Synchronous information processing method and system for multiple e-commerce platforms based on artificial intelligence

Abstract

The invention discloses a synchronous information processing method and system for multiple e-commerce platforms based on artificial intelligence. The method comprises the following steps: collecting historical data of each e-commerce platform; constructing a multi-task learning model; dividing the preprocessed historical data into a training set and a test set; training an LSTM model by using the training set data, then combining an LSTM prediction result with actual data, inputting the combined result into a multi-task learning model for training, outputting a likelihood for each anomaly by the multi-task learning model, and evaluating the performance of the multi-task learning model by using the test set data; and carrying out real-time anomaly detection, and carrying out synchronization strategy adjustment and inventory data synchronization. By means of the method, automatic data integration and synchronization of multiple e-commerce platforms are achieved, the processing efficiency and accuracy are improved, and meanwhile the method has the real-time data analysis capacity and the function of dynamically adjusting the synchronization strategy.

Description

TECHNICAL FIELD

[0001] The present application belongs to the technical field of e-commerce information processing, and particularly relates to a multi-e-commerce platform synchronization information processing method and system based on artificial intelligence. BACKGROUND

[0002] With the rapid development and popularization of Internet technology, e-commerce has been widely applied in the global scope. More and more merchants choose to operate on multiple e-commerce platforms to improve sales and expand market share. In the process of multi-platform operation, merchants need to monitor key indicators such as orders, inventory, return rate, cancellation rate in real time to ensure data consistency between platforms and adjust synchronization strategies in time to avoid the problem of insufficient or excessive inventory.

[0003] However, in the actual operation process, multi-platform operation brings many challenges. First, the data format and interface between different platforms may differ, making data integration and synchronization difficult. To solve this problem, merchants usually need to develop and maintain multiple customized data extraction and integration programs, which undoubtedly increases cost and complexity.

[0004] In addition, merchants need to develop flexible synchronization strategies based on real-time data to respond to market changes and abnormal situations. This may include real-time adjustment of synchronization frequency, optimization of inventory allocation, etc. However, traditional manual methods are difficult to meet these needs, especially in the case of multi-platform parallel operation. At this time, merchants need a multi-e-commerce platform synchronization information processing method based on artificial intelligence to improve data processing efficiency, reduce labor costs, and achieve real-time synchronization strategy adjustment.

[0005] In existing technologies, some methods can be used to process data of multiple e-commerce platforms. For example, data mining and statistical analysis methods can be used to analyze order data to discover potential trends and anomalies. However, these methods usually require manual setting of thresholds and rules, making it difficult to adapt to changing market environments. In addition, traditional methods have limited ability to process real-time data and dynamically adjust synchronization strategies, and traditional processing methods often focus on a single task such as predicting order quantity or detecting inventory anomalies, ignoring the correlation between multiple tasks. In addition, existing methods still have limited ability to process real-time data and dynamically adjust synchronization strategies, and cannot meet the needs of real-time monitoring and synchronization of multiple e-commerce platforms.

[0006] To solve these problems, a new multi-e-commerce platform synchronization information processing method and system need to be designed to improve data processing efficiency, reduce labor costs, and achieve real-time synchronization strategy adjustment, which has important practical significance and application value. SUMMARY

[0007] In view of the defects in the prior art, the present application provides a multi-ecomerce platform synchronization information processing method based on artificial intelligence, which comprises:

[0008] Collecting historical data of each e-commerce platform;

[0009] Building a multi-task learning model including a shared input layer, at least one shared hidden layer and independent task-specific output layers, setting different output layers for different anomaly types, and setting an independent output layer for each anomaly type;

[0010] Dividing the preprocessed historical data into a training set and a test set;

[0011] Training the LSTM model using the training set data, then combining the LSTM prediction results with the actual data, inputting them into the multi-task learning model for training, the multi-task learning model outputting a likelihood for each anomaly, and using the test set data to evaluate the performance of the multi-task learning model;

[0012] And real-time anomaly detection and synchronization strategy adjustment and inventory data synchronization.

[0013] Among them, the real-time anomaly detection and synchronization strategy adjustment and inventory data synchronization comprise:

[0014] Collecting real-time data from each e-commerce platform through distributed crawler technology, then combining the preprocessed real-time data and the prediction results of the LSTM model to build features, and inputting them into the trained multi-task learning model, the multi-task learning model outputting the likelihood of four anomaly types;

[0015] According to the likelihood of the four anomaly types, assigning a real-time requirement level to each anomaly type, and adjusting the synchronization strategy according to the real-time requirement level;

[0016] Under the adjusted synchronization strategy, realizing the synchronization of inventory data between each e-commerce platform.

[0017] Among them, the combination of preprocessed real-time data and LSTM model prediction results to build features comprises:

[0018] Using the trained LSTM model to predict the number of orders in the future period;

[0019] Collecting real-time data, including actual order quantity, inventory change, return and cancellation order data;

[0020] Calculating a plurality of features, and combining the plurality of features into a feature vector as an input vector of the multi-task learning model;

[0021] The plurality of features include: a percentage error between a predicted order quantity and an actual order quantity, a percentage of inventory change, a return rate, and a cancellation rate.

[0022] The assigning of the real-time requirement level to each of the four abnormal types according to the likelihood of the four abnormal types comprises:

[0023] The mapping of the likelihood of the four abnormal types to a real-time requirement level comprises mapping a higher likelihood to a higher real-time requirement level, and the mapping is using linear mapping or piecewise mapping.

[0024] A comprehensive real-time requirement level R is calculated, and the real-time requirement level R is a sum of the real-time requirement levels corresponding to the four abnormal types.

[0025] The synchronization frequency is adjusted according to the comprehensive real-time requirement level R.

[0026] The synchronization frequency=g(R), wherein g is a function for mapping the comprehensive real-time requirement level to the synchronization frequency, and a higher real-time requirement level is mapped to a higher synchronization frequency.

[0027] g(R)=base_frequency*(exp_factor^(R-1))

[0028] The base_frequency is a basic synchronization frequency, i.e., the synchronization frequency when the real-time requirement level is 1.

[0029] The exp_factor is an exponential factor for controlling the speed of increasing the synchronization frequency with the increase of the real-time requirement level, and the exp_factor>1.

[0030] The base_frequency is adjusted according to the predicted order quantity output by the trained LSTM model.

[0031] The base_frequency is calculated according to the following formula:

[0032] base_frequency=a*P_t^b*exp(c*delta_t)+d

[0033] P_t is the order quantity predicted by the LSTM model at time t; delta_t is the order quantity change trend predicted at time t; a, b, c, and d are adjustable parameters; and exp(x) is an exponential function representing e raised to the power of x.

[0034] Among them, the impact of different synchronization frequencies on actual business scenarios is simulated. In each experiment, different combinations of parameters a, b, c, and d are selected. By traversing the parameter space, a set of parameters a, b, c, and d is found to make the evaluation index optimal.

