Cross-border e-commerce intelligent product selection system based on artificial intelligence

Abstract

The invention discloses a cross-border e-commerce intelligent product selection system based on artificial intelligence, and the system comprises a data collection and preprocessing module which is used for collecting original data of a cross-border e-commerce platform and carrying out the processing operation; the parameter sampling module is used for sampling and generating a structure parameter coding vector set; the candidate model construction module is used for constructing a deformable attention neural network model; the model output module is used for outputting a commodity sales prediction value and a selected commodity recommendation probability; the multi-task loss evaluation module is used for calculating a multi-task fitness score; the fitness driving optimization module is used for selecting elite individuals and updating evolutionary strategy parameters; the evolution control module is used for executing an iterative optimization process; the optimal model extraction module is used for constructing a final deformable attention neural network model; and the recommended selected product output module is used for outputting a commodity recommendation result set. According to the invention, collaborative output and performance optimization of commodity sales prediction and intelligent commodity selection judgment results in a cross-border e-commerce scene are realized.

Description

Technical Field

[0001] The present invention relates to the field of artificial intelligence and e-commerce technology, and in particular to an artificial intelligence-based cross-border e-commerce intelligent product selection system. Background Art

[0002] With the rapid development of global e-commerce, cross-border e-commerce has become a crucial component of global trade. Numerous e-commerce platforms connect consumers and suppliers across different countries and regions, enabling international collaboration in terms of products, traffic, logistics, and payments. Faced with a vast array of product categories, rapidly evolving user demands, a complex and diverse market structure, and highly competitive product listing strategies, platforms and merchants are increasingly relying on intelligent and automated product selection technologies. In particular, in scenarios such as personalized marketing, targeted recommendations, and supply chain pre-positioning, intelligently predicting a product's future sales performance and determining whether it should be recommended for listing based on multi-dimensional data such as historical sales data, market trends, user behavior, and product attributes has become a critical component of current cross-border e-commerce intelligent operations systems.

[0003] Currently, mainstream cross-border e-commerce product selection strategies primarily include rule-based expert systems, statistical model-driven evaluation mechanisms, and, in recent years, predictive product selection models based on machine learning and deep learning. Rule-based systems rely on manually formulated selection criteria, such as category popularity, rating thresholds, sales cycles, and return rates, statically evaluating products using thresholds based on selected indicators. While logically clear and easy to understand, these approaches are less responsive to dynamic market environments and fail to effectively capture nonlinear relationships between features. Statistical modeling methods, such as those based on linear regression, logistic regression, or Bayesian analysis, can incorporate more variables for comprehensive evaluation, but their expressive power is limited, resulting in poor performance with heterogeneous data and complex behavioral sequences.

[0004] In recent years, with the widespread application of deep neural networks, researchers have gradually introduced multi-layer perceptrons (MLPs), convolutional neural networks (CNNs), long short-term memory networks (LSTMs), and attention-based models (such as the Transformer) into e-commerce product selection tasks. By constructing nonlinear mappings between product features and historical behavior, they aim to improve the accuracy of sales forecasts and recommendation decisions. These methods are often based on supervised learning frameworks, using product features as input to build sales regression models and recommendation classification models, respectively. However, existing methods generally suffer from three shortcomings when faced with multi-task collaboration in real-world business scenarios.

[0005] First, most current product selection models use a single-task modeling approach, constructing separate sales prediction models and product selection judgment models, each trained and deployed independently. This approach results in a lack of shared representations between models, making it difficult to leverage sales features to enhance recommendation judgments or optimize sales modeling using recommendation tags. More critically, the two models lack coordination during the optimization process, potentially achieving high performance on one task while performing poorly on the other, reducing the stability and explainability of the system's overall decision-making. In particular, in scenarios involving "high sales but no recommendations" or "low sales but strong recommendations," the lack of a collaborative optimization mechanism can lead to conflicting model judgments, impacting the actual effectiveness of the product selection system.

[0006] Secondly, the structure of traditional deep models is typically fixed, and key structural parameters such as model depth, number of attention heads, and feature fusion strategies are manually set before deployment based on experience. This manual structure setting method lacks adaptability and is difficult to dynamically optimize based on data distribution, sample characteristics, and task complexity. Some studies have attempted to use automatic adjustment strategies such as hyperparameter search and network structure search (NAS), but most remain within single-task models and have yet to develop a unified structure search mechanism suitable for multi-task and multimodal fusion scenarios. For tasks such as cross-border e-commerce product selection, which involve complex input structures such as text categories, numerical attributes, and behavioral sequences, the ability of the model structure to adapt is particularly critical.

[0007] Third, existing multi-task modeling methods often employ static loss weighting strategies, such as combining classification and regression losses with fixed weights for joint training. This ignores the nonlinear coupling characteristics of the objective functions between tasks. The contributions of different tasks to the network gradient may conflict with each other, easily leading to gradient fluctuations and learning instability during training, ultimately impacting the model's generalization and deployment reliability. Furthermore, the model's final output is often used directly for decision-making, lacking an error feedback mechanism. This can lead to deviations from actual demand, particularly in situations of sample imbalance, severe target offset, or significant market fluctuations, leading to biased recommendations and even commercial risks.

[0008] To alleviate these issues, some studies have introduced swarm intelligence optimization algorithms (such as genetic algorithms and particle swarm optimization) to assist in model parameter optimization and structural adjustment. Although these algorithms have certain advantages in solving complex optimization problems, they suffer from issues such as high randomness in evolutionary paths, slow convergence, and sensitive parameter adjustments, making them difficult to adapt to the high-frequency iteration and multi-objective real-time optimization requirements of e-commerce. In contrast, the covariance matrix adaptation evolutionary strategy (CMA-ES), a continuous space optimization method, demonstrates good robustness and adaptability in structural parameter adjustment and multi-objective model evolution, making it suitable for intelligent model structure search and performance adjustment in e-commerce. However, there is currently a lack of systematic methods that organically combine CMA-ES with deep attention models for cross-border e-commerce product selection tasks.

[0009] Furthermore, the current deformable attention mechanism is primarily used in computer vision and image modeling. It can dynamically adjust the attention area within the feature space, enabling location awareness and multi-scale modeling capabilities. In cross-border e-commerce data, different product attributes, user behaviors, and market feedback signals are heterogeneous and time-sensitive. Introducing a deformable attention mechanism can help improve the model's ability to focus on key feature areas and enhance the model's understanding of dynamic product selection patterns. However, no research has yet deeply integrated this mechanism with product selection prediction tasks, especially technical solutions combining structural parameter optimization with dual-branch task modeling.

