A Method and Device for Product Selection and Advertising Push in Cross-border E-commerce Based on Regional Heat Analysis

Abstract

This application relates to the field of information technology and discloses a method and apparatus for product selection and advertising in cross-border e-commerce based on regional popularity analysis. The method integrates multi-channel data on regions, products, and users to construct a multi-granularity regional market knowledge graph and extract regional popularity vectors for candidate products. It mines user characteristics and calculates geographical distance weights to quantify the compatibility of products with regional culture, mines product association strength, and calculates dynamic recommendation scores. A decision-making system is built to generate regional product selection lists and personalized recommendation lists, and strategies are dynamically adjusted through anomaly detection. This method abandons subjective experience and the limitations of single data sources, accurately captures differences in regional market popularity, achieves deep matching between product selection and regional needs and user preferences, solves the problem of supply and demand imbalance, ensures accurate and flexible product selection, helps enterprises improve capital turnover efficiency and market competitiveness, and promotes the large-scale and refined operation of cross-border e-commerce.

Description

Technical Field

[0001] This application relates to the field of information technology, and more specifically, to a method and apparatus for cross-border e-commerce product selection and advertising push based on regional heat analysis. Background Technology

[0002] In cross-border e-commerce operations, product selection is a core element determining operational efficiency, and the accuracy of product selection directly impacts a company's cash flow efficiency and market competitiveness. Currently, most cross-border e-commerce companies rely primarily on the subjective experience of their operations staff or simply use historical sales data from a single platform as a reference when selecting products. They fail to conduct targeted analysis based on the differences in market demand across different target regions. This results in a mismatch between product selection strategies and actual consumer demand in various regions, easily leading to supply and demand imbalances such as unsold goods in some areas and stockouts in others. This hinders accurate product selection and restricts the large-scale and refined operational development of cross-border e-commerce companies. Summary of the Invention

[0003] The main purpose of this application is to provide a method for product selection and advertising in cross-border e-commerce based on regional popularity analysis. This method aims to solve the technical problem that product selection in cross-border e-commerce relies on subjective experience or single data and does not analyze regional market popularity, resulting in a disconnect between product selection and regional demand and making it difficult to achieve accurate product selection.

[0004] The first aspect of this application proposes a method for cross-border e-commerce product selection and advertising based on regional popularity analysis, including: By integrating multi-channel data on regions, products, and users, a multi-granularity regional market knowledge graph containing multi-granularity regional, product, and user nodes and their relationships is constructed, and regional popularity vectors of candidate products are extracted based on graph neural networks. The user association subgraph is extracted from the knowledge graph, the user graph structure features are extracted through graph neural network, the user consumption cycle features are mined by temporal convolutional network, and the logistic regression network is input to predict the cross-regional consumption-related features of users and calculate the location distance weight. The regional differentiation popularity index is calculated based on the regional popularity vector. The cultural fit is quantified by the semantic similarity between the product and the regional culture. The two are then combined to generate a comprehensive cultural fit score. Construct a product association graph, use graph neural networks to mine the association strength between products, and integrate the regional popularity vector, cross-regional consumption-related features, comprehensive cultural adaptability score and association strength to calculate dynamic recommendation score; A decision-making system is built based on knowledge graphs, which integrates the aforementioned data to generate regional product selection lists and personalized recommendation lists for users. The product selection and push strategies are dynamically adjusted through anomaly detection.

[0005] Furthermore, the steps of extracting user-related subgraphs from the knowledge graph, extracting user graph structure features through graph neural networks, mining user consumption cycle features by combining temporal convolutional networks, and inputting the data into a logistic regression network to predict user cross-regional consumption-related features and calculate location distance weights include: Extract user-related subgraphs from the knowledge graph, and define the dimensions of subgraph structural feature extraction and the types of node association relationships; A graph neural network is used to perform feature aggregation and quantization on the user-related subgraph to generate a user graph structure feature vector. By organizing users' historical consumption data into a time series, and then using a time series convolutional network to mine cyclical patterns and extract consumption cycle characteristics; By integrating the user graph structure feature vector and consumption cycle features, the data is input into a logistic regression network to predict cross-regional consumption-related features of users and calculate the geographical distance weights.

[0006] Furthermore, the step of using a graph neural network to perform feature aggregation and quantization on the user-related subgraph to generate a user graph structure feature vector includes the following steps: Define the node types, structural feature types, and core extraction dimensions of the user-related subgraph, and formulate feature extraction priorities, fusion rules, and aggregation order; A graph neural network is used to extract shallow and deep structural features from the user association subgraph in stages, and node attributes and association features are aggregated according to priority and order rules; The aggregated features are normalized to eliminate dimensional differences and redundant information, and then integrated into a unified structural feature set according to the fusion rules. The integrated unified structural feature set is quantized and converted into a standardized vector form to generate user graph structural feature vectors.

[0007] Furthermore, the step of using a graph neural network to extract shallow and deep structural features from the user-related subgraph in stages, and aggregating node attributes and association features according to priority and order rules, includes: Define the node attribute dimensions and relationship types of the user association subgraph, divide the shallow and deep structural feature categories and definition standards, and formulate feature extraction priorities and hierarchical aggregation rules; The first stage of graph neural network is used to extract shallow structural features, covering basic node attributes and direct relationships, and key features are selected according to priority. Based on the aforementioned key shallow features, deep structural features are extracted through the second stage of the graph neural network to mine multi-node association-derived and implicit association features; Based on the aforementioned hierarchical aggregation rules, first aggregate contractual hierarchical features, then fuse shallow key features and deep derived features, and integrate node attributes and relationship features.

[0008] Furthermore, the step of calculating a regionally differentiated popularity index based on a regional popularity vector, quantifying cultural fit through the semantic similarity between products and regional culture, and fusing the two to generate a comprehensive cultural fit score includes: Collect relevant popularity data of products in various regions, clarify the calculation dimensions and construction elements of regional differentiated popularity index, and determine the text source and type of multilingual semantic analysis; By comparing the popularity of products in each region with the overall level, we can extract the unique popularity characteristics of each region and generate a regionally differentiated popularity index by combining the core analysis dimensions of popularity differences. Extract core information and descriptions of products in multiple languages, perform semantic matching with multilingual user text in the target region, and quantify cultural compatibility. According to preset integration rules and logic, the regional differentiation popularity index and cultural compatibility are integrated to generate a comprehensive cultural compatibility score.

[0009] Furthermore, the steps of comparing the popularity of products in each region with the overall level, extracting unique regional popularity characteristics, and generating a regionally differentiated popularity index by combining core analysis dimensions of popularity differences include: Collect product popularity data from various regions and global popularity benchmark data, sort out the core analysis dimensions of popularity differences, and clarify the priority ranking and hierarchical division rules of each dimension; According to the dimensional priority and hierarchical division rules, the heat data of each region and the global heat benchmark data are compared in layers to capture the changing trends and abnormal change points. Based on the hierarchical comparison results, the unique heat change characteristics of each region at different priorities and levels are extracted and integrated with dimensional correlation to form a set of regional heat features. Based on the priority of the core analysis dimensions of heat difference and the set of regional heat characteristics, a generation logic is constructed to output a regional differentiated heat index.

[0010] Furthermore, the step of performing hierarchical comparison of regional heat data and global heat benchmark data according to the aforementioned dimensional priority and hierarchical division rules to capture changing trends and abnormal change points includes: Clearly define the priority order of the core analytical dimensions of heat difference, formulate hierarchical division standards and rules for each dimension, and determine the criteria for capturing change trends and abnormal change points; Based on the aforementioned dimensional priority, and combined with the hierarchical division standards and rules, the regional heat data and global heat benchmark data are dimensionally classified and hierarchically split to form corresponding hierarchical subsets. Compare subsets of data at the same dimension and level, track the relative global trend of regional heat based on the capture criteria, and identify and verify abnormal change points; The comparison results of the sub-datasets are recorded level by level according to the priority of dimensions and the order of hierarchical division. The verified change trends and abnormal change points are marked to form a complete comparison record.

[0011] Furthermore, the step of comparing subsets of data at the same dimension and level, tracking the relative global trend of regional heat intensity based on the capture criteria, and identifying and verifying anomalous change points includes: Extract regional and global heat maps of the same dimension and level, and clarify the core indicators, sequence, anomaly identification logic and verification process of trend tracking in the capture standard; Based on the core indicators, the two sets of subsets are compared synchronously in a predetermined order to track the relative global trend of regional popularity and record the trend details and direction. Based on the aforementioned capture criteria and identification logic, identify suspected abnormal changes that deviate from the norm from the trends, analyze the abnormal features and their occurrence scenarios, and complete the annotation. Following the verification process, and combining the abnormal features, scenarios, and the original data of the two sets of subsets, each suspected anomaly is verified to confirm the final abnormal change point.

[0012] Furthermore, the steps of constructing a product association graph, utilizing graph neural networks to mine the association strength between products, and integrating the regional popularity vector, cross-regional consumption-related features, comprehensive cultural suitability score, and association strength to calculate a dynamic recommendation score include the following steps: Collect multi-dimensional information, interaction data and related clues about products, clarify the node attributes, association types and construction logic of the product association graph, sort out the potential associations between products and construct the product association graph; The product association graph is input into a graph neural network for hierarchical multi-round mining to extract association features at different levels between products, and then quantified and integrated to form data on the strength of association between products. Extract regional popularity vectors, cross-regional consumption-related features, and comprehensive cultural suitability scores to determine the weight allocation methods, calculation, and fusion rules for each dimension. Based on the weight allocation method and fusion rules, the weights of each dimension are integrated, and combined with the correlation strength data between products, a dynamic recommendation score for products is calculated and generated.

[0013] The second aspect of this application also proposes a cross-border e-commerce product selection and advertising push device based on regional popularity analysis, including: The graph construction module is used to integrate multi-channel data on regions, products, and users to build a multi-granularity regional market knowledge graph containing multi-granularity regional, product, and user nodes and their relationships. It also extracts regional popularity vectors for candidate products based on graph neural networks. The consumption prediction module is used to extract user-related subgraphs from the knowledge graph, extract user graph structure features through graph neural networks, mine user consumption cycle features by combining temporal convolutional networks, and input logistic regression networks to predict user cross-regional consumption-related features and calculate the location distance weights. The cultural scoring module is used to calculate the regional differentiation popularity index based on the regional popularity vector, measure the cultural fit by the semantic similarity between the product and the regional culture, and integrate the two to generate a comprehensive cultural fit score. The recommendation rating module is used to construct a product association graph, use graph neural networks to mine the association strength between products, and integrate the regional popularity vector, cross-regional consumption-related features, comprehensive cultural suitability score and association strength to calculate dynamic recommendation rating; The decision adjustment module is used to build a decision system based on knowledge graphs, integrate the aforementioned data to generate regional product selection lists and user-personalized recommendation lists, and dynamically adjust product selection and push strategies through anomaly detection.

[0014] The first aspect of this plan brings the following benefits: This application integrates multi-channel data and constructs a multi-granularity knowledge graph. It relies on algorithms to objectively extract regional popularity vectors, accurately capturing differences in market popularity across different regions. Simultaneously, it integrates cultural compatibility, user consumption characteristics, and product association strength to achieve a deep match between product selection and regional needs, cultural habits, and user preferences, fundamentally changing the current supply-demand imbalance. Furthermore, dynamic recommendation scoring and anomaly adjustment mechanisms ensure the accuracy and flexibility of product selection, helping enterprises improve capital turnover efficiency, strengthen market competitiveness, and promote the transformation of cross-border e-commerce towards large-scale and refined operations. Attached Figure Description

[0015] Figure 1 This is a flowchart illustrating a cross-border e-commerce product selection and advertising push method based on regional popularity analysis, according to an embodiment of this application. Figure 2 This is a schematic diagram of the structure of a cross-border e-commerce product selection and advertising push device based on regional heat analysis according to an embodiment of this application; Figure 3 This is a schematic diagram of the structure of a computer device according to an embodiment of this application; The purpose, features, and advantages of this application will be further explained in conjunction with the embodiments and with reference to the accompanying drawings. Detailed Implementation

[0016] To make the objectives, technical solutions, and advantages of this application clearer, the following detailed description is provided in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the scope of this application.

[0017] Those skilled in the art will understand that, unless explicitly stated otherwise, the singular forms “a,” “an,” “the,” and “the” used herein may also include the plural forms. It should be further understood that the term “comprising” as used in the specification of this application means the presence of features, integers, steps, operations, elements, modules, and / or components, but does not exclude the presence or addition of one or more other features, integers, steps, operations, elements, modules, components, and / or groups thereof. It should be understood that when an element is “connected” or “coupled” to another element, it may be directly connected or coupled to the other element, or there may be intermediate elements. Furthermore, “connected” or “coupled” as used herein may include wireless connections or wireless coupling. The term “and / or” as used herein includes all or any modules and all combinations of one or more associated listed items.

[0018] Those skilled in the art will understand that, unless otherwise defined, all terms used herein (including technical and scientific terms) have the same meaning as commonly understood by one of ordinary skill in the art to which this application pertains. It should also be understood that terms such as those defined in general dictionaries should be understood to have the same meaning as in the context of the prior art, and should not be interpreted in an idealized or overly formal sense unless specifically defined as herein.

[0019] Reference Figure 1 This application provides a method for cross-border e-commerce product selection and advertising push based on regional popularity analysis, including: S1: Integrate multi-channel data on regions, products, and users to construct a multi-granularity regional market knowledge graph containing multi-granularity regional, product, and user nodes and their relationships, and extract regional popularity vectors for candidate products based on graph neural networks; S2: Extract user-related subgraphs from the knowledge graph, extract user graph structure features through graph neural networks, mine user consumption cycle features by combining temporal convolutional networks, input logistic regression networks to predict user cross-regional consumption-related features and calculate the location distance weights. S3: Calculate the regional differentiated popularity index based on the regional popularity vector, measure the cultural fit by the semantic similarity between the product and the regional culture, and integrate the two to generate a comprehensive cultural fit score; S4: Construct a product association graph, use graph neural networks to mine the association strength between products, and integrate the regional popularity vector, cross-regional consumption-related features, comprehensive cultural adaptability score and association strength to calculate dynamic recommendation score; S5: Build a decision-making system based on knowledge graphs, integrate the aforementioned data to generate regional product selection lists and user-personalized recommendation lists, and dynamically adjust product selection and push strategies through anomaly detection.