[0035] This invention also discloses a synchronous information processing system for multiple e-commerce platforms based on artificial intelligence, the system comprising:

[0036] The data acquisition module is used to crawl raw data from various e-commerce platforms using distributed crawler technology;

[0037] The data preprocessing module is used to receive raw data from the data acquisition module and perform preprocessing operations.

[0038] The data processing and synchronization module is used to synchronize data and store the processed and synchronized data in a relational database or a non-relational database.

[0039] The prediction and adjustment module is used to process the results obtained by the synchronization module based on the data and adjust the synchronization strategy accordingly. It is also used to train LSTM models and multi-task learning models, as well as to construct input features for multi-task learning models.

[0040] The API interface layer provides external access APIs for other systems or users to query data and adjust synchronization strategies.

[0041] The AI-based synchronous information processing method for multiple e-commerce platforms of this invention effectively learns and predicts sales data and trends from different e-commerce platforms by using LSTM and a multi-task learning model. This prediction method can more accurately predict future sales data based on historical data, thereby improving the targeting of synchronization strategies. Simultaneously, the multi-task learning model can effectively distinguish different types of anomalies, outputting a likelihood for each anomaly type, thus improving the accuracy of anomaly detection.

[0042] This invention assigns a real-time requirement level to each anomaly type based on the likelihood of the anomaly based on prediction results and real-time data. These real-time requirement levels allow for automatic adjustment of the synchronization strategy, achieving optimization under different conditions. This adaptive adjustment of the synchronization strategy can effectively cope with fluctuations in sales data across different e-commerce platforms, ensuring that the total inventory remains consistent across all platforms and reducing the risk of stockouts and overstocking.

[0043] This invention employs error handling to ensure stability and security during data synchronization. Error handling effectively addresses anomalies such as packet loss and delays during data transmission, preventing data inconsistencies or other problems. Attached Figure Description

[0044] The above and other objects, features, and advantages of exemplary embodiments of the present disclosure will become readily apparent upon reading the following detailed description with reference to the accompanying drawings. In the drawings, several embodiments of the present disclosure are illustrated by way of example and not limitation, and like or corresponding reference numerals denote like or corresponding parts, wherein:

[0045] Figure 1 This is a flowchart illustrating a method for synchronous information processing across multiple e-commerce platforms based on artificial intelligence, according to an embodiment of the present invention;

[0046] Figure 2 This is a schematic diagram illustrating a synchronous information processing system for multiple e-commerce platforms based on artificial intelligence according to an embodiment of the present invention. Detailed Implementation

[0047] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to the accompanying drawings. Obviously, the described embodiments are only a part of the embodiments of this invention, and not all of them. Based on the embodiments of this invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this invention.

[0048] The terminology used in the embodiments of this invention is for the purpose of describing particular embodiments only and is not intended to limit the invention. The singular forms “a,” “the,” and “the” as used in the embodiments of this invention and the appended claims are also intended to include the plural forms, and “multiple” generally includes at least two unless the context clearly indicates otherwise.

[0049] It should be understood that although the terms first, second, third, etc., may be used to describe... in the embodiments of the present invention, these... should not be limited to these terms. These terms are only used to distinguish... For example, first... may also be referred to as second... without departing from the scope of the embodiments of the present invention, and similarly, second... may also be referred to as first...

[0050] It should be understood that the term "and / or" used in this article is merely a description of the relationship between related objects, indicating that three relationships can exist. For example, A and / or B can represent: A existing alone, A and B existing simultaneously, and B existing alone. Additionally, the character " / " in this article generally indicates that the preceding and following related objects have an "or" relationship.

[0051] Depending on the context, the words “if” or “suppose” as used here can be interpreted as “when” or “in response to determination” or “in response to detection.” Similarly, depending on the context, the phrases “if determination” or “if detection (of the stated condition or event)” can be interpreted as “when determination” or “in response to determination” or “when detection (of the stated condition or event)” or “in response to detection (of the stated condition or event).”

[0052] It should also be noted that the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that an article or device that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such an article or device. Without further limitation, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the article or device that includes said element.

[0053] like Figure 1 As shown, this invention discloses a method for synchronous information processing across multiple e-commerce platforms based on artificial intelligence, comprising the following steps:

[0054] Step 1, Data Collection and Preprocessing: Collect historical data from various e-commerce platforms, such as key indicators like actual order quantity, inventory changes, returns, and order cancellations. Preprocess this data, including missing value imputation, outlier handling, and data scaling (normalization or standardization).

[0055] Step 2, Construct a multi-task learning model: Build a multi-task learning model, including a shared input layer, several shared hidden layers, and independent task-specific output layers. The shared layers are responsible for extracting general features, while the task-specific output layers are responsible for handling different types of anomalies. This includes setting different output layers for different anomaly types such as order surges, inventory crashes, abnormal returns, and abnormal cancellations, with an independent output layer for each anomaly type.

[0056] Step 3, Training the LSTM Model and the Multi-Task Learning Model: The preprocessed historical data is divided into training and test sets. First, the LSTM model is trained using the training set data. Then, the LSTM prediction results are combined with the preprocessed actual data to construct features, which are then input into the multi-task learning model for training. During training, the model learns to distinguish different types of anomalies and outputs a likelihood for each anomaly. To ensure the continuity of the time series, a rolling window method is used to divide the training and test sets.

[0057] Step 4, Multi-task learning model evaluation: Evaluate the performance of the multi-task learning model using test set data. The prediction accuracy and generalization ability for each task can be calculated to assess the model's performance on different anomaly types.

[0058] Step 5: Real-time anomaly detection, synchronization strategy adjustment, and inventory data synchronization:

[0059] Real-time data, including key metrics such as actual order quantity, inventory changes, returns, and order cancellations, is collected from various e-commerce platforms using distributed web crawling technology. The preprocessed real-time data is then combined with LSTM prediction results to construct features. These features are input into a trained multi-task learning model, which outputs likelihood scores for four anomaly types. A real-time requirement level is assigned to each anomaly type based on these likelihood scores, and the synchronization strategy is adjusted accordingly. Finally, under the adjusted synchronization strategy, inventory data is synchronized across the e-commerce platforms, ensuring consistent total inventory across all platforms and implementing error handling.

[0060] Long Short-Term Memory (LSTM) is a type of recurrent neural network designed to address the long-term dependency problem inherent in general RNNs (Recurrent Neural Networks). All RNNs have a chain-like structure of repeating neural network modules.

[0061] Among them, features are constructed based on LSTM prediction results and real-time data, including calculating the percentage difference between predicted and actual values, which is used as input features for anomaly detection.

[0062] The AI-based synchronous information processing method for multiple e-commerce platforms of this invention effectively learns and predicts sales data and trends from different e-commerce platforms by using LSTM and a multi-task learning model. This prediction method can more accurately predict future sales data based on historical data, thereby improving the targeting of synchronization strategies. Simultaneously, the multi-task learning model can effectively distinguish different types of anomalies, outputting a likelihood for each anomaly type, thus improving the accuracy of anomaly detection.

