A cross-border product selection strategy generation method and device based on a dynamic knowledge graph

Abstract

The application discloses a cross-border product selection strategy generation method and device based on a dynamic knowledge graph, and relates to the field of e-commerce product selection.The method comprises the following steps: collecting multi-modal behavior data from multiple data sources in real time; mapping semantic feature vectors and visual feature vectors to a shared embedding space to obtain unified intention feature vectors, and determining structured consumer intentions based on the unified intention feature vectors; constructing a dynamic demand knowledge graph based on the consumer intentions; performing time series analysis on the dynamic demand knowledge graph, identifying the heat change trend of demand expression nodes, determining newly emerging demand hotspots and demand migration paths; predicting candidate commodity attribute combinations that will become hotspots in a future preset time period, and generating a predicted hotspot product selection direction; and performing differential adaptation on the current hotspot product selection direction and the predicted hotspot product selection direction, and outputting product selection strategy suggestions for different cross-border e-commerce platforms. The technical scheme provided by the application can solve the problem of low product selection accuracy.

Description

Technical Field

[0001] This application relates to the field of e-commerce product selection, specifically to a method and apparatus for generating cross-border product selection strategies based on dynamic knowledge graphs. Background Technology

[0002] Product selection is a core decision-making process in cross-border e-commerce, and the accuracy and timeliness of product selection strategies directly affect sales conversion and inventory efficiency.

[0003] Currently, when making product selection decisions, cross-border sellers mainly rely on operations staff to manually browse, record, and judge based on experience on the aforementioned platforms on a regular basis, or use simple keyword popularity monitoring tools for assistance. However, due to the differences in user structure and recommendation mechanisms among different cross-border e-commerce platforms, it is difficult to adapt product selection to each platform, resulting in a disconnect between product selection strategies and the actual consumption environment of the platform, which leads to a problem of low product selection accuracy.

[0004] Therefore, there is an urgent need for a method and apparatus for generating cross-border product selection strategies based on dynamic knowledge graphs. Summary of the Invention

[0005] This application provides a method and apparatus for generating cross-border product selection strategies based on dynamic knowledge graphs, which solves the problem of low product selection accuracy.

[0006] This application provides a method for generating cross-border product selection strategies based on dynamic knowledge graphs in its first aspect. The method includes: collecting multimodal consumer behavior data in real time from multiple data sources, including cross-border e-commerce platforms, social media platforms, and search engines; extracting semantic feature vectors from the text data using a multilingual large language model and visual feature vectors from the image data using a computer vision model; mapping the semantic and visual feature vectors to a shared embedding space to obtain a unified intent feature vector; determining structured consumer intent based on the unified intent feature vector; and extracting consumer identifiers, demand descriptions, product attributes, and usage scenarios based on the consumer intent; mapping consumer identifiers to consumer group nodes, demand descriptions to demand expression nodes, product attributes to product attribute nodes, and usage scenarios to usage scenario nodes; and establishing edges between corresponding nodes based on the relationships in the consumer intent to construct a [method / structure]. A dynamic demand knowledge graph is used, where the relation weights of edges are dynamically updated based on a time decay function. Temporal analysis is performed on the relation weights of edges representing demand nodes in the dynamic demand knowledge graph to calculate the popularity values ​​of these nodes across different preset time windows. Based on the changes in popularity values ​​over time, the popularity trends of demand nodes are identified, and new demand hotspots and demand migration paths are determined. Based on the existing relation edges between demand nodes and product attribute nodes, new demand hotspots are mapped to candidate product attribute combinations associated with the demand nodes, generating current hot product selection directions. Combined with demand migration paths, candidate product attribute combinations that will become hotspots within a preset time period are predicted, generating predicted hot product selection directions. User profile data, best-selling category data, and recommendation algorithm parameters from different cross-border e-commerce platforms are acquired to differentiate and adapt current and predicted hot product selection directions, outputting product selection strategy suggestions for different cross-border e-commerce platforms.

[0007] By adopting the above technical solutions, through cross-modal semantic alignment and dynamic graph construction, the semantic fragmentation between multilingual and multimodal data is eliminated, enabling consumer intent to be uniformly expressed as four types of computable entities and a time-sensitive relationship network. Graph-based temporal trend analysis and hotspot mapping achieve a leap from passive screening to forward-looking prediction, simultaneously outputting product selection directions in both current and future time dimensions by combining demand migration paths. The introduction of platform-differentiated adaptation addresses the pain point of a single product selection solution being unable to adapt to the user characteristics and recommendation mechanisms of different platforms, improving the accuracy of product selection across different platforms.

[0008] Optionally, semantic feature vectors of text data are extracted using a multilingual large language model, and visual feature vectors of image data are extracted using a computer vision model. The semantic and visual feature vectors are then mapped to a shared embedding space to obtain a unified intent feature vector. Based on this unified intent feature vector, a structured consumer intent is determined. Specifically, this involves: inputting text data into a multilingual large language model; mapping text from different languages ​​to a shared embedding space using a cross-language semantic encoder within the multilingual large language model to obtain semantic feature vectors; inputting image data into a computer vision model; extracting features of visual elements related to the product using an object detection and attribute recognition network to obtain visual feature vectors; using a cross-modal attention mechanism to perform a weighted sum of the semantic and visual feature vectors to obtain a unified intent feature vector; and inputting the unified intent feature vector into an intent classifier to output the consumer intent, which includes consumer identification, demand description, product attributes, and usage scenarios.

[0009] By adopting the above technical solution, cross-language encoding, visual attribute extraction, and cross-modal attention are integrated into a unified pipeline, enabling textual demands and image preferences to mutually reinforce each other in a shared space. This avoids misjudgment of intent caused by missing or ambiguous information in a single modality. The structured output of consumer intent directly carries all the fields required for subsequent mapping, reducing accuracy loss during information transmission.

