Cross-border e-commerce marketing strategy intelligent management system based on big data

Abstract

The invention discloses a cross-border e-commerce marketing strategy intelligent management system based on big data, and relates to the technical field of cross-border e-commerce marketing strategy management.The cross-border e-commerce marketing strategy intelligent management system utilizes a heterogeneous graph neural network to construct a user-commodity-culture relation graph, and dynamic weights are given to nodes through an attention mechanism; differences of consumers in each region in historical transactions, click rates and culture similarities are effectively captured, so that the problem of insufficient adaptability of a traditional fixed rule model in a multi-culture market is solved; carrying out emotion and semantic feature extraction on a multi-language comment text by adopting a Transform structure, and realizing rapid self-adaption of the model under the condition of data scarcity in combination with a meta-learning strategy; according to the combination, deep semantic information reflecting emotions and preferences of different regions can be accurately extracted, and cross-culture adaptability recommendation results are generated by fusing image embedding and text processing results, so that marketing contents better meet personalized requirements of consumers in various regions.

Description

Technical Field

[0001] The present invention relates to the technical field of cross - border e - commerce marketing strategy management, and particularly to an intelligent management system for cross - border e - commerce marketing strategies based on big data. Background Art

[0002] Existing cross - border e - commerce marketing systems mostly rely on data such as transaction records, click - through rates, and user browsing to construct a unified user profile and then achieve personalized recommendations through statistical learning.

[0003] However, the global market is becoming increasingly diverse, and consumers in different regions have different aesthetics, consumption habits, and cultural heritages. Most traditional systems process data from various countries in the same model, resulting in deviations in recommendation strategies in markets with different cultural backgrounds. For example, the same product emphasizes design concepts more in the European and American markets, while in Asia, more attention is paid to cost - effectiveness and practicality. However, the single - rule or manual - partitioning method is difficult to capture the emotional changes of consumers towards products and local hotspots on social platforms in a timely manner, leading to a disconnection between marketing content and actual needs.

[0004] In response to this, some enterprises add regional rules, manual optimization, etc. However, such methods have a lag in data update and are difficult to balance real - time performance and accuracy, still resulting in an obvious gap between the recommended effect and the actual preferences of users. At the same time, the constantly changing market environment and dynamic feedback on social media in the process of globalization further expose the deficiencies of traditional systems in cross - cultural adaptation ability, affecting conversion rates and customer loyalty. Against this background, how to utilize more - dimensional data resources and deeply explore the cultural information behind user behavior has become an urgent problem to be solved. Summary of the Invention

[0005] In view of the above - mentioned existing problems, the present invention is proposed.

[0006] The present invention provides an intelligent management system for cross - border e - commerce marketing strategies based on big data to solve the problem that traditional recommendation algorithms ignore the subtle behaviors and emotional preferences of consumers under different cultural backgrounds, resulting in low recommendation hit rates in markets such as Europe, America, and Southeast Asia.

[0007] To solve the above - mentioned technical problems, the present invention provides the following technical solutions:

[0008] An embodiment of the present invention provides an intelligent management system for cross - border e - commerce marketing strategies based on big data, which includes,

[0009] An e - commerce data collection module that collects original e - commerce data and constructs a heterogeneous data set including users, products, and cultural characteristics;

[0010] A data pre - processing module for pre - processing the heterogeneous data set;

[0011] Personalized recommendation and localization optimization module, which is constructed based on a multi-task joint learning model. The model includes:

[0012] An heterogeneous graph neural network, which is used to establish a relationship graph among users, commodities and cultural tags, and assigns dynamic weights to the relationships among nodes in the relationship graph through an attention mechanism;

[0013] A Transformer structure, which is used to process multi-language texts and social comment information, and extract semantic information reflecting regional sentiment and preferences from the original e-commerce data;

[0014] A meta-learning sub-module, which is used to adaptively adjust model parameters in the case of scarce data;

[0015] A feedback adjustment module, which collects the output of the management system and real-time market feedback data in real time, and online updates and optimizes the model parameters in the personalized recommendation and localization optimization module according to a predetermined strategy.

[0016] As a preferred solution of the intelligent management system for cross-border e-commerce marketing strategies based on big data according to the present invention, wherein: the original e-commerce data includes e-commerce platform transaction records, user behavior logs and social media text data;

[0017] The e-commerce data collection module includes an interface for cross-platform data connection and a big data engine for real-time data transmission. The engine is used to ensure the timeliness and continuity of data collection, so that the user-commodity-culture feature matrix has the ability of dynamic update.