[0063] This invention assigns a real-time requirement level to each anomaly type based on the likelihood of the anomaly based on prediction results and real-time data. These real-time requirement levels allow for automatic adjustment of the synchronization strategy, achieving optimization under different conditions. This adaptive adjustment of the synchronization strategy can effectively cope with fluctuations in sales data across different e-commerce platforms, ensuring that the total inventory remains consistent across all platforms and reducing the risk of stockouts and overstocking.

[0064] This invention employs error handling to ensure stability and security during data synchronization. Error handling effectively addresses anomalies such as packet loss and delays during data transmission, preventing data inconsistencies or other problems.

[0065] In one embodiment, various errors and anomalies may occur during the process of synchronizing inventory data across multiple e-commerce platforms, including packet loss, delays, or inconsistencies in data formats between platforms. Error handling methods include retry mechanisms and data verification.

[0066] In one embodiment, missing values ​​and outliers need to be processed during the preprocessing stage, including missing value processing and outlier processing.

[0067] Missing values ​​refer to values ​​in certain fields of a dataset that are unknown or missing. Methods for handling missing values ​​include imputation or prediction.

[0068] Among them, filling in missing values ​​is based on the distribution characteristics of existing data. Filling includes mean filling, median filling, mode filling, and interpolation.

[0069] (1) Mean imputation: Fill missing values ​​with the mean of the data for this feature. This is suitable for continuous features, such as the price feature of the product.

[0070] (2) Median imputation: Fill missing values ​​with the median of the feature. This is suitable for features that are not sensitive to outliers, such as product attribute information (e.g., size information of clothing products).

[0071] (3) Mode imputation: Fill missing values ​​with the mode of the feature. Suitable for discrete data (e.g., color and style selection information for clothing products).

[0072] (4) Interpolation method: Interpolation calculations are performed based on known data to fill in missing values. It is suitable for continuous data, such as time series data of product orders.

[0073] This includes using existing data to build models to predict missing values, including using regression analysis for prediction.

[0074] Outliers are values ​​that deviate from the normal data range. The main methods for handling outliers are the replacement window method and the quantile method.

[0075] Replacement involves replacing outliers with values ​​within the normal range. Replacement methods can include the mean, median, and mode.

[0076] For time series data, a sliding window method can be used to detect outliers. The mean and standard deviation of the data are calculated within a fixed-size window. If the difference between a data point and the mean exceeds a certain multiple of the standard deviation (e.g., 3 times), it is considered an outlier.

[0077] Among them, the quantile method calculates the upper and lower quantiles of data (such as the 5th and 95th quantiles), and values ​​outside the quantiles are considered outliers.

[0078] In one embodiment, during training, to ensure the continuity of the time series, a rolling window method can be used to divide the training set and the test set. The rolling window method is a method for dividing time series data into training and test sets, ensuring the continuity and temporal order of the data. By dividing the time series data into multiple fixed-size windows, the rolling window method can train the model's performance in different time periods. It includes: (1) Selecting the window size (window_size): The window size refers to the length of the time series data contained in each window. The choice of window size depends on the problem itself and the characteristics of the data. Larger windows can capture longer time dependencies, but may lead to increased computational complexity; (2) Dividing the dataset: Dividing the entire time series data into multiple consecutive windows. The data in each window will be used as part of the training set and the test set; (3) Training and testing: For each window, the model is trained using the data within the window and tested on the data in the next window. This allows the model's generalization ability to be evaluated in different time periods; (4) Rolling the window: Scrolling the window backward by a certain time step, repeating steps (2) and (3) until the rolling window covers the entire time series.

[0079] The rolling window method can be implemented as follows: Use the pandas data processing library or other data processing libraries to read the time series data. Divide the data into multiple windows according to the selected window size. For example, you can use pandas' `rolling` function to create a rolling window view. For each window, use the data within that window as the training set, and the data in the next window as the test set. Maintain the temporal order. Train and test the model on each window, recording the model's performance metrics (such as MSE or MAE). Analyze the overall model performance across all windows to evaluate the model's overall generalization ability. The rolling window method allows for better evaluation of the time series model's performance across different time periods, avoiding the overfitting problem that may result from using fixed training and test sets.

[0080] In one embodiment, to convert the training and test set data into a format suitable for LSTM input, a suitable sliding window size T needs to be selected based on the characteristics of the problem and the length of the time dependencies that the model needs to capture. The time series data is traversed, extracting a subsequence of length T each time. For each subsequence, the first T-1 data points are used as input features, and the T-th data point is used as the target value; this process is repeated until the entire time series has been traversed.

[0081] To adapt to the input requirements of LSTM, the input features need to be adjusted to a 3D format. Specifically, the shape of the input feature matrix should be (n_samples, window_size, 1), where n_samples is the number of subsequences, window_size is the number of input features in each subsequence, and the last dimension is the number of feature variables (1 in this example, because only the order quantity feature is considered). In this embodiment, a subsequence is a set of data points within a continuous time period of length T. By extracting subsequences using a sliding window approach, the time series data can be converted into a format suitable for LSTM model input. The aforementioned process is applied to the training and test sets to generate a format suitable for LSTM input.

[0082] In one embodiment, recent order data is input into a trained LSTM model to predict the number of orders in the future. This includes using the trained LSTM model to predict the number of orders in the future after evaluating its performance. For this purpose, the recent order data needs to be preprocessed to meet the model's input requirements. The preprocessed recent order data is then input into the trained LSTM model. The model generates predictions based on the long-term dependencies learned from the time-series data. The predicted number of orders in the future (e.g., the number of orders in the next 7 days) is extracted from the model output. The prediction results need to be inversely scaled (e.g., inverse normalization or inverse standardization) to restore the scale of the original data. Through these steps, the trained LSTM model can be used to predict the number of orders in the future.

[0083] In one embodiment, the LSTM prediction results are combined with real-world data and input into a multi-task learning model for training. This includes constructing features based on the LSTM prediction results and real-time data, comprising the following steps:

[0084] Step s1: Use the trained LSTM model to predict the number of orders for a future period. For example, predict the number of orders for the next 7 days.

[0085] Step s2: Collect real-time data, such as the actual number of orders, inventory changes, returns, and order cancellations for the current 7 days.

[0086] Step s3, calculate the following features:

[0087] a. Percentage error between predicted order quantity and actual order quantity: error_order = (actual_order - predicted_order) / actual_order;

[0088] in,

[0089] actual_order: The actual number of orders, representing the total number of orders that actually occurred within a given time period;

[0090] predicted_order: The number of predicted orders, representing the total number of orders predicted based on the model within a given time period;

[0091] error_order: The percentage error between the actual number of orders and the predicted number of orders, used to measure the model's predictive performance.

