Big data management system based on cross-border e-commerce

Abstract

The invention relates to the technical field of e-commerce data management, in particular to a big data management system based on cross-border e-commerce, which comprises a data acquisition module, a multi-layer graph construction module, a multi-layer graph processing module and a multi-layer graph application module, and is characterized in that the data acquisition module collects commodity data used by users, including the users, commodities, operation behaviors and user emotions; the multi-layer graph construction module is used for forming a commodity recommendation multi-layer graph and constructing nodes and edges in emotion, behavior and time layer graphs so as to realize fine modeling of a complex relation; the multi-layer graph processing module calculates embedded vectors of all layers of graphs, and the embedded vectors are fused into aggregation embedded vectors through a full connection layer; and the multi-layer graph application module optimizes the training of a recommendation model by using an aggregation embedded vector, identifies potential interested commodities of the user by predicting a recommendation list, and particularly improves the relevance and individuation of recommendation in combination with an emotion layer graph weight, so that the problem that a recommendation system cannot timely and accurately respond to user requirements is solved, and the user experience is improved. And the user satisfaction and the platform viscosity are improved.

Description

Technical Field

[0001] The present invention relates to the technical field of e-commerce data management, and in particular to a big data management system based on cross-border e-commerce. Background Art

[0002] Traditional big data management systems based on cross-border e-commerce mainly rely on user behavior data for product recommendations, usually building recommendation models based on user browsing, clicking, and purchasing behaviors. This approach has several technical problems.

[0003] First, single-dimensional behavioral data cannot fully capture the complex interactive relationship between users and products, especially in a diversified market environment. Since user emotions and time dynamics are not incorporated, models built only by behavioral frequency may lead to insufficient relevance and personalization of recommendation results.

[0004] Secondly, in the process of data processing, traditional systems often have large amounts of data, single types, and incomplete data cleaning, which leads to noise data interfering with model results and further affecting the accuracy of recommendations. In addition, due to the failure to incorporate user emotional feedback, the system cannot effectively distinguish between users who may have significantly different emotional needs under similar operations, and cannot flexibly respond to dynamic changes in user preferences.

[0005] Finally, due to the neglect of time factors, traditional systems usually lack the ability to capture the recent changes in user behavior in real time, resulting in the recommendation content lagging behind the changes in user interests and being unable to dynamically respond to changes in buyer market demand. These problems make the system obviously insufficient in capturing multi-level user interaction relationships, real-time and personalized recommendations, making it difficult to carry increasingly complex user needs and provide a satisfactory user experience, which in turn restricts the user growth and user loyalty of e-commerce platforms. Summary of the invention

[0006] The purpose of the present invention is to provide a big data management system based on cross-border e-commerce to solve the problems raised in the above background technology.

[0007] To achieve the above purpose, the present invention provides the following technical solution: a big data management system based on cross-border e-commerce, including a data acquisition module, a multi-layer graph construction module, a multi-layer graph processing module and a multi-layer graph application module, wherein:

[0008] The data acquisition module acquires the user's commodity usage data, wherein the user's commodity usage data includes the user, commodity, operation behavior and user emotion;

[0009] The multi-layer graph construction module is used to construct a product recommendation multi-layer graph, wherein the product recommendation multi-layer graph includes an emotion layer graph, a behavior layer graph and a time layer graph; nodes of each layer graph in the product recommendation multi-layer graph are formed according to users and products; edges between middle nodes of the emotion layer graph are constructed according to user emotions; edges between nodes in the behavior layer graph and the time layer graph are constructed according to operation behaviors;

[0010] The multi-layer graph processing module calculates the embedding vector of each layer of the graph according to the nodes and corresponding edges of each layer of the product recommendation multi-layer graph, which includes the embedding vector of the emotion layer graph, the embedding vector of the behavior layer graph and the embedding vector of the time layer graph; and transforms and fuses the embedding vector of each layer of the graph using the fully connected layer to form an aggregated embedding vector;

[0011] The multi-layer graph application module trains the recommendation model according to the aggregated embedding vector; uses the trained recommendation model to predict the recommendation list, and selects the edge corresponding to the sentiment layer graph in the product recommendation multi-layer graph according to the products in the recommendation list, and uses it as the user's tendency score for the product.