[0010] Optionally, the relational weights of the edges are dynamically updated according to a time decay function. Specifically, this includes: configuring an initial weight and decay coefficient for each edge, with the decay coefficient adaptively determined based on the demand cycle characteristics of the product attributes associated with the edge; applying an increment to the current weight of the edge each time a new consumer intent is detected corresponding to an edge, and recording the matching timestamp; traversing all edges in the dynamic demand knowledge graph at a preset frequency, calculating the current weight of each edge using an exponential decay function based on the time interval between the current time and the matching timestamp, and updating it to the dynamic demand knowledge graph.

[0011] By adopting the above technical solution, an adaptive decay coefficient and incremental update mechanism are introduced for each edge, ensuring that the change rhythm of relationship weights matches the demand cycle characteristics of different product categories. This avoids the excessive retention of FMCG signals or the premature dilution of durable goods signals. The synergy between incremental triggering and timed decay ensures that the graph always maintains an accurate representation of the current market state.

[0012] Optionally, based on the trend of popularity changes, newly emerging demand hotspots are identified, specifically including: obtaining the popularity values ​​of the same demand expression node in multiple consecutive time-series snapshots to form a popularity value sequence; detecting the popularity value sequence, and when the first-order difference of the popularity value sequence changes from negative to positive and exceeds a preset difference threshold, the demand expression node is identified as a newly emerging demand hotspot.

[0013] By employing the above technical solution, and using the positive-to-negative transformation of the first-order difference of the heat value sequence as the basis for hotspot determination, the identification of emerging demand is transformed from subjective experience-based judgment into an objectively calculable statistical indicator. This method filters out random fluctuation interference and only triggers hotspot identification when the upward trend is continuously confirmed, thus reducing the probability of false alarms.

[0014] Optionally, user profile data, best-selling category data, and recommendation algorithm parameters of different cross-border e-commerce platforms are acquired to differentiate and adapt current and predicted hot product selection directions, outputting product selection strategy suggestions for different cross-border e-commerce platforms. Specifically, this includes: for each cross-border e-commerce platform, obtaining the platform adaptation parameters of the target platform, including user profile data, best-selling category data, and recommendation algorithm parameters. User profile data includes price sensitivity, decision-making cycle, and content preference; extracting product attribute combinations of products to be adapted from current and predicted hot product selection directions; calculating the matching weight of each attribute in the product attribute combination with the target platform based on the platform adaptation parameters, removing attributes with matching weights lower than a preset trimming threshold, and obtaining product selection feature combinations adapted to the target platform; matching the product selection feature combinations with a preset strategy template to generate product selection strategy suggestions for the target platform.

[0015] By adopting the above technical solution, multi-dimensional platform parameters are used to perform item-by-item matching weight calculation and tailoring of product attributes, enabling the combination of product selection features to proactively adapt to the user decision-making logic and content distribution mechanism of the target platform. Platform differences are transformed from qualitative experience into quantifiable adaptation decisions, overcoming the homogenization problem of cross-platform product selection strategies.

[0016] Optionally, after generating the predicted hot product selection direction, the method further includes: matching the current hot product selection direction and the candidate product attribute combination corresponding to the predicted hot product selection direction with a preset supplier database to obtain candidate factories that meet the production conditions of the target product, and calculating the modification cost and production cycle. The preset supplier database includes factory process capabilities, minimum order quantity and historical sampling cycle, and the target product is the product corresponding to the candidate product attribute combination; adding the candidate factory information, modification cost and production cycle to the product selection strategy suggestion.

[0017] By adopting the above technical solution, the candidate attribute combinations of the predicted output are automatically matched directly with the supplier database, thus streamlining the decision-making process from trend perception to supply chain feasibility verification. Production assessment information is included with the product selection direction at the time of proposal, avoiding decision failures caused by a disconnect between direction and capacity.

[0018] Optionally, after generating the predicted hot product selection direction, the method further includes: dividing consumer group nodes according to geographic market labels to obtain the consumer group set corresponding to each geographic market; aggregating the heat values ​​of demand expression nodes associated with the consumer group set according to different preset time windows to obtain the demand heat evolution curve of each geographic market; taking the market where demand heat rises first as the source market and the market where demand heat rises later as the target market, calculating the time difference between the start of demand heat rise in the source market and the target market to obtain the demand transmission delay; when the demand transmission delay is lower than the preset delay threshold and the demand heat value of the source market is higher than the preset heat threshold, marking the demand hotspot in the source market as a transmittable hotspot; adding the transmittable hotspot to the predicted hot product selection direction of the target market.

[0019] By employing the aforementioned technical solutions, and through geographical market segmentation and time-series comparison of demand intensity, the demand transmission delay between different markets is quantified, capturing the spatial diffusion patterns of hotspots. Validated hotspots in the source market are pre-marked as transmittable signals and incorporated into target market forecasts, expanding the information sources relied upon for forecasting from single markets to cross-market prior signals, thereby improving the lead time and accuracy of forecasts.

[0020] In a second aspect, this application provides a cross-border product selection strategy generation device based on dynamic knowledge graphs, the device including an acquisition unit and a processing unit;

[0021] The acquisition unit is used to collect multimodal consumer behavior data in real time from multiple data sources, including cross-border e-commerce platforms, social media platforms, and search engines. The multimodal behavior data includes text data and image data.