[0020] In step S1, through multi-channel data integration and knowledge graph construction technology, combined with graph neural networks (GNN), heterogeneous data such as regions, products, and users are transformed into structured relational networks, and regional popularity vectors of candidate products at different regional granularities are extracted, providing core data support for subsequent regional adaptation analysis. Taking the expansion of cross-border e-commerce platforms into the European market as an example, their multi-granularity regional division is divided into "country-state-city". The standard node association relationship includes the sales association between Munich, Germany (city) and outdoor camping equipment (product), and the association between local users and the consumption preferences of this category. After the platform collects sales data, user browsing history and merchant registration information from countries such as Germany, France and Italy, the system first standardizes and cleans the data, removes invalid and redundant data, and generates a standardized dataset. Then, it constructs a knowledge graph - the nodes cover Germany (country), Bavaria (state), Munich (city), outdoor camping equipment (product), user ID123 (user), etc., and labels the sales intensity, preference weight and other association attributes between each node. Finally, it extracts the cross-granularity association subgraph of outdoor camping equipment, and generates the regional heat vector of the product in Germany (country granularity), Bavaria (state granularity), and Munich (city granularity) through multi-layer aggregation operation of GNN. The vector values ​​are 0.85, 0.92 and 0.96 respectively, which intuitively reflects that its heat is the highest in Munich. This step breaks through the limitations of single data sources, accurately captures market differences through multi-granularity regional segmentation, visualizes data associations using knowledge graphs, and ensures accurate extraction of popularity vectors using GNNs. It does not rely on subjective experience, providing an objective data foundation for subsequent cultural adaptation scoring and dynamic recommendations. At the same time, it dynamically maintains the strength of graph associations to adapt to real-time changes in the cross-border e-commerce market, helping to refine product selection decisions.

[0021] In step S2, by integrating graph neural networks (GNN), temporal convolutional networks (TCN), and logistic regression networks, user association subgraphs are extracted from the knowledge graph and multi-dimensional features are mined to predict cross-regional consumption-related features of users and calculate the location distance weight, providing user-dimensional support for accurate push notifications. Taking user ID123, a consumer of outdoor camping equipment in Munich, Germany, as an example, its associated subgraph nodes include the user, the city of Munich, the outdoor camping equipment category, and related social user nodes. The system first defines the subgraph structure feature extraction dimensions as betweenness centrality, clustering coefficient, etc., and uses GNN to extract shallow (user basic information, direct consumption association) and deep (social derivative preferences, cross-category implicit needs) features in stages. After normalization, a structural feature vector is generated. Then, the user's consumption data for the past 12 months is organized into a time series. Through TCN, it is found that there is a concentrated consumption pattern at the end of each quarter (before the holiday), and the consumption cycle features are extracted. Finally, the two types of features are integrated into the logistic regression network to predict that the user's cross-regional consumption type is "leisure and vacation type" and the consumption capacity range is mid-to-high-end. Through the migration path tracking model, it is identified that the user may travel to Paris, France for 15 days in the next 3 months, and the weight of the geographical distance is calculated to be 0.7 (the closer the distance, the higher the weight). This step integrates graph structure and temporal features to accurately characterize users' cross-regional consumption attributes, breaking through the limitations of single feature analysis. It provides core user parameters for subsequent dynamic recommendation scoring, improves the targeting of product selection and push notifications, and adapts to the liquidity needs of cross-border consumption scenarios.

[0022] In step S3, the regional differentiation heat index is calculated by the hierarchical comparison method, and the cultural adaptability is quantified by combining the multilingual semantic similarity measure. The two are then integrated to generate a comprehensive cultural adaptability score, providing a core basis for determining the regional adaptability of products. Taking the outdoor camping equipment and user ID 123 mentioned above as examples, the system first collects sales, search, and social media popularity data for the product in Germany, France, Italy, and other countries, as well as global benchmark data. The core analysis dimensions are identified as sales growth rate and search frequency, which are then prioritized. After stratified comparison, it is found that Munich (city granularity) has a 35% higher popularity growth rate than the global average, generating a regional differentiated popularity index of 0.91. Next, the system extracts the product's multilingual core description (e.g., "lightweight waterproof camping tent") and performs semantic matching with the search keywords "tente de camping légère et étanche" and the high-frequency evaluation term "pratique pour lesvoyages" from users in Paris, France. The basic compatibility score is 0.83. Since camping equipment is not a culturally sensitive product, the cultural compatibility coefficient is set to 1.0, resulting in a comprehensive semantic compatibility score. Finally, the system merges the index and compatibility score according to preset rules, generating a comprehensive cultural compatibility score of 0.87 for the product in the Paris region. This step accurately captures regional differences in popularity and cultural adaptation needs, breaking through the limitations of single-dimensional assessment, providing quantitative basis for subsequent dynamic recommendations, and improving the matching degree between product selection and regional needs and cultural habits.

[0023] In step S4, a product association graph is constructed, and a graph neural network (GNN) is used to mine the association strength. Regional popularity vectors, cross-regional consumption characteristics, and comprehensive cultural adaptability scores are integrated to calculate a dynamic recommendation score, providing a quantitative standard for product selection and push priority. Taking outdoor camping equipment, user ID 123, and the Paris region as examples, the system first collects multi-dimensional information about the products. Using camping equipment, sleeping bags, and portable cooking utensils as nodes, it establishes association edges based on co-purchase relationships and complementary attributes to construct a product association graph. This graph is then input into a GNN for hierarchical multi-round mining to extract association features at different levels. The association strength between camping equipment and portable cooking utensils is quantified as 0.86. Next, the system extracts the product's regional popularity vector value in Paris (0.88), the user's cross-regional consumption type (leisure and vacation), and the comprehensive cultural suitability score (0.87). The weights for each dimension are determined as follows: popularity 0.3, consumption characteristics 0.25, suitability 0.25, and association strength 0.2. Finally, the system integrates and calculates according to the fusion rules, generating a dynamic recommendation score of 0.865 for the camping equipment for user ID 123. This step integrates multi-dimensional core data, mines the value of product associations, quantifies recommendation priorities, breaks through the limitations of single-factor decision-making, provides support for accurate product selection and cross-selling, and improves conversion efficiency.

[0024] In step S5, a knowledge graph decision-making system is built to integrate the multi-dimensional data from the previous steps to generate a targeted list. An anomaly detection mechanism is embedded to dynamically adjust the strategy, achieving intelligent implementation of product selection and recommendation. Taking outdoor camping equipment, user ID123, and the European market as examples, the system first integrates data such as regional popularity vectors, user cross-regional consumption characteristics, and comprehensive cultural compatibility scores to build an architecture including data storage, correlation analysis, and recommendation generation modules. Based on user ID123's predicted stay in Paris and mid-to-high-end spending power, a personalized recommendation list is generated, including camping equipment + portable cooking utensils combinations and accessories suitable for local campsites. Simultaneously, product selection lists are generated for France (country level) and Paris (city level), prioritizing the inclusion of this camping equipment and related products. The system monitors data in real time and discovers that sales of this equipment in the Paris region have declined by 20% within three weeks (an anomaly). Analysis shows this is due to price reductions by competitors, automatically lowering its product selection priority and supplementing with high-value alternatives, thus completing dynamic strategy optimization. This step integrates end-to-end data to generate an accurate list, and the anomaly detection mechanism ensures the flexibility of the strategy, breaking through the limitations of static decision-making, helping cross-border e-commerce achieve large-scale and refined operations, and enhancing market competitiveness.

[0025] In one embodiment, the step of integrating multi-channel data on regions, products, and users to construct a multi-granularity regional market knowledge graph containing multi-granularity regional, product, and user nodes and their relationships, and extracting regional popularity vectors of candidate products based on graph neural networks, includes: S10: Define the multi-granularity region division criteria and core node attributes, and clarify the types and representation methods of the relationships between product and user nodes; S11: Integrate heterogeneous data from multiple channels, classify and collect them according to node attributes, and perform standardization and cleaning to form a standardized data set; S12: Construct a knowledge graph based on the standardized data, establish a mapping between node attributes and relationships, and dynamically maintain the relationship strength and attribute updates; S13: Extract cross-granularity correlation subgraphs of candidate products, mine and integrate subgraph features through graph neural networks, and generate candidate product regional heat vectors at each regional granularity.

[0026] In this embodiment, firstly, multi-granularity regional division standards and core node attributes are defined, clarifying the types of relationships and representation methods. Taking the expansion of cross-border e-commerce into the European market as an example, regions are divided into three levels of granularity: "country-state-city." The country granularity includes Germany, France, etc.; the state granularity includes Bavaria, Germany, Île-de-France, France, etc.; and the city granularity includes Munich, Paris, etc. Among the core node attributes, product nodes include category, material, price range, etc., while user nodes include age, consumption level, social relationship, etc. The relationship type between product and user nodes is consumption preference relationship (represented as preference weight), and the relationship between region and product nodes is sales relationship (represented as sales volume percentage). Secondly, heterogeneous data from multiple channels is integrated and processed to form a standardized dataset. Sales data from e-commerce platforms in cities such as Munich, Germany, and Paris, France (e.g., 500 units of camping equipment sold monthly in Munich), user browsing and collection records, merchant onboarding information, and social media popularity data (e.g., discussion volume of camping equipment-related topics) are collected and categorized according to node attributes. Duplicate user behavior data and sales records with incorrect formats are removed, and data units and coding standards are unified to form a standardized dataset.

[0027] Then, a knowledge graph is constructed and dynamically maintained based on standardized data. A framework is built according to node classification and mapping rules, and standardized data is populated into the corresponding node attributes. For example, the attributes of the "Munich" city node are supplemented with population size, consumption capacity level, etc., establishing a sales association mapping between "Munich-outdoor camping equipment". Based on node data interaction, the initial association strength is calculated; for example, the association strength between Munich users and camping equipment consumption preferences is 0.8. When camping equipment sales in Munich increase by 30% month-on-month, this association strength is updated to 0.89. Finally, cross-granularity association subgraphs of candidate products are extracted, and regional heat vectors are generated. A cross-granularity relational subgraph of outdoor camping equipment is extracted, including three levels of regional nodes: Germany (country), Bavaria (state), and Munich (city), as well as associated user and merchant nodes. The subgraph is input into a graph neural network, and through multi-layer aggregation operations, subgraph features are mined and integrated. Data such as sales association strength and user preference weights are fused to generate regional popularity vectors for the product at each granularity: Germany (0.85), Bavaria (0.92), and Munich (0.96), consistent with the popularity data mentioned earlier. This embodiment constructs a knowledge graph and extracts popularity vectors through a standardized process, ensuring data standardization and accurate association. A dynamic maintenance mechanism adapts to market changes, providing a reliable data foundation for subsequent analysis and improving the scientific nature of product selection decisions.

[0028] In one embodiment, the steps of constructing a knowledge graph based on the standardized data, establishing a mapping between node attributes and relationships, and dynamically maintaining the relationship strength and attribute updates include: S120: Organize and standardize the classification, attribute system, and relationship categories of nodes in the data, and clarify the mapping rules between node attributes and relationships; S121: Construct a knowledge graph framework based on the mapping rules and node classification, and fill the corresponding standardized data into the node attributes and relationships; S122: Establish a correlation strength assessment system, calculate the initial correlation strength based on node data interaction and importance, and associate it with the corresponding relationship; S123: Track data updates dynamically, adjust node attribute information synchronously, and recalculate and update association strength based on new interaction data.

[0029] In this embodiment, firstly, the classification, attribute system, and relationship categories of nodes in the standardized data are sorted out, and the mapping rules are clarified. Taking the data related to cross-border e-commerce and outdoor camping equipment in the European market mentioned above as an example, the nodes are classified into four categories: region (country / state / city), product, user, and merchant. In the attribute system, the region node includes geocode, purchasing power level, cultural tags, etc., the product node includes SKU, category, price range, etc., and the user node includes ID, purchase frequency, social media account, etc. The relationship categories include region-product sales association, user-product preference association, and user-regional location association. The mapping rules are clear: sales association needs to be bound to regional granularity and product sales data, and preference association needs to be linked to user ID and product browsing / purchase records. Secondly, a knowledge graph framework is constructed based on the mapping rules and node classification, and the standardized data is populated. A two-layer framework of "node storage - association connection" is built. The regional node storage module records information such as Germany (country code DE), Bavaria (state code DE-BY), and Munich (city code DE-MUC). The product node storage module records attributes such as material and price range of outdoor camping equipment (SKU: CAMP-001). Relationship data is filled in the association connection module according to the mapping rules. For example, the sales association of "Munich - Outdoor Camping Equipment" is bound to the data of "monthly sales of 500 units", and the preference association of "User ID123 - Outdoor Camping Equipment" is bound to the data of "viewed 4 times and purchased 1 time in the past 3 months".