[0012] As a further improvement of the technical solution, the multi-layer graph construction module includes a node construction unit, and the process of the node construction unit constructing a node specifically includes:

[0013] Extract all user identifiers and product identifiers according to the user and product data in the historical user product data, and generate a user identifier set and a product identifier set;

[0014] Through data cleaning and deduplication operations, duplicate items in the user identifier set and the product identifier set are removed. The unique user identifier and product identifier represent the nodes that need to be created in the graph and correspond to an independent user node and product node.

[0015] As a further improvement of the technical solution, the multi-layer graph construction module includes an edge construction unit, and the process of constructing an edge by the edge construction unit specifically includes:

[0016] Based on user sentiment data, obtain each user's sentiment score for each product, and for each user and product combination, check whether the sentiment score exists; if the sentiment score exists, add a sentiment edge between the user node and the product node; use the sentiment score as the sentiment edge weight, the weight of the sentiment edge is expressed as a continuous value, and use the sentiment edge and the corresponding node as the sentiment layer graph;

[0017] According to the operation behavior, obtain each user's interaction behavior for each product, and check each user's interaction behavior for each product. By setting the minimum interaction number threshold, determine whether to add a behavior edge; if the interaction behavior reaches the minimum interaction number threshold, create a behavior edge between the corresponding user node and the product node; and accumulate weights according to the behavior frequency; use the behavior edge and the corresponding node as the behavior layer graph;

[0018] Based on the operational behavior, the behavior records in a fixed time period are extracted as the recent behavior; and the recent behavior of each user and the product is checked. If there is a behavior greater than the average number of purchases in the fixed time period, a time edge is added between the user node and the product node; and the number of behaviors in the time period is used as the weight of the time edge; the time edge and the corresponding node are used as the time layer graph.

[0019] As a further improvement of the technical solution, the multi-layer graph processing module includes an embedding vector calculation unit, and the process of the embedding vector calculation unit calculating the embedding vector of each layer of the product recommendation multi-layer graph specifically includes:

[0020] The node sequence is generated by performing controlled random walks in the sentiment layer graph, and the parameters P and Q are used to adjust the walk strategy to balance the proximity and distance between nodes. The node sequence is input into the Skip-gram model for training, and the embedding representation of the node in the low-dimensional vector space is optimized to form the sentiment layer graph embedding vector. The parameter P is used to control the probability of the walk returning to the previous node; the parameter Q is used to control the tendency of the walk to deeply explore new nodes in the network. The Skip-gram model is a neural network model for word embedding learning, which is used to capture the semantic similarity of neighboring nodes.

[0021] For each node in the behavior layer graph, a portion of neighboring nodes are randomly sampled, and the feature information of the sampled neighboring nodes is aggregated through a learnable aggregation function, and the representation of the current node is updated in combination with its own features. This process is performed layer by layer, and after multiple updates, a stable embedding vector that can represent the node features and its neighborhood structure is finally generated and used as the embedding vector of the behavior layer graph.

[0022] It traverses each node and corresponding edge in the temporal graph, and learns and calculates the temporal graph embedding vector of each node by mapping the local structure of the node and its global position into a low-dimensional space.

[0023] As a further improvement of the technical solution, the multi-layer graph processing module includes an aggregation unit, and the process of the aggregation unit forming an aggregated embedding vector specifically includes:

[0024] The embedding vectors of the emotion layer graph, behavior layer graph, and time layer graph are input into the fully connected layer in sequence. Through parameterized linear transformation combined with nonlinear activation function, the dimension of each embedding vector is adjusted and features are extracted. The transformed vectors of each layer are concatenated and fused to obtain an aggregated embedding vector.

[0025] As a further improvement of the technical solution, the process of the multi-layer graph application module predicting the recommendation list specifically includes:

[0026] The recommendation model is trained using the aggregated embedding vector. By designing a recommendation model, the aggregated embedding vector is input into the recommendation model, and the rules of user-product interaction are learned through supervised learning. During the training process, the recommendation model adjusts its internal parameters to minimize the prediction error.

[0027] After the recommendation model training is completed, the trained recommendation model is used to generate a recommendation list, which includes different products and predicted propensity. The predicted propensity is calculated as follows:

[0028] For each item in the recommendation list, search the sentiment layer graph in the item recommendation multi-layer graph and find the sentiment edge between the corresponding user node and item node;

[0029] For each matching sentiment edge, extract its weight value; use the weight on the sentiment edge as part of the user's sentiment tendency score for the product, and directly use the sentiment weight as the tendency score; reorder the products in the recommendation list according to the calculated tendency score.