[0022] The processing unit is used to extract semantic feature vectors from text data using a multilingual large language model and visual feature vectors from image data using a computer vision model. It maps the semantic and visual feature vectors to a shared embedding space to obtain a unified intent feature vector, and determines structured consumer intent based on this vector. It is also used to extract consumer identifiers, demand descriptions, product attributes, and usage scenarios based on consumer intent. Consumer identifiers are mapped to consumer group nodes, demand descriptions to demand expression nodes, product attributes to product attribute nodes, and usage scenarios to usage scenario nodes. Edges are established between corresponding nodes based on the relationships in the consumer intent, constructing a dynamic demand knowledge graph. The edge weights are dynamically updated according to a time decay function. Furthermore, it is used to manage the edges of the demand expression nodes in the dynamic demand knowledge graph. The system performs time-series analysis based on weights, calculates the popularity value of demand expression nodes in different preset time windows, identifies the popularity trend of demand expression nodes based on the change of popularity value over time, and determines newly emerging demand hotspots and demand migration paths based on the popularity trend. It is also used to map newly emerging demand hotspots to candidate product attribute combinations associated with demand expression nodes based on existing relationship edges between demand expression nodes and product attribute nodes, generating current hot product selection directions, and predicting candidate product attribute combinations that will become hotspots in the future within a preset time period based on demand migration paths, thus generating predicted hot product selection directions. Furthermore, it is used to acquire user profile data, best-selling category data, and recommendation algorithm parameters from different cross-border e-commerce platforms, and to perform differentiated adaptation of current and predicted hot product selection directions, outputting product selection strategy suggestions for different cross-border e-commerce platforms.

[0023] In a third aspect, this application provides an electronic device including a processor, a memory, a user interface, and a network interface. The memory is used to store instructions, the user interface and the network interface are used to communicate with other devices, and the processor is used to execute the instructions stored in the memory to cause the electronic device to perform the first aspect or any possible implementation of the first aspect.

[0024] In a fourth aspect, this application provides a computer-readable storage medium storing a computer program, which is executed by a processor as described in the first aspect or any possible implementation thereof.

[0025] In summary, one or more technical solutions provided in the embodiments of this application have at least the following technical effects or advantages: 1. By employing cross-modal semantic alignment and dynamic graph construction, the semantic fragmentation between multilingual and multimodal data is eliminated, enabling consumer intent to be uniformly expressed as four types of computable entities and a time-sensitive relationship network. Graph-based temporal trend analysis and hotspot mapping achieve a leap from passive selection to forward-looking prediction, simultaneously outputting product selection directions in both current and future time dimensions by combining demand migration paths. The introduction of platform-differentiated adaptation addresses the pain point of a single product selection solution being unable to adapt to the user characteristics and recommendation mechanisms of different platforms, thus improving the accuracy of product selection across various platforms.

[0026] 2. By using the positive and negative transformation of the first-order difference of the heat value sequence as the basis for hotspot determination, the identification of emerging demand is transformed from subjective experience-based judgment into an objectively calculable statistical indicator. This method filters out random fluctuation interference and only triggers hotspot identification when the upward trend is continuously confirmed, thus reducing the probability of false alarms.

[0027] 3. The predicted candidate attribute combinations are automatically matched directly with the supplier database, streamlining the decision-making process from trend perception to supply chain feasibility verification. Production assessment information is included with the product selection direction at the time of proposal, avoiding decision failures caused by a disconnect between direction and capacity. Attached Figure Description

[0028] Figure 1 This is a flowchart illustrating a method for generating cross-border product selection strategies based on dynamic knowledge graphs, as provided in an embodiment of this application.

[0029] Figure 2 This is a schematic diagram of a cross-border product selection strategy generation device based on dynamic knowledge graphs provided in an embodiment of this application.

[0030] Figure 3 This is a schematic diagram of the structure of an electronic device provided in an embodiment of this application.

[0031] Explanation of reference numerals in the attached drawings: 201, acquisition unit; 202, processing unit; 300, electronic device; 301, processor; 302, communication bus; 303, user interface; 304, network interface; 305, memory. Detailed Implementation

[0032] To enable those skilled in the art to better understand the technical solutions in this specification, the technical solutions in the embodiments of this specification will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, and not all embodiments.

[0033] In the description of the embodiments of this application, the words "for example" or "for instance" are used to indicate examples, illustrations, or explanations. Any embodiment or design that is described as "for example" or "for instance" in the embodiments of this application should not be construed as being more preferred or advantageous than other embodiments or design options. Rather, the use of the words "for example" or "for instance" is intended to present the relevant concepts in a specific manner.

[0034] In the description of the embodiments of this application, the term "multiple" means two or more. For example, multiple systems means two or more systems, and multiple screen terminals means two or more screen terminals. Furthermore, the terms "first" and "second" are used for descriptive purposes only and should not be construed as indicating or implying relative importance or implicitly specifying the indicated technical features. Thus, a feature defined with "first" or "second" may explicitly or implicitly include one or more of that feature. The terms "comprising," "including," "having," and variations thereof all mean "including but not limited to," unless otherwise specifically emphasized.

[0035] When making product selection decisions, cross-border sellers primarily rely on operations staff regularly browsing, recording, and judging based on experience on the aforementioned platforms, or using simple keyword popularity monitoring tools for assistance. However, due to differences in user structures and recommendation mechanisms across different cross-border e-commerce platforms, it is difficult to adapt product selection to each platform, leading to a disconnect between product selection strategies and the actual consumption environment of the platform, resulting in low product selection accuracy. Therefore, this embodiment provides a method and apparatus for generating cross-border product selection strategies based on dynamic knowledge graphs.

[0036] This application provides a method for generating cross-border product selection strategies based on dynamic knowledge graphs, which can be used as a reference. Figure 1 , Figure 1 This is a flowchart illustrating a method for generating cross-border product selection strategies based on dynamic knowledge graphs, provided in an embodiment of this application, and applied to a server. The method includes steps S101 to S106.

[0037] S101. Collect multimodal consumer behavior data in real time from multiple data sources, including cross-border e-commerce platforms, social media platforms, and search engines. The multimodal behavior data includes text data and image data.