[0030] Next, an association strength assessment system was established to calculate the initial association strength. The assessment system uses data interaction frequency (weight 0.6) and node importance (weight 0.4) as core indicators. In node importance assessment, Munich (city node) is assigned a value of 0.9 due to its high purchasing power, and outdoor camping equipment (product node) is assigned a value of 0.85 due to its large market potential. Combining the interaction data (Munich users interact with this equipment 1200 times per month), the initial association strength of "Munich - Outdoor Camping Equipment" was calculated to be 0.82 and "User ID123 - Outdoor Camping Equipment" to be 0.78, according to the formula "Association Strength = (Interaction Frequency / Industry Average) × 0.6 + (Node Importance Product) × 0.4", thus linking them to the corresponding relationships. Finally, data updates were tracked, and attributes and association strengths were adjusted accordingly. When the monthly sales of Munich outdoor camping equipment increased to 650 units (a 30% increase month-over-month), the "Monthly Sales" attribute of the product node was updated accordingly. New interactive data showed that user ID123 had 2 additional views and 1 repeat purchase, recalculating the association strength: "Munich - Outdoor Camping Equipment" was updated to 0.91, and "User ID123 - Outdoor Camping Equipment" was updated to 0.86, ensuring that the knowledge graph data is synchronized with market dynamics in real time. This implementation method constructs and dynamically maintains the knowledge graph through a standardized process, ensuring the accuracy and timeliness of node attributes and associations, providing high-quality data support for subsequent feature extraction and predictive analysis, and improving the reliability of decision-making.

[0031] In one embodiment, the step of constructing a knowledge graph framework based on the mapping rules and node classification, and filling the standardized data into the node attributes and relationships, includes: S1210: Clearly define and standardize the classification criteria, core attributes, and categories of nodes in the data, and formulate two-way mapping rules for node attributes and relationships; S1211: Based on the classification criteria and bidirectional mapping rules, construct a knowledge graph framework and set up a node attribute storage module and an association relationship connection module; S1212: Filter standardized data by node category, enter the corresponding attribute data into the node attribute storage module, and decompose and populate the node attribute fields; S1213: Following the bidirectional mapping rules, mine the data association basis, fill the association clues into the association relationship connection module, and complete the knowledge graph construction.

[0032] In this embodiment, firstly, the classification standards, core attributes, and scope of nodes in the data are clearly defined, and two-way mapping rules are established. Taking the data related to the European cross-border e-commerce market and outdoor camping equipment mentioned above as an example, the node classification standards are set according to "functional dimension + business association," divided into four categories: region, product, user, and merchant. The scope is limited to the core European market (Germany, France, Italy, etc.), outdoor product categories, active consumer users, and compliant merchants. In terms of core attributes, regional nodes include administrative codes, purchasing power levels, and cultural labels; product nodes include SKUs, categories, price ranges, and materials; user nodes include IDs, purchase frequency, and preferred categories; and merchant nodes include qualification numbers and main product categories. The two-way mapping rules clearly state that regional nodes and product nodes are bidirectionally bound according to "administrative code-SKU," and user nodes and product nodes are bidirectionally associated according to "user ID-SKU," ensuring the consistency of data interaction. Secondly, a knowledge graph framework is built based on the classification standards and two-way mapping rules, and dedicated modules are set up. A framework structure of "four-layer node storage + one-layer association connection" is built. The node attribute storage module is divided into regional storage sub-modules, product storage sub-modules, user storage sub-modules, and merchant storage sub-modules according to categories. Each sub-module reserves interfaces for core attribute fields. The association connection module is set up with three sub-units: sales association, preference association, and onboarding association, which correspond to the core business relationships between nodes. Each unit supports the invocation of bidirectional mapping rules and data verification.

[0033] Next, standardized data is filtered by node category and entered into the node attribute storage module. Regional data such as Germany (administrative code DE), Bavaria (DE-BY), and Munich (DE-MUC) are filtered from the standardized data and entered into the regional storage sub-module. Attributes such as "Consumer Power Level A+" and "Cultural Tag: Outdoor Leisure Preference" are broken down and filled into the corresponding fields. Data such as "Material: Waterproof Oxford Cloth" and "Price Range: Mid-to-High-End" for outdoor camping equipment (SKU: CAMP-001) are filtered and entered into the product storage sub-module. Data such as "Purchase Frequency: 3 times per month" and "Preferred Category: Outdoor Equipment" for user ID123 are filtered and entered into the user storage sub-module, ensuring accurate matching between attribute data and module fields. Finally, following the two-way mapping rules, correlation evidence is mined and the correlation connection module is populated. Based on the "administrative code-SKU" mapping rule, the sales data correlation between the Munich (DE-MUC) region and CAMP-001 (monthly sales of 500 units) was discovered, and this correlation clue was populated into the sales correlation sub-unit. Based on the "user ID-SKU" mapping rule, the browsing and purchase records correlation between user ID123 and CAMP-001 (4 views and 1 purchase in the past 3 months) was discovered, and this was populated into the preference correlation sub-unit, completing the full construction of the knowledge graph and maintaining consistency with the node correlation logic mentioned above. This embodiment ensures a clear knowledge graph structure and accurate data through standardized classification, modular construction, and precise population. The bidirectional mapping rule guarantees the effectiveness of the correlation, laying a solid foundation for subsequent dynamic maintenance and feature extraction.

[0034] In one embodiment, the steps of extracting user association subgraphs from a knowledge graph, extracting user graph structure features through a graph neural network, mining user consumption cycle features using a temporal convolutional network, and inputting the data into a logistic regression network to predict user cross-regional consumption-related features and calculate location distance weights include: S20: Extract user-related subgraphs from the knowledge graph, and define the subgraph structural feature extraction dimensions and node association types; S21: Use a graph neural network to perform feature aggregation and quantization on the user-related subgraph to generate a user graph structure feature vector; S22: Organize users' historical consumption data into a time series, and use a time series convolutional network to mine cyclical patterns and extract consumption cycle characteristics; S23: Integrate the user graph structure feature vector and consumption cycle features, input them into the logistic regression network, predict user cross-regional consumption-related features, and calculate the location distance weight.

[0035] In this embodiment, firstly, user-related subgraphs are extracted from the knowledge graph, defining the subgraph structure feature extraction dimensions and node association types. Taking user ID123 from the European cross-border e-commerce example above, its related subgraph is extracted from the constructed knowledge graph, including the user ID123 node, the Munich (DE-MUC) region node, the outdoor camping equipment (CAMP-001) product node, the user's social association ID456 node, and the merchant M78 node specializing in outdoor equipment. The subgraph structure feature extraction dimensions are defined as betweenness centrality, clustering coefficient, and degree centrality, and the node association types include user-regional location association, user-product preference association, user-user social association, and user-merchant transaction association. Secondly, a graph neural network is used to aggregate and quantify the features of the user-related subgraphs, generating a user graph structure feature vector. Features are extracted in stages using a graph neural network. First, shallow features (basic attributes such as age and consumption level of user ID123, and direct purchase association with CAMP-001) are captured. Then, deep features (potential preferences for outdoor cooking utensils derived from social association ID456, and implicit needs for cross-merchant comparative consumption) are mined. After features are aggregated according to priority, they are normalized to eliminate dimensional differences and finally quantized into a standardized vector form to generate user graph structure feature vectors [0.82, 0.75, 0.91], corresponding to betweenness centrality, clustering coefficient, and degree centrality features, respectively.

[0036] Then, according to preset feature processing rules, the user's historical consumption data is organized into a time series. A temporal convolutional network is used to mine cyclical patterns and extract consumption cycle features. The preset fusion rule is weighted summation, and the weights can be adjusted according to regional market characteristics. The consumption records of user ID123 for the past 18 months are collected and organized into a time series, including data such as the purchase of CAMP-001 in March 2023, the purchase of an outdoor sleeping bag in June 2023, the purchase of portable cookware in September 2023, and the repurchase of CAMP-001 accessories in March 2024. The time series is input into the temporal convolutional network, which captures the time-dimensional dependencies through multiple convolutional kernels. The consumption cycle pattern is found to be "concentrated consumption at the end of each quarter, coinciding with holiday travel needs," and the consumption cycle features are extracted as "quarterly high-frequency consumption + targeted preference for outdoor products." Finally, the user graph structure feature vector and consumption cycle features are fused and input into a logistic regression network to predict cross-regional consumption-related features and calculate the geographical distance weights. The feature vector [0.82, 0.75, 0.91] is fused with consumption cycle features, supplemented with user static attributes (35 years old, mid-to-high-end consumption level), and then input into the logistic regression network. Based on the output of the trained model, the user's cross-regional consumption type is predicted to be "leisure and vacation driven," and the consumption capacity range is "mid-to-high-end." Through the user migration path tracking model, combined with their social dynamics and travel records, it is predicted that they will travel to Paris, France (FR-PAR) for 20 days within the next 4 months. According to the rule of "the closer the distance, the higher the weight," the location distance weight is calculated to be 0.75, forming a closed loop with the Paris region analysis above. This embodiment mines multi-dimensional user features through multi-model fusion, accurately predicts cross-regional consumption attributes and location weights, provides core basis for subsequent suitability scoring and recommendation, and improves the targeting of product selection and push.

[0037] In one embodiment, the step of using a graph neural network to perform feature aggregation and quantization on the user-related subgraph to generate a user graph structure feature vector includes the following steps: S210: Define the node types, structural feature types, and core extraction dimensions of the user-related subgraph, and formulate feature extraction priorities, fusion rules, and aggregation order; S211: Use graph neural networks to extract shallow and deep structural features from the user association subgraph in stages, and aggregate node attributes and association features according to priority and order rules; S212: The aggregated features are normalized to eliminate dimensional differences and redundant information, and integrated into a unified structural feature set according to the fusion rules; S213: Quantize the integrated unified structural feature set into a standardized vector form to generate user graph structural feature vectors.

[0038] In this embodiment, firstly, the node types, structural feature types, and core extraction dimensions of the user association subgraph are defined, and the feature extraction priority, fusion rules, and aggregation order are established. Taking the association subgraph of user ID123 above as an example, the node types include user nodes (ID123, ID456), region nodes (Munich DE-MUC), product nodes (CAMP-001, outdoor sleeping bag), and merchant nodes (M78); the structural feature types are divided into basic attribute features and association relationship features, and the core extraction dimensions are betweenness centrality, clustering coefficient, and degree centrality; the feature extraction priority is set as "association relationship features > basic attribute features", the fusion rule is "weighted summation (association relationship weight 0.6, basic attribute weight 0.4)", and the aggregation order is "aggregate features of the same type first, then fuse features across types".

[0039] Secondly, a graph neural network is used to extract shallow and deep structural features from the user association subgraph in stages, and node attributes and relationship features are aggregated according to priority and order rules. The first stage extracts shallow structural features, covering basic attributes such as user ID123's age (35 years old) and consumption level (mid-to-high-end), as well as direct relationships such as direct purchase association with CAMP-001 and geographical association with the Munich region. Key features such as "mid-to-high-end consumption level" and "CAMP-001 purchase association" are selected according to priority. The second stage mines deep structural features based on key shallow features, such as the potential preference for outdoor cooking utensils derived from the social association between ID123 and ID456, and the brand loyalty reflected by multiple transactions with merchant M78. Then, contract type features are aggregated first, and then the shallow key features and deep derived features are integrated according to fusion rules to form a preliminary feature set.

[0040] Then, the aggregated features are normalized to eliminate dimensional differences and redundant information, and integrated into a unified structural feature set according to the fusion rules. Using the Min-Max normalization method, features of different magnitudes, such as "transaction frequency (5 times)" and "number of social connections (3)," are mapped to the [0,1] interval to eliminate dimensional differences. Through feature correlation analysis, redundant features with extremely low correlation to consumption behavior, such as "user registration time," are removed. Then, according to the preset fusion rules, the standardized core features, such as "betweenness centrality (0.82)," "clustering coefficient (0.75)," and "degree centrality (0.91)," are weighted and summed to form a unified structural feature set {0.82, 0.75, 0.91}. Finally, the integrated unified structural feature set is quantized and converted into a standardized vector form to generate a user graph structure feature vector. Vector normalization is employed to convert the unified structural feature set {0.82, 0.75, 0.91} ​​into a standardized vector [0.82, 0.75, 0.91] that meets the model input requirements. Each dimension of this vector corresponds to betweenness centrality, clustering coefficient, and degree centrality features, consistent with the user feature analysis logic described earlier. This provides standardized graph structure feature data support for subsequent consumption cycle feature fusion and cross-regional consumption prediction. This embodiment extracts, aggregates, and quantifies user features through a standardized process, ensuring the accuracy and consistency of the feature vector, laying the foundation for subsequent model input, and improving the reliability of user feature analysis.

[0041] In one embodiment, the step of using a graph neural network to extract shallow and deep structural features from the user association subgraph in stages, and aggregating node attributes and association features according to priority and order rules, includes: S2110: Define the node attribute dimensions and relationship types of the user association subgraph, divide the shallow and deep structural feature categories and definition standards, and formulate feature extraction priorities and hierarchical aggregation rules; S2111: The first stage of graph neural network is used to extract shallow structural features, covering the basic attributes of nodes and direct relationships, and key features are selected according to priority. S2112: Based on the key shallow features, deep structural features are extracted through the second stage of the graph neural network to mine multi-node association-derived and implicit association features; S2113: Based on the hierarchical aggregation rules, first aggregate the contractual hierarchical features, then fuse the shallow key features and deep derived features, and integrate the node attributes and association features.

[0042] In this embodiment, firstly, the node attribute dimensions and relationship types of the user association subgraph are clarified, and the shallow and deep structural feature categories and definition standards are divided. Feature extraction priorities and hierarchical aggregation rules are then established. Taking the association subgraph of user ID123 above as an example, in the node attribute dimensions, user nodes include age, consumption level, and social media account; region nodes include consumption power level and cultural tags; product nodes include category and price range; and merchant nodes include main product category and qualification level. Relationship types include user-regional location association, user-product preference association, user-user social association, and user-merchant transaction association. The shallow structural feature category is "node basic attributes + direct relationship," defined as "directly obtainable without multi-node derivation"; the deep structural feature category is "multi-node relationship derivation + implicit relationship features," defined as "requires derivation through two or more node relationships or cannot be directly observed." The feature extraction priority is set as "direct association > basic node attributes > implicit association features > multi-node derived features", and the hierarchical aggregation rule is "features at the same level are aggregated by weight, and cross-level features are fused by shallow layer with a ratio of 0.6 and deep layer with a ratio of 0.4".

