the TensorFlow Lite Converter tool is used to convert the original model to obtain a lightweight prediction model. The number of parameters of the model is reduced from the original 10 million to 5 million, and the model size is reduced by 50%. For the hardware configuration of the edge node, the lightweight prediction model is further compressed using the pruning technology in the TensorFlow ModelOptimization Toolkit. By setting a sparsity of 30%, 30% of the weight parameters are pruned while retaining important connections. The number of parameters of the compressed model is further reduced to 3.5 million, and the model size is only 35% of the original model. To ensure the security of model transmission, the HTTPS protocol is used to encrypt and transmit the compressed model to the edge node deployed in the factory workshop. After the edge node receives the real-time data from the sensor, it first cleans the data to filter out noise and outliers. Then, predefined feature extraction algorithms, such as Fourier transform and wavelet analysis, are used to extract the key features of the data. Based on the extracted features, the lightweight prediction model deployed on the edge node is used to predict the real-time data. During the prediction process, the confidence of the prediction result is calculated at the same time. When the confidence is lower than the preset threshold of 8, the data is transmitted to the cloud node through the AES encrypted channel for secondary prediction. The cloud node uses a more complex prediction model, such as a deep neural network, to predict low-confidence data. The cloud prediction result is returned to the edge node through the same encrypted channel. The edge node comprehensively considers the local prediction results and the cloud prediction results, and uses the weighted average method to assign weights of 6 and 4 to the local and cloud results respectively to obtain the final prediction result. According to the prediction results and predefined judgment conditions, such as the predicted failure probability exceeding 9, the corresponding control instructions are generated and sent to the PLC control device through the RS-485 interface to achieve real-time adjustment of the production equipment. The cloud continuously collects new data samples, retrains the prediction model every week, uses the Adam optimizer and the cross entropy loss function, trains 50 epochs, and continuously updates and optimizes the model parameters. The updated model parameters are distributed to each edge node through the MQTT protocol to achieve online update of the lightweight prediction model and maintain continuous improvement of prediction performance.

[0040] Step S105, on the edge node, an online learning algorithm is used to update and optimize the lightweight prediction model in real time to adapt to changes in sales data.

[0041] According to the computing power of the edge node, the deployment of the XGBoost algorithm is determined. For the acquired sales data, a data cleaning tool is used for preprocessing, and the preprocessing includes removing outliers, filling missing values ​​and standardization. The processed data set is divided into a training set and a validation set, wherein the training set is used for initial training of the XGBoost model, and the validation set is used to evaluate the model performance. When new sales data is received, the model is updated using an incremental learning method. Specifically, the new data is added to the training set, and the model is retrained using the updated training set, while retaining the original model structure and parameters as initial values. In order to adapt to the storage limitations of the edge node, the model is compressed using a model compression technology, and the model compression technology includes pruning. The compressed model is obtained and deployed to the edge node through an incremental update mechanism. During the model deployment process, double buffering technology is used to achieve smooth switching between the old and new models. After deployment, the model performance and resource usage are tracked in real time by setting a monitoring system.

[0042] Exemplarily, when deploying the XGBoost algorithm on the edge node, the acquired sales data is first preprocessed. Use Python's Pandas library to clean the data, identify and remove outliers through methods such as box plots, and use the mean filling strategy for missing values. Then use the Min-Max standardization method to normalize the data and scale the value range to between 0 and 1. Then divide the training set and the validation set in a ratio of 8:2, use the training set to perform initial training on the XGBoost model, set the maximum depth of the tree to 6, the learning rate to 1, and the subsampling ratio to 8. Evaluate the model performance on the validation set, and optimize the model by calculating indicators such as the root mean square error (RMSE) and the mean absolute error (MAE). When new sales data is received, add it to the training set, and use the gradient boosting algorithm to incrementally learn the model, update the tree structure and the weights of the leaf nodes, and retain the original model parameters as regularization terms to avoid overfitting. To meet the storage limit of edge nodes, the model is compressed using threshold-based pruning technology to remove features and nodes with importance less than 05. The size of the compressed model is reduced by 40%. By setting up a double buffer, the compressed model is incrementally updated to the edge node without affecting the prediction service. After deployment, a monitoring system is built using Prometheus and Grafana to collect indicators such as the prediction accuracy, inference latency, and CPU and memory usage of the model in real time. When anomalies are found, alarms are triggered in time to ensure the stable operation of the model.

[0043] Step S106, establish a cloud-edge collaborative data transmission and synchronization mechanism to reduce the amount of data transmission between the cloud and the edge, and ensure that the prediction models on the cloud and edge nodes are consistent.