[0038] In the above steps, the server acquires multimodal consumer behavior data in real time from multiple heterogeneous data sources. These data sources include cross-border e-commerce platforms such as Amazon and eBay, social media platforms such as TikTok, Instagram, and Pinterest, and search engines such as Google Trends. The collected data types cover two main categories: text data and image data. Text data includes search term sequences entered by consumers on e-commerce platforms, product reviews, and text and hashtags from social media posts. Image data includes product outfit photos, unboxing photos, and keyframes from short videos shared publicly on social media. The acquisition method prioritizes using the application programming interfaces (APIs) provided by each platform. For platforms without public APIs, the server obtains anonymized, standardized data streams from compliant third-party data service providers. The server operates the acquisition process using a streaming processing architecture, uniformly encapsulating raw data from different sources into event messages containing timestamps, geographic market tags, and content type fields. These messages are buffered and distributed through a distributed message queue to ensure that subsequent processing stages consume data in chronological and geographic order. Before the data enters downstream processing, the server performs deduplication and noise filtering on the original event messages. Deduplication is based on the joint matching of content hash value and publisher identifier. Noise filtering removes low-quality samples containing pure advertising copy, repeated emoji stacking, or image resolution below a preset threshold.

[0039] S102. Extract semantic feature vectors from text data using a multilingual large language model, and extract visual feature vectors from image data using a computer vision model. Map the semantic feature vectors and visual feature vectors to a shared embedding space to obtain a unified intent feature vector. Determine the structured consumer intent based on the unified intent feature vector.

[0040] In the above steps, after obtaining multimodal behavioral data, the server performs feature extraction and semantic alignment on both text and image data. For text data, the server inputs it into a multilingual large language model. This model uses a Transformer architecture and is pre-trained on a cross-lingual corpus covering more than fifty languages. Its internal cross-lingual semantic encoder can map consumer expressions from different languages ​​such as English, Spanish, Japanese, and Thai to the same shared embedding space, outputting fixed-dimensional semantic feature vectors. This results in a cosine similarity of over 0.9 between "beach wedding guest dress" and "vestidopara invitada a boda en la playa" in the vector space. For image data, the server inputs it into a computer vision model. This model uses a Swing Transformer as its backbone network and is fine-tuned on a publicly available fashion and home furnishing product attribute annotation dataset. It locates product regions in the image through an object detection branch and then outputs the visual element categories and their confidence levels for those regions through an attribute recognition branch, forming visual feature vectors. For example, it can identify attributes such as "square neckline," "floral print," and "puff sleeves" from an outfit image. After feature extraction, the server performs missing modality detection on the semantic and visual feature vectors. If a record contains only text data without corresponding image data, the visual feature vector is filled with all zeros and marked as a single-modality sample, reducing its weight in the loss function during subsequent fusion. The server then uses a cross-modal attention mechanism to perform weighted fusion of the semantic and visual feature vectors. The attention weights are calculated by linearly projecting the feature vectors of the two modalities, calculating their inner product similarity, and then normalizing them using softmax to obtain the contribution ratio of each modality. The weighted sum is then used to obtain a unified intent feature vector. Finally, the server inputs this unified intent feature vector into an intent classifier, which consists of a two-layer fully connected network and a softmax output layer. The classifier outputs a structured consumer intent, including consumer identifiers such as anonymous user IDs or behavioral clustering labels, demand descriptions such as "summer outdoor camping chair," product attributes such as "foldable," "aluminum alloy," and "120kg load capacity," and usage scenarios such as "car camping" and "beach."

[0041] In one possible implementation, text data is input into a multilingual large language model, and the cross-lingual semantic encoder in the multilingual large language model maps texts from different languages ​​to a shared embedding space to obtain semantic feature vectors. Image data is input into a computer vision model, and the object detection and attribute recognition network extracts features of visual elements related to the product to obtain visual feature vectors. A cross-modal attention mechanism is used to perform a weighted summation of the semantic feature vectors and the visual feature vectors to obtain a unified intent feature vector. The unified intent feature vector is input into an intent classifier to output the consumer intent, which includes consumer identification, demand description, product attributes, and usage scenario.

[0042] Specifically, the cross-lingual semantic encoder of the multilingual large language model adopts the XLM-RoBERTa architecture. The server further pre-trains it on a hybrid dataset containing CommonCrawl multilingual snapshots and e-commerce vertical corpora, enabling it to encode e-commerce domain terms on the basis of general semantic understanding. The e-commerce vertical corpora are constructed from publicly available product detail pages and review data from platforms such as Amazon and eBay, covering five languages: English, Spanish, Japanese, German, and French, with a total size of approximately five billion tokens. During training, the server adopts a contrastive learning strategy, using multilingual descriptions of the same product as positive sample pairs and descriptions of different products as negative sample pairs. The optimization objective is to minimize the distance between positive sample pairs in the shared embedding space and maximize the distance between negative sample pairs. The object detection branch of the computer vision model adopts a Deformable DETR structure, and the attribute recognition branch performs multi-label classification within the object detection bounding box. The server uses a subset of product-related annotations from DeepFashion and OpenImages for training, covering fine-grained attributes under three major categories: apparel, home furnishings, and 3C electronics, including color, material, style, pattern, and functional interface type. In the cross-modal attention mechanism, semantic and visual feature vectors are first projected onto the same dimension through their respective fully connected layers. The server calculates the element-wise product of the two vectors as the attention weight, and then the features from the two modalities are weighted and summed to obtain a unified intent feature vector. The training data for the intent classifier is constructed from manually labeled consumer intent samples. Each sample contains a text-image pair and corresponding four-tuple labels: consumer identifier, demand description, product attribute, and usage scenario. Annotators comprehensively judge the consumer's actual purchase intent based on the image content and text semantics and fill in the four-tuple fields. The classifier achieves an F1 score of 92% on the training set.

[0043] S103. Based on consumer intent, extract consumer identifiers, demand descriptions, product attributes, and usage scenarios. Map consumer identifiers to consumer group nodes, demand descriptions to demand expression nodes, product attributes to product attribute nodes, and usage scenarios to usage scenario nodes. Establish edges between corresponding nodes according to the relationships in consumer intent to construct a dynamic demand knowledge graph. The relationship weights of the edges are dynamically updated according to the time decay function.