[0092] b. Percentage change in inventory: inventory_change = (current_inventory - previous_inventory) / previous_inventory;

[0093] in,

[0094] current_inventory: Current inventory, representing the total inventory at a given moment;

[0095] previous_inventory: The inventory at the previous moment, representing the total inventory at the previous moment;

[0096] inventory_change: percentage change in inventory, indicating the degree of change in inventory between adjacent time points.

[0097] c. Return rate: return_rate=total_returned_orders / total_orders;

[0098] in,

[0099] total_returned_orders: Total number of returned orders, representing the total number of returned orders that occurred within a given time period;

[0100] total_orders: Total number of orders, representing the total number of orders placed within a given time period;

[0101] return_rate: Return rate, which represents the proportion of returned orders to the total number of orders.

[0102] d. Cancellation rate: cancel_rate=total_canceled_orders / total_orders;

[0103] in,

[0104] total_canceled_orders: Total number of canceled orders, representing the total number of orders canceled within a given time period;

[0105] total_orders: Total number of orders, representing the total number of orders placed within a given time period;

[0106] cancel_rate: Cancellation rate, representing the proportion of canceled orders to the total number of orders.

[0107] Step s4: Finally, these features are combined into a feature vector, which serves as the input vector x for the multi-task learning model.

[0108] x=[error_order, inventory_change, return_rate, cancel_rate].

[0109] Through the above steps, feature vectors are constructed based on LSTM prediction results and real-time data, which are used to train and predict multi-task learning models.

[0110] In one embodiment, the input feature vector x includes four features: order quantity error (error_order), inventory change percentage (inventory_change), return rate (return_rate), and cancellation rate (cancel_rate). A multi-task learning model is constructed based on the above input feature vector x, and the multi-task learning model includes:

[0111] (1) Input layer: The input layer receives the feature vector x. The size of the input layer is 4, corresponding to 4 features.

[0112] (2) Shared Hidden Layers: Two shared hidden layers are set up. The first hidden layer has 8 neurons, and the second hidden layer has 4 neurons. The ReLU activation function is used. The formula for calculating the hidden layers is as follows:

[0113] Fully connected layer 1: h1 = ReLU(W1*x + b1);

[0114] Fully connected layer 2: h2 = ReLU(W2*h1 + b2);

[0115] Where W1 and W2 are weight matrices, and b1 and b2 are bias terms.

[0116] (3) Task-specific output layer: Set up an independent output layer for each type of exception. A total of 4 output layers are set up, corresponding to order surge, inventory drop, abnormal return and abnormal cancellation respectively.

[0117] Output layer 1 (order surge): y1 = sigmoid(W3*h2+b3);

[0118] Output layer 2 (inventory crash): y2 = sigmoid(W4*h2 + b4);

[0119] Output layer 3 (abnormal returns): y3 = sigmoid(W5*h2+b5);

[0120] Output layer 4 (cancellation rate): y4 = sigmoid(W6*h2 + b6);

[0121] In this matrix, W3, W4, W5, and W6 are weight matrices, and b3, b4, b5, and b6 are bias terms. The sigmoid activation function is used to restrict the output value to between 0 and 1, representing the probability of an anomaly occurring.

[0122] The ReLU function is a piecewise linear function. When the input is greater than or equal to 0, the output equals the input; when the input is less than 0, the output is 0. Its main function is to map negative numbers to 0, thus introducing nonlinearity. The formula for the ReLU function is ReLU(x) = max(0,x).

[0123] The sigmoid function is a smooth, S-shaped curve function that maps input values ​​to the interval (0,1). Because the output range of the sigmoid function is between (0,1), it is often used to represent the output in probability or binary classification problems. The formula for the sigmoid function is sigmoid(x) = 1 / (1 + exp(-x)).

[0124] The fully connected layer 1 (h1) of the multi-task model has 8 neurons, so the weight matrix W1 is 8x4 (8 rows and 4 columns), and the bias term b1 is 8x1 (8 rows and 1 column). The input vector x is multiplied by W1, then b1 is added, and then the ReLU activation function is applied to obtain the output h1 of the fully connected layer 1.

[0125] The output h1 of fully connected layer 1 is used as the input of fully connected layer 2. Fully connected layer 2 has 4 neurons, so the weight matrix W2 is 4x8 (4 rows and 8 columns), and the bias term b2 is 4x1 (4 rows and 1 column). Multiplying h1 by W2 and then adding b2, followed by applying the ReLU activation function, yields the output h2 of fully connected layer 2.

[0126] The output of the shared hidden layer 2 (h2) is used as the input to each task-specific output layer. For each task-specific output layer (order surge, inventory collapse, abnormal returns, and abnormal cancellations), a weight matrix (W3, W4, W5, and W6) and a bias term (b3, b4, b5, and b6) are required. Assuming each task-specific output layer has one neuron, the weight matrix is ​​1x4 (1 row, 4 columns), and the bias term is 1x1 (1 row, 1 column). Multiplying h2 by the weight matrix, adding the corresponding bias term, and applying the sigmoid activation function yields the outputs y1, y2, y3, and y4 for each task-specific output layer. These outputs can be interpreted as the likelihood of anomalies occurring. The specific output layers for each task are: Output Layer 1 (Order Surge): y1 = sigmoid(W3*h2+b3); Output Layer 2 (Inventory Plunge): y2 = sigmoid(W4*h2+b4); Output Layer 3 (Abnormal Returns): y3 = sigmoid(W5*h2+b5); Output Layer 4 (Cancellation Rate): y4 = sigmoid(W6*h2+b6).

[0127] The sigmoid(Wi*h²+bi) process begins with a matrix multiplication of Wi*h². h² is a 4x1 (4 rows, 1 column) vector, and Wi is a 1x4 (1 row, 4 columns) matrix. The result of Wi*h² is a 1x1 (1 row, 1 column) scalar. Next, a bias term bi is added to this scalar. The bias term bi is also a 1x1 (1 row, 1 column) scalar. Finally, the sigmoid function is applied to this scalar. The sigmoid function takes this as input and outputs a value between 0 and 1. This value can be interpreted as the likelihood of the anomaly type corresponding to the task. In this example, there are four different output layers, each corresponding to one of the four anomaly types, resulting in likelihood values ​​for each anomaly type.

[0128] The output of the multi-task model is a vector containing four elements, each element representing the probability of an anomaly type. For example, the output vector [0.1, 0.8, 0.05, 0.15] represents a 10% probability of an order surge, an 80% probability of an inventory crash, a 5% probability of abnormal returns, and a 15% probability of a cancellation rate.