[0018] As a preferred solution of the intelligent management system for cross-border e-commerce marketing strategies based on big data according to the present invention, wherein: the preprocessing includes cleaning, normalizing and data fusion of the data in the heterogeneous data set to form a user-commodity-culture label feature matrix in a unified format, and the culture feature matrix is labeled according to an internationally recognized culture index system.

[0019] As a preferred solution of the intelligent management system for cross-border e-commerce marketing strategies based on big data according to the present invention, wherein: in the personalized recommendation and localization optimization module, the heterogeneous graph neural network constructs a graph embedding according to the historical transaction data, click-through rate and cultural similarity of users in different regions; the Transformer structure performs sentiment and semantic modeling based on multi-language review texts, and the two are combined to output a cross-cultural adaptation recommendation result, which is used to reflect different attention focuses and consumption preferences of the same commodity in different markets; product recommendations are made based on the cross-cultural adaptation recommendation result.

[0020] As an optimal solution of the intelligent management system for cross-border e-commerce marketing strategies based on big data of the present invention, wherein: in the personalized recommendation and localization optimization module, the steps for constructing graph embeddings by the heterogeneous graph neural network based on the historical transaction data, click-through rate, and cultural similarity of users in different regions are as follows,

[0021] Calculate the association scores between nodes based on the historical transaction data, click-through rate, and cultural similarity between each node. The calculation formula is:

[0022]

[0023] where, e ij represents the association score between node i and node j, a represents the attention weight vector, and its transpose is used to map the concatenated vector to the score space, b represents the feature transformation matrix, x i and x j are the initial feature vectors of node i and node j respectively, ‖ represents concatenating the two vectors, c1 represents the coefficient for transaction data T ij c2 represents the coefficient for click-through rate data C ij c3 represents the coefficient for cultural similarity data R ij T ij C ij and R ij represent the historical transaction data, click-through rate index, and cultural similarity index between node i and node j respectively. i and j are node indices;

[0024] Calculate the normalized attention weights between nodes. The formula is:

[0025]

[0026] where, α ij represents the normalized attention weight between node i and node j, exp(·) is the exponential function, represents the neighbor set of node i, k is the neighbor node index, e ij and e ik are the association scores between nodes respectively;

[0027] Update the node embeddings by aggregating neighbor information. The update formula is:

[0028]

[0029] where, h i represents the graph embedding of node i, α ij is the normalized attention weight between node i and neighbor node j, b is the feature transformation matrix, x jis the initial feature vector of neighbor node j, where j is the neighbor node index, is the neighbor set of node i.

[0030] As a preferred solution of the intelligent management system for cross-border e-commerce marketing strategies based on big data according to the present invention, wherein: the Transformer structure performs sentiment and semantic modeling based on multilingual review texts, and the steps of combining the two to output cross-cultural adaptability recommendation results are as follows,

[0031] For multilingual review texts, use the Transformer structure to extract sentiment and semantic features, represent the preprocessed text as an embedding vector z0, and then perform projections of query, key, and value. The formula is:

[0032] Q = z0d, K = z0e, V = z0f,

[0033] where Q represents the query vector in the Transformer, K represents the key vector in the Transformer, V represents the value vector in the Transformer, z0 is the text embedding after preprocessing, d is the query projection matrix, e is the key projection matrix, and f is the value projection matrix;

[0034] Use the scaled dot-product attention mechanism for feature fusion. The fusion formula is:

[0035]

[0036] where A represents the vector output by the attention mechanism, represents the transpose of the key vector K, g is the scaling factor, softmax(·) represents the standard normalization function, and V is the value vector;

[0037] Normalize and fuse the attention output with the initial embedding to obtain the semantic representation:

[0038] z = LayerNorm(z0 + A),

[0039] where z represents the semantic representation output by the Transformer structure, LayerNorm(·) represents the layer normalization operation, z0 is the initial text embedding, and A is the attention output.

[0040] As a preferred solution of the intelligent management system for cross-border e-commerce marketing strategies based on big data according to the present invention, wherein: the steps of the Transformer structure performing sentiment and semantic modeling based on multilingual review texts and combining the two to output cross-cultural adaptability recommendation results further include:

[0041] Perform sentiment analysis through a fully connected layer:

[0042] s = σ(iz + j),

[0043] where s represents the sentiment score, σ(·) represents the activation function, i is the sentiment analysis weight matrix, z is the semantic representation output by the Transformer, and j is the sentiment analysis bias;

[0044] Let the heterogeneous graph embedding be h, and fuse h, the Transformer semantic representation z, and the sentiment score s:

[0045]

[0046] where r represents the final cross - cultural adaptability recommendation result, k is the fusion weight matrix,

[0047] represents the vector formed by vertically concatenating the graph embedding h, the semantic representation z, and the sentiment score s,

[0048] l is the fusion bias, and h is the node embedding.

