Cross-border e-commerce intelligent product selection system and method based on multi-platform data fusion

Abstract

The invention relates to the field of data processing, and provides a cross-border e-commerce intelligent product selection system and method based on multi-platform data fusion, and the method comprises the steps: carrying out the data collection and preprocessing of a plurality of cross-border e-commerce platforms, constructing a neural network model through optimal hyper-parameters, training, and generating a selected product recommendation list. Meanwhile, the structural parameters and the search direction of the selected product model are continuously updated, and the selected product recommendation list is dynamically updated. According to the system, multi-platform commodity data are integrated, versioning feature snapshots are constructed, hyper-parameter selection of a multi-snapshot version evaluation strategy optimization model with a consistent time sequence is adopted, the generalization ability of a commodity selection model to a time-varying data environment is enhanced, the possibility that the accuracy of the model is reduced due to the change of data distribution is reduced, and the accuracy of the commodity selection model is improved. Meanwhile, dynamic updating of the selected product recommendation list is achieved, and it is ensured that the recommendation strategy can be fully matched with preference migration of the selected product trend.

Description

TECHNICAL FIELD

[0001] The present application relates to the field of data processing, in particular to a cross-border e-commerce intelligent product selection system based on multi-platform data fusion. BACKGROUND

[0002] In the development of cross-border e-commerce, data-driven intelligent product selection has gradually become a key technical means to improve operational efficiency. The field generally adopts a method of multi-platform data fusion analysis combined with machine learning, which integrates product information, user behavior and transaction data from multiple e-commerce platforms, builds feature engineering and establishes a prediction model to assist merchants in product selection decision-making.

[0003] The existing scheme is generally based on multi-source heterogeneous data fusion, and after time series feature extraction, a prediction model is built by using statistical learning or neural network method. Its essence depends on the assumption of static data distribution, and its model optimization process is based on data snapshots of a specific time section, but due to the characteristics of data distribution evolution over time in the actual business environment, the product selection model faces the problem of degradation of generalization ability.

[0004] In a non-stationary data environment, model parameter optimization is easily affected by data noise in a specific time period, resulting in overfitting phenomenon, and the model selection mechanism lacks evaluation dimensions for time domain robustness, relying only on single time point validation set performance. The traditional hyperparameter search method neither considers the time sequence change characteristics of data distribution in the model structure optimization level, nor can it find a solution with time domain robustness in a high-dimensional parameter space. Therefore, a technical solution is needed that can explicitly model the change of data time distribution and guide the optimization direction based on it to improve the robustness and adaptability of the product selection model in a non-stationary environment. SUMMARY

[0005] The purpose of the present application is to propose a cross-border e-commerce intelligent product selection system based on multi-platform data fusion to solve one or more technical problems existing in the prior art, and at least to provide a beneficial choice or create conditions.

[0006] The present application provides a cross-border e-commerce intelligent product selection system and method based on multi-platform data fusion, which collects and preprocesses data from multiple cross-border e-commerce platforms, uses optimal hyperparameters to build a neural network model and trains it, generates a product selection recommendation list, and continuously updates the structure parameters and search direction of the product selection model to dynamically update the product selection recommendation list. The system integrates multi-platform product data and builds versioned feature snapshots, uses a time-consistent multi-snapshot version evaluation strategy to optimize the selection of model hyperparameters, enhances the generalization ability of the product selection model to time-varying data environment, reduces the possibility of decline in model accuracy due to changes in data distribution, and simultaneously realizes dynamic updating of the product selection recommendation list, ensuring that the recommendation strategy can fully match the preference migration of product selection trends.

[0007] In order to achieve the above-mentioned purpose, according to an aspect of the present disclosure, a cross-border e-commerce intelligent selection system based on multi-platform data fusion is provided, which comprises a processor, a memory and a computer program stored in the memory and running on the processor, wherein the processor executes the modules running by the computer program, and the modules comprise:

[0008] A data acquisition module is configured to acquire data from multiple cross-border e-commerce platforms and preprocess the acquired data to obtain a structured data set.

[0009] A data processing module is configured to perform feature engineering on the structured data set and generate a global snapshot according to a time version and store it in a feature database.

[0010] A model presetting module is configured to initialize an optimizer and output optimal hyperparameters based on the feature database.