[0010] Therefore, how to provide an artificial intelligence-based cross-border e-commerce intelligent product selection system is an urgent problem that technicians in this field need to solve. Summary of the Invention

[0011] One purpose of the present invention is to propose a cross-border e-commerce intelligent product selection system based on artificial intelligence. The present invention fully integrates the modeling advantages of the covariance matrix adaptive evolution strategy and the deformable multi-head attention mechanism. By constructing a multi-task deep neural network model with evolving structural parameters and feedback of task outputs, the collaborative learning and optimization feedback of commodity sales prediction and product selection judgment are realized. The method has an end-to-end closed-loop design in key links such as structural parameter sampling, model construction, performance evaluation, fitness feedback, strategy update and final model reasoning, supporting the automatic evolution of model structure and the adaptive balance of task weights, and providing a deployable, iterative and generalizable intelligent product selection solution for e-commerce platforms.

[0012] An artificial intelligence-based cross-border e-commerce intelligent product selection system according to an embodiment of the present invention includes:

[0013] Data collection and preprocessing module, used to collect raw data from cross-border e-commerce platforms and process the raw data;

[0014] Parameter sampling module, used to sample and generate a set of structural parameter encoding vectors based on the covariance matrix adaptation evolution strategy;

[0015] Candidate model building module for building deformable attention neural network models;

[0016] Model output module, used to output product sales forecast value and product recommendation probability;

[0017] Multi-task loss evaluation module, used to construct a joint fitness function and calculate multi-task fitness scores;

[0018] Fitness-driven optimization module, used to select elite individuals and update the covariance matrix to adapt to the evolution strategy parameters;

[0019] Evolution control module, used to execute the iterative optimization process;

[0020] The optimal model extraction module is used to select the optimal structural parameter encoding vector for fitness and construct the final deformable attention neural network model;

[0021] The recommended product selection output module is used to load the final deformable attention neural network model and output a set of product recommendation results.

[0022] Optionally, modules can be connected using the following methods:

[0023] S1. Collect cross-border e-commerce data, construct the original data set, pre-process the original data set, and extract the product selection input feature set;

[0024] S2. Initialize the covariance matrix to adapt to the evolutionary strategy optimizer, set the initial parameters, and generate a structural parameter encoding population based on the initial parameter sampling to construct a set of structural parameter encoding vectors;

[0025] S3. Based on the set of structural parameter encoding vectors, a deformable attention neural network model is constructed. The product selection input feature set is input into the deformable attention neural network model, and the sales prediction results and product selection judgment results are output.

[0026] S4. Based on the sales forecast results and product selection judgment results, the prediction error and classification accuracy are calculated respectively, and a joint fitness function is constructed;

[0027] S5. Update the covariance matrix based on the joint fitness function to adapt the evolutionary strategy optimizer and sample a new generation of structural parameter encoding vector sets;

[0028] S6. Based on the new generation of structural parameter encoding vector sets, the model construction, model training, result output and fitness score calculation processes are repeatedly executed until the fitness change amplitude meets the convergence condition, and the final optimal structural parameter encoding vector is output;

[0029] S7. Based on the final optimal structural parameter encoding vector, select the final deformable attention neural network model and extract the intelligent product selection output results.

[0030] Optionally, the original data set includes structured attribute data, category information data, market environment data, and user behavior data.

[0031] Optionally, the preprocessing includes normalizing the structured attribute data, embedding the category information data, numerically quantifying and feature-normalizing the market environment data, and slicing the user behavior data in chronological order.

[0032] Optionally, the S2 specifically includes:

[0033] S21. Setting the initial parameters of the covariance matrix adaptation evolution strategy optimizer, including population size, dimension of the structural parameter vector, initial mean vector, initial covariance matrix, and step size factor;

[0034] S22. Based on the initial mean vector, covariance matrix and step factor, sampling is performed from a multivariate normal distribution centered on the initial mean and with covariance as the distribution structure to generate multiple structural parameter encoding vectors to constitute a set of structural parameter encoding vectors.

[0035] Optionally, the S3 specifically includes:

[0036] S31. Extracting a structure control parameter set based on each structure parameter encoding vector in the structure parameter encoding vector set to obtain a structure control parameter set;

[0037] S32. Construct a deformable attention neural network model, including a feature encoding module, a deformable attention module, a feature fusion module, and an output branch module;

[0038] S33. Based on the feature encoding module, feature transformation is performed on the product selection input feature set, and linear mapping and dimension unification are performed on the structured attributes, category embeddings, market characteristics, and user behavior sequences to obtain an embedded feature set;

[0039] S34. Input the embedded feature set into the deformable multi-head attention module to perform cross-dimensional feature interaction and position offset modeling operations:

[0040]

[0041] Among them, A is the attention feature output set, is the feature aggregation operation in the multi-head attention mechanism, h is the number of attention heads, j is the attention head index, p is the position index, P j is the offset sampling position set of the jth attention head, α j,p is the attention weight assigned by the jth attention head at position p, W j is the linear transformation weight matrix of the j-th attention head, is the product embedding vector after offset correction, Δp j,p is the offset of the jth attention head at position p;

[0042] S35. Based on the feature fusion module, the embedded feature set is fused with the attention feature output set A:

[0043]

[0044] Among them, F is the fusion feature set, E is the embedded feature set, [E; A] is the feature splicing operation, λ is the feature fusion weight coefficient, and f is the fusion method parameter;

[0045] S36. Based on the fusion feature set F, a dual-task output structure is constructed, including a regression output branch for predicting product sales and a classification output branch for judging whether the product is recommended as a product selection candidate. The fusion representation F is input into the regression prediction branch through a set of fully connected layers to output the sales prediction result. The fusion representation F is input into the classification judgment branch, through a set of fully connected layers and a sigmoid activation function, to output the product selection judgment result.