[0030] Compared with the prior art, the present invention has the following beneficial effects:

[0031] 1. This big data management system based on cross-border e-commerce realizes in-depth analysis of the complex relationship between users and products by constructing a multi-layer graph for product recommendation. First, the system not only relies on traditional user behavior data, but also comprehensively utilizes user emotional data and time information to construct a multi-layer graph including emotional layer graph, behavioral layer graph and time layer graph. In the emotional layer graph, the system collects user comment data and calculates emotional scores to convert user emotional feedback into emotional edges to reflect the user's attitude towards the product. For the behavioral layer graph, the user's interactive behavior is recorded and weights are generated by the frequency of behavior to ensure more accurate behavior modeling. At the same time, the time layer graph emphasizes the importance of recent behavior by recording the number of interactions within a specific time period, helping the system capture the latest dynamics of user interests. This multi-dimensional and multi-level graph structure successfully captures the rich relationship between users and products.

[0032] 2. This big data management system based on cross-border e-commerce greatly improves the training effect and prediction accuracy of the recommendation model by calculating the aggregate embedding vector through the multi-layer graph of product recommendation; the system first processes and fuses the embedding vectors of different layers of graphs to generate a comprehensive aggregate embedding vector, which combines the information of three levels: emotion, behavior and time, and provides a comprehensive and detailed feature representation; in the training stage, these features are input into the recommendation model for supervised learning to guide the model to identify the deep-level rules of user-product interaction; then, the recommendation list generated based on the trained model ensures that the recommended products match the basic interests of users through preliminary screening; more importantly, the system introduces the emotional edge weights in the emotional layer graph when sorting the recommendation list, and evenly considers the user's emotional tendency for each product, thereby achieving more accurate personalized recommendations. This not only improves the relevance and accuracy of the recommendation results, but also enhances user satisfaction and loyalty to the platform by better capturing user emotional resonance, effectively reducing user churn. BRIEF DESCRIPTION OF THE DRAWINGS

[0033] Figure 1 It is a schematic diagram of the overall module of the present invention;

[0034] Figure 2 A schematic diagram of a multi-layer graph building module unit of the present invention;

[0035] Figure 3 It is a schematic diagram of a multi-layer image processing module unit of the present invention.

[0036] In the figure: 100, data acquisition module; 200, multi-layer graph construction module; 201, node construction unit; 202, edge construction unit; 300, multi-layer graph processing module; 301, embedding vector calculation unit; 302, aggregation unit; 400, multi-layer graph application module. DETAILED DESCRIPTION

[0037] The following will be combined with the accompanying drawings in the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0038] Next, see Figure 1-Figure 3 The present invention provides a technical solution: a big data management system based on cross-border e-commerce, including a data acquisition module 100, a multi-layer graph construction module 200, a multi-layer graph processing module 300 and a multi-layer graph application module 400.

[0039] The data acquisition module 100 acquires the user's commodity usage data, wherein the user's commodity usage data specifically includes:

[0040] Users and products: Collect unique identifiers for each user and product, which are used to construct nodes in the subsequent product recommendation multi-layer graph;

[0041] Operational behavior: records various interactive behaviors of users on products, including browsing, clicking, purchasing, and behavior frequency data. The specific number of each behavior will be used to construct the edges between nodes in the behavior layer graph; the behavior frequency data represents the frequency and specific time of tracking behavior, so as to be used when constructing the edges between nodes in the behavior layer graph and the time layer graph; in particular, the number of times the product is purchased and clicked by users in different time periods (such as the last month) is obtained to construct the edges between nodes in the time layer graph.

[0042] User sentiment data: collects user comments on products and uses natural language processing technology to calculate sentiment scores. The sentiment scores are used to construct edges between nodes in the sentiment layer graph to reflect users' attitudes towards products.

[0043] The node construction unit 201 in the multi-layer graph construction module 200 is used to construct a product recommendation multi-layer graph, wherein the product recommendation multi-layer graph includes an emotion layer graph, a behavior layer graph, and a time layer graph; wherein, the nodes of each layer graph in the product recommendation multi-layer graph are formed according to users and products, specifically including:

[0044] Extract all user identifiers and product identifiers according to the user and product data in the historical user product data, and generate a user identifier set and a product identifier set;

[0045] Through data cleaning and deduplication operations, it is ensured that there are no duplicates in the extracted user identifier set and product identifier set. The unique user identifier and product identifier represent the nodes that need to be created in the graph and correspond to an independent user node and product node; these nodes are recorded in the product recommendation multi-layer graph to ensure that each user has his or her own unique representation in the product recommendation multi-layer graph; this process not only creates user nodes and product nodes, but also provides basic elements for other layer graphs in the subsequent product recommendation multi-layer graph.