[0049] As a preferred solution of the intelligent management system for cross - border e - commerce marketing strategies based on big data according to the present invention, wherein: the way for the meta - learning sub - module to adaptively adjust the model parameters in the case of scarce data is:

[0050] Adopt a meta - learning strategy based on gradient update to perform inner - loop update for each task τ, and the formula is:

[0051]

[0052] where m′ τ represents the model parameters obtained after the update within task τ, m represents the initial model parameters, p is the inner - update step size, represents the gradient calculation of the model parameters m, and l τ (m) represents the loss function of task τ, and τ is the task index, representing different tasks,

[0053] In the outer loop, perform global parameter update according to the loss information of all tasks, and the update formula is:

[0054]

[0055] where m represents the updated model parameters, q is the outer - update step size, regulating the amplitude of global parameter update, and ∑ τ (·) represents the summation operation of the losses of all tasks, represents the gradient operation with respect to the parameter m, and l τ (m′ τ ) represents the loss value under each task τ after the inner - loop update.

[0056] As a preferred solution of the intelligent management system for cross-border e-commerce marketing strategies based on big data according to the present invention, wherein: the feedback adjustment module includes a monitoring sub-module for real-time monitoring of the effectiveness of marketing strategies and a decision-making sub-module for feedback-driven adaptive update of the model. The monitoring sub-module uses real-time data to statistically analyze the performance indicators of the management system, and the decision-making sub-module automatically tunes the model parameters according to the preset safety margin and benefit threshold; realizing the instant response of the system to price fluctuations, inventory status, and regional promotion strategies in the cross-border e-commerce environment;

[0057] The output data of the management system includes recommendation results, content display data, and system self-diagnosis logs generated by the personalized recommendation and localization optimization module;

[0058] The market real-time feedback data includes real-time transaction data, user click-through rate, conversion rate, online reviews, social media sentiment analysis results, inventory status, exchange rate fluctuations, and competitor price dynamics from e-commerce platforms;

[0059] The real-time data used by the monitoring sub-module refers to the data set updated on a millisecond-to-minute time scale in the output data of the management system and the market real-time feedback data;

[0060] The performance indicators are quantitative indicators reflecting the overall operation efficiency and recommendation effect of the system, including click-through rate, conversion rate, user stay time, recommendation accuracy, inventory turnover rate, and return on advertising expenses;

[0061] The output data of the management system here reflects the internal self-diagnosis of the system and the execution effect of the strategy, while the market real-time feedback data is for the instant dynamic monitoring of the external environment. Both exist in the form of high-frequency real-time data, and the statistical and analysis results constitute the system performance indicators, which are used to drive the decision-making sub-module to adaptively update and optimize the model parameters according to the preset safety margin and benefit threshold.

[0062] As a preferred solution of the intelligent management system for cross-border e-commerce marketing strategies based on big data according to the present invention, wherein: in the feedback adjustment module, the steps of online updating and optimizing the model parameters in the personalized recommendation and localization optimization module according to the predetermined strategy are as follows,

[0063] The feedback adjustment module constructs a comprehensive performance indicator by real-time collecting the output data of the management system and the market real-time feedback data, and online updates the model parameters according to the preset strategy. The calculation formula of the comprehensive performance indicator is:

[0064] P = γX + (1 - γ)Y,

[0065] Among them, P represents the comprehensive performance index, γ is the weight coefficient of the output data of the management system, X represents the output data of the management system, Y represents the real-time market feedback data, and 1 - γ is the supplementary weight, representing the contribution of the market feedback data.

[0066] According to the performance index, the preset safety margin, and the benefit threshold, a deviation function is defined:

[0067] If P < u, then I(P) = u - P.

[0068] If P > v, then I(P) = P - v.

[0069] In other cases, then I(P) = 0.

[0070] Among them, I(P) represents the performance deviation function, u is the preset safety margin, v is the preset benefit threshold, and P is the comprehensive performance index.

[0071] The online update formula for the final model parameters is:

[0072]

[0073] Among them, m represents the current model parameters, w is the online update control weight, I(P) is the deviation value obtained from the performance deviation function, represents the gradient of the loss function with respect to the model parameters m, and l(m) is the loss function of the model.

[0074] The beneficial effects of the present invention are as follows: The present invention integrates global e-commerce transaction records, user behavior logs, and social media text data, and annotates the data according to the international cultural index system, realizing the heterogeneous data fusion of users, commodities, and cultural tags. It constructs a user-commodity-culture relationship graph using a heterogeneous graph neural network, and assigns dynamic weights to each node through an attention mechanism, effectively capturing the differences in historical transactions, click-through rates, and cultural similarities of consumers in each region, thus breaking through the problem of insufficient adaptability of traditional fixed-rule models in multi-cultural markets.