[0046] Optionally, the S4 specifically includes:

[0047] S41. Obtain sales prediction results and product selection judgment results output by the deformable attention neural network model;

[0048] S42. Obtain the actual sales records and manually annotated and rule-generated product selection labels for each product from the historical business database of the cross-border e-commerce platform, and construct a set of real sales labels and a set of real product selection labels;

[0049] S43. Construct a sales prediction loss function based on relative error and importance weight:

[0050]

[0051] Among them, L reg is the sales prediction loss function value, n is the number of product samples, i is the product sample index, ω i is the sample importance weight of the i-th product, is the sales forecast value output by the model for the i-th product, is the real historical sales label of the i-th product, δ is the normalized stability factor;

[0052] S44. Construct a product selection loss function with positive and negative sample penalty distinction:

[0053]

[0054] Among them, L cls is the product selection judgment loss function, is the real selection label of the i-th product, is the recommendation probability of the i-th product, λ1 is the loss weight of positive samples, and λ2 is the loss weight of negative samples;

[0055] S45. The sales forecast loss function L reg And the product selection loss function L cls Perform weighted combination to construct joint fitness function Ltotal .

[0056] Optionally, the S5 specifically includes:

[0057] S51, based on the joint fitness function L total Perform fitness evaluation on each vector in the current structural parameter encoding vector set to obtain a corresponding fitness score set;

[0058] S52, sorting the set of structural parameter encoding vectors and their corresponding fitness score sets, retaining the first a encoding vectors with the best fitness, and obtaining an elite set;

[0059] S53, updating the covariance matrix according to the elite set to adapt the mean vector, covariance matrix and step size factor in the evolution strategy optimizer:

[0060]

[0061] Among them, μ t+1 is the mean vector of the t+1th structural parameter encoding vector, t is the iteration round, a is the number of elite individuals, i is the elite individual index, w i is the weight coefficient of the i-th elite structure encoding vector, x (i) is the encoding vector of the structure parameters of the i-th elite body, and T is the matrix transpose operation;

[0062] S54. Based on the updated mean vector, covariance matrix and step size factor, sample from the new multivariate normal distribution to generate a new generation of structural parameter encoding vector set.

[0063] Optionally, the S6 specifically includes:

[0064] S61, setting the maximum number of evolutionary iterations, the fitness convergence threshold, and the fitness convergence judgment window length, and initializing the iteration rounds;

[0065] S62, based on the new generation structure parameter encoding vector set, repeatedly executing the model construction, model training, result output and fitness score calculation process;

[0066] S63, recording the change in the optimal fitness value from the t-ω+1th round to the tth round of iteration, calculating the maximum difference between the minimum fitness values of two adjacent rounds in consecutive ω iterations as the fitness change amplitude, and when the fitness change amplitude is less than or equal to the preset convergence threshold, it is determined that the convergence condition is met and the iteration is terminated;

[0067] S64. Select the individual with the smallest fitness score from all previous structural parameter encoding vectors as the final optimal structural parameter encoding vector.

[0068] Optionally, the S7 specifically includes:

[0069] S71. Based on the set of structural control parameters in the optimal structural parameter encoding vector, construct a final deformable attention neural network model;

[0070] S72: Input the selected input feature set into the final deformable attention neural network model, and generate the final fused feature set through multi-head deformable attention processing and fusion operations;

[0071] S73. Input the final fused feature set into the sales prediction branch and the product selection judgment branch of the final deformable attention neural network model, respectively, to obtain a final sales prediction result set and a final product selection judgment result set;

[0072] S74. The final sales forecast result set and the final product selection judgment result set are jointly output to construct an intelligent product selection output result.

[0073] The beneficial effects of the present invention are:

[0074] The present invention constructs a cross-border e-commerce intelligent product selection system that integrates a covariance matrix adaptive evolutionary strategy and a deformable multi-head attention neural network, establishing a closed-loop linkage mechanism from structural parameter optimization, model construction, task output to fitness feedback, and realizing the joint learning and dynamic optimization of product sales forecasts and product selection judgment results. In terms of system structure, the present invention continuously optimizes network structure parameters such as the number of attention heads, offset range, embedding dimension, and fusion method through evolutionary strategies, so that the model can automatically adapt to the optimal architecture under different data distributions and product selection strategy preferences, avoiding the accuracy bottleneck and manual parameter adjustment burden brought by the rigid structure of traditional models.

[0075] At the same time, the present invention introduces a dual-branch output structure, modeling sales forecasting and recommendation judgment tasks respectively. It accurately characterizes the two tasks by designing a normalized relative error loss function and a category-weighted cross-entropy loss function. Combining regularization terms with task weight coefficients, it constructs a unified multi-task fitness function, effectively improving the model's generalization performance and stability under conditions of complex product samples, category imbalance, and prediction offset. During model training, all output results can be reverse-optimized based on fitness feedback, ensuring that both diagnostic and prediction results undergo a performance adjustment process, enhancing the credibility and controllability of recommendation results.

[0076] During the evolutionary optimization process, the CMA-ES evolutionary strategy employed in this invention automatically updates the sampling mean and covariance matrix based on each round of model evaluation, guiding the evolution of structural parameters toward optimal fitness. When convergence criteria are met or the maximum number of iterations is reached, the optimal parameter vector is extracted to construct the final model, enabling the system to approach global optimality through trial and error. During the deployment phase, the resulting model exhibits optimal structure, streamlined parameters, and balanced tasks. It can be directly loaded to predict sales of new products and determine whether to recommend them, enabling automated evaluation and intelligent screening of massive quantities of products.

[0077] The technical solution proposed by the present invention not only improves the product selection system's ability to optimize the structure and collaboratively model tasks, but also significantly enhances the model's adaptability and intelligence. In the face of frequently changing market demands, complex product feature spaces, and diverse user behavior patterns in cross-border e-commerce, the system can continuously output high-quality sales forecasts and recommendation results, thereby assisting the platform in fine-grained product management, optimizing shelf decisions, and improving overall operational efficiency. In summary, the present invention realizes the intelligent closed-loop evolution of the cross-border e-commerce product selection system from "static rule screening" to "evolvable structure, collaborative tasks, and adjustable results," with beneficial effects such as high accuracy, high stability, low manual dependence, and strong generalization capabilities. BRIEF DESCRIPTION OF THE DRAWINGS

[0078] The accompanying drawings are used to provide a further understanding of the present invention and constitute a part of the specification. Together with the embodiments of the present invention, they are used to explain the present invention and do not constitute a limitation of the present invention. In the accompanying drawings:

[0079] Figure 1 This is a flowchart of a method for a cross-border e-commerce intelligent product selection system based on artificial intelligence proposed by the present invention;

[0080] Figure 2 This is a system diagram of an artificial intelligence-based cross-border e-commerce intelligent product selection system proposed by the present invention;

[0081] Figure 3 This is a flow chart of the evolution control module of the artificial intelligence-based cross-border e-commerce intelligent product selection system proposed by the present invention. DETAILED DESCRIPTION

[0082] The present invention will now be described in further detail with reference to the accompanying drawings, which are simplified schematic diagrams that illustrate the basic structure of the present invention in a schematic manner.