[0129] In one embodiment, training a multi-task learning model requires a training set containing a large number of labeled samples. The training process involves dividing the input data into a training set and a test set. The training set is used to train the model, while the test set is used to evaluate the model's performance. The multi-task learning model is trained using the training set data. In each iteration step, the model parameters are updated to minimize the loss function. During training, the model performance is monitored using a validation set. Training is stopped when the model performance no longer improves. The model performance is evaluated using the test set. Evaluation metrics such as accuracy, precision, recall, and F1 score are calculated between the predicted results and the actual labels.

[0130] In one embodiment, a loss function is defined for each task, including using a cross-entropy loss function to measure the difference between the predicted probability and the actual label. During training, the loss function for each task needs to be minimized. A multi-task learning model is trained using training set data. In each iteration step, the model parameters are updated using an optimization algorithm (such as stochastic gradient descent) to minimize the loss function. During training, model performance is periodically evaluated using validation set data. Training is stopped when the model performance no longer improves on the validation set (e.g., no improvement for N consecutive epochs). This prevents the model from overfitting.

[0131] In one embodiment, for a multi-class classification problem, the cross-entropy loss function is expressed as:

[0132] Loss = -∑[y_i*log(p_i)];

[0133] Where y_i is the actual label (0 or 1) of the i-th category, and p_i is the probability predicted by the model for the i-th category.

[0134] In multi-task learning, each task has an independent cross-entropy loss. To calculate the total loss, the losses from each task need to be summed.

[0135] During training, this total loss needs to be minimized. This can be achieved using the gradient descent algorithm (or its variants, such as stochastic gradient descent, Adam, etc.). The gradient descent algorithm adjusts the parameters based on the gradient of the loss function with respect to the model parameters. The parameter update formula is:

[0136]

[0137] Where θ represents the model parameters, and α is the learning rate (a hyperparameter that controls the step size for parameter updates). It is the gradient of the loss function L with respect to the parameter θ.

[0138] To calculate the gradients, the backpropagation algorithm can be used. Backpropagation starts from the output layer and propagates the error back along the network to the input layers, calculating the parameter gradients for each layer. After calculating all gradients, the parameters can be updated using the parameter update formula described above.

[0139] In multi-task learning, the loss function for each task produces a gradient with respect to the model parameters. The total gradient is the sum of the gradients from all tasks.

[0140] In one embodiment, in each training iteration, input data is passed to the model, and the predicted output for each task is calculated. Based on the predicted output and the actual labels, the cross-entropy loss for each task is calculated. The losses for each task are summed to obtain the total loss. The gradient of the model parameters is calculated based on the total loss. The model parameters are updated using the gradient descent algorithm. This process is repeated until the model converges or the number of iterations is reached.

[0141] In one embodiment, the activation values ​​h1 and h2 of the shared hidden layer, and the activation values ​​y1, y2, y3, and y4 of the task-specific output layer are calculated. For each task (order surge, inventory crash, abnormal returns, cancellation rate), the cross-entropy loss is calculated. Then, the losses of all tasks are summed to obtain the total loss L.

[0142] In one embodiment, backpropagation can be used to calculate the gradients. After calculating all gradients, the parameters are updated using a parameter update formula. The backpropagation and parameter update process is as follows:

[0143] Calculate the gradients of the task-specific output layers. For each output layer (y1, y2, y3, and y4), calculate its gradient with respect to the weight matrices (W3, W4, W5, and W6) and the bias terms (b3, b4, b5, and b6). Update the model parameters (W1, b1, W2, b2, W3, b3, W4, b4, W5, b5, and W6, b6) using gradient descent.

[0144] W_i=W_i-α*dL / dW_i b_i=b_i-α*dL / db_i

[0145] Where i represents the parameter index (1 to 6), and α is the learning rate.

[0146] Repeat this process (forward propagation, loss calculation, back propagation, parameter update) until the model converges or the number of iterations is reached.

[0147] In one embodiment, the gradient descent algorithm for updating model parameters is implemented using the deep learning framework PyTorch. The PyTorch framework provides corresponding tools and functions for forward propagation, loss calculation, backpropagation, and parameter updates. In PyTorch, various layers (nn.Linear), loss functions (nn.MSELoss), and optimizers (optim.SGD) are used to build and train multi-task learning models. PyTorch can automatically handle gradient calculation and parameter updates.

[0148] In one embodiment, a real-time requirement level is assigned to each of the four exception types based on the likelihood of those exceptions. The real-time requirement level indicates how quickly a response is required for each exception type. The synchronization strategy is then adjusted based on these real-time requirement levels.

[0149] The following real-time requirement levels are assigned to each exception type:

[0150] Order surge real-time requirement level r1: r1 = f(likelihood_order_increase);

[0151] Real-time requirement level r2 for inventory drop: r2 = f(likelihood_inventory_drop);

[0152] Real-time requirement level for abnormal returns r3: r3 = f(likelihood_return_rate);

[0153] Cancellation rate real-time requirement level r4: r4 = f(likelihood_cancellation_rate).

[0154] The parameters in f(likelihood_xxx) represent the likelihood values ​​output by the multi-task model for four anomaly types (order surge, inventory drop, abnormal returns, and abnormal cancellations). These likelihood values ​​measure the probability of each anomaly type occurring given real-time data. These likelihood outputs can be used to assign a real-time requirement level to each anomaly type. likelihood_order_increase, likelihood_inventory_drop, likelihood_return_rate, and likelihood_cancellation_rate represent the likelihood values ​​output by the multi-task model for the order surge, inventory drop, abnormal returns, and abnormal cancellation anomaly types, respectively.

[0155] Here, f is a function that maps likelihood to a real-time requirement level. Linear or piecewise mapping can be used to map higher likelihood to higher real-time requirement levels.

[0156] The synchronization strategy is adjusted according to the real-time requirement level, including calculating the comprehensive real-time requirement level R: R = r1 + r2 + r3 + r4.

[0157] The synchronization frequency is adjusted according to the overall real-time requirement level R: Synchronization frequency = g(R), where g is a function that maps the overall real-time requirement level to the synchronization frequency. Linear mapping, piecewise mapping, or nonlinear mapping can be used to map higher real-time requirement levels to higher synchronization frequencies.

[0158] The aforementioned synchronization frequency adjustment method considers the likelihood of each anomaly type and its impact on inventory and sales, adjusting the synchronization strategy by adjusting the real-time requirement level. In this way, when the likelihood of a certain anomaly type is high, the response to that anomaly type will be faster, thereby allowing for timely adjustments to the synchronization strategy to adapt to the actual situation.

[0159] In one embodiment, a piecewise linear mapping is used as the function f to map the likelihood to a level of real-time requirement. Here is a specific example:

[0160] Suppose there are 5 real-time requirement levels (1-5), where 1 represents the lowest real-time requirement and 5 represents the highest real-time requirement. The likelihood range (0 to 1) can be divided into 4 intervals, and then the likelihood can be mapped to the corresponding real-time requirement level according to the interval it belongs to.