[0075] The present invention uses the Transformer structure to extract emotional and semantic features from multi-language review texts, and combines the meta-learning strategy to achieve fast adaptation of the model in the case of scarce data; this combination can accurately extract deep semantic information reflecting emotions and preferences in different regions, and generate cross-cultural adaptable recommendation results by fusing graph embedding and text processing results, making the marketing content more in line with the personalized needs of consumers in each region.

[0076] The present invention provides a real-time feedback adjustment module, which collects the output of the management system and market dynamic data in real time, including trading volume, click-through rate, inventory status, exchange rate fluctuations, and competitor price dynamics, to construct a comprehensive performance index. According to the preset safety margin and benefit threshold, the system deviation is automatically determined, and the gradient descent method is used to update the model parameters online, so as to achieve an immediate response to price fluctuations and promotion strategies, thereby improving the overall operation efficiency and stability of the system.

[0077] In summary, the present invention not only solves the limitations of traditional cross-border e-commerce systems in dealing with cross-cultural differences, data dynamic updates, and real-time feedback adjustments, but also provides reliable technical support for enterprises to establish an efficient and agile marketing system in the global diversified market, with significant economic benefits and application value for promotion. Brief Description of the Drawings

[0078] In order to more clearly illustrate the technical solutions of the embodiments of the present invention, the drawings required for the description of the embodiments will be briefly introduced below. Obviously, the drawings in the following description are only some embodiments of the present invention, and those of ordinary skill in the art can obtain other drawings based on these drawings without creative efforts.

[0079] Figure 1 It is a schematic framework diagram of the intelligent management system for cross-border e-commerce marketing strategies based on big data of the present invention. Detailed Embodiments

[0080] To make the above objects, features, and advantages of the present invention more obvious and understandable, the detailed embodiments of the present invention will be described below with reference to the drawings in the specification.

[0081] In the following description, many specific details are set forth to fully understand the present invention. However, the present invention can also be implemented in other ways different from those described herein. Those skilled in the art can make similar extensions without departing from the connotation of the present invention. Therefore, the present invention is not limited by the specific embodiments disclosed below.

[0082] Secondly, the so-called "one embodiment" or "embodiment" herein refers to a specific feature, structure, or characteristic that can be included in at least one implementation of the present invention. The "in one embodiment" appearing in different places in this specification does not necessarily refer to the same embodiment, nor is it a separate or alternative embodiment that mutually excludes other embodiments.

[0083] Embodiment 1, referring to Figure 1 , this embodiment provides an intelligent management system for cross-border e-commerce marketing strategies based on big data, including:

[0084] An e-commerce data collection module that collects original e-commerce data and constructs a heterogeneous data set containing user, product, and cultural features;

[0085] The original e-commerce data includes e-commerce platform transaction records, user behavior logs, and social media text data;

[0086] The e-commerce data collection module includes an interface for cross-platform data connection and a big data engine for real-time data transmission. This engine is used to ensure the timeliness and continuity of data collection, enabling the user-product-cultural feature matrix to have the ability to be dynamically updated;

[0087] A data preprocessing module for preprocessing the heterogeneous data set;

[0088] The preprocessing includes cleaning, normalizing, and fusing the data in the heterogeneous data set to form a user-product-cultural label feature matrix in a unified format, where the cultural feature matrix is labeled according to internationally recognized cultural indicator systems;

[0089] A personalized recommendation and localization optimization module, which is constructed based on a multi-task joint learning model. This model includes:

[0090] A heterogeneous graph neural network for establishing a relationship graph between users, products, and cultural labels, and assigning dynamic weights to the relationships between nodes in the relationship graph through an attention mechanism;

[0091] A Transformer structure for processing multi-language texts and social comment information, and extracting semantic information reflecting regional sentiment and preferences from the original e-commerce data;

[0092] A meta-learning sub-module for adaptively adjusting model parameters in the case of scarce data;

[0093] In the personalized recommendation and localization optimization module, the heterogeneous graph neural network constructs a graph embedding based on the user's historical transaction data, click-through rate, and cultural similarity in different regions; the Transformer structure performs sentiment and semantic modeling based on multi-language review texts. The two are combined to output a cross-cultural adaptive recommendation result, which is used to reflect different market's different focus and consumption preferences for the same product; product recommendations are made based on the cross-cultural adaptive recommendation result;