[0011] A model training module is configured to use the optimal hyperparameters to construct and train a neural network model to obtain a selection model.

[0012] A model calculation module is configured to input feature data into the selection model for forward calculation and output a selection recommendation list.

[0013] A model updating module is configured to continuously update the structure parameters and search direction of the selection model and dynamically update the selection recommendation list.

[0014] The system is run in a computing device of a host, a desktop computer, a notebook computer, a palm computer and a cloud data center.

[0015] Further, the step of acquiring data from multiple cross-border e-commerce platforms comprises using a web crawler to acquire product data from public web pages of cross-border e-commerce platforms or obtaining product data through official API interfaces of cross-border e-commerce platforms; wherein the product data comprises product basic data, dynamic transaction data, user interaction data and reputation evaluation data.

[0016] Further, the data acquisition from multiple cross-border e-commerce platforms is in a continuous manner.

[0017] Further, the step of preprocessing the acquired data to obtain a structured data set comprises data cleaning of the acquired product data, unification of data mode and scale of the cleaned data, construction of derived features and extraction of text semantic features, and generation of a structured data set for model training.

[0018] Further, the step of feature engineering on the structured dataset and generating global snapshots by time version and storing into the feature database comprises: feature construction, feature screening and feature coding on the structured dataset based on business rules and model requirements, integrating the processed feature data by fixed time period or event trigger, generating global feature snapshots with time stamp, and persistently storing into the feature database.

[0019] The feature database continuously receives and stores the newly generated global feature snapshots for supporting time series analysis, model training and incremental update.

[0020] Further, the step of initializing the optimizer and outputting the optimal hyperparameters based on the feature database comprises: selecting a hyperparameter optimization algorithm, using the historical global feature snapshots in the feature database as training data, and using the prediction accuracy of the selection target as the objective function, and outputting the hyperparameter vector combination that optimizes the performance of the objective function through multiple rounds of iterative search.

[0021] Optionally, Bayesian optimization is used as the hyperparameter optimization algorithm, the past 12 months of global feature snapshots stored in the feature database are used as training and validation data, the F1-Score of the selection model on the validation set is maximized as the objective function, the preset hyperparameter space includes neural network learning rate, number of hidden layers, number of neurons per layer, Dropout ratio and batch training size, and the optimizer performs iterative search within the preset hyperparameter space to output a set of optimal hyperparameter vectors.

[0022] Further, the step of constructing a neural network model and training using the optimal hyperparameters to obtain a selection model comprises: configuring the model structure and training parameters of the neural network model with the optimal hyperparameter vector, instantiating a multilayer perceptron model, using the historical global feature snapshots in the feature database as the training set, and training the model to minimize the cross-entropy loss between the predicted value and the true value; during model training, the predicted value is calculated by forward propagation, the network weights are updated by back propagation, overfitting is prevented by early stopping method, and the trained selection model is output after training convergence.

[0023] Further, the step of inputting feature data into the selection model for forward calculation and outputting a selection recommendation list comprises: preprocessing the feature data of the target product list and inputting it into the trained selection model, calculating the selection score of each product through the multilayer forward propagation of the selection model, sorting all products in the target product list in descending order according to the selection score, and selecting the top ranked specified number of products to generate a structured selection recommendation list.

[0024] Further, the step of dynamically updating the product selection recommendation list by continuously updating the structure parameters and search direction of the product selection model is as follows: K latest feature snapshots are taken from the feature database, a set of N individuals (N individuals in the set) is obtained, each individual is a set of hyperparameter vectors, the fitness score of each individual is calculated in turn, a plurality of individuals with the highest fitness score are marked as secondary individuals, the structure parameters and search direction of the product selection model are updated using the secondary individuals, and the product selection recommendation list is regenerated (trained) using the updated product selection model to realize dynamic updating of the product selection recommendation list.

[0025] Optionally, the number of selected secondary individuals is an integer in the interval [N / 4, N / 2].

[0026] Since the performance of different hyperparameters on each feature snapshot is different, selecting secondary individuals only according to the comprehensive fitness score may lead the search direction to a local optimum, so the distribution of individual sub-scores on different features needs to be considered.