[0046] The edge construction unit 202 in the multi-layer graph construction module 200 constructs edges between nodes in the emotion layer graph according to user emotions; and constructs edges between nodes in the behavior layer graph and the time layer graph according to operation behaviors, specifically including:

[0047] Based on user sentiment data, obtain each user's sentiment score for each product, and check whether the sentiment score exists for each combination of user and product (i.e., each comment record); if the sentiment score exists, add a sentiment edge between the user node and the product node; use the sentiment score as the sentiment edge weight, and the sentiment edge weight is expressed as a continuous value in the range of [-1,1], reflecting the user's positive or negative sentiment towards the product, and use the sentiment edge and the corresponding node as the sentiment layer graph;

[0048] According to the operation behavior, obtain each user's interaction behavior for each product, and check each user's interaction behavior for each product. By setting the minimum interaction number threshold, determine whether to add a behavior edge; if the interaction behavior reaches the minimum interaction number threshold, create a behavior edge between the corresponding user node and the product node; and accumulate weights according to the behavior frequency (different behaviors such as browsing, clicking, and purchasing can be set with different weights); for example, the click weight is 1, the purchase weight is 3, and the total weight of the behavior edge is the sum of the weights of different behaviors; the behavior edge and the corresponding nodes are used as the behavior layer graph;

[0049] Based on the operational behavior, the behavior records in a fixed time period are extracted as the recent behavior; and the recent behavior of each user and the product is checked. If there is a behavior greater than the average number of purchases in the fixed time period, a time edge is added between the user node and the product node; and the number of behaviors in the time period is used as the weight of the time edge, which highlights the importance of recent interactions; the time edge and the corresponding node are used as a time layer graph.

[0050] The embedding vector calculation unit 301 in the multi-layer graph processing module 300 calculates the embedding vector of each layer of the graph according to the nodes and corresponding edges of each layer of the product recommendation multi-layer graph, which includes the sentiment layer graph embedding vector, the behavior layer graph embedding vector and the time layer graph embedding vector, specifically including:

[0051] The node sequence is generated by performing controlled random walks in the sentiment layer graph, and the parameters P and Q are used to adjust the walk strategy to balance the proximity and distance between nodes. The node sequence is input into the Skip-gram model for training, and the embedding representation of the node in the low-dimensional vector space is optimized to form the sentiment layer graph embedding vector. The parameter P is used to control the probability of the walk returning to the previous node. Specifically, it determines the probability of backing off during the walk. The parameter Q is used to control the tendency of the walk to deeply explore new nodes in the network. Specifically, it affects the breadth of the walk. The Skip-gram model is a neural network model for word embedding learning, which is used to capture the semantic similarity of neighboring nodes.

[0052] For each node in the behavior layer graph, randomly sample some neighboring nodes, aggregate the feature information of the sampled neighboring nodes through a learnable aggregation function (such as mean or pooling), and update its representation in combination with the current node’s own features; this process is performed layer by layer, and after multiple updates, a stable embedding vector that can represent the node features and its neighborhood structure is finally generated and used as the behavior layer graph embedding vector;

[0053] By applying graph embedding technology, we traverse the structure of the time-layer graph, map the local structure of the node and its global position into a low-dimensional space, and learn and calculate the time-layer graph embedding vector of each node. The time-layer graph embedding vector can efficiently capture and represent the diverse and complex relationship characteristics between users and products. The time-layer graph embedding vector is used to predict the user's possible future behavior tendencies and help the system provide accurate personalized recommendations, thereby improving user experience and satisfaction.

[0054] The aggregation unit 302 in the multi-layer graph processing module 300 transforms and fuses the embedding vectors of each layer of the graph using a fully connected layer to form an aggregated embedding vector, specifically including:

[0055] The embedding vectors of the emotion layer graph, behavior layer graph, and time layer graph are input into the fully connected layer in sequence. The dimension of each embedding vector is adjusted and features are extracted through parameterized linear transformation combined with nonlinear activation function. The transformed vectors of each layer are concatenated or weightedly fused to obtain a comprehensive aggregate embedding vector that can represent the information of all layer graphs. This aggregate embedding vector not only retains the unique information of each layer graph, but also captures the potential relationship between different layers, providing a comprehensive and rich feature representation for subsequent user behavior prediction and personalized recommendation.