[0161] When 0 ≤ likelihood < 0.25, f(likelihood) = 1;

[0162] When 0.25 ≤ likelihood < 0.5, f(likelihood) = 2;

[0163] When 0.5 ≤ likelihood < 0.75, f(likelihood) = 3;

[0164] When 0.75 ≤ likelihood < 1, f(likelihood) = 4.

[0165] When likelihood = 1, f(likelihood) = 5.

[0166] The function f described above maps the likelihood to an integer real-time requirement level between 1 and 5. A lower likelihood corresponds to a lower real-time requirement level, and vice versa.

[0167] The number of segments and real-time requirement levels in the above piecewise linear mapping can be adjusted according to the actual situation.

[0168] In one embodiment, the function g is a nonlinear model. For example, the function g is an exponential function that maps the overall real-time requirement level R to the synchronization frequency.

[0169] g(R)=base_frequency*(exp_factor^(R-1))

[0170] in:

[0171] base_frequency: The basic synchronization frequency, which is the synchronization frequency when the real-time requirement level is 1.

[0172] exp_factor: An exponential factor used to control the rate at which the synchronization frequency increases with the level of real-time requirements. exp_factor > 1.

[0173] Assuming the base synchronization frequency (base_frequency) is set to 1 (unit: times / hour) and the exponential factor (exp_factor) is set to 2, the synchronization frequency g(R) is calculated as follows for different comprehensive real-time requirement levels R (1 to 5):

[0174] g(1) = 1 * (2^0) = 1 time / hour

[0175] g(2)=1*(2^1)=2 times / hour

[0176] g(3) = 1 * (2^2) = 4 times / hour

[0177] g(4) = 1 * (2^3) = 8 times / hour

[0178] g(5) = 1 * (2^4) = 16 times / hour

[0179] The synchronization frequency increases exponentially with the level of comprehensive real-time requirements. When the real-time requirements are high, the synchronization frequency will also increase accordingly.

[0180] The base_frequency and exp_factor parameters mentioned above can be adjusted according to actual needs. Exponential functions can ensure real-time performance while avoiding excessive consumption of system resources due to excessively high synchronization frequencies.

[0181] In one embodiment, the base_frequency can be adjusted based on the predicted order volume output by the trained LSTM model, including correlating the base synchronization frequency with the fluctuation or trend of the predicted order volume. If the prediction results indicate that the order volume will continue to rise in the future, the base_frequency can be increased to ensure better data real-time performance.

[0182] When the likelihood of all exception types is equal to 0, it means that no exception has occurred, and synchronization can be performed according to the base synchronization frequency. When the likelihood of one or more exception types, such as order surges, inventory crashes, abnormal returns, and abnormal cancellations, increases, the closer it is to 1, the more likely the corresponding exception is to occur, and the required synchronization frequency will increase accordingly.

[0183] In one embodiment, the relationship between the current order volume and the trend of order volume changes and the base synchronization frequency (base_frequency) is as follows:

[0184] Assuming the order quantity (P_t) predicted by the LSTM model is already obtained, the base_frequency is calculated using the following formula:

[0185] base_frequency=a*P_t^b*exp(c*delta_t)+d;

[0186] in,

[0187] P_t: The number of orders predicted by the LSTM model at time t.

[0188] delta_t: The predicted trend of order volume change at time t.

[0189] a, b, c, d: Adjustable parameters that can be adjusted according to actual needs and data characteristics.

[0190] exp(x): Exponential function, representing e raised to the power of x.

[0191] In this formula, P_t^b*exp(c*delta_t) describes the impact of the current order quantity and its changing trend on the base synchronization frequency (base_frequency). Parameters a and d can be used to adjust the overall synchronization frequency range.

[0192] To calculate the order volume trend delta_t, the following method can be used: Predict the order volume for a future period (e.g., the next 7 days; the time period is related to the unit of prediction, such as 7 days for LSTM prediction). Calculate the difference in order volume between adjacent time points (e.g., the difference between t+1 and t, or between two adjacent 7-day periods). Calculate the proportion of the order volume difference (the difference divided by the order volume at time t, e.g., the order volume of the previous 7 days) to obtain delta_t.

[0193] To obtain suitable parameters a, b, c, and d, adjustments can be made based on actual business needs and datasets. For example, the optimal parameter combination can be found by simulating different synchronization frequencies and performance under real-world business scenarios.

[0194] In one embodiment, a series of experiments are designed to simulate the impact of different synchronization frequencies on real-world business scenarios. In each experiment, different combinations of parameters a, b, c, and d are selected, and their impact on the synchronization frequency is observed. To measure the effectiveness of different parameter combinations, several evaluation metrics need to be defined. These metrics may include data real-time performance, system load, order processing efficiency, etc. The evaluation metrics should fully reflect business requirements and performance goals.

[0195] By traversing the parameter space, a set of parameters (a, b, c, and d) is found that optimizes the evaluation metric. This can be achieved by manually tuning the parameters or by using optimization algorithms such as grid search, random search, or Bayesian optimization.

[0196] In one embodiment, the trained LSTM model can predict the order quantity (P_t) and obtain the trend of order quantity changes (delta_t). The optimal parameter combination obtained after the optimization process is: a = 1.0, b = 0.5, c = 0.8, d = 0.2.

[0197] To calculate the base synchronization frequency (base_frequency) at a given time t, we first need to use an LSTM model to predict the order volume P_t at time t. Assume that at time t, the LSTM model predicts an order volume P_t = 1000 (order volume is measured in weekly orders).

[0198] Calculate the order volume trend delta_t: Predict the order volume over a future period (the next 7 days). Calculate the difference in order volume between adjacent time points (the difference between t+1 and t). Calculate the proportion of the order volume difference (the difference divided by the order volume at time t). This yields delta_t; for example, an order volume trend delta_t = 0.05 (indicating a projected 5% increase in order volume).

[0199] Substitute the predicted values ​​P_t and delta_t into the formula:

[0200] base_frequency=a*P_t^b*exp(c*delta_t)+d=1.0*1000^0.5*exp(0.8*0.05)+0.2;

[0201] The calculated result is base_frequency≈31.75

[0202] In this example, the base_frequency is approximately 31.75. Depending on the specific business scenario, this value can be rounded or rounded up to determine the basic synchronization frequency. For the current calculation result, it can be rounded to 32.

[0203] In one embodiment, such as Figure 2 As shown, this invention discloses an artificial intelligence-based system for synchronous information processing across multiple e-commerce platforms. This system can adjust synchronization strategies in a timely manner based on real-time and predicted data, resulting in more timely and accurate information processing outcomes. The system includes the following modules:

[0204] (1) A data acquisition module is used to crawl raw data from various e-commerce platforms. Distributed crawling technology can be used, and the data acquisition module is deployed on multiple servers to improve data acquisition efficiency.

[0205] (2) Data preprocessing module, used to receive raw data from data acquisition module and perform preprocessing operations, such as missing value handling, outlier handling and data type conversion.