[0094] In the personalized recommendation and localization optimization module, the steps for the heterogeneous graph neural network to construct a graph embedding based on the user's historical transaction data, click-through rate, and cultural similarity in different regions are as follows:

[0095] Calculate the association score between nodes based on the historical transaction data, click-through rate, and cultural similarity between each node. The calculation formula is:

[0096]

[0097] Among them, e ij represents the association score between node i and node j, a represents the attention weight vector, and its transpose is used to map the concatenated vector to the score space, b represents the feature transformation matrix, and x i and x j are the initial feature vectors of node i and node j respectively. ‖ represents concatenating the two vectors. c1 represents the coefficient for transaction data T ij c2 represents the coefficient for click-through rate data C ij c3 represents the coefficient for cultural similarity data R ij T ij C ij and R ij represent the historical transaction data, click-through rate index, and cultural similarity index between node i and node j respectively. i and j are node indices;

[0098] Calculate the normalized attention weight between nodes. The formula is:

[0099]

[0100] Among them, α ij represents the normalized attention weight between node i and node j, exp(·) is the exponential function, represents the neighbor set of node i, k is the neighbor node index, and e ij and e ik are the association scores between nodes respectively;

[0101] Update the node embedding by aggregating neighbor information. The update formula is:

[0102]

[0103] Among them, h i represents the graph embedding of node i, α ij is the normalized attention weight between node i and neighbor node j, b is the feature transformation matrix, and x j is the initial feature vector of neighbor node j, j is the neighbor node index, is the neighbor set of node i;

[0104] Specifically, here, the heterogeneous graph composed of user, commodity, and cultural tags is regarded as each node in the network. By integrating historical transaction data, click-through rate, and cultural similarity, the association strength between nodes is measured. The LeakyReLU function is used to perform non-linear mapping on the concatenated features, and combined with the normalized attention mechanism, the information of each neighbor node is involved in the update of node embedding with different weights, realizing the transformation from multi-source data to graph embedding;

[0105] The steps of the Transformer structure for sentiment and semantic modeling based on multi-language review texts and combining the two to output cross-cultural adaptability recommendation results are as follows:

[0106] For multi-language review texts, the Transformer structure is used to extract sentiment and semantic features. The preprocessed text is represented as an embedding vector z0, and then projections of query, key, and value are performed. The formula is:

[0107] Q = z0d, K = z0e, V = z0f,

[0108] where Q represents the query vector in the Transformer, K represents the key vector in the Transformer, V represents the value vector in the Transformer, z0 is the preprocessed text embedding, d is the query projection matrix, e is the key projection matrix, and f is the value projection matrix;

[0109] The scaled dot-product attention mechanism is used for feature fusion. The fusion formula is:

[0110]

[0111] where A represents the vector output by the attention mechanism, represents the transpose of the key vector K, g is the scaling factor, softmax(·) represents the standard normalization function, and V is the value vector;

[0112] The attention output and the initial embedding are normalized and fused to obtain the semantic representation:

[0113] z = LayerNorm(z0 + A),

[0114] where z represents the semantic representation output by the Transformer structure, LayerNorm(·) represents the layer normalization operation, z0 is the initial text embedding, and A is the attention output;

[0115] The steps of the Transformer structure for sentiment and semantic modeling based on multi-language review texts and combining the two to output cross-cultural adaptability recommendation results also include:

[0116] Performing sentiment analysis through a fully connected layer:

[0117] s = σ(iz + j),

[0118] where s represents the sentiment score, σ(·) represents the activation function, i is the sentiment analysis weight matrix, z is the semantic representation output by the Transformer, and j is the sentiment analysis bias;

[0119] Let the heterogeneous graph embedding be h, and fuse h, the Transformer semantic representation z, and the sentiment score s:

[0120]

[0121] where r represents the final cross - cultural adaptability recommendation result, k is the fusion weight matrix,

[0122] represents the vector formed by vertically concatenating the graph embedding h, the semantic representation z, and the sentiment score s,

[0123] l is the fusion bias, and h is the node embedding;

[0124] Specifically, here the Transformer structure is used to process multilingual review texts. By respectively projecting the text embeddings for query, key, and value, and then combining the scaled dot - product attention mechanism to extract semantic information, the LayerNorm operation ensures the output stability. Subsequently, the fully - connected layer and the activation function are used for sentiment score calculation. Finally, the text processing result is fused with the heterogeneous graph embedding to construct a recommendation result reflecting cross - cultural user preferences;

[0125] The way the meta - learning sub - module adaptively adjusts the model parameters in the case of scarce data is as follows:

[0126] Adopt a meta - learning strategy based on gradient update to perform inner - loop updates for each task τ, and the formula is:

[0127]