[0027] Preferably, the selection method of the secondary individuals can also be: let ind(j) represent the jth individual in the individual set, and let z(j, k) represent the sub-score of ind(j) on the kth feature snapshot, then the K sub-scores of ind(j) on the K feature snapshots are z(j, 1), z(j, 2), …, z(j, K), the K sub-scores are grouped into an array and used as the integrated array INA(j) of ind(j); each secondary individual corresponds to an integrated array.

[0028] The value obtained by (1÷S0(j))+(1÷S1(j)) is used as the aggregation score of ind(j), and a plurality of individuals with the highest aggregation score are marked as secondary individuals.

[0029] Wherein, the calculation method of S0(j) is: subtract each value in the array INA(j) from the maximum value in the array INA(j) in turn to obtain K first difference values, and accumulate the K first difference values as S0(j).

[0030] The calculation method of S1(j) is: Ave j represents the average value of all values in the array INA(j), b1(j) represents the number of values in the array INA(j) that are less than Ave j , Ave j ×(1+(b1(j)÷K)) is used as R1(j), and the sum of the values in the array INA(j) that belong to the interval [R0(j), R1(j)] is used as S1(j), and R0(j) is the minimum value in the array INA(j).

[0031] The beneficial effect of this step is that: since the timeliness of e-commerce platform data is strong, the data distribution changes greatly with time evolution, and the selection model is significantly affected by the dynamic data environment. For an individual that performs stably on multiple feature snapshots, its long-term performance and generalization ability are usually better than another individual that performs extremely well on a single feature snapshot but has greater performance fluctuations on other versions; therefore, the method of this step quantitatively evaluates the robustness of individual performance from two dimensions through the index of aggregation score. Obviously, an individual that performs well and stably on multiple feature snapshots will have a relatively small S0(j) of its integrated array (the value will be relatively large for an individual with large fluctuations), so the (1 ÷ S0(j)) term will convert this bias into a positive indicator. However, if an individual performs poorly and stably on multiple feature snapshots, its corresponding S0(j) will also be relatively small, and relying solely on S0(j) cannot effectively filter the target individual. Therefore, S1(j) is introduced to calculate the score sum of the individual in the lower performance interval. When the score sum is small, it means that there are fewer low scores in the integrated array of the individual. Therefore, (1 ÷ S1(j)) has a higher reward weight. By adding the two reciprocal values to obtain the aggregation score, individuals with low volatility and high consistency can be rewarded, thereby selecting a combination of secondary individuals that perform stably and excellently on multiple snapshot versions, so that the search direction of the iteration process can better adapt to changes in the data environment, guiding the model to continuously improve the accuracy and stability of product selection recommendations.

[0032] Further, the step of obtaining the individual set is: defining a local sampling range based on the optimal hyperparameters, and performing regular sampling in the local sampling range to obtain N groups of hyperparameter vectors (N individuals) to form the individual set.

[0033] Optionally, the regular sampling is Latin hypercube sampling.

[0034] Optionally, the method of defining the local sampling range is: for the optimal hyperparameters P = [p1, p2, …, pD], D is the dimension of the hyperparameters, define the perturbation ratio h as a value in the interval [0.1, 0.5], and take [P*(1-h), P*(1+h)] as the local sampling range.

[0035] Further, the step of calculating the fitness score of the individual is: training K multilayer perceptron models using the individual on K latest feature snapshots, each feature snapshot corresponds to a multilayer perceptron model, calculating the sub-score of the corresponding multilayer perceptron model through the validation set in the feature snapshot, and aggregating K sub-scores to obtain the fitness score of the individual.

[0036] Optionally, the sub-score is defined as the negative of the loss function value or the F1-Score, and the aggregation method for aggregating K sub-scores is weighted average or accumulation.

[0037] Further, the step of updating the structure parameters and search direction of the product selection model using the secondary individuals is: fusing the plurality of secondary individuals (corresponding to the hyperparameter vectors) to obtain a consensus hyperparameter vector, and using the consensus hyperparameter vector to reconfigure the model structure of the neural network model; taking all the secondary individuals as a sample set in the hyperparameter space, and adjusting the search direction of the product selection model (i.e. adjusting the local sampling range of the next iteration) according to the sample set, so that the performance of the product selection model is continuously improved.

[0038] Optionally, the plurality of secondary individuals are fused by weighted averaging, clipped averaging or weighted random sampling.