[0206] (3) Data processing and synchronization module, which is used to synchronize data from multiple e-commerce platforms according to the synchronization frequency in the synchronization strategy determined by the prediction and adjustment module, and store the processed and synchronized data in a relational database or a non-relational database.

[0207] (4) Prediction and adjustment module, used to organize the results obtained by the data synchronization module and adjust the synchronization strategy accordingly. The prediction and adjustment module is also used to train the LSTM model and the multi-task learning model. The prediction and adjustment module is also used to construct the input features of the multi-task learning model. The prediction and adjustment module is also used to calculate the basic synchronization frequency.

[0208] (5) API interface layer, which provides API interfaces for external access, allowing other systems or users to query data, adjust synchronization strategies, etc.

[0209] The entire system can adopt a microservice architecture, deploying different modules in independent containers, such as using Docker. This reduces coupling between modules and improves the system's scalability and fault tolerance. Furthermore, container orchestration tools like Kubernetes can be used for deployment and management, achieving automation, high availability, and elastic scaling.

[0210] The data acquisition module is used to perform... Figure 1 In step 1, the data collection part involves scraping raw data from various e-commerce platforms.

[0211] The data preprocessing module is used to perform... Figure 1 The data preprocessing part in step 1 performs preprocessing operations on the raw data.

[0212] The data processing and synchronization module is used to perform... Figure 1 Step 5 involves real-time anomaly detection, synchronization strategy adjustment, and inventory data synchronization. The data processing and synchronization module synchronizes the data based on the processing results of step 5 and stores the synchronized data in the database.

[0213] Among them, the prediction and adjustment module and Figure 1 In the process, steps 2, 3, and 4 correspond to the prediction and adjustment module, which trains an LSTM model based on historical data to predict key indicators such as order quantity, inventory, return rate, and cancellation rate. The prediction and adjustment module constructs features based on the LSTM prediction results and real-time data, and inputs these features into a multi-task learning model for training. The multi-task learning model can simultaneously handle different types of anomalies, including order surges, inventory crashes, abnormal returns, and abnormal cancellations. The prediction and adjustment module combines real-time data with the LSTM prediction results to construct features, which are then input into the trained multi-task learning model to obtain the likelihood of each anomaly type. Based on the output of the multi-task learning model and real-time data, the prediction and adjustment module assigns a real-time requirement level to each anomaly type. The synchronization strategy is adjusted according to these real-time requirement levels to ensure timely synchronization of inventory data between e-commerce platforms under various anomaly conditions.

[0214] The API interface layer provides an interface for other systems or users to access. Through the API interface, users can query data, adjust synchronization strategies, and so on. This module covers all parts of the entire system, enabling other systems or users to interact with this system.

[0215] In one embodiment, the prediction and adjustment module is used to adjust the synchronization strategy based on the results obtained by the data processing and synchronization module. The prediction and adjustment module includes the following sub-modules:

[0216] The LSTM prediction model submodule is used to predict the number of orders over a future period based on preprocessed data.

[0217] The anomaly detection and multi-task learning model submodule is used to detect different types of anomalies and output a likelihood for each anomaly.

[0218] The synchronization frequency adjustment model submodule is used to adjust the synchronization frequency based on the output of the LSTM prediction model and the anomaly detection model using the f and g(R) functions, and returns the adjusted synchronization frequency.

[0219] The distributed task scheduling and coordination module is used to coordinate and schedule the computing tasks of various modules, ensuring that no processing conflicts occur during multi-platform computing and adjustment. It receives task requests through API interfaces and distributes tasks to the corresponding modules.

[0220] In one embodiment, the distributed task scheduling and coordination module is deployed on a cloud server or local server, responsible for coordinating and scheduling the computing tasks of various modules to ensure that no processing conflicts occur during multi-platform computing and adjustment. An API interface is designed to receive task requests and distribute tasks to the appropriate modules. To avoid data collection and processing conflicts between multiple e-commerce platforms, the distributed task scheduling and coordination module can adopt the following design strategy:

[0221] Message Queue: Use message queues as a buffer for tasks on various platforms.

[0222] Distributed Lock: Use distributed locks to ensure that only one platform's data collection and processing tasks are executed at any given time.

[0223] Concurrency control: In the distributed task scheduling and coordination module, an appropriate number of worker threads or processes are set according to the number of platforms. Each worker thread or process is responsible for handling the data collection and processing tasks of one platform. By limiting the number of concurrent processes, processing conflicts can be avoided.

[0224] In one embodiment, the data structures and formats differ across platforms, necessitating data cleaning, integration, and standardization. To address these differences, a data preprocessing module is used to handle data cleaning, integration, and standardization. This module can be deployed as a standalone service on a cloud server or a local server. The distributed task scheduling and coordination module can call the data preprocessing module via API, passing the collected raw data for processing. After processing, the preprocessing module returns the cleaned, integrated, and standardized data to the scheduling module for further model training and prediction. To ensure that data processing on different platforms does not conflict, message queues, distributed locks, or concurrency control strategies can be used for task scheduling to guarantee that only one platform's data processing task is executed at any given time, thus avoiding processing conflicts.

[0225] In one embodiment, a data processing and synchronization module is used to synchronize data and store the processed and synchronized data in a relational database or a non-relational database. For time-series data, the data is sorted according to time order. The cleaned and processed data is synchronized to a unified data storage system, including data merging, deduplication, and data association operations. The synchronized data is then stored in a relational database (such as MySQL) or a non-relational database (such as Redis).

[0226] It should be noted that the computer-readable medium described in this disclosure can be a computer-readable signal medium or a computer-readable storage medium, or any combination thereof. A computer-readable storage medium can be, for example,—but not limited to—an electrical, magnetic, optical, electromagnetic, infrared, or semiconductor system, apparatus, or device, or any combination thereof. More specific examples of a computer-readable storage medium may include, but are not limited to: an electrical connection having one or more wires, a portable computer disk, a hard disk, random access memory (RAM), read-only memory (ROM), erasable programmable read-only memory (EPROM or flash memory), optical fiber, portable compact disk read-only memory (CD-ROM), optical storage device, magnetic storage device, or any suitable combination thereof. In this disclosure, a computer-readable storage medium can be any tangible medium containing or storing a program that can be used by or in connection with an instruction execution system, apparatus, or device. In this disclosure, a computer-readable signal medium can include a data signal propagated in baseband or as part of a carrier wave, carrying computer-readable program code. Such propagated data signals can take various forms, including but not limited to electromagnetic signals, optical signals, or any suitable combination thereof. A computer-readable signal medium can be any computer-readable medium other than a computer-readable storage medium, which can send, propagate, or transmit a program for use by or in connection with an instruction execution system, apparatus, or device. The program code contained on the computer-readable medium can be transmitted using any suitable medium, including but not limited to: wires, optical fibers, RF (radio frequency), etc., or any suitable combination thereof.