[0083] refer to Figure 1-3 , an AI-based cross-border e-commerce intelligent product selection system, including:

[0084] Data collection and preprocessing module, used to collect raw data from cross-border e-commerce platforms and process the raw data;

[0085] Parameter sampling module, used to sample and generate a set of structural parameter encoding vectors based on the covariance matrix adaptation evolution strategy;

[0086] Candidate model building module for building deformable attention neural network models;

[0087] Model output module, used to output product sales forecast value and product recommendation probability;

[0088] Multi-task loss evaluation module, used to construct a joint fitness function and calculate multi-task fitness scores;

[0089] Fitness-driven optimization module, used to select elite individuals and update the covariance matrix to adapt to the evolution strategy parameters;

[0090] Evolution control module, used to execute the iterative optimization process;

[0091] The optimal model extraction module is used to select the optimal structural parameter encoding vector for fitness and construct the final deformable attention neural network model;

[0092] The recommended product selection output module is used to load the final deformable attention neural network model and output a set of product recommendation results.

[0093] The present invention provides an artificial intelligence-based cross-border e-commerce intelligent product selection system that integrates multiple functional modules such as data collection and preprocessing, structural optimization, model construction, prediction output and evolutionary control, realizing the optimization of the entire process of intelligent product selection driven by cross-border e-commerce big data. The system constructs a unified feature input set by collecting structured product information, category labels, market environment and user behavior data, and introduces a covariance matrix adaptive evolution strategy to efficiently sample and optimize the structural parameters of the deformable attention model, thereby improving the adaptability and accuracy of the model structure. The model realizes product selection target evaluation through the dual-task output of sales forecast and recommendation judgment, and jointly drives the dynamic convergence of the optimization process with the fitness function to ensure the selection of the structural configuration with the best performance. The final system can stably output a list of recommended products with high accuracy and high conversion potential, significantly improving product selection efficiency and platform operating revenue. It is suitable for cross-platform, multi-language and multi-regional cross-border e-commerce product selection business scenarios, and has the advantages of strong structural adaptability, high prediction accuracy and good deployment flexibility.

[0094] In this embodiment, the modules are connected through the following methods:

[0095] S1. Collect cross-border e-commerce data, construct the original data set, pre-process the original data set, and extract the product selection input feature set;

[0096] S2. Initialize the covariance matrix to adapt to the evolutionary strategy optimizer, set the initial parameters, and generate a structural parameter encoding population based on the initial parameter sampling to construct a set of structural parameter encoding vectors;

[0097] S3. Based on the set of structural parameter encoding vectors, a deformable attention neural network model is constructed. The product selection input feature set is input into the deformable attention neural network model, and the sales prediction results and product selection judgment results are output.

[0098] S4. Based on the sales forecast results and product selection judgment results, the prediction error and classification accuracy are calculated respectively, and a joint fitness function is constructed;

[0099] S5. Update the covariance matrix based on the joint fitness function to adapt the evolutionary strategy optimizer and sample a new generation of structural parameter encoding vector sets;

[0100] S6. Based on the new generation of structural parameter encoding vector sets, the model construction, model training, result output and fitness score calculation processes are repeatedly executed until the fitness change amplitude meets the convergence condition, and the final optimal structural parameter encoding vector is output;

[0101] S7. Based on the final optimal structural parameter encoding vector, select the final deformable attention neural network model and extract the intelligent product selection output results.

[0102] The present invention provides an artificial intelligence-based cross-border e-commerce intelligent product selection method, which combines the covariance matrix adaptation evolution strategy and deep learning structure adaptation technology to achieve intelligent optimization and dynamic evolution of product selection strategies in cross-border e-commerce scenarios. The method collects multi-source heterogeneous product data to construct a product selection input feature set, uses an evolutionary strategy optimizer to efficiently sample and update the model structure parameters, and constructs a joint fitness function through the dual-task output of sales prediction and product selection judgment to guide the structural parameters to iteratively converge to the optimal solution. Finally, by selecting the optimal fitness structure, a high-performance deformable attention model is constructed to output accurate product selection recommendation results. This method has the advantages of strong structural adaptability, accurate prediction, and high optimization convergence efficiency. It significantly improves the product selection decision-making quality and automation level of cross-border e-commerce platforms in complex market environments, and is suitable for highly dynamic and data-imbalanced international e-commerce product selection tasks.

[0103] In this embodiment, the original data set includes structured attribute data, category information data, market environment data and user behavior data.

[0104] This paper presents an AI-based intelligent product selection method for cross-border e-commerce. It fully integrates structured attributes, category information, market environment, and user behavior data to construct high-dimensional product selection input features and implement in-depth modeling of multi-source heterogeneous information. By using an evolutionary strategy to drive adaptive optimization of the model structure, it significantly improves the accuracy of sales forecasts and product selection decisions, enhancing the model's robustness and generalization capabilities in multiple market environments. The method is suitable for large-scale intelligent product selection applications on cross-border e-commerce platforms.

[0105] In this embodiment, the preprocessing includes normalizing the structured attribute data, embedding the category information data, numerically quantifying and feature-normalizing the market environment data, and slicing the user behavior data in chronological order.

[0106] This method effectively improves the comparability and temporal expression of multi-source features by normalizing structured attribute data, embedding and encoding category information, quantifying and standardizing market environment data, and sharding user behavior data over time. This preprocessing approach enhances the model's ability to identify key product selection factors, optimizes data input quality, and significantly improves the accuracy and stability of subsequent models in sales forecasting and product selection tasks. It is suitable for large-scale intelligent product selection applications in complex cross-border e-commerce scenarios.