[0128] where m′ τ represents the model parameters obtained after the update within task τ, m represents the initial model parameters, p is the inner - update step size, represents the gradient calculation of the model parameters m, and l τ (m) represents the loss function of task τ, τ is the task index, representing different tasks,

[0129] In the outer loop, global parameter updates are performed according to the loss information of all tasks, and the update formula is:

[0130]

[0131] Among them, m represents the updated model parameters, q is the outer update step size, which regulates the amplitude of the global parameter update, and ∑ τ (·) represents the summation operation of all task losses, represents the gradient operation with respect to the parameter m, and l τ (m′ τ ) represents the loss value under each task τ after the inner loop update;

[0132] Specifically, here the meta - learning strategy of inner and outer loops is used to solve the problem of rapid model adaptation in the case of data scarcity. The inner loop uses single - task data to locally fine - tune the model parameters to form task - specific parameter updates, and the outer loop integrates the update information of all tasks and then corrects the global model parameters;

[0133] On the premise of ensuring the uniqueness and consistency of parameters among tasks, efficient parameter iteration under small - sample conditions is realized;

[0134] A feedback adjustment module, which collects the output of the management system and real - time market feedback data in real time, and online updates and optimizes the model parameters in the personalized recommendation and localization optimization module according to a predetermined strategy;

[0135] The feedback adjustment module includes a monitoring sub - module for real - time monitoring of the marketing strategy effect and a decision - making sub - module for feedback - driven model adaptive update. The monitoring sub - module uses real - time data to statistically analyze the performance indicators of the management system, and the decision - making sub - module automatically tunes the model parameters according to the preset safety margin and benefit threshold; realizing the system's instant response to price fluctuations, inventory status, and regional promotion strategies in the cross - border e - commerce environment;

[0136] The output data of the management system includes the recommendation results, content display data, and system self - diagnosis logs generated by the personalized recommendation and localization optimization module;

[0137] The real - time market feedback data includes real - time transaction data, user click - through rate, conversion rate, online reviews, social media sentiment analysis results, inventory status, exchange rate fluctuations, and competitor price dynamics from the e - commerce platform;

[0138] The real - time data used by the monitoring sub - module refers to the data set updated on the time scale of milliseconds to minutes in the output data of the management system and the real - time market feedback data;

[0139] The performance indicators are quantitative indicators reflecting the overall operation efficiency and recommendation effect of the system, including click - through rate, conversion rate, user stay time, recommendation accuracy, inventory turnover rate, and return on advertising cost;

[0140] The output data of the management system here reflects the internal self-diagnosis of the system and the effect of policy execution, while the real-time market feedback data is used for the immediate dynamic monitoring of the external environment. Both exist in the form of high-frequency real-time data, and the results of their statistics and analysis constitute the system performance indicators, which are used to drive the decision-making sub-module to adaptively update and optimize the model parameters according to the preset safety margin and benefit threshold;

[0141] In the feedback adjustment module, the steps of online updating and optimizing the model parameters in the personalized recommendation and localization optimization module according to the predetermined policy are as follows:

[0142] The feedback adjustment module constructs a comprehensive performance indicator by collecting the output data of the management system and the real-time market feedback data in real time, and online updates the model parameters according to the preset policy. The calculation formula of the comprehensive performance indicator is:

[0143] P = γX+(1 - γ)Y,

[0144] where P represents the comprehensive performance indicator, γ is the weight coefficient of the output data of the management system, X represents the output data of the management system, Y represents the real-time market feedback data, and 1 - γ is the supplementary weight, representing the contribution of the market feedback data.

[0145] According to the performance indicator, the preset safety margin and the benefit threshold, a deviation function is defined:

[0146] If P < u, then I(P) = u - P,

[0147] If P > v, then I(P) = P - v,

[0148] In other cases, then I(P) = 0;

[0149] where I(P) represents the performance deviation function, u is the preset safety margin, v is the preset benefit threshold, and P is the comprehensive performance indicator;

[0150] The online update formula for the final model parameters is:

[0151]

[0152] where m represents the current model parameter, w is the online update control weight, I(P) is the deviation value obtained from the performance deviation function, represents the gradient of the loss function with respect to the model parameter m, and l(m) is the loss function of the model;

[0153] Specifically, it is described here that the feedback adjustment module constructs comprehensive performance indicators by integrating the output of the management system and market feedback data, determines system deviations using preset safety margins and benefit thresholds, calculates the update amount according to the deviation function, and online adjusts the model parameters in a gradient descent manner to achieve the adaptive optimization of the system. This process makes full use of real-time data feedback to achieve precise correction of parameters, thereby ensuring that the cross-border e-commerce recommendation system can quickly respond to changes in the external environment.