[0227] The aforementioned computer-readable medium may be included in the aforementioned electronic device; or it may exist independently and not assembled into the electronic device.

[0228] Computer program code for performing the operations of this disclosure can be written in one or more programming languages ​​or a combination thereof, including object-oriented programming languages ​​such as Java, Smalltalk, and C++, and conventional procedural programming languages ​​such as the "C" language or similar programming languages. The program code can be executed entirely on the user's computer, partially on the user's computer, as a standalone software package, partially on the user's computer and partially on a remote computer, or entirely on a remote computer or server. In cases involving remote computers, the remote computer can be connected to the user's computer via any type of network—including a local area network (LAN) or a wide area network (WAN)—or can be connected to an external computer (e.g., via the Internet using an Internet service provider).

[0229] The flowcharts and block diagrams in the accompanying drawings illustrate the architecture, functionality, and operation of possible implementations of systems, methods, and computer program products according to various embodiments of this disclosure. In this regard, each block in a flowchart or block diagram may represent a module, segment, or portion of code containing one or more executable instructions for implementing a specified logical function. It should also be noted that in some alternative implementations, the functions indicated in the blocks may occur in a different order than those indicated in the drawings. For example, two consecutively indicated blocks may actually be executed substantially in parallel, and they may sometimes be executed in reverse order, depending on the functions involved. It should also be noted that each block in the block diagrams and / or flowcharts, and combinations of blocks in the block diagrams and / or flowcharts, can be implemented using a dedicated hardware-based system that performs the specified function or operation, or using a combination of dedicated hardware and computer instructions.

[0230] The units described in the embodiments of this disclosure can be implemented in software or hardware. The names of the units are not, in some cases, intended to limit the specific unit.

[0231] The preferred embodiments of the present invention have been described above to make the spirit of the present invention clearer and easier to understand, and are not intended to limit the present invention. All modifications, substitutions, and improvements made within the spirit and principles of the present invention should be included within the protection scope summarized by the appended claims.

Claims

1. A method for synchronous information processing across multiple e-commerce platforms based on artificial intelligence, the method comprising: Collect historical data from various e-commerce platforms; Construct a multi-task learning model, including a shared input layer, at least one shared hidden layer, and independent task-specific output layers. Set different output layers for different anomaly types, and set an independent output layer for each anomaly type. The preprocessed historical data is divided into training and testing sets. The LSTM network model is trained using training set data. The LSTM prediction results are then combined with the actual data and input into a multi-task learning model for training. The multi-task learning model outputs a likelihood for each anomaly. The performance of the multi-task learning model is evaluated using test set data. It also includes real-time anomaly detection and synchronization strategy adjustments, as well as inventory data synchronization.

2. The method for synchronous information processing across multiple e-commerce platforms based on artificial intelligence as described in claim 1, characterized in that, The real-time anomaly detection and synchronization strategy adjustment and inventory data synchronization include: Real-time data is collected from various e-commerce platforms using distributed crawling technology. Then, the preprocessed real-time data and the prediction results of the LSTM model are combined to construct features, which are then input into a trained multi-task learning model. The multi-task learning model will output the likelihood of four anomaly types. Assign a real-time requirement level to each of the four anomaly types based on their likelihood, and adjust the synchronization strategy according to the real-time requirement level. Under the adjusted synchronization strategy, inventory data synchronization between various e-commerce platforms is achieved.

3. The method for synchronous information processing across multiple e-commerce platforms based on artificial intelligence as described in claim 2, characterized in that, The process of combining preprocessed real-time data with the prediction results of the LSTM model to construct features includes: Use the trained LSTM model to predict the number of orders for a future period of time; Collect real-time data, including actual order quantity, inventory changes, returns, and order cancellations; Calculate multiple features and combine the multiple features into a feature vector, which is used as the input vector of a multi-task learning model; The features include: percentage error between predicted order quantity and actual order quantity, percentage change in inventory, return rate, and cancellation rate.

4. The method for synchronous information processing across multiple e-commerce platforms based on artificial intelligence as described in claim 2, characterized in that, The process of assigning a real-time requirement level to each of the four anomaly types based on their likelihood includes: Mapping the likelihood of the four anomaly types to a real-time requirement level includes mapping higher likelihoods to higher real-time requirement levels, wherein the mapping is performed using a linear mapping or a piecewise mapping. Calculate the overall real-time requirement level R, which is the sum of the real-time requirement levels corresponding to the four anomaly types.

5. The method for synchronous information processing across multiple e-commerce platforms based on artificial intelligence as described in claim 4, characterized in that, Adjust the synchronization frequency according to the comprehensive real-time requirement level R; Synchronization frequency = g(R), where g is a function used to map the overall real-time requirement level to the synchronization frequency; where a higher real-time requirement level is mapped to a higher synchronization frequency.

6. The method for synchronous information processing across multiple e-commerce platforms based on artificial intelligence as described in claim 5, characterized in that, g(R)=base_frequency*(exp_factor^(R-1)) Among them, base_frequency is the base synchronization frequency, which is the synchronization frequency when the real-time requirement level is 1; exp_factor is an exponential factor used to control the rate at which the synchronization frequency increases with the level of real-time requirements, wherein exp_factor>1.

7. The method for synchronous information processing across multiple e-commerce platforms based on artificial intelligence as described in claim 6, characterized in that, The base synchronization frequency (base_frequency) is adjusted based on the predicted order volume output by the trained LSTM model.

8. The method for synchronous information processing across multiple e-commerce platforms based on artificial intelligence as described in claim 7, characterized in that, The base frequency is calculated using the following formula: base_frequency=a*P_t^b*exp(c*delta_t)+d; Where P_t is the order quantity predicted by the LSTM model at time t; delta_t is the trend of the predicted order quantity at time t; a, b, c, and d are adjustable parameters; exp(x) is an exponential function, representing e raised to the power of x.

9. The method for synchronous information processing across multiple e-commerce platforms based on artificial intelligence as described in claim 8, characterized in that, The impact of different synchronization frequencies on real-world business scenarios is simulated. In each experiment, different combinations of parameters a, b, c, and d are selected. By traversing the parameter space, a set of parameters a, b, c, and d is found that optimizes the evaluation metrics.

10. A synchronous information processing system for multiple e-commerce platforms based on artificial intelligence, the system comprising: The data acquisition module is used to crawl raw data from various e-commerce platforms using distributed crawler technology; The data preprocessing module is used to receive raw data from the data acquisition module and perform preprocessing operations. The data processing and synchronization module is used to synchronize data and store the processed and synchronized data in a relational database or a non-relational database. The prediction and adjustment module is used to process the results obtained by the synchronization module based on the data and adjust the synchronization strategy accordingly. It is also used to train LSTM models and multi-task learning models, as well as to construct input features for multi-task learning models. The API interface layer provides external API interfaces for other systems or users to query data and adjust synchronization strategies.

Images