[0039] Preferably, all undefined variables in the present disclosure, if not specifically defined, can be threshold values set by humans.

[0040] The present application also provides a cross-border e-commerce intelligent product selection method based on multi-platform data fusion, which comprises the following steps:

[0041] Data is collected from a plurality of cross-border e-commerce platforms, and the collected data is preprocessed to obtain a structured data set;

[0042] Feature engineering is performed on the structured data set, and a global snapshot is generated according to the time version and stored in a feature database;

[0043] An optimizer is initialized, and the optimal hyperparameters are output based on the feature database;

[0044] A neural network model is constructed using the optimal hyperparameters and trained to obtain a product selection model;

[0045] Feature data is input into the product selection model for forward calculation, and a product selection recommendation list is output;

[0046] The structure parameters and search direction of the product selection model are continuously updated, and the product selection recommendation list is dynamically updated.

[0047] The present application has the following advantages: the system integrates multi-platform product data and constructs versioned feature snapshots, adopts a time-consistent multi-snapshot version evaluation strategy to optimize the selection of hyperparameters of the model, enhances the generalization ability of the product selection model to time-varying data environment, reduces the possibility of decline in model accuracy caused by changes in data distribution, simultaneously realizes dynamic updating of the product selection recommendation list, and ensures that the recommendation strategy can fully match the preference migration of the product selection trend. BRIEF DESCRIPTION OF DRAWINGS

[0048] The above and other features of the present invention will become more apparent from the detailed description of the embodiments shown in conjunction with the accompanying drawings. In the accompanying drawings, the same reference numerals denote the same or similar elements. Obviously, the drawings described below are merely some embodiments of the present invention. For those skilled in the art, other drawings can be obtained from these drawings without any creative effort. In the drawings:

[0049] Figure 1 The flowchart shown is a process for a cross-border e-commerce intelligent product selection method based on multi-platform data fusion.

[0050] Figure 2 The diagram shows the system architecture of a cross-border e-commerce intelligent product selection system based on multi-platform data fusion. Detailed Implementation

[0051] The following will provide a clear and complete description of the concept, specific structure, and technical effects of the present invention in conjunction with the embodiments and accompanying drawings, so as to fully understand the purpose, solution, and effects of the present invention. It should be noted that, unless otherwise specified, the embodiments and features described in this application can be combined with each other.

[0052] In the description of this invention, "several" means one or more, "multiple" means two or more, "greater than," "less than," and "exceeding" are understood to exclude the stated number, while "above," "below," and "within" are understood to include the stated number. The use of "first" and "second" in the description is merely for distinguishing technical features and should not be construed as indicating or implying relative importance, or implicitly indicating the number of indicated technical features, or implicitly indicating the order of the indicated technical features.

[0053] like Figure 1 The diagram shown is a flowchart of a cross-border e-commerce intelligent product selection method based on multi-platform data fusion according to the present invention. The following is a combination of... Figure 1 This paper describes an intelligent product selection method for cross-border e-commerce based on multi-platform data fusion according to an embodiment of the present invention.

[0054] This invention proposes a smart product selection method for cross-border e-commerce based on multi-platform data fusion, the method comprising the following steps:

[0055] Data was collected from multiple cross-border e-commerce platforms and preprocessed to obtain a structured dataset;

[0056] Perform feature engineering on the structured dataset and generate global snapshots by time version and store them in the feature database;

[0057] Initialize the optimizer and output the optimal hyperparameters based on the feature database;

[0058] The neural network model is constructed and trained using the optimal hyperparameters to obtain a product selection model;

[0059] The feature data is input into the product selection model for forward calculation, and a product selection recommendation list is output;

[0060] The structure parameters and search direction of the product selection model are continuously updated, and the product selection recommendation list is dynamically updated.

[0061] Further, the method for collecting data from multiple cross-border e-commerce platforms is as follows: using a web crawler to collect product data from the public web pages of the cross-border e-commerce platforms, or obtaining product data through the official API interface of the cross-border e-commerce platforms; wherein the product data includes product basic data, dynamic transaction data, user interaction data, and reputation evaluation data.

[0062] Further, the data collection from multiple cross-border e-commerce platforms is in a continuous manner.