[0044] In the above steps, based on the structured consumer intent output from the previous step, the server instantiates nodes and constructs relationships in the knowledge graph. The server extracts consumer identifiers from the consumer intent and maps them to consumer group nodes. This mapping is achieved by inputting the behavior vectors of individual consumers into a pre-trained clustering model, grouping consumers with similar needs and preferences into the same group node, with group labels such as "North American outdoor sports enthusiasts." The server extracts demand descriptions and maps them to demand expression nodes, with node content being standardized demand phrases. The server extracts product attributes and maps them to product attribute nodes, with each attribute existing as an independent node; for example, "foldable," "aluminum alloy," and "USB-C interface" are each a node. The server extracts usage scenarios and maps them to usage scenario nodes. Between nodes, the server establishes directed edges: a consumer group node pointing to a demand expression node indicates "this group has a certain demand," a demand expression node pointing to a product attribute node indicates "this demand expects a certain attribute," and a product attribute node pointing to a usage scenario node indicates "this attribute is suitable for a certain scenario." Each edge is configured with an initial weight and a decay coefficient. The initial weight is set based on the confidence level of the relationship in a single intent, while the decay coefficient is adaptively determined based on the historical demand cycle of the product attributes associated with the edge, making the edge weights of fast fashion attributes decay faster than those of furniture attributes. The underlying storage of the graph uses a graph database. The server organizes the data structure of nodes and edges using an attribute graph model, supporting multi-hop traversal and subgraph extraction based on graph query languages ​​such as Cypher or Gremlin. The server triggers persistent storage of graph snapshots every hour, with snapshots marked by time-series version numbers, supporting backtracking to the graph state at any historical point in time.

[0045] In one possible implementation, the relational weights of edges are dynamically updated according to a time decay function. Specifically, this includes: configuring an initial weight and decay coefficient for each edge, wherein the decay coefficient is adaptively determined based on the demand cycle characteristics of the product attributes associated with the edge; each time a new consumer intent is detected corresponding to an edge, an increment is applied to the current weight of the edge, and a matching timestamp is recorded; all edges in the dynamic demand knowledge graph are traversed at a preset frequency, and the current weight of each edge is calculated using an exponential decay function based on the time interval between the current time and the matching timestamp, and then updated to the dynamic demand knowledge graph.

[0046] Specifically, the adaptive determination of the decay coefficient is completed by the server's offline analysis module. This module counts the number of demand expression nodes associated with each product attribute each month over the past twelve months. If the number of demands associated with an attribute drops to less than half of its peak value within three months, it is determined to be a short-cycle attribute, with a decay coefficient of 0.05 per hour; if the number of demands remains above half of its peak value for more than six months, it is determined to be a long-cycle attribute, with a decay coefficient of 0.01 per hour; otherwise, it is determined to be a medium-cycle attribute, with a decay coefficient of 0.025 per hour. Incremental application adopts a saturated incremental strategy. During a single match, the server increases the edge weight by 10% of the edge's initial weight, but the total weight does not exceed twice the initial weight, avoiding excessive weight inflation caused by short-term concentrated matching. The server performs exponential decay calculation traversal on an hourly basis. At the top of the hour, it traverses the graph for edges that have not been matched since the last update and performs decay calculations. Edges whose weight is below 1% of their initial weight are marked and treated as dormant edges in subsequent graph snapshot generation to save computational overhead. Dormant edges do not participate in popularity value calculation and migration path analysis, but their node structure is retained in the graph. When a consumer intent matches the edge again, the server will reactivate the dormant edge and restore it to its initial weight.

[0047] S104. Perform time-series analysis on the relational weights of the edges of the demand expression nodes in the dynamic demand knowledge graph, calculate the popularity value of the demand expression nodes in different preset time windows, identify the popularity trend of the demand expression nodes based on the change of popularity value over time, and determine the newly emerging demand hotspots and demand migration paths based on the popularity trend.

[0048] In the above steps, the server continuously performs temporal analysis on the relational weights of the edges of demand expression nodes in the dynamic demand knowledge graph. The server generates a snapshot of the graph every 24 hours, using a preset time window. In each snapshot, the server calculates the popularity value of each demand expression node. The popularity value is determined by the sum of the weights of all incoming and outgoing edges of that node; the higher the weight and the more associated edges a demand expression node has, the greater its popularity value. The server arranges the popularity values ​​of the same demand expression node in consecutive snapshots in chronological order to form a popularity value sequence. The server traverses all popularity value sequences of demand expression nodes, identifies nodes with an upward trend in popularity through differential detection, and marks them as currently hot demand nodes. Simultaneously, the server constructs a transition matrix from consumer group nodes to demand expression nodes between consecutive snapshots, calculating the change in the associated demand expression nodes for each consumer group node in adjacent snapshots. The transition matrix is ​​constructed with consumer group nodes as rows and demand expression nodes as columns. The matrix elements represent the change in the group's attention to the demand in adjacent snapshots, with attention measured by the weight of the edge pointing from the group to the demand expression node. The server normalizes the transition matrix row-wise to obtain a transition probability matrix, and identifies paths with a transition probability exceeding 0.3 as demand migration paths. For example, if the server detects a shift in demand from the "outdoor beginner" group from "basic camping chair" to "moon chair with footrest," it records this as a migration path. This record includes the starting demand node, the ending demand node, the migration probability value, and the time window identifier where the migration occurred. The server can aggregate multiple demand migration paths into a demand evolution tree for visualization, showing the segmentation and iteration direction of demand within a category.

[0049] In one possible implementation, new demand hotspots are identified based on the trend of popularity changes. Specifically, this includes: obtaining the popularity values ​​of the same demand expression node in multiple consecutive time-series snapshots to form a popularity value sequence; detecting the popularity value sequence, and when the first-order difference of the popularity value sequence changes from negative to positive and exceeds a preset difference threshold, the demand expression node is identified as a new demand hotspot.