[0043] Secondly, the first stage of the graph neural network was used to extract shallow structural features, covering basic node attributes and direct relationships, and key features were selected according to priority. By traversing the subgraph through the graph neural network, the basic attributes of user ID123 (35 years old, mid-to-high-end consumption level), the basic attributes of the Munich region (consumption power level A+, cultural label: outdoor leisure preference), and the basic attributes of CAMP-001 products (mid-to-high-end price range, outdoor category) were extracted. Simultaneously, direct relationships were extracted, including user ID123's geographical association with Munich (3 years of residence), preference association with CAMP-001 (1 purchase, 4 browsing visits), and transaction association with merchant M78 (5 transactions). Four key shallow features were selected according to priority: "mid-to-high-end consumption level," "CAMP-001 preference association," "merchant M78 transaction association," and "Munich geographical association." Then, based on these key shallow features, the second stage of the graph neural network was used to extract deep structural features, uncovering multi-node relationship derivatives and implicit relationship features. Based on the "CAMP-001 preference association" and the "social association between user ID123 and ID456", we derived the multi-node derivative features of "potential preference for outdoor cookware"; combined with the attributes of "merchant M78 transaction association" and "M78 mainly sells all categories of outdoor products", we mined the implicit association features of "brand loyalty in outdoor categories"; through "Munich geographical association" and "Munich cultural label", we derived the deep feature of "high-frequency demand for outdoor leisure scenarios", forming 3 core deep structural features.

[0044] Finally, based on the hierarchical aggregation rules, first aggregate features at the same level, then fuse shallow key features and deep derived features, integrating node attributes and relationship features. In same-level aggregation, shallow features are weighted according to weights (preference association 0.3, transaction association 0.3, location association 0.2, basic attributes 0.2) to obtain a shallow aggregation result of 0.81; deep features are weighted according to weights (potential preference 0.4, brand loyalty 0.3, scenario demand 0.3) to obtain a deep aggregation result of 0.76. Then, according to the cross-level fusion rules, 0.81 × 0.6 + 0.76 × 0.4 = 0.79 is calculated, integrating to form a comprehensive feature set containing node attributes and relationships, maintaining a closed loop with the feature extraction logic described above. This embodiment accurately captures explicit and implicit user features through phased extraction and hierarchical aggregation, with priority and rules ensuring feature effectiveness, laying a solid foundation for subsequent normalization processing and vector generation.

[0045] In one embodiment, the step of calculating a regionally differentiated popularity index based on a regional popularity vector, quantifying cultural fit by measuring the semantic similarity between goods and regional culture, and fusing the two to generate a comprehensive cultural fit score includes: S30: Collect relevant popularity data of products in various regions, clarify the calculation dimensions and construction elements of regional differentiated popularity index, and determine the text source and type of multilingual semantic analysis; S31: Compare the popularity of products in each region with the overall level, extract the unique popularity characteristics of each region, and generate a regional differentiated popularity index by combining the core analysis dimensions of popularity differences. S32: Extract the core information and description of the product in multiple languages, perform semantic matching with the multilingual user text in the target region, and quantify the cultural compatibility. S33: According to the preset fusion rules and logic, integrate the regional differentiation popularity index and cultural compatibility to generate a comprehensive cultural compatibility score.

[0046] In this embodiment, firstly, relevant popularity data of the product in various regions is collected to clarify the calculation dimensions and construction elements of the regional differentiated popularity index, and to determine the text sources and types for multilingual semantic analysis. Taking the outdoor camping equipment (CAMP-001) mentioned above as an example, its popularity data in core European cities such as Munich, Germany, Paris, France, and Rome, Italy is collected, including monthly sales, search frequency, user interaction volume (favorites / comments), and social media topic discussion volume; the calculation dimensions are clarified as "sales growth rate, search share, and interaction popularity", and the construction elements include regional popularity data, global popularity benchmark value, and dimension weight; the text sources for multilingual semantic analysis are determined to be user search keywords in the target region, product reviews, and social media topic tags, with types covering multiple languages ​​such as German, French, and Italian.

[0047] Secondly, by comparing the popularity of products in each region with the overall level, we can extract the unique popularity characteristics of each region and generate a regionally differentiated popularity index by combining the core analysis dimensions of popularity differences. The global average monthly sales growth rate for outdoor camping equipment was calculated to be 20%, search share to be 5%, and average interaction intensity to be 0.6. In comparison, the Munich region showed a monthly sales growth rate of 35% (higher than the global average of 75%), a search share of 8% (higher than the global average of 60%), and an interaction intensity of 0.85 (higher than the global average of 42%), revealing unique popularity characteristics of "high growth rate, high search share, and high interaction intensity." Calculated using dimension weights (sales growth rate 0.4, search share 0.3, interaction intensity 0.3), the regional differentiated popularity index = (35% / 20%)×0.4 + (8% / 5%)×0.3 + (0.85 / 0.6)×0.3≈1.7×0.4 +1.6×0.3 +1.42×0.3≈0.68+0.48+0.43=1.59. The index for the Paris region was calculated similarly to be 1.32, consistent with the logic of regional popularity differences mentioned earlier.

[0048] Then, the core information and description of the product in multiple languages ​​were extracted and semantically matched with the text of users in multiple languages ​​in the target region to quantify cultural compatibility. The core multilingual description of CAMP-001 was extracted: German "leichte und wasserdichteZelten für outdoor-Aktivitäten" and French "tentes légères et étanches pour activités de plein air"; search keywords from users in the Paris region were collected: "tente de camping pratique" and "équipement de plein air résistant", and high-frequency evaluation words: "adapté aux voyages en France" and "facile à transporter"; through semantic similarity algorithm, the similarity between the French description and the search keywords was 0.82, and the similarity between the French description and the high-frequency evaluation words was 0.78. Since camping equipment is not a culturally sensitive product, the cultural compatibility coefficient was set to 1.0, and the quantified cultural compatibility was calculated as (0.82 + 0.78) / 2 × 1.0 = 0.8. Finally, following preset integration rules and logic, the regional differentiation popularity index and cultural compatibility are integrated to generate a comprehensive cultural compatibility score. The preset integration rule is "Comprehensive Score = Regional Differentiation Popularity Index × 0.5 + Cultural Compatibility × 0.5". Using data from the Paris region, the comprehensive cultural compatibility score is calculated as follows: Comprehensive Cultural Compatibility Score = 1.32 × 0.5 + 0.8 × 0.5 = 0.66 + 0.4 = 1.06; the comprehensive score for the Munich region is 1.59 × 0.5 + 0.88 × 0.5 = 0.795 + 0.44 = 1.235. The score results form a closed loop with the product regional compatibility logic described above, providing a quantitative basis for subsequent dynamic recommendations. This embodiment, through multi-dimensional data collection, comparative analysis, and semantic matching, accurately quantifies regional popularity differences and cultural compatibility, integrating them to generate a scientific score, providing reliable support for product selection and recommendation.

[0049] In one embodiment, the steps of comparing the popularity of goods in each region with the overall level, extracting unique regional popularity characteristics, and generating a regionally differentiated popularity index by combining core analysis dimensions of popularity differences include: S310: Collect product popularity data in various regions and global popularity benchmark data, sort out the core analysis dimensions of popularity differences, and clarify the priority ranking and hierarchical division rules of each dimension; S311: According to the dimensional priority and hierarchical division rules, perform hierarchical comparison of the heat data of each region and the global heat benchmark data to capture the changing trends and abnormal change points. S312: Based on the hierarchical comparison results, extract the unique heat change characteristics of each region at different priorities and levels, and integrate them with dimensional correlations to form a set of regional heat features; S313: Based on the priority of the core analysis dimensions of heat difference and the set of regional heat characteristics, construct the generation logic and output the regional differentiated heat index.

[0050] In this embodiment, firstly, product popularity data for each region and global popularity benchmark data are collected, core analytical dimensions for popularity differences are sorted out, and priority ranking and hierarchical division rules for each dimension are clarified. Taking the outdoor camping equipment (CAMP-001) mentioned above as an example, monthly sales, search frequency, interaction popularity (number of favorites + number of comments), and social media discussion volume data were collected from regions such as Munich, Germany, Paris, France, and Rome, Italy. At the same time, the average monthly sales (1,000 units), average search frequency (50,000 times / month), average interaction popularity (0.6), and average social media discussion volume (8,000 posts / month) of outdoor camping equipment were calculated as the global popularity benchmark data. The core analysis dimensions for the difference in popularity were sales growth rate, search share, interaction popularity, and social discussion popularity. The priority order of each dimension was "sales growth rate (0.4) > search share (0.3) > interaction popularity (0.2) > social discussion popularity (0.1)". The hierarchical division rule was set according to four levels: "excellent (30% higher than the global average), good (10%-30% higher than the global average), average (±10%), and poor (10% lower than the global average)".

[0051] Secondly, based on the priority and hierarchical classification rules, the regional popularity data and global benchmark data were compared in layers to capture trends and anomalies. Prioritizing the comparison of high-priority sales growth, the Munich region saw monthly sales of CAMP-001 reach 1350 units, a growth rate of 35% (higher than the global average of 30%, considered excellent), while the Paris region saw monthly sales of 1200 units, a growth rate of 20% (higher than the global average of 10%, considered good). Next, comparing search share, the Munich region had 75,000 search visits, accounting for 8% of global searches for similar products (higher than the global average of 3%, considered excellent), while the Paris region had 60,000 search visits, accounting for 6% (higher than the global average of 1%, considered good). Subsequently, interaction popularity and social media discussion popularity were compared, with Munich reaching the excellent level and Paris reaching the good level. The analysis revealed that the Munich region's sales growth rate had been consistently rising over the past three months, with no anomalies; however, the Paris region's social media discussion popularity decreased by 15% month-on-month, which is an anomaly but does not affect the core dimension judgment.

[0052] Then, based on the hierarchical comparison results, the unique heat change characteristics of each region at different priorities and levels were extracted, and integrated with the dimensional correlations to form a set of regional heat features. The Munich region is excellent in the high-priority dimensions of sales growth and search share, and the interaction heat and social discussion heat are growing in synergy, extracting the unique feature of "leading in core dimensions and climbing in all dimensions in synergy"; the Paris region is good in the core dimensions (sales growth and search share), with stable interaction heat but abnormal decline in social discussion heat, extracting the feature of "stable core dimensions and local fluctuations in secondary dimensions"; combined with the correlations of each dimension (sales growth and search share are positively correlated), the heat feature set of the Munich region is integrated to form {excellent core dimensions, synergy in all dimensions, no abnormal fluctuations}, and the feature set of the Paris region is {good core dimensions, local fluctuations in secondary dimensions, no core anomalies}.

[0053] Finally, based on the priority of the core analysis dimensions of heat difference and the set of regional heat characteristics, a generation logic is constructed to output the regional differentiated heat index. The generation logic is set as "the sum of the standardized score of each dimension × the corresponding priority weight, combined with the adjustment coefficient of the feature set", and the standardized score is converted according to the level (excellent = 1.5, good = 1.2, average = 1.0, poor = 0.8). Munich region: Sales growth rate 1.5 × 0.4 + Search share 1.5 × 0.3 + Interaction popularity 1.5 × 0.2 + Social discussion popularity 1.5 × 0.1 = 1.5. The feature set has no negative factors, the adjustment coefficient is 1.0, and the final index = 1.5 × 1.0 = 1.5. Paris region: Sales growth rate 1.2 × 0.4 + Search share 1.2 × 0.3 + Interaction popularity 1.2 × 0.2 + Social discussion popularity 0.8 × 0.1 = 1.16. Due to fluctuations in secondary dimensions, the adjustment coefficient is 0.95, and the final index = 1.16 × 0.95 ≈ 1.10, maintaining the logical loop of the previous regional popularity analysis. This example, through hierarchical comparison and feature extraction, accurately quantifies regional popularity differences. The generation logic balances priorities and actual characteristics, resulting in objective and reliable index results, providing a solid foundation for cultural compatibility scoring.

[0054] In one embodiment, the step of performing hierarchical comparison of regional heat data and global heat benchmark data according to the dimensional priority and hierarchical division rules, and capturing changing trends and abnormal change points, includes: S3110: Clearly define the priority order of the core analytical dimensions of heat difference, formulate hierarchical division standards and rules for each dimension, and determine the criteria for capturing change trends and abnormal change points; S3111: Sort according to the aforementioned dimensional priority, and combine with the hierarchical division standards and rules to perform dimensional classification and hierarchical splitting on the regional heat data and global heat benchmark data to form corresponding hierarchical subsets; S3112: Compare subsets of data at the same dimension and level, track the relative global trend of regional heat based on the capture criteria, and identify and verify abnormal change points; S3113: Record the comparison results of the sub-datasets layer by layer according to the priority of dimensions and the order of hierarchical division, and mark the verified change trend and abnormal change point information to form a complete comparison record.

[0055] In this embodiment, firstly, the priority ranking of the core analysis dimensions of heat difference is clarified, the hierarchical division standards and rules for each dimension are formulated, and the criteria for capturing change trends and abnormal change points are determined. Taking the outdoor camping equipment (CAMP-001) mentioned above as an example, the priority ranking of the core analysis dimensions is "sales growth rate (0.4) > search share (0.3) > interaction heat (0.2) > social discussion heat (0.1)"; the hierarchical division standard is set according to the difference from the global benchmark: excellent (above 30%), good (10%-30%), average (±10%), poor (below 10%); the change trend capture standard is "consistency of data fluctuation direction for 3 consecutive months", and the abnormal change point capture standard is "data in a single month deviates from the corresponding level threshold by 20% or more". Secondly, according to the dimension priority ranking, combined with the hierarchical division standards and rules, the heat data of each region and the global benchmark data are classified and hierarchically split to form corresponding hierarchical subsets. The global baseline data is: monthly sales growth of 20%, search share of 5%, interaction popularity of 0.6, and social discussion popularity of 8,000 posts / month. The sales growth dimension is first broken down by priority: Munich region 35% (excellent level), Paris region 20% (good level), forming excellent and good level subsets. Then, the search share dimension is broken down: Munich 8% (excellent), Paris 6% (good). The remaining dimensions are similarly broken down, with each dimension forming a corresponding level subset that matches the actual regional data.