[0063] Specifically, the product basic data includes product identification, category attributes, title description, image data, and specification parameters; the dynamic transaction data includes real-time price, historical sales, inventory status, and platform promotion information; the user interaction data includes exposure, click volume, purchase behavior, and conversion rate indicators; and the reputation evaluation data includes ratings, comment text, and merchant service indicators.

[0064] Further, the method for preprocessing the collected data to obtain a structured data set is as follows: cleaning the collected product data, unifying the data mode and scale of the cleaned data, constructing derived features, and extracting text semantic features to generate a structured data set for model training.

[0065] Further, the method for feature engineering on the structured data set and generating a global snapshot according to the time version and storing it in the feature database is as follows: based on business rules and model requirements, feature construction, feature selection, and feature coding are performed on the structured data set, the processed feature data is integrated according to a fixed time period or event trigger, a global feature snapshot with a time stamp is generated, and is persistently stored in the feature database.

[0066] The feature database continuously receives and stores the newly generated global feature snapshot, which is used to support time series analysis, model training, and incremental updating.

[0067] Specifically, the business rules and model requirements are obtained from an expert system, the business rules refer to the quantitative decision criteria defined by expert knowledge in the field of cross-border e-commerce product selection, and the model requirements refer to the technical requirements of the selected machine learning algorithm on the format and quality of the input data.

[0068] Further, the method for initializing the optimizer based on the output optimal hyperparameters in the feature database is specifically: selecting a hyperparameter optimization algorithm, using the historical global feature snapshots in the feature database as training data, and using the prediction accuracy of the selection target as the objective function, and outputting the hyperparameter vector combination that optimizes the performance of the objective function through multiple rounds of iterative search.

[0069] Further, the method for constructing and training the neural network model using the optimal hyperparameters to obtain the selection model is specifically: configuring the model structure and training parameters of the neural network model with the optimal hyperparameter vector, instantiating a multilayer perceptron model, using the historical global feature snapshots in the feature database as the training set, and performing model training to minimize the cross-entropy loss between the predicted value and the true value; during model training, the predicted value is calculated by forward propagation, the network weights are updated by back propagation, and the early stopping method is used to prevent overfitting, and the trained selection model is output after training convergence.

[0070] Specifically, the Adam optimization algorithm is used for model training.

[0071] Further, the method for inputting feature data into the selection model for forward calculation and outputting the selection recommendation list is: pre-processing the feature data of the target product list and inputting it into the trained selection model, calculating the selection score of each product through the multilayer forward propagation of the selection model, sorting all products in the target product list in descending order according to the selection score, and selecting the top ranked products to generate a structured selection recommendation list.

[0072] Further, the method for dynamically updating the selection recommendation list by continuously updating the structure parameters and search direction of the selection model is specifically: taking K latest feature snapshots from the feature database, obtaining a set of N individuals (the set contains N individuals), each individual being a set of hyperparameter vectors, calculating the fitness score of each individual in turn, selecting multiple individuals with the highest fitness score as secondary individuals, updating the structure parameters and search direction of the selection model using the secondary individuals, and generating a new selection recommendation list using the updated selection model to realize dynamic updating of the selection recommendation list.

[0073] The structure parameters of the selection model are the model structure of the selection model (neural network model), and the model structure is composed of multiple parameters.

[0074] Specifically, the number of secondary individuals selected is N / 4.

[0075] Since the performance of different hyperparameters on each feature snapshot is different, relying only on the comprehensive fitness score to select secondary individuals may lead to a local optimal search direction, so the distribution of individual sub-scores on different features needs to be considered.

[0076] Specifically, the selection method of the secondary individual can also be: let ind(j) represent the jth individual in the individual set, j is a serial number, j = 1, 2, …, N, let z(j, k) be the sub-score of ind(j) on the kth feature snapshot, then the K sub-scores of ind(j) on the K feature snapshots are z(j, 1), z(j, 2), …, z(j, K), and the K sub-scores are grouped into an array and taken as the integrated array INA(j) of ind(j); each secondary individual corresponds to an integrated array;

[0077] The sub-score z(j, k) of ind(j) on the kth feature snapshot refers to training a multilayer perceptron model on the kth feature snapshot by using ind(j), and calculating the sub-score of the multilayer perceptron model by the validation set in ind(j);

[0078] Taking the value obtained by (1÷S0(j))+(1÷S1(j)) as the degree score of ind(j), and marking the multiple individuals with the highest degree score as secondary individuals;