[0107] In this embodiment, S2 specifically includes:

[0108] S21. Setting the initial parameters of the covariance matrix adaptation evolution strategy optimizer, including population size, dimension of the structural parameter vector, initial mean vector, initial covariance matrix, and step size factor;

[0109] S22. Based on the initial mean vector, covariance matrix and step factor, sampling is performed from a multivariate normal distribution centered on the initial mean and with covariance as the distribution structure to generate multiple structural parameter encoding vectors to constitute a set of structural parameter encoding vectors.

[0110] The present invention introduces a covariance matrix to adapt to the evolutionary strategy optimizer, sets key parameters such as population size, structural parameter vector dimension, initial mean and covariance matrix, and constructs an efficient and stable structural search mechanism. The multivariate normal distribution is used to sample structural parameters and generate multiple sets of structural parameter encoding vectors with differences, providing a basis for subsequent diversity exploration and performance optimization of model structures. This method effectively improves the globality and convergence speed of model search, avoids falling into local optimality, and at the same time enhances the adaptability of cross-border e-commerce product selection models to different task scenarios and product feature dimensions. While improving prediction accuracy and product selection stability, it reduces the human intervention and trial-and-error costs of structural adjustment, and has high engineering practical value and algorithm promotion potential.

[0111] In this embodiment, S3 specifically includes:

[0112] S31. Extracting a structure control parameter set based on each structure parameter encoding vector in the structure parameter encoding vector set to obtain a structure control parameter set;

[0113] S32. Construct a deformable attention neural network model, including a feature encoding module, a deformable attention module, a feature fusion module, and an output branch module;

[0114] S33. Based on the feature encoding module, feature transformation is performed on the product selection input feature set, and linear mapping and dimension unification are performed on the structured attributes, category embeddings, market characteristics, and user behavior sequences to obtain an embedded feature set;

[0115] S34. Input the embedded feature set into the deformable multi-head attention module to perform cross-dimensional feature interaction and position offset modeling operations:

[0116]

[0117] Among them, A is the attention feature output set, is the feature aggregation operation in the multi-head attention mechanism, h is the number of attention heads, j is the attention head index, p is the position index, P j is the offset sampling position set of the jth attention head, α j,p is the attention weight assigned by the jth attention head at position p, W j is the linear transformation weight matrix of the j-th attention head, is the product embedding vector after offset correction, Δp j,p is the offset of the jth attention head at position p;

[0118] S35. Based on the feature fusion module, the embedded feature set is fused with the attention feature output set A:

[0119]

[0120] Among them, F is the fusion feature set, E is the embedded feature set, [E; A] is the feature splicing operation, λ is the feature fusion weight coefficient, and f is the fusion method parameter;

[0121] S36. Based on the fusion feature set F, a dual-task output structure is constructed, including a regression output branch for predicting product sales and a classification output branch for judging whether the product is recommended as a product selection candidate. The fusion representation F is input into the regression prediction branch through a set of fully connected layers to output the sales prediction result. The fusion representation F is input into the classification judgment branch, through a set of fully connected layers and a sigmoid activation function, to output the product selection judgment result.

[0122] The present invention extracts a set of structural control parameters based on the structural parameter encoding vector and dynamically constructs a deformable attention neural network model, thereby realizing the adaptive configuration and automatic construction of the model structure. The system design includes feature encoding, deformable multi-head attention, feature fusion and dual-task output modules, which support unified modeling and dynamic interaction of multi-dimensional features such as structured attributes, categories, markets and behavior sequences. By introducing offset position modeling and multi-head attention mechanism, the model's ability to capture complex nonlinear feature relationships is effectively improved. The fusion module realizes efficient information integration according to the fusion strategy, and finally outputs the commodity sales forecast value and product selection recommendation probability. This method significantly enhances the cross-border e-commerce product selection model's ability to express multimodal data and structural control flexibility, has strong adaptability and high reasoning accuracy, and can greatly improve product selection decision-making efficiency and recommendation effect.

[0123] In this embodiment, the S4 specifically includes:

[0124] S41. Obtain sales prediction results and product selection judgment results output by the deformable attention neural network model;

[0125] S42. Obtain the actual sales records and manually annotated and rule-generated product selection labels for each product from the historical business database of the cross-border e-commerce platform, and construct a set of real sales labels and a set of real product selection labels;

[0126] S43. Construct a sales prediction loss function based on relative error and importance weight:

[0127]

[0128] Among them, L reg is the sales prediction loss function value, n is the number of product samples, i is the product sample index, ω i is the sample importance weight of the i-th product, is the sales forecast value output by the model for the i-th product, is the real historical sales label of the i-th product, δ is the normalized stability factor;

[0129] S44. Construct a product selection loss function with positive and negative sample penalty distinction:

[0130]

[0131] Among them, L cls is the product selection judgment loss function, is the real selection label of the i-th product, is the recommendation probability of the i-th product, λ1 is the loss weight of positive samples, and λ2 is the loss weight of negative samples;

[0132] S45. The sales forecast loss function L reg And the product selection loss function L cls Perform weighted combination to construct joint fitness function L total .

[0133] The present invention realizes dual modeling and refined evaluation of cross-border e-commerce product selection tasks by constructing a sales prediction loss function based on relative error and sample weight, and introducing a product selection judgment loss function that introduces the difference in positive and negative sample penalties. The sales prediction loss function can dynamically adjust the error impact according to the importance of sample sales, thereby improving the prediction accuracy of the model on high-value goods; the product selection judgment loss function effectively alleviates the interference of sample imbalance on recommendation judgment by distinguishing the weights of positive and negative samples. Finally, by weighted fusion of the two types of losses, a joint fitness function is constructed to guide the coordinated evolution of structural search and model optimization, which significantly improves the model's product selection judgment accuracy and sales prediction stability, and enhances the system's practicality and robustness in actual product selection scenarios.

[0134] In this embodiment, the S5 specifically includes:

[0135] S51, based on the joint fitness function L total Perform fitness evaluation on each vector in the current structural parameter encoding vector set to obtain a corresponding fitness score set;

[0136] S52, sorting the set of structural parameter encoding vectors and their corresponding fitness score sets, retaining the first a encoding vectors with the best fitness, and obtaining an elite set;

[0137] S53, updating the covariance matrix according to the elite set to adapt the mean vector, covariance matrix and step size factor in the evolution strategy optimizer:

[0138]

[0139] Among them, μ t+1 is the mean vector of the t+1th structural parameter encoding vector, t is the iteration round, a is the number of elite individuals, i is the elite individual index, w i is the weight coefficient of the i-th elite structure encoding vector, x (i) is the encoding vector of the i-th elite structure parameter, and T is the matrix transpose operation;

[0140] S54. Based on the updated mean vector, covariance matrix and step size factor, sample from the new multivariate normal distribution to generate a new generation of structural parameter encoding vector set.