[0056] Then, the datasets of the same dimension and level were compared, and the changing trends were tracked according to the capture criteria to identify and verify abnormal changes. In the sales growth dimension, the Munich excellent level dataset showed a growth rate of 28%, 32%, and 35% in the past three months, showing a continuous upward trend without any anomalies; the Paris good level dataset showed a growth rate of 18%, 20%, and 22%, showing steady growth. In the social discussion popularity dimension, the Paris good level dataset had 9,000 and 8,500 entries in the first two months, respectively, but dropped sharply to 6,800 entries in the third month, deviating from the good level threshold (7,200 entries), which met the anomaly criteria; verification revealed that this was due to a competitor's new marketing campaign diverting topics, confirming this abnormal change. Finally, the comparison results were recorded level by level according to dimension priority and level division order, and the trends and anomaly information were marked to form a complete comparison record. Following the order of "sales growth rate → search share → interaction popularity → social discussion popularity," the specific values ​​of each sub-dataset at each level in each region are recorded, along with their differences from the global benchmark. Information such as "continuous rise across all dimensions" in Munich and "steady growth in core dimensions, with an abnormal decline in social discussion popularity in the third month" in Paris is noted. The values, deviations, and verification reasons for anomalies are fully recorded, providing clear data support for subsequent feature extraction and index generation, and maintaining a logical loop with the preceding text. This embodiment, through a standardized hierarchical comparison process, accurately captures regional popularity trends and anomalies, recording complete and logically coherent data, providing a reliable basis for subsequent index generation, and improving the accuracy of regional popularity analysis.

[0057] In one embodiment, the step of comparing subsets of data at the same dimension and level, tracking the relative global trend of regional heat intensity based on the capture criteria, and identifying and verifying anomalous change points includes: S31120: Extract regional and global heat subsets of the same dimension and level, and clarify the core indicators, sequence, anomaly identification logic and verification process of trend tracking in the capture standard; S31121: Based on the core indicators, compare the two sets of subset datasets synchronously in a predetermined order to track the relative global trend of regional popularity and record the trend details and direction; S31122: Based on the aforementioned capture criteria and identification logic, identify suspected abnormal change points that deviate from the norm from the trend, sort out the abnormal features and occurrence scenarios, and complete the annotation; S31123: Following the verification process, and combining the abnormal features, scenarios, and the original data of the two sets of subsets, verify each suspected abnormal point to confirm the final abnormal change point.

[0058] In this embodiment, firstly, regional and global popularity subsets of the same dimension and level are extracted to clarify the core indicators, order, and anomaly identification logic and verification process. Taking the social discussion popularity dimension of outdoor camping equipment (CAMP-001) mentioned above as an example, a good-level subset of data from the Paris region (9000 and 8500 entries in the first two months, and 6800 entries in the third month) and a global good-level subset of data (stable at 8000-8500 entries per month) are extracted. The core indicator for trend tracking is "monthly growth rate," and the predetermined order is "comparing month by month along the time axis." The anomaly identification logic is "a single month's month-on-month growth rate deviating from the average of the same level by 20% or more," and the verification process is "first checking for data errors, then analyzing market factors, and finally confirming the nature of the anomaly." Secondly, the core indicators are compared synchronously in a predetermined order to track the changing trends and record details. The month-on-month growth rates for the Paris region and the overall global growth rate were calculated on a timeline: Paris's first month saw a month-on-month decrease of 5.56%, while the overall growth rate remained at 0%; the second month saw a month-on-month decrease of 12.94%, while the overall growth rate was 0.63%; and the third month saw a month-on-month decrease of 20%, while the overall growth rate was 0.5%. A simultaneous comparison revealed that the growth rate in the Paris region continued to decline and the rate of decline widened, with the decline in the third month far exceeding the overall stable level. The recorded trend detail was that "social discussion activity in the Paris region decreased month by month, entering a rapid decline in the third month, which is significantly different from the overall stable trend."

[0059] Then, based on the capture criteria and identification logic, suspected anomalies were identified and their features and scenarios were labeled. According to the anomaly identification logic, the month-on-month growth rate of social media discussion in the Paris region in the third month was -20%, deviating from the average of the same level (-5%) by 15 percentage points, exceeding the 20% anomaly threshold, and was judged as a suspected anomaly. The anomaly characteristics were identified as "sudden drop in a single month, with the drop exceeding the threshold", and the scenario was "a competitor launching new camping equipment and initiating social media marketing activities", thus completing the feature and scenario labeling. Finally, the anomaly was confirmed by verifying it one by one according to the verification process and combining multi-dimensional information. First, data errors were checked, and the original statistical data of social media discussion in the Paris region in the third month was checked to confirm that there were no statistical errors. Then, market factors were analyzed, and the timeline of the competitor's marketing activities was retrieved. It was found that the time of their new product promotion and the decline in popularity in the Paris region completely overlapped. Combined with the original data of the subset dataset (6800 records in Paris in the third month vs. 8400 records globally), it was confirmed that the suspected anomaly was a real market change caused by the diversion of traffic by competitors, and was finally judged as a valid anomaly, forming a closed loop with the stratified comparison logic mentioned above. This embodiment accurately identifies real abnormal change points through standardized comparison, identification, and verification processes, eliminates data error interference, provides a reliable basis for extracting regional heat characteristics, and improves the accuracy of analysis results.

[0060] In one embodiment, the steps of constructing a product association graph, utilizing graph neural networks to mine the association strength between products, and fusing the regional popularity vector, cross-regional consumption-related features, comprehensive cultural suitability score, and association strength to calculate a dynamic recommendation score include the following steps: S40: Collect multi-dimensional information, interaction data and related clues about products, clarify the node attributes, association types and construction logic of the product association graph, sort out the potential associations between products and construct the product association graph; S41: Input the product association graph into a graph neural network, perform hierarchical multi-round mining, extract the association features between products at different levels, quantify and integrate them to form product association strength data; S42: Extract regional popularity vectors, cross-regional consumption-related features, and comprehensive cultural suitability scores; determine the weight allocation method and calculation and integration rules for each dimension. S43: Based on the weight allocation method and fusion rules, integrate the weights of each dimension, combine the correlation strength data between products, and calculate and generate a dynamic recommendation score for the products.

[0061] In this embodiment, firstly, multi-dimensional information, interaction data, and related clues of the products are collected to clarify the node attributes, association types, and construction logic of the product association graph, and potential associations are sorted out and the graph is constructed. Taking the outdoor camping equipment (CAMP-001) mentioned above as an example, its multi-dimensional information (category, material, price range), user interaction data (co-purchase records with sleeping bags and portable cookware), and related clues (keywords such as "camping set" and "used together" in user reviews) are collected; the node attributes are set as SKU, category, price range, and user interaction frequency; the association types are divided into complementary association, substitution association, and co-purchase association; the construction logic is "deriving association relationships based on user behavior and product attributes"; the potential associations between CAMP-001 and portable cookware (complementary), high-end sleeping bags (co-purchase), and simple tents (substitute) are sorted out, and the product association graph is constructed with products as nodes and association relationships as edges. Secondly, the product association graph is input into a graph neural network, and the association features are mined in layers and multiple rounds to form association strength data. Following a hierarchical priority of "core association (complementary / co-purchase) → secondary association (substitution)," the first round of data mining was initiated to extract co-purchase interaction features between CAMP-001 and portable cookware. The second round of mining extracted scenario complementarity features with high-end sleeping bags, and the third round extracted functional substitution features with simple tents. The features at each level were quantified and processed. Combined with indicators such as co-purchase frequency and evaluation association degree, the association strength between CAMP-001 and portable cookware was calculated to be 0.86, with high-end sleeping bags 0.78, and with simple tents 0.65. These were then integrated to form an association strength dataset.

[0062] Next, multi-dimensional core data is extracted to determine the weight allocation method and fusion rules. The following data is extracted: CAMP-001's popularity vector value in the Paris region (0.88), user ID123's cross-regional consumption characteristics (leisure and vacation type, mid-to-high-end consumption), and comprehensive cultural suitability score (1.06). The weight allocation is determined as follows: regional popularity 0.3, consumption characteristics 0.25, cultural suitability 0.25, and association strength 0.2. The fusion rule is "standardized score of each dimension × sum of corresponding weights". Finally, based on the weights and fusion rules, and combined with the association strength data, a dynamic recommendation score is calculated. The data across all dimensions is standardized (heat vector 0.88→0.88, consumer feature matching degree 0.9→0.9, cultural adaptability 1.06→1.06, association strength takes the highest value 0.86→0.86); calculated according to the rules: 0.88×0.3 + 0.9×0.25 + 1.06×0.25 + 0.86×0.2 = 0.264 + 0.225 + 0.265 +0.172 = 0.926, generating a dynamic recommendation score for user ID123, thus closing the logical loop with the previous region and user analysis. This embodiment, through graph construction, feature mining, and multi-dimensional fusion, accurately quantifies the priority of product recommendations, providing a scientific basis for product selection and push, and improving the conversion efficiency and user experience of cross-border e-commerce.

[0063] In one embodiment, the step of inputting the product association graph into a graph neural network, performing hierarchical multi-round mining, extracting association features at different levels between products, quantifying and integrating them to form product association strength data includes: S410: Obtain the product association graph, clarify the dimensions, priorities and division criteria of the graph neural network hierarchical mining, and determine the order, triggering conditions and feature extraction standards of each level of multi-round mining; S411: Input the product association graph into the graph neural network, start the first round of mining according to hierarchical priority, extract the initial association features of high priority levels, and record the mining results and details; S412: Based on the triggering conditions, conduct multiple rounds of hierarchical mining, extract low-priority hierarchical features in sequence, verify, supplement, and correct deviations of all initial associated features, and integrate the associated features of each level; S413: Formulate quantitative rules to quantify the correlation characteristics of each level after integration, sort out the quantitative data and integrate them systematically to form a data set of correlation strength between products.

[0064] In this embodiment, firstly, a product association graph is obtained, clarifying the dimensions, priorities, and division criteria for hierarchical mining using the graph neural network, and determining the order, triggering conditions, and feature extraction standards for multiple rounds of mining. Taking the product association graph of the outdoor camping equipment (CAMP-001) mentioned above as an example, the mining dimensions are divided into functional association, scene association, and user behavior association, with the priority ranking as "user behavior association (0.4) > scene association (0.3) > functional association (0.3)", and the division criteria as "the degree of influence of association on consumption decisions"; the order of multiple rounds of mining is "high priority → low priority", and the triggering condition is "the completeness of the features mined in the previous round ≥ 90%"; the feature extraction standards for each level are clearly defined: user behavior association must include the frequency of co-purchase and the overlap of browsing paths, scene association must match the consistency of usage scenarios, and functional association must satisfy the attributes of functional complementarity or substitution. Secondly, the product association graph is input into the graph neural network, and the first round of mining is started according to priority, extracting high-priority initial association features and recording details. The first round of analysis focused on user behavior correlations, revealing that CAMP-001 was co-purchased with portable cookware 320 times per month with a 78% overlap in browsing paths, and with high-end sleeping bags 280 times per month with a 65% overlap. The analysis results showed that "CAMP-001 has a strong user behavior correlation with portable cookware and high-end sleeping bags," with details including specific frequencies and overlap values ​​to ensure data traceability.

[0065] Then, based on the triggering conditions, multiple rounds of hierarchical mining were conducted to extract low-priority features and verify and supplement them. Because the first round of feature completeness reached 95%, a second round of scene association mining was triggered, revealing that CAMP-001, portable cookware, and high-end sleeping bags are all suitable for "outdoor camping scenarios," with scene consistency of 85% and 76%, respectively. A third round of functional association mining confirmed that the three are functionally complementary, with complementarity scores of 82 and 73, respectively. Verification of the initial features revealed that the substitution association between CAMP-001 and simple tents was not captured; after supplementary mining, the bias was corrected, and a complete set of associated features was integrated. Finally, quantification rules were formulated to quantify each associated feature, integrating them into a set of association strength data. The quantification rules were set as follows: co-purchase frequency weight 0.3, browsing overlap 0.2, scene consistency 0.3, and functional complementarity 0.2, calculated as "sum of each score × weight". The quantitative score for CAMP-001 and portable cooking utensils is: (320 / 400)×0.3 + 0.78×0.2 + 0.85×0.3 + 0.82×0.2 = 0.24 + 0.156 + 0.255 + 0.164 = 0.815, rounded to 0.82. The score is 0.76 with high-end sleeping bags and 0.63 with simple tents (a substitute association). These scores are integrated to form the association strength data set {0.82, 0.76, 0.63}, creating a logical loop with the association analysis above. This embodiment, through hierarchical multi-round mining and standardized quantification, accurately extracts product association features and transforms them into quantifiable data, ensuring the objectivity and accuracy of association strength and providing reliable support for dynamic recommendation scoring.

[0066] In one embodiment, the steps of obtaining the product association graph, clarifying the dimensions, priorities, and division criteria of the graph neural network hierarchical mining, and determining the order, triggering conditions, and feature extraction standards for each level of mining include the following steps: S4101: Obtain the product association graph, sort out the node types, association attributes, association patterns and path features in the graph, and extract key information and core content of the graph; S4102: Based on the key information and core content of the graph, define the core dimensions of graph neural network hierarchical mining, and clarify the division logic, basis and core connotation of each dimension; S4103: Based on the actual value, correlation, and mining difficulty of each mining dimension, determine the dimension priority, plan the order of multiple mining rounds, and set the trigger conditions for each mining round. S4104: Based on the aforementioned hierarchical mining dimensions, dimension priorities, and division criteria, define the boundaries for feature extraction at each level, refine the extraction requirements and judgment criteria, and formulate a unified standard for feature extraction.