[0079] Wherein, the calculation method of S0(j) is: subtract each value in the array INA(j) from the maximum value in the array INA(j) in turn, to obtain K first difference values, and accumulate the K first difference values as S0(j);

[0080] The calculation method of S1(j) is: Ave j represents the average value of all values in the array INA(j), and b1(j) represents that the array INA(j) contains b1(j) values less than Ave j , Ave j ×(1+(b1(j)÷K)) is taken as R1(j), and the sum of the values in the array INA(j) belonging to the interval [R0(j), R1(j)] is taken as S1(j), and R0(j) is the minimum value in the array INA(j).

[0081] Further, the method for obtaining the individual set is: taking the optimal hyperparameter as a reference point, defining a local sampling range, and performing regular sampling in the local sampling range, thereby obtaining N groups of hyperparameter vectors (N individuals) to constitute the individual set.

[0082] Further, the method for calculating the fitness score of the individual is: training K multilayer perceptron models using the individual on K latest feature snapshots respectively, each feature snapshot corresponds to a multilayer perceptron model, calculating the sub-score of the corresponding multilayer perceptron model by the validation set in the feature snapshot, and aggregating the K sub-scores to obtain the fitness score of the individual.

[0083] Specifically, a sub-score is defined as the negative of the loss function value, and the aggregation method used to aggregate K sub-scores is a weighted average.

[0084] Furthermore, the method for updating the structural parameters and search direction of the product selection model using secondary individuals is as follows: multiple secondary individuals (corresponding hyperparameter vectors) are fused to obtain a set of consensus hyperparameter vectors, and the model structure of the neural network model is reconfigured using the consensus hyperparameter vectors; all secondary individuals are used as a sample set in the hyperparameter space, and the search direction of the product selection model is adjusted according to the sample set (i.e., the local sampling range of the next iteration is adjusted) so that the performance of the product selection model is continuously improved.

[0085] Specifically, the method for adjusting the search direction of the product selection model based on the sample set is as follows: construct a covariance matrix C with all secondary individuals, and the eigenvectors and eigenvalues ​​of matrix C determine the new search direction;

[0086]

[0087] Where μ is the number of second-order individuals, x i Let m represent the i-th second-level individual. new This represents the consensus hyperparameter vector.

[0088] The eigenvectors of the covariance matrix C represent the main directions of the second-order individual distribution, and the magnitude of the eigenvalues ​​corresponds to the step size that needs to be explored in these directions.

[0089] The intelligent product selection system for cross-border e-commerce based on multi-platform data fusion includes: a processor, a memory, and a computer program stored in the memory and executable on the processor. When the processor executes the computer program, it implements the steps in the above-described embodiment of the intelligent product selection method for cross-border e-commerce based on multi-platform data fusion. The intelligent product selection system for cross-border e-commerce based on multi-platform data fusion can run on computing devices such as desktop computers, laptops, mobile phones, handheld phones, tablets, PDAs, and cloud data centers. The runnable system may include, but is not limited to, processors, memory, and server clusters.

[0090] The embodiments of the present invention provide a cross-border e-commerce intelligent product selection system based on multi-platform data fusion, such as... Figure 2 As shown, the intelligent product selection system for cross-border e-commerce based on multi-platform data fusion in this embodiment includes: a processor, a memory, and a computer program stored in the memory and executable on the processor. When the processor executes the computer program, it implements the steps in the above-described embodiment of the intelligent product selection method for cross-border e-commerce based on multi-platform data fusion. The processor executes the computer program in the following system units:

[0091] The data collection module is configured to collect data from multiple cross-border e-commerce platforms and preprocess the collected data to obtain a structured data set.

[0092] The data processing module is configured to perform feature engineering on the structured data set and generate a global snapshot according to a time version and store the global snapshot in a feature database.

[0093] The model presetting module is configured to initialize an optimizer and output optimal hyperparameters based on the feature database.

[0094] The model training module is configured to construct a neural network model and train the neural network model using the optimal hyperparameters to obtain a product selection model.

[0095] The model calculation module is configured to input feature data into the product selection model for forward calculation and output a product selection recommendation list.

[0096] The model updating module is configured to continuously update structure parameters and search directions of the product selection model and dynamically update the product selection recommendation list.