[0141] The present invention accurately scores and ranks the structural parameter encoding vectors based on the joint fitness function, constructs a fitness-driven elite set, and updates the mean vector, covariance matrix and step size factor of the evolutionary strategy optimizer with the elite individuals as the core, thereby guiding the optimizer to converge towards a structural space with better performance. This method fully inherits the historical high-quality structural characteristics when sampling a new generation of structural parameter encoding vectors, achieving high efficiency and goal orientation in structural parameter search. During the evolutionary process, excellent structural features are retained and invalid disturbances are suppressed, which improves the stability and convergence speed of structural search and significantly enhances the model's adaptability. This strategy effectively improves the structural optimization accuracy and operating efficiency of the cross-border e-commerce product selection system, and is suitable for large-scale and diverse product recommendation scenarios.

[0142] In this embodiment, S6 specifically includes:

[0143] S61, setting the maximum number of evolutionary iterations, the fitness convergence threshold, and the fitness convergence judgment window length, and initializing the iteration rounds;

[0144] S62, based on the new generation structure parameter encoding vector set, repeatedly executing the model construction, model training, result output and fitness score calculation process;

[0145] S63, recording the change in the optimal fitness value from the t-ω+1th round to the tth round of iteration, calculating the maximum difference between the minimum fitness values of two adjacent rounds in consecutive ω iterations as the fitness change amplitude, and when the fitness change amplitude is less than or equal to the preset convergence threshold, it is determined that the convergence condition is met and the iteration is terminated;

[0146] S64. Select the individual with the smallest fitness score from all previous structural parameter encoding vectors as the final optimal structural parameter encoding vector.

[0147] The present invention introduces a dynamic convergence control mechanism by setting the maximum number of iterations, the fitness convergence threshold and the judgment window length, thereby realizing stability evaluation and adaptive termination during the evolution process. In each round of iteration, the system repeatedly executes model construction and fitness evaluation based on the new generation of structural parameter encoding vectors, records the trend of optimal fitness changes for multiple consecutive rounds, calculates the amplitude of the change and compares it with the convergence threshold, so as to determine whether the optimization process has reached the convergence condition. This mechanism effectively avoids the waste of resources caused by invalid iterations, and improves the search efficiency and reliability of structural optimization convergence. Finally, the optimal structural parameters for fitness are selected from all generations of individuals to ensure that the final model has the global optimal structural performance, thereby improving the practicality and decision-making accuracy of the cross-border e-commerce product selection system in a large-scale heterogeneous data environment.

[0148] In this embodiment, the S7 specifically includes:

[0149] S71. Based on the set of structural control parameters in the optimal structural parameter encoding vector, construct a final deformable attention neural network model;

[0150] S72: Input the selected input feature set into the final deformable attention neural network model, and generate the final fused feature set through multi-head deformable attention processing and fusion operations;

[0151] S73. Input the final fused feature set into the sales prediction branch and the product selection judgment branch of the final deformable attention neural network model, respectively, to obtain a final sales prediction result set and a final product selection judgment result set;

[0152] S74. The final sales forecast result set and the final product selection judgment result set are jointly output to construct an intelligent product selection output result.

[0153] The present invention constructs the final deformable attention neural network model based on the optimal structural parameter encoding vector, realizing a model customization process with deep coordination between structural parameters and task objectives. The model integrates the multi-head deformable attention mechanism and the feature fusion strategy, which can effectively capture the dynamic relationship between multi-source heterogeneous features in cross-border e-commerce data and generate a fusion feature set with high expressiveness. By inputting the fusion features into the sales prediction branch and the product selection judgment branch, high-precision sales prediction results and recommendation judgment results are output respectively, and finally a unified intelligent product selection output result is formed, which effectively supports the accurate screening and recommendation of cross-border commodities. This method enhances the business adaptability of the system while improving the model prediction performance, significantly optimizes the product selection process, and has strong commercial landing value and promotion prospects.

[0154] Example 1:

[0155] In order to verify the feasibility of the present invention in implementation, the present invention is applied to an intelligent product selection optimization project jointly carried out by a large cross-border e-commerce platform and a well-known domestic digital supply chain research institute. The project aims to solve the core problems of the platform in the product selection process for Europe, the United States, the Middle East, and Southeast Asian markets, such as insufficient accuracy, large prediction deviation, and delayed model updates. The test period is from October 2024 to January 2025, covering multiple commodity cycle nodes and peak season sales windows. The test system is deployed in the platform's Dongguan product selection center and overseas virtual machine GPU nodes, using the heterogeneous computing architecture of Intel Xeon 6258R+2×A10080G. The platform data comes from historical cross-border e-commerce transaction data from Q4 2023 to Q3 2024.

[0156] Traditional product selection systems mainly use static indicator ranking methods and LightGBM models to predict and recommend sales. However, due to the lack of consideration of nonlinear interactions between multimodal features of products and the lack of structural adaptive mechanisms, the following problems often arise in product selection decisions: First, faced with sparse product data on Southeast Asian minority language platforms, sales forecasts frequently miss predictions; second, faced with products with strong seasonal fluctuations, such as outdoor heaters and swimwear, the classification accuracy rate drops significantly; third, the model has weak generalization and is difficult to adapt to countries in Europe, the United States, and the Middle East, where consumer preferences are completely different.

[0157] To solve the above problems, the research team adopted the "cross-border e-commerce intelligent product selection method based on evolutionary strategy and deformable attention network" proposed in this invention. The actual application process is as follows: First, the platform collects a data set containing 740,000 product records, covering structured attributes, category information, market environment data and user behavior sequences to construct the original data set; then the data is preprocessed, the dimensions are unified and the input feature set is constructed. The covariance matrix adaptation evolutionary strategy (CMA-ES) is used to initialize the model structure parameter population, and sample and generate a set of structural parameter encodings with different numbers of attention heads, offset ranges, and fusion strategies. Each set of structural parameters controls the construction of a deformable attention neural network structure and predicts sales and product selection results.