[0067] In this embodiment, firstly, a product association graph is obtained, and the node types, association attributes, association patterns, and path features within the graph are analyzed to extract key information and core content. Taking the product association graph of outdoor camping equipment (CAMP-001) mentioned above as an example, the graph node types include outdoor category products (CAMP-001, portable cookware, high-end sleeping bags, etc.) and consumption scenario tags (outdoor camping, leisure vacation); the association attributes include co-purchase frequency, functional complementarity, and scenario matching degree; the association patterns are divided into "product-product direct association" and "product-scenario-product indirect association"; the path features present high-frequency paths such as "CAMP-001→portable cookware" and "CAMP-001→outdoor camping scenario→high-end sleeping bag", and the core content extracted is "outdoor category products are mainly associated through co-purchase and scenario matching". Secondly, based on the key information and core content of the graph, the core dimensions of layered mining are defined, and the logic, basis, and core connotation of each dimension are clarified. The core dimensions defined by the association logic are user behavior association, scenario association, and function association. The user behavior association classification logic is "derived based on actual user interaction data," based on the frequency of co-purchase and the degree of overlap in browsing paths, and its core connotation is the strength of product association driven by user behavior. The scenario association classification logic is "derived based on the consistency of usage scenarios," based on the scenario matching degree, and its core connotation is the adaptability of product association under scenario requirements. The function association classification logic is "derived based on the complementary / substitutable attributes of product functions," based on the degree of functional complementarity, and its core connotation is the product association value determined by functional attributes.

[0068] Then, based on the actual value, correlation, and mining difficulty of each dimension, the dimension priority was determined, the mining sequence was planned, and trigger conditions were set. The actual value and correlation were ranked as "user behavior correlation > scenario correlation > function correlation," and the mining difficulty was ranked as "function correlation > scenario correlation > user behavior correlation." The overall priority was determined as "user behavior correlation (0.4) > scenario correlation (0.3) > function correlation (0.3)." The multi-round mining sequence was planned as "user behavior correlation → scenario correlation → function correlation." The trigger conditions were set as "feature integrity of the previous round of mining ≥ 90%" and "no missing key data" to ensure the orderly progress of the mining process and avoid deviations in results due to insufficient data. Finally, based on the hierarchical mining dimensions, priorities, and division criteria, the extraction boundaries were defined, the requirements and criteria were refined, and a unified extraction standard was formulated. User behavior association extraction boundaries are limited to "user interaction data from the past 12 months," requiring a co-purchase frequency of ≥50 times / month and a browsing path overlap of ≥50%, with the criterion being "meeting any one of these conditions determines an association." Scene association extraction boundaries are limited to "core consumption scene tags," requiring a scene matching degree of ≥60%, with the criterion being "functions directly related to scene requirements." Function association extraction boundaries are limited to "core functional attributes," requiring a functional complementarity of ≥55%, with the criterion being "no functional conflicts and improved user experience." This forms a unified standard for association feature extraction, creating a logical closed loop with the aforementioned graph. This embodiment, through systematically organizing graph information and scientifically defining dimensions and standards, provides clear guidance for subsequent layered multi-round mining, ensuring the accuracy and consistency of association feature extraction and improving the reliability of association strength quantification.

[0069] In one embodiment, the steps of building a decision-making system based on a knowledge graph, integrating the aforementioned data to generate a regional product selection list and a personalized user recommendation list, and dynamically adjusting the product selection and push strategy through anomaly detection include: S50: Collect multi-dimensional relationship data, clarify the functional requirements, core modules and construction logic of the knowledge graph decision system, and build the system infrastructure; S51: Filter, clean, classify and integrate the multi-dimensional relationship data, establish a data association model and mapping relationship, and import the processed association data into the knowledge graph decision system to complete the data filling. S52: Based on the data integrated by the knowledge graph decision system, mine users’ personalized needs, preference characteristics, regional market characteristics, and product selection pain points to generate regional product selection lists and user-personalized recommendation lists; S53: An embedded anomaly detection mechanism monitors the rationality of the product selection list in the region and system data anomalies, captures anomaly information and analyzes the causes of deviations, and dynamically optimizes product selection and push strategies based on the analysis results.

[0070] In this embodiment, firstly, multi-dimensional relationship data is collected to clarify the functional requirements, core modules, and construction logic of the knowledge graph decision-making system, and to build the system's basic architecture. Taking the European cross-border e-commerce and outdoor camping equipment (CAMP-001) scenario mentioned above as an example, the collected multi-dimensional relationship data covers all the core data mentioned above, including regional popularity vectors (Munich 0.96, Paris 0.88), regional differentiation popularity index (Munich 1.5, Paris 1.10), comprehensive cultural adaptability score (Munich 1.235, Paris 1.06), user graph structure feature vector [0.82, 0.75, 0.91], cross-regional consumption characteristics (leisure and vacation type, mid-to-high-end), geographical distance weight 0.75, and product association strength data (such as portable cookware 0.82). The system's functional requirements are clearly defined as "regional product selection decision-making, personalized user push, and dynamic strategy optimization." The core modules are divided into a data storage module, a correlation analysis module, a list generation module, an anomaly detection module, and a strategy adjustment module. The construction logic is "to use knowledge graphs as the core carrier, integrate data across the entire chain, and achieve closed-loop operation from data input to strategy output." A four-layer infrastructure architecture of "data access layer - core processing layer - output layer - optimization layer" is built to ensure that each module works in synergy and adapts to the actual business needs of cross-border e-commerce product selection and push.

[0071] Secondly, the multi-dimensional relational data is screened, cleaned, and categorized for integration. A data association model and mapping relationship are established to populate the system with data. Invalid and redundant data are removed, while core valid data is retained. Data biases are cleaned and corrected (e.g., data quantification standards are standardized to the [0,1] range). Data is then categorized and integrated into four main types: regional data, user data, product data, and relational data. A data association model is established to clarify core relationships such as the positive correlation between regional popularity data and cultural suitability scores, and the matching mapping between user consumption characteristics and product association strength. For example, a precise mapping is established between user ID123's mid-to-high-end consumption characteristics and CAMP-001's price range attributes, and the cultural suitability score of the Paris region is bound to the product association strength. The processed standardized relational data is then imported into the decision-making system according to module division of labor. The data storage module retains the original data, while the association analysis module loads the mapping relationships, providing high-quality data support for subsequent list generation and strategy adjustments, ensuring data consistency with the preceding analysis logic.

[0072] Then, based on the data integrated by the decision-making system, the system mines users' personalized needs, preferences, regional market characteristics, and product selection pain points to generate corresponding lists. For example, it mines users ID123's personalized needs: combining their leisure and vacation-oriented cross-regional consumption characteristics, outdoor product preferences, and Paris-related planning, identifying their needs as "lightweight, waterproof, and suitable mid-to-high-end gear combinations for short-distance camping." It also mines regional market characteristics: Munich has a strong outdoor leisure atmosphere, with high popularity and suitability, but the product selection pain point is "a lack of high-end accessory combinations"; Paris has good suitability but slightly lower popularity, with the pain point being "the need to improve the accuracy of product scenario adaptation." Based on this, a regional product selection list is generated: the Munich list prioritizes CAMP-001 and related high-end accessories (portable cookware, high-end sleeping bags), while the Paris list includes CAMP-001, lightweight suitable models, and complementary equipment. Finally, a personalized recommendation list is generated: ID123 is recommended a CAMP-001 + portable cookware combination, supplemented with accessories suitable for popular camping scenarios in Paris, achieving precise matching between product selection and users / regions.

[0073] Finally, an anomaly detection mechanism is embedded to monitor the rationality of the list and data anomalies, analyze the causes, and dynamically optimize the strategy. The anomaly detection mechanism sets monitoring indicators: product suitability compliance rate ≥ 85%, sales fluctuation ≤ 20% month-on-month, and user recommendation click-through rate ≥ 15%. Monitoring revealed that the CAMP-001 recommendation click-through rate in the Paris region was only 10% (below the threshold), which is considered an anomaly. The analysis showed that the deviation was due to "the recommendation not combining the core demand of Parisian users for 'portability'." Simultaneously, it was detected that the sales of high-end accessories in the Munich list declined by 22% month-on-month, with the anomaly being "high pricing." Based on the analysis results, dynamic optimization was implemented: the recommendation list for ID123 was adjusted, adding lightweight accessories and removing high-priced redundant products; the pricing of high-end accessories in the Munich list was lowered, and high-value alternatives were added; the product selection priority was updated simultaneously, forming a closed loop of "monitoring-analysis-optimization," consistent with the overall process logic described above. This embodiment solves the problem of accuracy in product selection and push notifications by standardizing the decision-making system, accurately mining needs to generate the list, and combining anomaly detection to achieve dynamic strategy optimization, ensuring operational flexibility and facilitating refined operations in cross-border e-commerce.

[0074] In one embodiment, the step of using a graph neural network to perform feature aggregation and quantization on the user-related subgraph to generate a user graph structure feature vector includes the following steps: S510: Organize the original data of the multi-dimensional relationship, clarify the core indicators, conditions and priorities of data screening, and screen effective target data according to these requirements, and eliminate invalid, redundant and irrelevant data; S511: Clean and verify the selected valid target data, correct data deviations and errors, fill in missing information, unify data format and standards, and form an accurate and complete standardized data set. S512: Classify and integrate the standardized data according to category, related attributes and business logic, sort out the data association rules, construct the data association model, and clarify the corresponding mapping relationship between various types of data; S513: In accordance with the data interface specifications and import requirements of the knowledge graph decision system, import the classified and integrated data with completed association mapping into the system, simultaneously verify the data integrity, and complete the system data filling.

[0075] In this embodiment, firstly, the original data of multi-dimensional relationships is organized, the core indicators, conditions, and priorities for data screening are clarified, and effective target data is screened while invalid and redundant information is eliminated. Taking the European cross-border e-commerce and outdoor camping equipment (CAMP-001) scenario mentioned above as an example, the original data of multi-dimensional relationships organized covers the entire chain data mentioned above, including the regional dimension (multi-granular regional information of countries such as Germany and France, popularity vector, and differentiated popularity index), the user dimension (ID123 related subgraph data, graph structure feature vector, consumption records, and cross-regional consumption characteristics), the product dimension (attributes of CAMP-001 and related products, and association strength), and the association relationship dimension (user-product preference association, region-product sales association, etc.). The core indicators for data screening are clarified as "data completeness, business relevance, and accuracy," and the screening conditions are "data missing rate ≤5%, directly related to product selection / push business, and no obvious abnormalities in quantitative data." The priority order is "user core feature data > regional popularity and suitability data > product association data > auxiliary association data." Based on the requirements, valid data was filtered out, retaining core data such as user ID123's mid-to-high-end consumption level and leisure and vacation-type cross-regional consumption characteristics. Redundant information such as user registration time and irrelevant product attributes were removed, as well as invalid data with abnormally deviating regional popularity. This ensured that the filtered data was consistent with the core logic of the previous case, laying the foundation for subsequent processing.

[0076] Secondly, the selected valid target data is cleaned and verified to correct deviations and errors, fill in missing information, unify the format and standards, and form a standardized data set. The filtered data underwent cleaning and verification: data bias was corrected by refining the original data of user ID123's consumption frequency ("2-3 times per month") to the precise value of "3 times per month"; the original data of CAMP-001's association strength in Paris ("0.81-0.83") was refined to a standardized value of "0.82"; missing information was supplemented by deriving the missing value of user ID123's location distance weight through association data, and supplemented to "0.75" based on its Paris stay plan; scenario matching data for some of CAMP-001's associated products was supplemented; data errors were corrected by correcting the regional coding errors to ensure the accuracy of the "Munich DE-MUC, Paris FR-PAR" codes; and data format and caliber were standardized by mapping all quantitative data to the [0,1] interval, converting qualitative data such as user consumption characteristics and product attributes into standardized codes, and uniformly retaining three decimal places for data such as regional popularity and suitability scores. This resulted in an accurate, complete, and standardized dataset with a unified format, eliminating data dimensional differences and biases, and ensuring data usability.

[0077] Then, the standardized data is categorized and integrated according to type, related attributes, and business logic. Association rules are clarified, a data association model is constructed, and the mapping relationships between various types of data are defined. The data is integrated into four main categories: regional data, user data, product data, and relationship data. Regional data is integrated into heat vectors, differentiated heat indices, and cultural suitability scores for core regions such as Munich and Paris. User data is integrated into the core characteristics of user ID123, consumption cycles, and cross-regional consumption attributes. Product data is integrated into the attributes and association strength of CAMP-001 and related products. Relationship data is integrated into the association information between various nodes. We analyzed the data association rules and clarified core rules such as "positive correlation between user consumption characteristics and product attributes", "positive correlation between regional popularity and product suitability score", and "positive correlation between product association strength and recommendation priority". We constructed a data association model to clarify the corresponding mapping relationships between various types of data. For example, the mid-to-high-end consumption characteristics of user ID123 are mapped one-to-one with the mid-to-high-end price range attribute of CAMP-001. The comprehensive cultural suitability score of the Paris region is positively correlated with the recommendation weight of CAMP-001. The cross-regional consumption characteristics of users are matched with the regional product selection list, ensuring that the association model fits the business logic described above.