[0097] The cross-border e-commerce intelligent product selection system based on multi-platform data fusion can run in a host computer, a desktop computer, a notebook computer, a palm computer, a cloud data center and the like. The cross-border e-commerce intelligent product selection system based on multi-platform data fusion includes, but is not limited to, a processor and a memory. Those skilled in the art can understand that the example is only an example of the cross-border e-commerce intelligent product selection method and system based on multi-platform data fusion, and does not constitute a limitation on the cross-border e-commerce intelligent product selection method and system based on multi-platform data fusion. The cross-border e-commerce intelligent product selection method and system based on multi-platform data fusion can include more or fewer components, or combine certain components, or different components, for example, the cross-border e-commerce intelligent product selection system based on multi-platform data fusion can also include an input / output device, a network access device, a bus and the like.

[0098] The processor can be a central processing unit (CPU), and can also be other general-purpose processors, digital signal processors (DSP), application specific integrated circuits (ASIC), field programmable gate arrays (FPGA) or other programmable logic devices, discrete component gate circuits or transistor logic devices, discrete hardware components, etc. The general-purpose processor can be a microprocessor or any conventional processor, etc. The processor is a control center of the cross-border e-commerce intelligent product selection system based on multi-platform data fusion, and is connected to each sub-region of the cross-border e-commerce intelligent product selection system based on multi-platform data fusion through various interfaces and lines.

[0099] The memory can be used to store the computer program and / or modules, and the processor realizes various functions of the cross-border e-commerce intelligent selection method and system based on multi-platform data fusion by running or executing the computer program and / or modules stored in the memory and calling the data stored in the memory. The memory can mainly include a program storage area and a data storage area, wherein the program storage area can store an operating system, at least one application program required by a function (such as a sound playing function, an image playing function, etc.), and the like; and the data storage area can store data (such as audio data, a phonebook, etc.) created according to the use of the mobile phone, and the like. In addition, the memory can include a high-speed random access memory, and can also include a non-volatile memory, for example, a hard disk, a memory, a plug-in hard disk, a smart media card (SMC), a secure digital (SD) card, a flash card, at least one disk storage device, a flash memory device, or other volatile solid-state memory devices.

[0100] The present application provides a cross-border e-commerce intelligent selection system and method based on multi-platform data fusion. The system collects and preprocesses data from multiple cross-border e-commerce platforms, uses optimal hyperparameters to construct and train a neural network model, generates a selection recommendation list, and continuously updates the structure parameters and search direction of the selection model to dynamically update the selection recommendation list. The system integrates multi-platform product data and constructs versioned feature snapshots, uses a time-consistent multi-snapshot version evaluation strategy to optimize the selection of model hyperparameters, enhances the generalization ability of the selection model to time-varying data environments, reduces the possibility of a decline in model accuracy due to changes in data distribution, and simultaneously realizes dynamic updating of the selection recommendation list to ensure that the recommendation strategy can fully match the preference migration of selection trends. Although the description of the present application has been quite detailed and specifically described the embodiments, it is not intended to be limited to any of these details or embodiments or any special embodiment, thereby effectively covering the intended scope of the present application. In addition, the present application is described above in embodiments that the inventors can foresee, and the purpose is to provide a useful description, and non-substantial modifications to the present application that have not yet been foreseen can still represent equivalent modifications to the present application.

Claims

1. A cross-border e-commerce intelligent product selection system based on multi-platform data fusion, characterized in that: The system includes: a processor, a memory, and a computer program stored in the memory and running on the processor. The module on the processor that executes the computer program includes: The data acquisition module is used to collect data from multiple cross-border e-commerce platforms and preprocess the collected data to obtain a structured dataset. The data processing module is used to perform feature engineering on structured datasets and generate global snapshots according to time versions and store them in the feature database. The model pre-configuration module initializes the optimizer and outputs the optimal hyperparameters based on the feature database. The model training module is used to build and train a neural network model using optimal hyperparameters to obtain the product selection model. The model calculation module is used to input feature data into the product selection model for forward calculation and output a product recommendation list; The model update module continuously updates the structural parameters and search direction of the product selection model, and dynamically updates the product recommendation list. The system runs on mainframes, laptops, handheld computers, and cloud data centers.