[0050] Specifically, the first-order difference calculation of the heat value sequence is performed by the server by subtracting the heat value of the previous snapshot from the current snapshot heat value. In the judgment logic, the requirement for consecutive negative values ​​to turn positive is that at least two consecutive positive differences appear after at least three consecutive negative differences, and the absolute value of the most recent positive difference exceeds a preset difference threshold of 0.15. The preset difference threshold is determined by the server through historical data statistics, taking the 95th percentile of the first-order differences of the heat values ​​of all demand expression nodes over the past six months. This threshold is recalculated every thirty days to adapt to changes in the overall market volatility. When a demand expression node is determined to be a newly emerging demand hotspot, the server adds it to the current hotspot list and continuously monitors it for at least two subsequent snapshot periods. If the heat value continues to rise, the hotspot label is maintained; if the heat value falls below the threshold in two consecutive snapshots, the label is removed, thus filtering out short-term pulse fluctuations. Each entry in the hotspot list is accompanied by a hotspot intensity value, which is the ratio of the positive difference value at the judgment time to the preset difference threshold. When multiple hotspots exist simultaneously, the server prioritizes them based on this intensity value.

[0051] S105. Based on the existing relationship edges between the demand expression node and the product attribute node, the newly emerging demand hotspots are mapped to the candidate product attribute combinations associated with the demand expression node to generate the current hot product selection direction. Combined with the demand migration path, the candidate product attribute combinations that will become hotspots in the future within a preset time period are predicted to generate the predicted hot product selection direction.

[0052] In the above steps, the server performs hotspot mapping based on the existing relationship edges between demand expression nodes and product attribute nodes. For newly emerging demand hotspot nodes identified in the previous step, the server searches for all connected product attribute nodes along the outgoing edges of the demand expression node in the graph, collecting them to form candidate product attribute combinations. For example, if the outgoing edge of the demand expression node "lightweight camping chair" points to the four attribute nodes "aluminum alloy," "foldable," "weight less than 1.5kg," and "with armrests," then the candidate product attribute combination is a product definition containing all or some of the above attributes. The server outputs this candidate product attribute combination as the current hotspot product selection direction. At the same time, the server obtains the demand migration path determined in the previous step, extracts the endpoint demand expression node of the migration path and its associated product attribute combination, as candidate directions that may become hotspots in the future within a preset time period, and generates predicted hotspot product selection directions. The preset time period is seven to fourteen days by default, and the specific parameters are adaptively adjusted by the server according to the historical conversion speed of the migration path. If the historical average conversion cycle of similar demands on the migration path is five days, the predicted time period is automatically shortened to five to ten days. The current hot product selection direction and the predicted hot product selection direction are output together. Each item in the predicted hot product selection direction is accompanied by a prediction confidence score. The confidence score is calculated by the server based on the transition probability value of the migration path and the current popularity value of the destination demand expression node.

[0053] In one possible implementation, after generating predicted hot product selection directions, the method further includes: matching the current hot product selection directions and the candidate product attribute combinations corresponding to the predicted hot product selection directions with a preset supplier database to obtain candidate factories that meet the production conditions of the target product, and calculating the modification cost and production cycle. The preset supplier database includes factory process capabilities, minimum order quantity, and historical sampling cycle. The target product is the product corresponding to the candidate product attribute combination. The information of the candidate factories, modification costs, and production cycles are added to the product selection strategy suggestions.

[0054] Specifically, the server maintains each supplier record in a pre-defined supplier database, including the factory's geographical location, a list of processable materials, minimum order quantity, average sampling cycle over the past six months, and capacity utilization rate over the past three months. The server performs both rule-based and semantic matching to verify the material, process requirements, complexity level, and supplier processing capabilities in the candidate product attribute combinations, filtering out candidate factories that meet at least one requirement. Rule-based matching is based on precise field matching of materials and processes, while semantic matching encodes the attribute description text and the supplier's capability tag text into vectors and calculates the cosine similarity; similarities exceeding 0.8 are considered a match. The cost of the modified product is calculated by adding the basic mold opening fee or pattern fee to the sum of the incremental costs corresponding to each attribute. The server obtains the incremental cost of each attribute by querying an attribute-cost mapping table, which is updated monthly based on price data from historical sampling records. For attributes not present in the mapping table, the server uses the median of incremental costs for similar attributes as the default value. The server estimates the production cycle based on the weighted average of the candidate factory's current capacity utilization rate and historical sampling cycles, extending the estimated production cycle of factories with a capacity utilization rate exceeding 80% by 20%.

[0055] In one possible implementation, after generating the predicted hot product selection direction, the method further includes: dividing consumer group nodes according to geographic market tags to obtain consumer group sets corresponding to each geographic market; aggregating the heat values ​​of demand expression nodes associated with the consumer group sets according to different preset time windows to obtain the demand heat evolution curve of each geographic market; taking the market where demand heat rises first as the source market and the market where demand heat rises later as the target market, calculating the time difference between the start of demand heat rise in the source market and the target market to obtain the demand transmission delay; when the demand transmission delay is lower than a preset delay threshold and the demand heat value of the source market is higher than the preset heat threshold, marking the demand hotspot in the source market as a transmittable hotspot; adding the transmittable hotspot to the predicted hot product selection direction of the target market.

[0056] Specifically, the server determines the geographic market segmentation of consumer groups based on a comprehensive analysis of three factors: user registration location, logistics destination country, and content publishing IP address. Consumers from the same region are grouped into the corresponding geographic market consumer group set. The demand heat evolution curve for each market is generated by the server aggregating the heat values ​​of all demand expression nodes within that market at a 24-hour granularity, forming an independent time series curve for each market. The demand transmission delay is calculated by the server performing inflection point detection on the heat value sequence of the same demand expression node in both markets. The inflection point determination uses the logic that the first-order difference of the heat value sequence changes from negative to positive and exceeds a preset difference threshold. The difference between the inflection point time in the source market and the inflection point time in the target market is taken as the demand transmission delay. The preset delay threshold is seven days by default. When the heat value in the source market exceeds the preset heat threshold and the transmission delay is less than seven days, the server marks the current hot spot in the source market as a transmittable hot spot and pre-includes it in the predicted hot spot product selection direction in the target market, while attaching the heat curve and transmission delay of the source market as a reference for prediction confidence. The prediction confidence of a transmittable hotspot is negatively correlated with the transmission delay. The shorter the transmission delay, the higher the confidence. The server uses the inverse of the transmission delay as the confidence weight.