[0044] On the edge node, the sliding window method is used to obtain new data, and the new data is transmitted to the cloud through the SSL protocol. After the cloud receives the new data, the random forest algorithm is used for incremental learning to obtain updated prediction model parameters. In order to protect data privacy, the homomorphic encryption algorithm is used to encrypt the original data collected by the edge node. The cloud receives the encrypted data and updates the model parameters using the additive homomorphic property. The parameter update can be completed without decryption. The updated encrypted parameters are transmitted back to the edge node, which is decrypted by the edge node and updates the local model.

[0045] For example, on the edge node, Python's Pandas library is used to obtain the latest sales data from the Kafka message queue, and the sliding window size is set to 30 minutes and the step length is 5 minutes. A secure communication channel with the cloud is established through the SSL protocol, and the data in the window is transmitted to the cloud in JSON format. The cloud uses the Flask framework to receive data, and uses the RandomForestRegressor class in the MLlib library of PySpark for incremental learning. The number of trees is set to 100, the maximum depth is 8, and the feature sampling ratio is 8. The original data collected by the edge node is encrypted using the Paillier homomorphic encryption algorithm, and the encrypted data size increases by 10 times. After the cloud receives the encrypted data, the model parameters are updated in the ciphertext space, and the gradient and weight are updated using homomorphic addition operations without decrypting the data. The updated encrypted parameters are transmitted back to the edge node through the SSL protocol, and the edge node uses the private key to decrypt and obtain the model parameters in plain text. The updated model parameters are applied to the local random forest model to achieve online update of the model. In order to ensure the security of the model update process, the SHA-256 algorithm is used to perform integrity verification on the transmitted data and parameters to prevent data tampering. At the same time, a VPN tunnel is established between the edge node and the cloud to encrypt the transmitted data to prevent data eavesdropping. The security of the system is improved by regularly rotating SSL certificates and VPN keys. Finally, the ELK log analysis system is used to monitor the operation of the system, collect system logs, performance indicators, security events and other data from edge nodes and the cloud in real time, display the system's operating status through a visual dashboard, and set alarm thresholds to promptly discover and deal with potential security risks.

[0046] It should also be noted that the various specific technical features described in the above specific embodiments can be combined in any suitable manner without contradiction. In order to avoid unnecessary repetition, the present invention will not further describe various possible combinations. In addition, the various different embodiments of the present invention can also be combined arbitrarily, as long as they do not violate the concept of the present invention, they should also be regarded as the contents disclosed by the present invention.

Claims

1. An e-commerce platform account management system based on cloud-edge collaboration, characterized in that: The e-commerce platform account splitting management system includes: Perform data cleaning and feature extraction on the historical sales data of the e-commerce platform to obtain a standardized sales data set with multi-dimensional features including user behavior, market trends, and promotional activities; Based on the normalized sales data set, a prediction model capable of capturing the inherent laws of sales data is constructed using the XGBoost algorithm; The prediction model is optimized by using model compression technology to obtain a lightweight prediction model to reduce the number of model parameters and the amount of calculation; Wherein, for the prediction model, the TensorFlowPruningAPI is used to perform pruning operations, the weight threshold is set to 0.1, and the weights in the prediction model whose absolute values ​​are less than the threshold are removed; Convert the pruned model to TensorFlowLite format through a TensorFlowLite converter, apply dynamic range quantization to it, and quantize the floating-point weights of the pruned model to 8-bit integers; The quantized prediction model is used as the student model, and the original uncompressed prediction model is used as the teacher model. The knowledge of the teacher model is transferred to the student model through the knowledge distillation API in TensorFlowModelOptimizationToolkit to obtain the compressed prediction model. Use the TensorFlowProfiler tool to analyze the computational graph of the compressed prediction model to identify computationally intensive operations and memory bottlenecks. According to the analysis results, the calculation graph of the compressed prediction model is optimized using TensorFlowGraphTransformTool, and unnecessary storage and transmission of intermediate results are reduced through operator fusion and constant folding to obtain an optimized model; Use TensorFlowLiteBenchmarkTool to perform performance tests on the optimized model on the target device to obtain its latency and memory usage indicators; According to the test results, the compression parameters of the prediction model are adjusted, and the best balance between model size, inference speed and accuracy is found through multiple rounds of iterative optimization; Sending the lightweight prediction model to edge nodes to perform real-time prediction at locations close to data sources; Wherein, a lightweight prediction model deployed on the edge node is used to perform local prediction on the real-time data to obtain a local prediction result; Calculating the confidence of the local prediction result, and if the confidence is lower than a preset threshold, transmitting the data to the cloud node through an encrypted channel; After receiving the low-confidence data, the cloud node uses a more complex and accurate prediction model to perform a secondary prediction to obtain a cloud prediction result; Returning the cloud prediction result to the edge node through an encrypted channel; The edge node obtains the final prediction result by using a weighted average method based on the local prediction result and the cloud prediction result; On the edge node, an online learning algorithm is used to update and tune the lightweight prediction model in real time to adapt to changes in sales data; Establish a cloud-edge collaborative data transmission and synchronization mechanism to reduce the amount of data transmission between the cloud and the edge, and ensure that the prediction models on the cloud and edge nodes are consistent.