[0078] Finally, following the data interface specifications and import requirements of the knowledge graph decision-making system, the categorized and integrated data with completed association mappings was imported into the system. Integrity was simultaneously verified, and system data population was completed. The data interface specifications for the knowledge graph decision-making system were clearly defined: data was imported according to the division of labor among core modules. The data storage module received the raw, standardized data; the association analysis module received the association model and mapping relationship data; and the list generation module received the categorized and integrated core data. The import requirements were "data format compatibility, accurate association mapping, and no data omissions." Data was imported step-by-step according to the requirements. Standardized data on regions, users, products, and related relationships were imported into the corresponding core modules of the decision-making system according to the interface specifications. Data integrity was simultaneously verified, confirming 100% coverage of the imported data, with no missing core data and no association mapping errors. After successful verification, data population of the knowledge graph decision-making system was completed, ensuring that the imported data was consistent with the data in the previous case study and compatible with the functions of each module in the system. This provided high-quality data support for the subsequent generation of regional product selection lists, personalized user recommendation lists, and strategy adjustments, forming a closed loop of data processing and system population. This embodiment ensures the accuracy, completeness, and consistency of decision-making system data through standardized data screening, cleaning, integration, and import processes, providing solid support for the efficient operation of the system and improving the scientific nature of product selection and push decisions.

[0079] refer to Figure 2 A cross-border e-commerce product selection and advertising push device based on regional heat analysis, comprising: The graph construction module 100 is used to integrate multi-channel data of regions, products, and users to construct a multi-granularity regional market knowledge graph containing multi-granularity regional, product, and user nodes and their relationships, and to extract regional popularity vectors of candidate products based on graph neural networks. The consumption prediction module 200 is used to extract user association subgraphs from knowledge graphs, extract user graph structure features through graph neural networks, mine user consumption cycle features by combining temporal convolutional networks, and input logistic regression networks to predict user cross-regional consumption-related features and calculate the location distance weights. The cultural scoring module 300 is used to calculate the regional differentiation popularity index based on the regional popularity vector, measure the cultural fit by measuring the semantic similarity between the product and the regional culture, and integrate the two to generate a comprehensive cultural fit score. The recommendation scoring module 400 is used to construct a product association graph, use graph neural networks to mine the association strength between products, and integrate the regional popularity vector, cross-regional consumption-related features, comprehensive cultural suitability score and association strength to calculate dynamic recommendation scores. The decision adjustment module 500 is used to build a decision system based on knowledge graphs, integrate the aforementioned data to generate regional product selection lists and user-personalized recommendation lists, and dynamically adjust product selection and push strategies through anomaly detection.

[0080] Furthermore, the aforementioned map construction module 100 includes: Multi-granularity and node attribute definition unit, used to define the multi-granularity region division standard and node core attributes, and to clarify the type and representation method of the relationship between product and user nodes; The multi-channel data integration and processing unit is used to integrate heterogeneous data from multiple channels, classify and collect data according to node attributes, and perform standardization and cleaning to form a standardized data set. The knowledge graph construction and maintenance unit is used to construct a knowledge graph based on the standardized data, establish a mapping between node attributes and relationships, and dynamically maintain the relationship strength and attribute updates. The regional heat vector generation unit is used to extract cross-granularity correlation subgraphs of candidate products, and to generate regional heat vectors of candidate products at each regional granularity by mining and integrating subgraph features through graph neural networks.

[0081] Furthermore, the aforementioned knowledge graph construction and maintenance unit includes: The node and association rule sorting unit is used to sort out the classification, attribute system and association category of nodes in the standardized data, and clarify the mapping rules between node attributes and association relationships; The knowledge graph framework construction and data filling unit is used to construct a knowledge graph framework based on the mapping rules and node classification, and to fill the standardized data into the node attributes and relationships. The initial association strength calculation unit is used to establish an association strength evaluation system, calculate the initial association strength based on the node data interaction and importance, and associate it with the corresponding relationship; The node attribute and association strength update unit is used to track data update dynamics, synchronously adjust node attribute information, and recalculate and update association strength based on new interaction data.

[0082] Furthermore, the aforementioned map framework construction and data filling unit includes: The node rule and bidirectional mapping formulation unit is used to clarify the classification criteria, core attributes and categories of nodes in the data, and to formulate bidirectional mapping rules for node attributes and relationships. The knowledge graph framework construction and module setting unit is used to build a knowledge graph framework based on the classification criteria and bidirectional mapping rules, and to set up a node attribute storage module and an association relationship connection module. The node attribute data filtering and input unit is used to filter standardized data by node category, input the attribute data into the corresponding node attribute storage module, and decompose and fill the node attribute fields. The association clue mining and filling unit is used to mine data association basis according to the bidirectional mapping rules, fill the association clues into the association relationship connection module, and complete the knowledge graph construction.

[0083] Furthermore, the aforementioned consumption forecasting module 200 includes: The user association subgraph extraction unit is used to extract user association subgraphs from the knowledge graph and define the subgraph structural feature extraction dimensions and node association relationship types. The user graph structure feature vector generation unit is used to perform feature aggregation and quantization on the user-related subgraphs using a graph neural network to generate user graph structure feature vectors. The user consumption cycle feature extraction unit is used to organize historical user consumption data into a time series, and to mine cycle patterns and extract consumption cycle features through a time series convolutional network. The cross-regional consumption feature prediction unit is used to fuse the user graph structure feature vector and consumption cycle features, input them into the logistic regression network, predict the cross-regional consumption-related features of users, and calculate the location distance weight.

[0084] Furthermore, the aforementioned user graph structure feature vector generation unit includes: The feature extraction rule formulation unit is used to clarify the node types, structural feature types, and core extraction dimensions of the user-related subgraph, and to formulate feature extraction priorities, fusion rules, and aggregation order. The shallow and deep feature extraction and aggregation unit is used to extract shallow and deep structural features of the user-related subgraph in stages using a graph neural network, and aggregate node attributes and relationship features according to priority and order rules; The aggregation feature normalization processing unit is used to normalize the aggregated features, eliminate dimensional differences and redundant information, and integrate them into a unified structural feature set according to the fusion rules. The standardized vector generation unit is used to quantize and convert the integrated unified structural feature set into a standardized vector form to generate user graph structural feature vectors.

[0085] Furthermore, the aforementioned shallow-deep feature extraction aggregation unit includes: The feature extraction rules and category definition unit is used to clarify the node attribute dimensions and relationship types of the user-related subgraph, divide the shallow and deep structural feature categories and definition standards, and formulate feature extraction priorities and hierarchical aggregation rules. The shallow structure feature extraction and filtering unit is used to extract shallow structure features using the first stage of the graph neural network, covering the basic attributes of nodes and direct relationships, and filtering key features according to priority. The deep structure feature mining unit is used to extract deep structure features based on the key shallow features through the second stage of the graph neural network, and to mine multi-node association-derived and implicit association features. The feature hierarchical aggregation and integration unit is used to first aggregate features of the same hierarchical level according to the hierarchical aggregation rules, then fuse shallow key features and deep derived features, and integrate node attributes and relationship features.

[0086] Furthermore, the aforementioned cultural scoring module 300 includes: The heat data collection and parameter determination unit is used to collect relevant heat data of products in various regions, clarify the calculation dimensions and construction elements of the regional differentiated heat index, and determine the text source and type of multilingual semantic analysis; The regional differentiation popularity index generation unit is used to compare the popularity of products in each region with the overall level, extract the unique popularity characteristics of each region, and generate a regional differentiation popularity index by combining the core analysis dimensions of popularity differences. The cultural adaptation quantification unit is used to extract the core information and description of the product in multiple languages, perform semantic matching with the multilingual user text in the target region, and quantify the degree of cultural adaptation. The comprehensive cultural suitability score generation unit is used to integrate the regional differentiation popularity index and cultural suitability according to preset integration rules and logic to generate a comprehensive cultural suitability score.

[0087] Furthermore, the aforementioned regionally differentiated heat index generation unit includes: The unit for collecting popularity data and defining dimensional rules is used to collect popularity data of products in various regions and global popularity benchmark data, sort out the core analysis dimensions of popularity differences, and clarify the priority ranking and hierarchical division rules of each dimension. The regional and global heat level comparison unit is used to perform a layered comparison of heat level data of each region and global heat level benchmark data according to the dimensional priority and hierarchical division rules, and to capture the changing trends and abnormal change points. The regional heat feature extraction and integration unit is used to extract the unique heat change features of each region at different priorities and levels based on the hierarchical comparison results, and integrate them with dimensional correlation to form a set of regional heat features. The regional differentiation heat index output unit is used to construct generation logic and output the regional differentiation heat index based on the priority of the core analysis dimensions of heat difference and the set of regional heat characteristics.

[0088] Furthermore, the aforementioned regional and global heat map stratification comparison units include: The comparison standard and rule formulation unit is used to clarify the priority ranking of the core analysis dimensions of heat difference, formulate the hierarchical division standards and rules of each dimension, and determine the capture standards of change trends and abnormal change points; The heat data dimension classification and splitting unit is used to sort according to the priority of the dimension, and combine the hierarchical division standards and rules to perform dimension classification and hierarchical splitting on the heat data of each region and the global heat benchmark data to form corresponding hierarchical subsets. The subset comparison and anomaly verification unit is used to compare subsets of the same dimension and level, track the relative global trend of regional heat based on the capture criteria, and identify and verify abnormal change points. The comparison result recording annotation unit is used to record the comparison results of the sub-datasets layer by layer according to the priority of dimensions and the hierarchical division order, and to annotate the verified change trend and abnormal change point information to form a complete comparison record.

[0089] Furthermore, the aforementioned subset comparison and anomaly verification unit includes: Subset extraction and standard definition unit, used to extract regional and global heat subsets of the same dimension and level, and to define the core indicators, order and anomaly identification logic and verification process in the trend tracking standard; The subset dataset synchronous comparison and trend tracking unit is used to synchronously compare two sets of subset datasets in a predetermined order around the core indicator, track the relative global trend of regional popularity, and record trend details and directions; The suspected anomaly identification and labeling unit is used to identify suspected abnormal change points that deviate from the norm from the trend based on the capture criteria and identification logic, sort out the abnormal characteristics and occurrence scenarios, and complete the labeling. The suspected anomaly point verification and confirmation unit is used to verify each suspected anomaly point one by one according to the verification process, combined with the anomaly characteristics, the scenario and the original data of the two sets of subsets, and to confirm the final abnormal change point.

[0090] Furthermore, the aforementioned recommendation rating module 400 includes: The product association graph construction unit is used to collect multi-dimensional information, interaction data and association clues of products, clarify the node attributes, association types and construction logic of the product association graph, sort out the potential associations between products and construct the product association graph; The product association strength quantification unit is used to input the product association map into the graph neural network, perform hierarchical multi-round mining, extract different levels of association features between products, quantify and integrate them to form product association strength data; The multi-dimensional data extraction and rule determination unit is used to extract regional popularity vectors, cross-regional consumption-related features and comprehensive cultural suitability scores, and to determine the weight allocation method and calculation and fusion rules for each dimension. The product dynamic recommendation score calculation unit is used to integrate the weights of each dimension according to the weight allocation method and fusion rules, and combine the product correlation strength data to calculate and generate a product dynamic recommendation score.

[0091] Furthermore, the aforementioned commodity association strength quantification unit includes: The mining parameters and standards are defined in a specific unit, which is used to obtain the product association graph, define the dimensions, priorities and division criteria of the graph neural network hierarchical mining, and determine the order, triggering conditions and feature extraction standards of each level of mining. The high-priority initial association feature mining unit is used to input the product association map into the graph neural network, start the first round of mining according to the hierarchical priority, extract the initial association features of the high-priority level, and record the mining results and details. The multi-round hierarchical association feature mining and integration unit is used to carry out multi-round hierarchical mining according to the triggering conditions, extract low-priority hierarchical features in sequence, verify, supplement and correct deviations of all initial association features, and integrate association features of each level. The correlation feature quantification and data integration unit is used to formulate quantification rules to quantify each level of correlation feature after integration, sort out the quantified data and integrate it into a data set of correlation strength between products.

[0092] Furthermore, the aforementioned mining parameters and standards define specific units, including: The product association graph key information extraction unit is used to obtain the product association graph, sort out the node types, association attributes, association patterns and path features in the graph, and extract key information and core content of the graph. The core dimension definition unit for hierarchical mining is used to define the core dimensions of hierarchical mining of graph neural networks based on the key information and core content of the graph, and to clarify the division logic, basis and core connotation of each dimension. The mining priority and trigger condition setting unit is used to determine the priority of each mining dimension by combining its actual value, correlation and mining difficulty, plan the order of multiple mining rounds, and set the trigger conditions for each mining round. The association feature extraction standard formulation unit is used to define the boundaries of association feature extraction at each level, refine the extraction requirements and judgment criteria, and formulate a unified association feature extraction standard based on the hierarchical mining dimensions, dimension priorities and division criteria.

[0093] Furthermore, the aforementioned decision adjustment module 500 includes: The decision system architecture building unit is used to collect multi-dimensional relationship data, clarify the functional requirements, core modules and construction logic of the knowledge graph decision system, and build the system's basic architecture. The multi-dimensional data processing and import unit is used to filter, clean, classify and integrate the multi-dimensional relational data, establish a data association model and mapping relationship, and import the processed relational data into the knowledge graph decision system to complete data filling. The product selection recommendation list generation unit is used to mine user personalized needs, preference characteristics, regional market characteristics, and product selection pain points based on the data integrated by the knowledge graph decision system, and generate regional product selection lists and user personalized recommendation lists. The product selection and push strategy optimization unit is used to embed an anomaly detection mechanism to monitor the rationality of the product selection list in the region and system data anomalies, capture anomaly information and analyze the causes of deviations, and dynamically optimize the product selection and push strategy based on the analysis results.

[0094] Furthermore, the aforementioned multi-dimensional data processing and import unit includes: The multi-dimensional raw data filtering unit is used to organize the multi-dimensional relationship raw data, clarify the core indicators, conditions and priorities of data filtering, filter effective target data according to these requirements, and eliminate invalid, redundant and irrelevant data. The effective data cleaning and verification unit is used to clean and verify the filtered effective target data, correct data deviations and errors, fill in missing information, unify data format and standards, and form an accurate and complete standardized data set. The standardized data classification and integration unit is used to classify and integrate the standardized data according to category, related attributes and business logic, sort out data association rules, build data association models, and clarify the corresponding mapping relationship between various types of data; The data import and integrity verification unit is used to import classified and integrated data with completed association mapping into the system in accordance with the data interface specifications and import requirements of the knowledge graph decision system, and simultaneously verify the data integrity and complete the system data filling.