[0057] S106. Obtain user profile data, best-selling category data, and recommendation algorithm parameters from different cross-border e-commerce platforms. Perform differentiated adaptation on current and predicted hot product selection directions, and output product selection strategy suggestions for different cross-border e-commerce platforms.

[0058] In the above steps, the server obtains platform adaptation parameters for different cross-border e-commerce platforms, including user profile data such as price sensitivity distribution, average decision-making cycle, and content format preferences; platform best-selling category data such as the percentage of transaction volume for each category in the past thirty days; and platform recommendation algorithm parameters such as the ratio of relevance score to sales conversion score in search weight. For each target platform, the server extracts product attribute combinations of products to be adapted from current and predicted hot product selection directions, calculates the degree of adaptation of each attribute to the target platform, and removes attributes with an adaptation degree lower than a preset threshold from the combinations, thus obtaining the product feature combination adapted to that platform. The strategy template library stores corresponding title structures, weights of five-point description points, pricing range suggestions, and visual material shooting directions according to platform type and category. The templates for different platforms differ in the presentation order and weight allocation of content elements. Taking Amazon as an example, its strategy template requires placing core functional parameters at the beginning of the title, the first two points of the five-point description covering the main usage scenarios and compatibility information, and the pricing range suggestion referring to the price range distribution of the top ten best-sellers in the same category. Taking the TikTok Shop platform as an example, its strategy template requires extracting visual differentiators into short hashtags of no more than three words, and specifying that the first three seconds of the video should showcase the product's core usage effects rather than static images of its appearance. After server matching, a structured product selection strategy suggestion is generated. Each strategy suggestion item includes an adapted attribute list, recommended title text, point-by-point descriptive copy, upper and lower price limits, and a summary of the shooting script.

[0059] In one possible implementation, user profile data, best-selling category data, and recommendation algorithm parameters from different cross-border e-commerce platforms are acquired. Differential adaptations are then performed on current and predicted trending product selection directions, outputting product selection strategy suggestions for different cross-border e-commerce platforms. Specifically, this includes: for each cross-border e-commerce platform, acquiring platform adaptation parameters, including user profile data, best-selling category data, and recommendation algorithm parameters. User profile data includes price sensitivity, decision-making cycle, and content preferences. From the current and predicted trending product selection directions, product attribute combinations to be adapted are extracted. Based on the platform adaptation parameters, the matching weight of each attribute in the product attribute combination with the target platform is calculated. Attributes with matching weights lower than a preset trimming threshold are removed to obtain product selection feature combinations adapted to the target platform. These product selection feature combinations are then matched with a preset strategy template to generate product selection strategy suggestions for the target platform.

[0060] Specifically, the server periodically pulls and aggregates platform adaptation parameters from public data reports of various platforms, the audience insight module of the advertising backend, and the trend insight interface provided by the platforms. The matching weight is calculated by the server mapping product attributes to 256-dimensional attribute embedding vectors. These vectors are obtained by inputting the attribute text into a pre-trained attribute encoder, which is trained on the product attribute-annotated corpus using a masked language model task. The server also maps platform user profile features to 256-dimensional platform feature vectors. The inner product of these two vectors is then normalized to the zero-to-one range using the sigmoid function. A preset clipping threshold of 0.3 is set; attributes below this threshold are considered significantly mismatched with the target platform's user preferences and are removed. The strategy template library is stored by platform. Amazon templates emphasize search keyword density and long-tail keyword coverage, requiring the main keyword to appear within the first five words of the title template. TikTokShop templates focus on visual hooks and hashtag combinations in the first three seconds of short videos, specifying at least one close-up shot and one usage scenario shot in the material template. Temu templates emphasize price anchoring and cost-effectiveness rhetoric, requiring the embedding of a "market price comparison" field in the description. Independent website templates emphasize brand storytelling and series product recommendations, with content templates including brand value paragraphs and cross-selling recommendation slots. The server fills in the trimmed attribute set with the attribute slots reserved in the template to generate a complete product selection strategy document.

[0061] This application also provides a cross-border product selection strategy generation device based on dynamic knowledge graphs, referring to... Figure 2 The device is a server, which includes an acquisition unit 201 and a processing unit 202.

[0062] The acquisition unit 201 is used to collect multimodal behavioral data of consumers in real time from multiple data sources, including cross-border e-commerce platforms, social media platforms and search engines. The multimodal behavioral data includes text data and image data.

[0063] Processing unit 202 is used to extract semantic feature vectors from text data using a multilingual large language model and visual feature vectors from image data using a computer vision model. It maps the semantic and visual feature vectors to a shared embedding space to obtain a unified intent feature vector, and determines structured consumer intent based on this vector. It is also used to extract consumer identifiers, demand descriptions, product attributes, and usage scenarios based on the consumer intent. Consumer identifiers are mapped to consumer group nodes, demand descriptions to demand expression nodes, product attributes to product attribute nodes, and usage scenarios to usage scenario nodes. Edges are established between corresponding nodes based on the relationships in the consumer intent to construct a dynamic demand knowledge graph. The relationship weights of the edges are dynamically updated according to a time decay function. Furthermore, it is used to process the demand expression nodes in the dynamic demand knowledge graph. The algorithm performs time-series analysis on the relational weights of edges to calculate the popularity values ​​of demand expression nodes in different preset time windows. Based on the changes in popularity values ​​over time, it identifies the popularity trends of demand expression nodes and determines newly emerging demand hotspots and demand migration paths. It also maps newly emerging demand hotspots to candidate product attribute combinations associated with demand expression nodes based on existing relational edges between demand expression nodes and product attribute nodes, generating current hot product selection directions. Combined with demand migration paths, it predicts candidate product attribute combinations that will become hotspots within a preset time period, generating predicted hot product selection directions. Furthermore, it acquires user profile data, best-selling category data, and recommendation algorithm parameters from different cross-border e-commerce platforms to differentiate and adapt current and predicted hot product selection directions, outputting product selection strategy suggestions for different cross-border e-commerce platforms.