[0158] By setting a joint fitness function that includes a weighted combination of sales forecast error and recommendation classification error, and then performing iterative optimization, the model finally converged after 37 iterations, generating the optimal structural encoding vector, and thus determining the final model. The test platform applied the final model to three representative markets (the United States, the United Arab Emirates, and Thailand) to evaluate its forecast accuracy, product recommendation precision, and commercial return.

[0159] During the test, the system evaluated 128 structural configurations and trained a total of 490 candidate models. The system ran for a total of 38 hours. The final model structure included 6 deformable attention heads, 3 offset modeling layers, residual fusion, and the embedding dimension was set to 96.

[0160] The test results show that compared with the existing models of the platform, the model of the present invention has achieved significant improvements in key indicators such as product selection accuracy, average absolute error of sales forecast, and click-through rate of preferred products.

[0161] Table 1 Comparative experimental data of the selection effect of the present invention and the traditional method

[0162]

[0163]

[0164] As can be seen from the above table, the present invention has achieved a comprehensive improvement in core business indicators in multiple major cross-border e-commerce target markets compared to the platform's existing product selection model. Specifically, in the sales forecasting task, the mean absolute error (MAE) of the present invention in the US market was significantly reduced from 17.2 of the traditional method to 11.3, with an error reduction of 34.3%; in the UAE market, MAE dropped from 21.5 to 13.7, a decrease of 36.3%; in the Thai market, where data is more sparse and user behavior is more dispersed, MAE dropped from 25.1 to 15.9, with an error reduction of 36.7%. These results show that the structural adaptive product selection model proposed in the present invention has stronger cross-market migration capabilities and sales forecast stability.

[0165] In terms of product selection accuracy, the present invention also shows obvious advantages. In the US market, the product selection accuracy of the traditional method is 81.6%, while the present invention has increased it to 89.2%, an increase of 7.6 percentage points; in the UAE market, it has increased from 75.1% to 83.5%, an increase of 8.4 percentage points; in the Thai market, it has increased from 69.8% to 79.1%, an increase of 9.3 percentage points. This result proves that the present invention can also make stable and accurate judgments in non-European and American language markets, especially in cases where multimodal features are incomplete or the behavioral data missing rate is high, it can still effectively model the product selection probability distribution, avoiding the common problems of biased judgment and missed judgment in traditional methods.

[0166] At the same time, in terms of actual user conversion, the model of the present invention also performs well in the click-through conversion rate (CTR) indicator. In the US market, CTR increased from 4.91% of the original model to 6.58%, a relative increase of 33.9%; in the UAE market, it increased from 3.88% to 5.33%, an increase of 37.4%; and in the Thai market, it increased from 2.51% to 3.45%, an increase of 37.5%. This shows that the products recommended by the present invention are more likely to arouse user interest and purchase behavior in actual operations, and the recommendation results have higher commercial value and implementation significance.

[0167] From the overall result analysis, it can be seen that the present invention adaptively adjusts the model structure by introducing the covariance matrix adaptive evolution strategy, so that the model can maintain strong prediction and judgment capabilities under data from different regions and different feature structures; at the same time, the deformable attention mechanism enhances the model's ability to model the interaction of heterogeneous features, effectively alleviating the performance degradation problem of traditional methods when faced with inconsistent data dimensions or nonlinear information expression. Compared with the platform's existing shallow models based on fixed structures, the method proposed by the present invention is superior to traditional solutions in many key indicators such as product selection accuracy, sales forecast reliability, and user conversion effect, verifying its strong adaptability and high practicality in actual cross-border e-commerce product selection systems. Especially in emerging markets such as Southeast Asia or the Middle East, facing challenges such as diverse category systems, complex behavior patterns, and dispersed user preferences, the present invention can still maintain highly robust prediction performance, reflecting its technical potential and application prospects as the next generation of intelligent product selection core algorithms.

[0168] The above description is only a preferred specific embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any technician familiar with the technical field, within the technical scope disclosed by the present invention, who makes equivalent replacements or changes based on the technical solution and inventive concept of the present invention, should be covered by the scope of protection of the present invention.

Claims

1. An artificial intelligence-based cross-border e-commerce intelligent product selection system, characterized by: include: Data collection and preprocessing module, used to collect raw data from cross-border e-commerce platforms and process the raw data; Parameter sampling module, used to sample and generate a set of structural parameter encoding vectors based on the covariance matrix adaptation evolution strategy; Candidate model building module for building deformable attention neural network models; Model output module, used to output product sales forecast value and product recommendation probability; Multi-task loss evaluation module, used to construct a joint fitness function and calculate multi-task fitness scores; Fitness-driven optimization module, used to select elite individuals and update the covariance matrix to adapt to the evolution strategy parameters; Evolution control module, used to execute the iterative optimization process; The optimal model extraction module is used to select the optimal structural parameter encoding vector for fitness and construct the final deformable attention neural network model; The recommended product selection output module is used to load the final deformable attention neural network model and output a set of product recommendation results.

2. The cross-border e-commerce intelligent product selection system based on artificial intelligence according to claim 1 is characterized in that: The modules are implemented as follows: S1. Collect cross-border e-commerce data, construct the original data set, pre-process the original data set, and extract the product selection input feature set; S2. Initialize the covariance matrix to adapt to the evolutionary strategy optimizer, set the initial parameters, and generate a structural parameter encoding population based on the initial parameter sampling to construct a set of structural parameter encoding vectors; S3. Based on the set of structural parameter encoding vectors, a deformable attention neural network model is constructed. The product selection input feature set is input into the deformable attention neural network model, and the sales prediction results and product selection judgment results are output. S4. Based on the sales forecast results and product selection judgment results, the prediction error and classification accuracy are calculated respectively, and a joint fitness function is constructed; S5. Update the covariance matrix based on the joint fitness function to adapt the evolutionary strategy optimizer and sample a new generation of structural parameter encoding vector sets; S6. Based on the new generation of structural parameter encoding vector sets, the model construction, model training, result output and fitness score calculation processes are repeatedly executed until the fitness change amplitude meets the convergence condition, and the final optimal structural parameter encoding vector is output; S7. Based on the final optimal structural parameter encoding vector, select the final deformable attention neural network model and extract the intelligent product selection output results.

3. The cross-border e-commerce intelligent product selection system based on artificial intelligence according to claim 2 is characterized in that: The original data set includes structured attribute data, category information data, market environment data and user behavior data.