[0154] It should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and not to limit them. Although the present invention has been described in detail with reference to the preferred embodiments, those of ordinary skill in the art should understand that the technical solutions of the present invention can be modified or equivalently replaced without departing from the spirit and scope of the technical solutions of the present invention, and they should all be covered within the scope of the claims of the present invention.

Claims

1. An intelligent management system for cross-border e-commerce marketing strategies based on big data, characterized in that: including an e-commerce data collection module that collects original e-commerce data and constructs a heterogeneous data set containing user, commodity, and cultural features a data preprocessing module for preprocessing the heterogeneous data set a personalized recommendation and localization optimization module, which is constructed based on a multi-task joint learning model, and the model includes a heterogeneous graph neural network for establishing a relationship graph between users, commodities, and cultural tags, and assigning dynamic weights to the relationships between nodes in the relationship graph through an attention mechanism a Transformer structure for processing multi-language text and social comment information, and extracting semantic information reflecting regional sentiment and preferences from the original e-commerce data a meta-learning sub-module for adaptively adjusting model parameters in the case of scarce data a feedback adjustment module that collects the output of the management system and real-time market feedback data in real time, and online updates and optimizes the model parameters in the personalized recommendation and localization optimization module according to a predetermined strategy 2. The intelligent management system for cross-border e-commerce marketing strategies based on big data according to claim 1, characterized in that: The original e-commerce data includes e-commerce platform transaction records, user behavior logs, and social media text data The e-commerce data collection module includes an interface for cross-platform data connection and a big data engine for real-time data transmission 3. The intelligent management system for cross-border e-commerce marketing strategies based on big data according to claim 2, characterized in that: The preprocessing includes cleaning, normalizing, and data fusion of the data in the heterogeneous data set to form a user-commodity-cultural tag feature matrix in a unified format 4. The intelligent management system for cross-border e-commerce marketing strategies based on big data according to claim 3, characterized in that: In the personalized recommendation and localization optimization module, the heterogeneous graph neural network constructs a graph embedding based on the historical transaction data, click-through rate, and cultural similarity of users in different regions; the Transformer structure performs sentiment and semantic modeling based on multi-language review texts, and the two are combined to output a cross-cultural adaptation recommendation result, which is used to reflect different market attention focuses and consumption preferences for the same commodity; product recommendations are made based on the cross-cultural adaptation recommendation result 5. The intelligent management system for cross-border e-commerce marketing strategies based on big data according to claim 4, characterized in that: In the personalized recommendation and localization optimization module, the steps for the heterogeneous graph neural network to construct a graph embedding based on the historical transaction data, click-through rate, and cultural similarity of users in different regions are Calculating the association score between nodes according to the historical transaction data, click-through rate, and cultural similarity between nodes, and the calculation formula is Among them, e ij represents the association score between node i and node j, a represents the attention weight vector, and its transpose is used to map the concatenated vector to the score space, b represents the feature transformation matrix, x i and x j are the initial feature vectors of node i and node j respectively, ‖ represents concatenating the two vectors, c1 represents the coefficient for transaction data T ij of, c2 represents the coefficient for click-through rate data C ij of, c3 represents the coefficient for cultural similarity data R ij of, T ij 、C ij and R ij represent the historical transaction data, click-through rate index, and cultural similarity index between node i and node j respectively, and i and j are node indices; Calculating the normalized attention weight between nodes, and the formula is Among them, α ij represents the normalized attention weight between node i and node j, and exp(·) is the exponential function, represents the neighbor set of node i, k is the neighbor node index, and e ij and e ik are the association scores between nodes respectively; Updating the node embedding by aggregating neighbor information, and the update formula is Among them, h i represents the graph embedding of node i, α ij is the normalized attention weight between node i and neighbor node j, b is the feature transformation matrix, x j is the initial feature vector of neighbor node j, j is the neighbor node index, is the neighbor set of node i.

6. The intelligent management system for cross-border e-commerce marketing strategies based on big data according to claim 5, characterized in that: The steps for the Transformer structure to perform sentiment and semantic modeling based on multi-language review texts and combine the two to output a cross-cultural adaptation recommendation result are For multi-language review texts, use the Transformer structure to extract sentiment and semantic features, represent the preprocessed text as an embedding vector z0, and then perform projections of query, key, and value, and the formula is Q = z0d, K = z0e, V = z0f where Q represents the query vector in the Transformer, K represents the key vector in the Transformer, V represents the value vector in the Transformer, z0 is the preprocessed text embedding, d is the query projection matrix, e is the key projection matrix, and f is the value projection matrix Feature fusion is performed using the scaled dot - product attention mechanism, and the fusion formula is: Among them, A represents the vector output by the attention mechanism, represents the transpose of the key vector K, g is the scaling factor, softmax(·) represents the standard normalization function, and V is the value vector; The attention output and the initial embedding are normalized and fused to obtain the semantic representation: z = LayerNorm(z0 + A), where z represents the semantic representation output by the Transformer structure, LayerNorm(·) represents the layer normalization operation, z0 is the initial text embedding, and A is the attention output.