[0056] The multi-layer graph application module 400 trains the recommendation model according to the aggregated embedding vector; uses the trained recommendation model to predict the recommendation list, and selects the edge corresponding to the sentiment layer graph in the product recommendation multi-layer graph according to the products in the recommendation list, and uses it as the user's tendency score for the product, specifically including:

[0057] The recommendation model is trained using the aggregated embedding vector. By designing a recommendation model, the aggregated embedding vector is input into the recommendation model, and the rules of user-product interaction are learned through supervised learning. During the training process, the recommendation model adjusts its internal parameters to minimize the prediction error, thereby improving the prediction accuracy.

[0058] After the recommendation model training is completed, the trained recommendation model is used to generate a recommendation list, which includes different products and predicted propensity. The predicted propensity is calculated as follows:

[0059] For each product in the recommendation list, search the sentiment layer graph in the product recommendation multi-layer graph, and find the sentiment edge between the corresponding user node and the product node. The sentiment edge is established based on the user sentiment score when the sentiment layer graph was previously constructed.

[0060] For each matching sentiment edge, extract its weight value; this weight value represents the user's sentiment score for the product in the historical records, usually a continuous value from -1 to 1, reflecting the user's positive or negative sentiment towards the product;

[0061] The weight of the sentiment edge is used as part of the user's sentiment propensity score for the product. This can be done by directly using the sentiment weight or by combining it with other factors (such as user behavior data) to calculate a more comprehensive propensity score.

[0062] Based on the calculated sentiment scores, the items in the recommendation list are reordered, and items with high sentiment tendencies will be given priority in the list, thereby improving the fit of the recommendation results to the user's interests.

[0063] The multi-layer graph application module 400 can quickly and efficiently screen out potential products of interest to users by using a trained recommendation model, thereby reducing the processing scope and computing overhead, while ensuring that the recommended products have higher relevance and personalization, which enhances user satisfaction and user stickiness of the platform; and introduces weights on the emotional edges for sorting, so that recommendations are more in line with user emotions and preferences, improving the accuracy of recommendations and user satisfaction, and at the same time, by capturing the user's deep emotional resonance, increasing the user's trust and loyalty to the platform, thereby reducing the user churn rate.

[0064] The above shows and describes the basic principles, main features and advantages of the present invention. Those skilled in the art should understand that the present invention is not limited by the above embodiments. The above embodiments and descriptions are only preferred examples of the present invention and are not intended to limit the present invention. Without departing from the spirit and scope of the present invention, the present invention may have various changes and improvements, which fall within the scope of the present invention. The scope of protection of the present invention is defined by the attached claims and their equivalents.

Claims

1. A big data management system based on cross-border e-commerce, characterized in that: The system comprises a data acquisition module (100), a multi-layer graph construction module (200), a multi-layer graph processing module (300) and a multi-layer graph application module (400), wherein: The data acquisition module (100) acquires user commodity usage data, wherein the user commodity usage data includes the user, commodity, operation behavior and user emotion; The multi-layer graph construction module (200) is used to construct a commodity recommendation multi-layer graph, wherein the commodity recommendation multi-layer graph includes an emotion layer graph, a behavior layer graph and a time layer graph; forming nodes of each layer graph in the commodity recommendation multi-layer graph according to users and commodities; constructing edges between nodes in the emotion layer graph according to user emotions; and constructing edges between nodes in the behavior layer graph and the time layer graph according to operation behaviors; The multi-layer graph processing module (300) calculates the embedding vector of each layer of the graph according to the nodes and corresponding edges of each layer of the commodity recommendation multi-layer graph, which includes the embedding vector of the emotion layer graph, the embedding vector of the behavior layer graph and the embedding vector of the time layer graph; and transforms and fuses the embedding vector of each layer of the graph using a fully connected layer to form an aggregated embedding vector; The multi-layer graph application module (400) trains a recommendation model based on the aggregated embedding vector; uses the trained recommendation model to predict a recommendation list, and selects the edge corresponding to the sentiment layer graph in the commodity recommendation multi-layer graph based on the commodities in the recommendation list, and uses it as a user's inclination score for the commodity.

2. The big data management system based on cross-border e-commerce according to claim 1 is characterized in that: The multi-layer graph construction module (200) comprises a node construction unit (201), and the process of constructing a node by the node construction unit (201) specifically comprises: Extract all user identifiers and product identifiers according to the user and product data in the historical user product data, and generate a user identifier set and a product identifier set; Through data cleaning and deduplication operations, duplicate items in the user identifier set and the product identifier set are removed. The unique user identifier and product identifier represent the nodes that need to be created in the graph and correspond to an independent user node and product node.