2. The intelligent product selection system for cross-border e-commerce based on multi-platform data fusion as described in claim 1, characterized in that, The specific steps for performing feature engineering on structured datasets and generating global snapshots based on time versions and storing them in the feature database are as follows: Based on business rules and model requirements, feature construction, feature filtering and feature encoding are performed on the structured datasets. The processed feature data is integrated according to a fixed time period or event triggering to generate a global feature snapshot with a timestamp, and persistently stored in the feature database.

3. The intelligent product selection system for cross-border e-commerce based on multi-platform data fusion as described in claim 1, characterized in that, The steps for initializing the optimizer and outputting the optimal hyperparameters based on the feature database are as follows: using historical global feature snapshots in the feature database as training data, using the prediction accuracy of the product selection target as the objective function, and searching through multiple rounds to output the hyperparameter vector combination that optimizes the performance of the objective function.

4. The intelligent product selection system for cross-border e-commerce based on multi-platform data fusion as described in claim 1, characterized in that, The specific steps for constructing and training a neural network model using optimal hyperparameters to obtain a product selection model are as follows: configure the structure and training parameters of the neural network model with the optimal hyperparameter vector, instantiate a multilayer perceptron model, use historical global feature snapshots in the feature database as the training set, and train the model with the goal of minimizing the cross-entropy loss between the predicted value and the true value. During model training, the predicted values ​​are calculated through forward propagation, and the network weights are updated through backpropagation. At the same time, the early stopping method is used to prevent overfitting. After the training converges, the product selection model is output.

5. The intelligent product selection system for cross-border e-commerce based on multi-platform data fusion as described in claim 1, characterized in that, The steps to input feature data into the product selection model for forward computation and output a product recommendation list are as follows: After preprocessing the feature data of the target product list, input it into the trained product selection model. Through multi-layer forward propagation computation of the product selection model, obtain the product selection score of each product. Sort all products in the target product list in descending order according to the product selection score, and select a specified number of products with the highest ranking to generate a structured product recommendation list.

6. The intelligent product selection system for cross-border e-commerce based on multi-platform data fusion as described in claim 1, characterized in that, The steps for continuously updating the structural parameters and search direction of the product selection model and dynamically updating the product recommendation list are as follows: retrieve the K latest feature snapshots from the feature database, obtain a set of individuals of length N, where each individual in the set is a set of hyperparameter vectors, select secondary individuals, use the secondary individuals to update the structural parameters and search direction of the product selection model, and use the updated product selection model to regenerate the product recommendation list, thereby realizing the dynamic updating of the product recommendation list.

7. The intelligent product selection system for cross-border e-commerce based on multi-platform data fusion as described in claim 6, characterized in that, The steps for selecting secondary individuals are as follows: calculate the fitness score of each individual in turn, and select the individuals with the highest fitness scores to be marked as secondary individuals; The steps for calculating an individual's fitness score are as follows: train K multilayer perceptron models on K latest feature snapshots using the individual, with each feature snapshot corresponding to a multilayer perceptron model. Calculate the sub-score of the corresponding multilayer perceptron model using the validation set in the feature snapshot, and aggregate the K sub-scores to obtain the individual's fitness score.

8. The intelligent product selection system for cross-border e-commerce based on multi-platform data fusion according to claim 6, characterized in that, The steps for updating the structural parameters and search direction of the product selection model using secondary individuals are as follows: merge multiple secondary individuals to obtain a set of consensus hyperparameter vectors, and use the consensus hyperparameter vectors to reconfigure the model structure of the neural network model; treat all secondary individuals as a sample set in the hyperparameter space, and adjust the search direction of the product selection model according to the sample set.

9. A method for intelligent product selection in cross-border e-commerce based on multi-platform data fusion, applicable to any one of claims 1-8, characterized in that, The method includes: Data was collected from multiple cross-border e-commerce platforms and preprocessed to obtain a structured dataset; Perform feature engineering on the structured dataset and generate global snapshots by time version and store them in the feature database; Initialize the optimizer and output the optimal hyperparameters based on the feature database; The optimal hyperparameters are used to construct and train a neural network model to obtain the product selection model; Input the feature data into the product selection model for forward calculation and output a product recommendation list; We continuously update the structural parameters and search direction of the product selection model, and dynamically update the product recommendation list.

Images