2. The e-commerce platform account management system according to claim 1, wherein: The historical sales data of the e-commerce platform is cleaned and features are extracted to obtain a standardized sales data set containing multi-dimensional features of user behavior, market trends, and promotional activities, including: Obtain sales data for the past year from the database of the e-commerce platform, wherein the sales data includes product information, transaction amount, and user behavior; Importing the sales data into at least one data processing module, processing missing values ​​using the fillna function, and identifying and removing outliers using the IQR method to obtain preliminarily cleaned sales data; For the sales data after preliminary cleaning, the OneHotEncoder method is used to encode the category features of the product category and the promotion activity type to obtain the encoded sales data; Building a sales prediction model based on the encoded sales data, wherein the sales prediction model uses an XGBoost algorithm to process nonlinear relationships and feature interactions in the sales data; The sales forecasting model is deployed to the forecasting server, and a real-time sales forecasting function is provided through a RESTful API interface. According to the requested real-time sales data, the sales forecasting model is used to obtain the corresponding sales forecasting results, and the sales forecasting results are returned to the requesting end.

3. The e-commerce platform account management system according to claim 1, wherein: Based on the normalized sales data set, the XGBoost algorithm is used to build a prediction model that can capture the inherent laws of sales data, including: After obtaining the sales data set, preprocessing is performed on the data set, and the preprocessing includes: processing missing values ​​by using mean filling or deletion methods, identifying and processing outliers through box plots, and normalizing the data by using standardization methods; Use the trained model to predict the test set and evaluate the accuracy, precision, recall and F1 score of the model; If the performance of the model meets the preset requirements, it will be saved as the final model and applied to the actual sales forecasting task. Sales strategies will be formulated based on the forecast results, and the forecasting effect will be continuously tracked.

4. The e-commerce platform account splitting management system according to claim 1, wherein: The step of sending the lightweight prediction model to the edge node and performing real-time prediction at a location close to the data source includes: According to the final prediction result, a predefined judgment condition is used to generate a corresponding control instruction; The edge node sends the control instruction to the execution device through a dedicated interface; The cloud end periodically retrains the prediction model using newly added samples to update model parameters; The updated model parameters are distributed to each of the edge nodes for updating the lightweight prediction model.

5. The e-commerce platform account splitting management system according to claim 1, wherein: On the edge node, an online learning algorithm is used to update and tune the lightweight prediction model in real time to adapt to changes in sales data, including: Determine the deployment of the XGBoost algorithm according to the computing power of the edge node; The acquired sales data is preprocessed using a data cleaning tool, the preprocessing includes removing outliers, filling missing values ​​and standardizing; Divide the processed data set into a training set and a validation set, wherein the training set is used for initial training of the XGBoost model, and the validation set is used to evaluate the performance of the XGBoost model; When new sales data is received, the model is updated using an incremental learning method, specifically, the new data is added to the training set, and the model is retrained using the updated training set, while retaining the original model structure and parameters as initial values; In order to adapt to the storage limitation of the edge node, the model is compressed by using a model compression technology, and the model compression technology includes pruning; Obtaining the compressed model, and deploying it to the edge node through an incremental update mechanism; During model deployment, double buffering technology is used to achieve smooth switching between the old and new models; After deployment, a monitoring system is set up to track model performance and resource usage in real time.

6. The e-commerce platform account splitting management system according to claim 1, wherein: The data transmission and synchronization mechanism for cloud-edge collaboration is established to reduce the amount of data transmission between the cloud and the edge, and ensure that the prediction models on the cloud and edge nodes are consistent, including: On the edge node, a sliding window method is used to obtain new data, and the new data is transmitted to the cloud via the SSL protocol; After receiving the new data, the cloud uses the random forest algorithm to perform incremental learning to obtain updated prediction model parameters; In order to protect data privacy, a homomorphic encryption algorithm is used to encrypt the original data collected by the edge node; The cloud receives the encrypted data and uses the additive homomorphic property to update the model parameters without decryption. The updated encryption parameters are transmitted back to the edge node, which decrypts the parameters and updates the local model.

Images