3. The big data management system based on cross-border e-commerce according to claim 2 is characterized in that: The multi-layer graph construction module (200) includes an edge construction unit (202), and the process of constructing an edge by the edge construction unit (202) specifically includes: Based on user sentiment data, obtain each user's sentiment score for each product, and for each user and product combination, check whether the sentiment score exists; if the sentiment score exists, add a sentiment edge between the user node and the product node; use the sentiment score as the sentiment edge weight, the weight of the sentiment edge is expressed as a continuous value, and use the sentiment edge and the corresponding node as the sentiment layer graph; According to the operation behavior, obtain each user's interaction behavior for each product, and check each user's interaction behavior for each product. By setting the minimum interaction number threshold, determine whether to add a behavior edge; if the interaction behavior reaches the minimum interaction number threshold, create a behavior edge between the corresponding user node and the product node; and accumulate weights according to the behavior frequency; use the behavior edge and the corresponding node as the behavior layer graph; Based on the operational behavior, the behavior records in a fixed time period are extracted as the recent behavior; and the recent behavior of each user and the product is checked. If there is a behavior greater than the average number of purchases in the fixed time period, a time edge is added between the user node and the product node; and the number of behaviors in the time period is used as the weight of the time edge; the time edge and the corresponding node are used as the time layer graph.

4. The big data management system based on cross-border e-commerce according to claim 3 is characterized in that: The multi-layer graph processing module (300) comprises an embedding vector calculation unit (301), and the process of the embedding vector calculation unit (301) calculating the embedding vector of each layer of the commodity recommendation multi-layer graph specifically comprises: The node sequence is generated by performing controlled random walks in the sentiment layer graph, and the parameters P and Q are used to adjust the walk strategy to balance the proximity and distance between nodes. The node sequence is input into the Skip-gram model for training, and the embedding representation of the node in the low-dimensional vector space is optimized to form the sentiment layer graph embedding vector. The parameter P is used to control the probability of the walk returning to the previous node; the parameter Q is used to control the tendency of the walk to deeply explore new nodes in the network. The Skip-gram model is a neural network model for word embedding learning, which is used to capture the semantic similarity of neighboring nodes. For each node in the behavior layer graph, a portion of neighboring nodes are randomly sampled, and the feature information of the sampled neighboring nodes is aggregated through a learnable aggregation function, and the representation of the current node is updated in combination with its own features. This process is performed layer by layer, and after multiple updates, a stable embedding vector that can represent the node features and its neighborhood structure is finally generated and used as the embedding vector of the behavior layer graph. It traverses each node and corresponding edge in the temporal graph, and learns and calculates the temporal graph embedding vector of each node by mapping the local structure of the node and its global position into a low-dimensional space.

5. The big data management system based on cross-border e-commerce according to claim 4 is characterized in that: The multi-layer graph processing module (300) includes an aggregation unit (302), and the process of forming an aggregated embedding vector by the aggregation unit (302) specifically includes: The embedding vectors of the emotion layer graph, behavior layer graph, and time layer graph are input into the fully connected layer in sequence. Through parameterized linear transformation combined with nonlinear activation function, the dimension of each embedding vector is adjusted and features are extracted. The transformed vectors of each layer are concatenated and fused to obtain an aggregated embedding vector.

6. The big data management system based on cross-border e-commerce according to claim 5 is characterized in that: The process of the multi-layer graph application module (400) predicting the recommendation list specifically includes: The recommendation model is trained using the aggregated embedding vector. By designing a recommendation model, the aggregated embedding vector is input into the recommendation model, and the rules of user-product interaction are learned through supervised learning. During the training process, the recommendation model adjusts its internal parameters to minimize the prediction error. After the recommendation model training is completed, the trained recommendation model is used to generate a recommendation list, which includes different products and predicted propensity. The predicted propensity is calculated as follows: For each item in the recommendation list, search the sentiment layer graph in the item recommendation multi-layer graph and find the sentiment edge between the corresponding user node and item node; For each matching sentiment edge, extract its weight value; use the weight on the sentiment edge as part of the user's sentiment tendency score for the product, and directly use the sentiment weight as the tendency score; reorder the products in the recommendation list according to the calculated tendency score.

Images