[0095] Reference Figure 3 This application also provides a computer device, which may be a server, and its internal structure may be as follows: Figure 3As shown, the computer device includes a processor, memory, network interface, and database connected via a bus. The processor provides computing and control capabilities. The memory includes non-volatile storage media and internal memory. The non-volatile storage media stores operations, computer programs, and the database. The internal memory provides an environment for the operation and execution of the computer programs stored in the non-volatile storage media. The database stores data such as cross-border e-commerce product selection and advertising methods based on regional heatmap analysis. The network interface is used for communication with external terminals via a network connection. When executed by a processor, this computer program implements a cross-border e-commerce product selection and advertising push method based on regional popularity analysis. The method includes: integrating multi-channel data on regions, products, and users; constructing a multi-granularity regional market knowledge graph containing multi-granularity regional, product, and user nodes and their relationships; extracting regional popularity vectors for candidate products based on graph neural networks; extracting user relationship subgraphs from the knowledge graph; extracting user graph structure features through graph neural networks; mining user consumption cycle features using temporal convolutional networks; inputting the data into a logistic regression network to predict user cross-regional consumption-related features and calculating location distance weights; calculating a regional differentiated popularity index based on regional popularity vectors; quantifying cultural fit through the semantic similarity between products and regional culture; and fusing the two to generate a comprehensive cultural fit score; constructing a product relationship graph; mining the relationship strength between products using graph neural networks; fusing the regional popularity vectors, cross-regional consumption-related features, comprehensive cultural fit score, and relationship strength to calculate a dynamic recommendation score; building a decision-making system based on the knowledge graph; integrating the aforementioned data to generate regional product selection lists and user-personalized recommendation lists; and dynamically adjusting product selection and push strategies through anomaly detection.

[0096] One embodiment of this application also provides a computer-readable storage medium storing a computer program thereon. When executed by a processor, the computer program implements a method for cross-border e-commerce product selection and advertising push based on regional heat analysis, including: integrating multi-channel data on regions, products, and users; constructing a multi-granularity regional market knowledge graph containing multi-granularity regional, product, and user nodes and their relationships; extracting regional heat vectors of candidate products based on graph neural networks; extracting user relationship subgraphs from the knowledge graph; extracting user graph structure features through graph neural networks; mining user consumption cycle features by combining temporal convolutional networks; and inputting the data into a logistic regression network for prediction. The system analyzes user cross-regional consumption characteristics and calculates location distance weights; it calculates a regional differentiated popularity index based on regional popularity vectors, quantifies cultural fit by measuring the semantic similarity between products and regional culture, and integrates the two to generate a comprehensive cultural fit score; it constructs a product association graph, uses graph neural networks to mine the association strength between products, and integrates the aforementioned regional popularity vectors, cross-regional consumption characteristics, comprehensive cultural fit score, and association strength to calculate a dynamic recommendation score; it builds a decision-making system based on a knowledge graph, integrates the aforementioned data to generate regional product selection lists and user personalized recommendation lists, and dynamically adjusts product selection and push strategies through anomaly detection.

[0097] Those skilled in the art will understand that all or part of the processes in the methods of the above embodiments can be implemented by a computer program instructing related hardware. The computer program can be stored in a computer-readable storage medium, and when executed, it can include the processes of the embodiments of the methods described above. Any references to memory, storage, databases, or other media provided in this application and used in the embodiments can include non-volatile and / or volatile memory. Non-volatile memory can include read-only memory (ROM), programmable ROM (PROM), electrically programmable ROM (EPROM), electrically erasable programmable ROM (EEPROM), or flash memory. Volatile memory can include random access memory (RAM) or external cache memory. By way of illustration and not limitation, RAM is available in various forms, such as static RAM (SRAM), dynamic RAM (DRAM), synchronous DRAM (SDRAM), dual-speed SDRAM (SSRSDRAM), enhanced SDRAM (ESDRAM), synchronous link DRAM (SLDRAM), Rambus direct RAM (RDRAM), direct memory bus dynamic RAM (DRDRAM), and memory bus dynamic RAM (RDRAM), etc.

[0098] The above description is only a preferred embodiment of this application and does not limit the patent scope of this application. Any equivalent structural or procedural changes made based on the content of this application's specification and drawings, or direct or indirect applications in other related technical fields, are similarly included within the patent protection scope of this application.

Claims

1. A method for product selection and advertising push in cross-border e-commerce based on regional popularity analysis, characterized in that, include: By integrating multi-channel data on regions, products, and users, a multi-granularity regional market knowledge graph containing multi-granularity regional, product, and user nodes and their relationships is constructed, and regional popularity vectors of candidate products are extracted based on graph neural networks. The user association subgraph is extracted from the knowledge graph, the user graph structure features are extracted through graph neural network, the user consumption cycle features are mined by temporal convolutional network, and the logistic regression network is input to predict the cross-regional consumption-related features of users and calculate the location distance weight. The regional differentiation popularity index is calculated based on the regional popularity vector. The cultural fit is quantified by the semantic similarity between the product and the regional culture. The two are then combined to generate a comprehensive cultural fit score. Construct a product association graph, use graph neural networks to mine the association strength between products, and integrate the regional popularity vector, cross-regional consumption-related features, comprehensive cultural adaptability score and association strength to calculate dynamic recommendation score; A decision-making system is built based on knowledge graphs, which integrates the aforementioned data to generate regional product selection lists and personalized recommendation lists for users. The product selection and push strategies are dynamically adjusted through anomaly detection.

2. The method for cross-border e-commerce product selection and advertising push based on regional heat analysis according to claim 1, characterized in that, The steps of extracting user-related subgraphs from the knowledge graph, extracting user graph structure features through graph neural networks, mining user consumption cycle features by combining temporal convolutional networks, and inputting the data into a logistic regression network to predict user cross-regional consumption-related features and calculate the location distance weights include: Extract user-related subgraphs from the knowledge graph, and define the dimensions of subgraph structural feature extraction and the types of node association relationships; A graph neural network is used to perform feature aggregation and quantization on the user-related subgraph to generate a user graph structure feature vector. By organizing users' historical consumption data into a time series, and then using a time series convolutional network to mine cyclical patterns and extract consumption cycle characteristics; By integrating the user graph structure feature vector and consumption cycle features, the data is input into a logistic regression network to predict cross-regional consumption-related features of users and calculate the geographical distance weights.

3. The method for cross-border e-commerce product selection and advertising push based on regional heat analysis according to claim 2, characterized in that, The step of using a graph neural network to aggregate and quantize features of the user-related subgraph to generate a user graph structure feature vector includes the following steps: Define the node types, structural feature types, and core extraction dimensions of the user-related subgraph, and formulate feature extraction priorities, fusion rules, and aggregation order; A graph neural network is used to extract shallow and deep structural features from the user association subgraph in stages, and node attributes and association features are aggregated according to priority and order rules; The aggregated features are normalized to eliminate dimensional differences and redundant information, and then integrated into a unified structural feature set according to the fusion rules. The integrated unified structural feature set is quantized and converted into a standardized vector form to generate user graph structural feature vectors.

4. The method for cross-border e-commerce product selection and advertising push based on regional heat analysis according to claim 3, characterized in that, The step of using a graph neural network to extract shallow and deep structural features from the user-related subgraph in stages, and aggregating node attributes and relationship features according to priority and order rules, includes: Define the node attribute dimensions and relationship types of the user association subgraph, divide the shallow and deep structural feature categories and definition standards, and formulate feature extraction priorities and hierarchical aggregation rules; The first stage of graph neural network is used to extract shallow structural features, covering basic node attributes and direct relationships, and key features are selected according to priority. Based on the aforementioned key shallow features, deep structural features are extracted through the second stage of the graph neural network to mine multi-node association-derived and implicit association features; Based on the aforementioned hierarchical aggregation rules, first aggregate contractual hierarchical features, then fuse shallow key features and deep derived features, and integrate node attributes and relationship features.

5. The method for cross-border e-commerce product selection and advertising push based on regional heat analysis according to claim 1, characterized in that, The steps of calculating a regionally differentiated popularity index based on a regional popularity vector, quantifying cultural fit by measuring the semantic similarity between goods and regional culture, and fusing the two to generate a comprehensive cultural fit score include: Collect relevant popularity data of products in various regions, clarify the calculation dimensions and construction elements of regional differentiated popularity index, and determine the text source and type of multilingual semantic analysis; By comparing the popularity of products in each region with the overall level, we can extract the unique popularity characteristics of each region and generate a regionally differentiated popularity index by combining the core analysis dimensions of popularity differences. Extract core information and descriptions of products in multiple languages, perform semantic matching with multilingual user text in the target region, and quantify cultural compatibility. According to preset integration rules and logic, the regional differentiation popularity index and cultural compatibility are integrated to generate a comprehensive cultural compatibility score.

6. The method for cross-border e-commerce product selection and advertising push based on regional heat analysis according to claim 5, characterized in that, The steps of comparing the popularity of products in each region with the overall level, extracting unique regional popularity characteristics, and generating a regionally differentiated popularity index by combining core analytical dimensions of popularity differences include: Collect product popularity data from various regions and global popularity benchmark data, sort out the core analysis dimensions of popularity differences, and clarify the priority ranking and hierarchical division rules of each dimension; According to the aforementioned dimensional priority and hierarchical division rules, the heat data of each region and the global heat benchmark data are compared in layers to capture the changing trends and abnormal change points. Based on the hierarchical comparison results, the unique heat change characteristics of each region at different priorities and levels are extracted and integrated with dimensional correlation to form a set of regional heat features. Based on the priority of the core analysis dimensions of heat difference and the set of regional heat characteristics, a generation logic is constructed to output a regional differentiated heat index.

7. The method for cross-border e-commerce product selection and advertising push based on regional heat analysis according to claim 6, characterized in that, The step of performing hierarchical comparison of regional heat data and global heat benchmark data according to the aforementioned dimensional priority and hierarchical division rules, and capturing changing trends and abnormal change points, includes: Clearly define the priority order of the core analytical dimensions of heat difference, formulate hierarchical division standards and rules for each dimension, and determine the criteria for capturing change trends and abnormal change points; Based on the aforementioned dimensional priority, and combined with the hierarchical division standards and rules, the regional heat data and global heat benchmark data are dimensionally classified and hierarchically split to form corresponding hierarchical subsets. Compare subsets of data at the same dimension and level, track the relative global trend of regional heat based on the capture criteria, and identify and verify abnormal change points; The comparison results of the sub-datasets are recorded level by level according to the priority of dimensions and the order of hierarchical division. The verified change trends and abnormal change points are marked to form a complete comparison record.

8. The method for cross-border e-commerce product selection and advertising push based on regional heat analysis according to claim 7, characterized in that, The steps of comparing subsets of data at the same dimension and level, tracking the relative global trend of regional heat based on the capture criteria, and identifying and verifying anomalous change points include: Extract regional and global heat maps of the same dimension and level, and clarify the core indicators, sequence, anomaly identification logic and verification process in the trend tracking standard; Based on the core indicators, the two sets of subsets are compared synchronously in a predetermined order to track the relative global trend of regional popularity and record the trend details and direction. Based on the aforementioned capture criteria and identification logic, identify suspected abnormal changes that deviate from the norm from the trends, analyze the abnormal features and their occurrence scenarios, and complete the annotation. Following the verification process, and combining the abnormal features, scenarios, and the original data of the two sets of subsets, each suspected anomaly is verified to confirm the final abnormal change point.

9. The method for cross-border e-commerce product selection and advertising push based on regional heat analysis according to claim 1, characterized in that, The steps of constructing a product association graph, using graph neural networks to mine the association strength between products, and integrating the regional popularity vector, cross-regional consumption-related features, comprehensive cultural suitability score, and association strength to calculate a dynamic recommendation score include the following steps: Collect multi-dimensional information, interaction data and related clues about products, clarify the node attributes, association types and construction logic of the product association graph, sort out the potential associations between products and construct the product association graph; The product association graph is input into a graph neural network for hierarchical multi-round mining to extract association features at different levels between products, and then quantified and integrated to form data on the strength of association between products. Extract regional popularity vectors, cross-regional consumption-related features, and comprehensive cultural suitability scores to determine the weight allocation methods, calculation, and integration rules for each dimension. Based on the weight allocation method and fusion rules, the weights of each dimension are integrated, and combined with the correlation strength data between products, a dynamic recommendation score for products is calculated and generated.

10. A cross-border e-commerce product selection and advertising push device based on regional heat analysis, characterized in that, include: The graph construction module is used to integrate multi-channel data of regions, products, and users to build a multi-granularity regional market knowledge graph containing multi-granularity regional, product, and user nodes and their relationships. It also extracts regional popularity vectors of candidate products based on graph neural networks. The consumption prediction module is used to extract user-related subgraphs from the knowledge graph, extract user graph structure features through graph neural networks, mine user consumption cycle features by combining temporal convolutional networks, and input logistic regression networks to predict user cross-regional consumption-related features and calculate the location distance weights. The cultural scoring module is used to calculate the regional differentiation popularity index based on the regional popularity vector, measure the cultural fit by the semantic similarity between the product and the regional culture, and integrate the two to generate a comprehensive cultural fit score. The recommendation rating module is used to construct a product association graph, use graph neural networks to mine the association strength between products, and integrate the regional popularity vector, cross-regional consumption-related features, comprehensive cultural suitability score and association strength to calculate dynamic recommendation rating; The decision adjustment module is used to build a decision system based on knowledge graphs, integrate the aforementioned data to generate regional product selection lists and user-personalized recommendation lists, and dynamically adjust product selection and push strategies through anomaly detection.

Images