4. The cross-border e-commerce intelligent product selection system based on artificial intelligence according to claim 2 is characterized in that: The preprocessing includes normalizing the structured attribute data, embedding the category information data, numerically quantifying and feature-normalizing the market environment data, and slicing the user behavior data in chronological order.

5. The cross-border e-commerce intelligent product selection system based on artificial intelligence according to claim 2 is characterized in that: The S2 specifically includes: S21. Setting the initial parameters of the covariance matrix adaptation evolution strategy optimizer, including population size, dimension of the structural parameter vector, initial mean vector, initial covariance matrix, and step size factor; S22. Based on the initial mean vector, covariance matrix and step factor, sampling is performed from a multivariate normal distribution centered on the initial mean and with covariance as the distribution structure to generate multiple structural parameter encoding vectors to constitute a set of structural parameter encoding vectors.

6. The cross-border e-commerce intelligent product selection system based on artificial intelligence according to claim 2 is characterized in that: The S3 specifically includes: S31. Extracting a structure control parameter set based on each structure parameter encoding vector in the structure parameter encoding vector set to obtain a structure control parameter set; S32. Construct a deformable attention neural network model, including a feature encoding module, a deformable attention module, a feature fusion module, and an output branch module; S33. Based on the feature encoding module, feature transformation is performed on the product selection input feature set, and linear mapping and dimension unification are performed on the structured attributes, category embeddings, market characteristics, and user behavior sequences to obtain an embedded feature set; S34. Input the embedded feature set into the deformable multi-head attention module to perform cross-dimensional feature interaction and position offset modeling operations: Among them, A is the attention feature output set, is the feature aggregation operation in the multi-head attention mechanism, h is the number of attention heads, j is the attention head index, p is the position index, P j is the offset sampling position set of the jth attention head, α j,p is the attention weight assigned by the jth attention head at position p, W j is the linear transformation weight matrix of the j-th attention head, is the product embedding vector after offset correction, Δp j,p is the offset of the jth attention head at position p; S35. Based on the feature fusion module, the embedded feature set is fused with the attention feature output set A: Among them, F is the fusion feature set, E is the embedded feature set, [E; A] is the feature splicing operation, λ is the feature fusion weight coefficient, and f is the fusion method parameter; S36. Based on the fusion feature set F, a dual-task output structure is constructed, including a regression output branch for predicting product sales and a classification output branch for judging whether the product is recommended as a product selection candidate. The fusion representation F is input into the regression prediction branch through a set of fully connected layers to output the sales prediction result. The fusion representation F is input into the classification judgment branch, through a set of fully connected layers and a sigmoid activation function, to output the product selection judgment result.

7. The cross-border e-commerce intelligent product selection system based on artificial intelligence according to claim 2 is characterized in that: The S4 specifically includes: S41. Obtain sales prediction results and product selection judgment results output by the deformable attention neural network model; S42. Obtain the actual sales records and manually annotated and rule-generated product selection labels for each product from the historical business database of the cross-border e-commerce platform, and construct a set of real sales labels and a set of real product selection labels; S43. Construct a sales prediction loss function based on relative error and importance weight: Among them, L reg is the sales prediction loss function value, n is the number of product samples, i is the product sample index, ω i is the sample importance weight of the i-th product, is the sales forecast value output by the model for the i-th product, is the real historical sales label of the i-th product, δ is the normalized stability factor; S44. Construct a product selection loss function with positive and negative sample penalty distinction: Among them, L cls is the product selection judgment loss function, is the real selection label of the i-th product, is the recommendation probability of the i-th product, λ1 is the loss weight of positive samples, and λ2 is the loss weight of negative samples; S45. The sales forecast loss function L reg And the product selection loss function L cls Perform weighted combination to construct joint fitness function L total .

8. The cross-border e-commerce intelligent product selection system based on artificial intelligence according to claim 2 is characterized in that: The S5 specifically includes: S51, based on the joint fitness function L total Perform fitness evaluation on each vector in the current structural parameter encoding vector set to obtain a corresponding fitness score set; S52, sorting the set of structural parameter encoding vectors and their corresponding fitness score sets, retaining the first a encoding vectors with the best fitness, and obtaining an elite set; S53, updating the covariance matrix according to the elite set to adapt the mean vector, covariance matrix and step size factor in the evolution strategy optimizer: Among them, μ t+1 is the mean vector of the t+1th structural parameter encoding vector, t is the iteration round, a is the number of elite individuals, i is the elite individual index, w i is the weight coefficient of the i-th elite structure encoding vector, x (i) is the encoding vector of the i-th elite structure parameter, and T is the matrix transpose operation; S54. Based on the updated mean vector, covariance matrix and step size factor, sample from the new multivariate normal distribution to generate a new generation of structural parameter encoding vector set.

9. The cross-border e-commerce intelligent product selection system based on artificial intelligence according to claim 2 is characterized in that: The S6 specifically includes: S61, setting the maximum number of evolutionary iterations, the fitness convergence threshold, and the fitness convergence judgment window length, and initializing the iteration rounds; S62, based on the new generation structure parameter encoding vector set, repeatedly executing the model construction, model training, result output and fitness score calculation process; S63, recording the change in the optimal fitness value from the t-ω+1th round to the tth round of iteration, calculating the maximum difference between the minimum fitness values of two adjacent rounds in consecutive ω iterations as the fitness change amplitude, and when the fitness change amplitude is less than or equal to the preset convergence threshold, it is determined that the convergence condition is met and the iteration is terminated; S64. Select the individual with the smallest fitness score from all previous structural parameter encoding vectors as the final optimal structural parameter encoding vector.

10. The cross-border e-commerce intelligent product selection system based on artificial intelligence according to claim 2, characterized in that: The S7 specifically includes: S71. Based on the set of structural control parameters in the optimal structural parameter encoding vector, construct a final deformable attention neural network model; S72: Input the selected input feature set into the final deformable attention neural network model, and generate the final fused feature set through multi-head deformable attention processing and fusion operations; S73. Input the final fused feature set into the sales prediction branch and the product selection judgment branch of the final deformable attention neural network model, respectively, to obtain a final sales prediction result set and a final product selection judgment result set; S74. The final sales forecast result set and the final product selection judgment result set are jointly output to construct an intelligent product selection output result.

Images