7. The intelligent management system for cross-border e-commerce marketing strategies based on big data according to claim 6, characterized in that: The steps of the Transformer structure for sentiment and semantic modeling based on multilingual review texts and combining them to output cross - cultural adaptability recommendation results also include: Performing sentiment analysis through a fully - connected layer: s = σ(iz + j), where s represents the sentiment score, σ(·) represents the activation function, i is the sentiment analysis weight matrix, z is the semantic representation output by the Transformer, and j is the sentiment analysis bias; Let the heterogeneous graph embedding be h, and fuse h, the Transformer semantic representation z, and the sentiment score s: where r represents the final output cross - cultural adaptability recommendation result, k is the fusion weight matrix, denotes the vector formed by vertically concatenating the graph embedding h, the semantic representation z, and the sentiment score s. l is the fusion bias, and h is the node embedding.

8. The intelligent management system for cross-border e-commerce marketing strategies based on big data according to claim 7, characterized in that: The way the meta - learning sub - module adaptively adjusts the model parameters in the case of scarce data is: Adopting a meta - learning strategy based on gradient update to perform inner - loop updates for each task τ, and the formula is: where m ′ τ represents the model parameters obtained after the update within task τ, m represents the initial model parameters, p is the inner update step size, represents the gradient calculation for the model parameters m, l τ (m) represents the loss function of task τ, τ is the task index, representing different tasks, In the outer loop, global parameter updates are performed according to the loss information of all tasks, and the update formula is: where m represents the updated model parameters, q is the outer update step size that regulates the amplitude of the global parameter update, and ∑ τ (·) represents the summation operation over all task losses, represents the gradient operation with respect to the parameter m, and l τ (m ′ τ ) represents the loss value under each task τ after the inner-loop update.

9. The intelligent management system for cross-border e-commerce marketing strategies based on big data according to claim 8, characterized in that: The feedback adjustment module includes a monitoring sub - module for real - time monitoring of the marketing strategy effect and a decision - making sub - module for feedback - driven model adaptive update. The monitoring sub - module uses real - time data to statistically analyze the performance metrics of the management system, and the decision - making sub - module automatically tunes the model parameters according to the preset safety margin and benefit threshold; The output data of the management system includes recommendation results, content display data, and system self - diagnosis logs generated by the personalized recommendation and localization optimization module; The market real - time feedback data includes real - time transaction data, user click - through rate, conversion rate, online reviews, social media sentiment analysis results, inventory status, exchange rate fluctuations, and competitor price dynamics from e - commerce platforms; The real - time data used by the monitoring sub - module refers to the data set updated on the time scale of milliseconds to minutes in the management system output data and the market real - time feedback data; The performance metrics are quantitative metrics reflecting the overall operation efficiency and recommendation effect of the system, including click - through rate, conversion rate, user stay time, recommendation accuracy, inventory turnover rate, and advertising cost return rate.

10. The intelligent management system for cross-border e-commerce marketing strategies based on big data according to claim 9, characterized in that: In the feedback adjustment module, the steps of online updating and optimizing the model parameters in the personalized recommendation and localization optimization module according to the predetermined strategy are, The feedback adjustment module constructs a comprehensive performance metric by real - time collecting the management system output data and the market real - time feedback data, and performs online updates on the model parameters according to the preset strategy. The formula for the comprehensive performance metric is: P = γX+(1 - γ)Y, Among them, P represents the comprehensive performance index, γ is the weight coefficient of the output data of the management system, X represents the output data of the management system, Y represents the real-time market feedback data, and 1 - γ is the supplementary weight, representing the contribution of the market feedback data. According to the performance index, the preset safety margin, and the benefit threshold, define the deviation function: If P < u, then I(P) = u - P. If P > v, then I(P) = P - v. In other cases, then I(P) = 0. Among them, I(P) represents the performance deviation function, u is the preset safety margin, v is the preset benefit threshold, and P is the comprehensive performance index. The online update formula for the final model parameters is: Among them, m represents the current model parameters, w is the online update control weight, I(P) is the deviation value obtained from the performance deviation function, represents the gradient of the loss function with respect to the model parameter m, and l(m) is the loss function of the model.

Images