[0064] In one possible implementation, the processing unit 202 is used to input text data into a multilingual large language model, and map texts of different languages ​​to a shared embedding space through a cross-language semantic encoder in the multilingual large language model to obtain semantic feature vectors; input image data into a computer vision model, and extract features of visual elements related to the product through an object detection and attribute recognition network to obtain visual feature vectors; use a cross-modal attention mechanism to perform a weighted summation of the semantic feature vectors and the visual feature vectors to obtain a unified intent feature vector; input the unified intent feature vector into an intent classifier to output consumer intent, which includes consumer identification, demand description, product attributes, and usage scenario.

[0065] In one possible implementation, the processing unit 202 is configured to assign an initial weight and a decay coefficient to each edge, wherein the decay coefficient is adaptively determined based on the demand cycle characteristics of the product attributes associated with the edge; each time a new consumer intent is detected that corresponds to an edge, an increment is applied to the current weight of the edge and a matching timestamp is recorded; all edges in the dynamic demand knowledge graph are traversed at a preset frequency, and the current weight of each edge is calculated using an exponential decay function based on the time interval between the current time and the matching timestamp, and then updated to the dynamic demand knowledge graph.

[0066] In one possible implementation, the acquisition unit 201 is used to acquire the heat value of the same demand expression node in multiple consecutive time-series snapshots to form a heat value sequence; the processing unit 202 is used to detect the heat value sequence, and when the first-order difference of the heat value sequence changes from negative to positive and exceeds a preset difference threshold, the demand expression node is identified as a newly emerging demand hotspot.

[0067] In one possible implementation, the acquisition unit 201 is used to acquire platform adaptation parameters for each cross-border e-commerce platform. The platform adaptation parameters include user profile data, platform best-selling category data, and platform recommendation algorithm parameters. The user profile data includes price sensitivity, decision-making cycle, and content preferences. The processing unit 202 is used to extract product attribute combinations of products to be adapted from current and predicted hot product selection directions; calculate the matching weight of each attribute in the product attribute combination with the target platform according to the platform adaptation parameters; remove attributes with matching weights lower than a preset clipping threshold to obtain product selection feature combinations adapted to the target platform; and match the product selection feature combinations with a preset strategy template to generate product selection strategy suggestions for the target platform.

[0068] In one possible implementation, the processing unit 202 is used to match the current hot product selection direction and the candidate product attribute combination corresponding to the predicted hot product selection direction with the preset supplier database to obtain candidate factories that meet the production conditions of the target product, and calculate the modification cost and production cycle. The preset supplier database includes factory process capabilities, minimum order quantity and historical sampling cycle, and the target product is the product corresponding to the candidate product attribute combination; the information of the candidate factory, modification cost and production cycle are added to the product selection strategy suggestion.

[0069] In one possible implementation, the processing unit 202 is used to divide consumer group nodes according to geographic market tags to obtain consumer group sets corresponding to each geographic market; according to different preset time windows, the heat values ​​of demand expression nodes associated with consumer group sets are aggregated to obtain demand heat evolution curves for each geographic market; the market where demand heat rises first is taken as the source market, and the market where demand heat rises later is taken as the target market, and the time difference between the start of demand heat rise in the source market and the target market is calculated to obtain the demand transmission delay; when the demand transmission delay is lower than a preset delay threshold and the demand heat value of the source market is higher than the preset heat threshold, the demand hotspot in the source market is marked as a transmittable hotspot; the transmittable hotspot is added to the predicted hotspot product selection direction of the target market.

[0070] It should be noted that the above embodiments of the apparatus are only illustrated by the division of the above functional modules. In practical applications, the above functions can be assigned to different functional modules as needed, that is, the internal structure of the device can be divided into different functional modules to complete all or part of the functions described above. In addition, the apparatus and method embodiments provided in the above embodiments belong to the same concept, and the specific implementation process can be found in the method embodiments, which will not be repeated here.

[0071] This application also provides an electronic device. (See reference...) Figure 3 , Figure 3 This is a schematic diagram of the structure of an electronic device provided in an embodiment of this application. The electronic device 300 may include: at least one processor 301, at least one communication bus 302, at least one user interface 303, a network interface 304, and a memory 305.

[0072] The communication bus 302 is used to enable communication between these components.

[0073] The user interface 303 may include a display screen and a camera. Optionally, the user interface 303 may also include a standard wired interface and a wireless interface.

[0074] The network interface 304 may include standard wired interfaces and wireless interfaces (such as Wi-Fi interfaces).

[0075] The processor 301 may include one or more processing cores. The processor 301 connects to various parts of the server using various interfaces and lines, and performs various server functions and processes data by running or executing instructions, programs, code sets, or instruction sets stored in memory 305, and by calling data stored in memory 305. Optionally, the processor 301 may be implemented using at least one hardware form of Digital Signal Processing (DSP), Field-Programmable Gate Array (FPGA), or Programmable Logic Array (PLA). The processor 301 may integrate one or a combination of several of the following: Central Processing Unit (CPU), Graphics Processing Unit (GPU), and modem. The CPU primarily handles the operating system, user interface, and applications; the GPU is responsible for rendering and drawing the content required for display; and the modem handles wireless communication. It is understood that the modem may also not be integrated into the processor 301 and may be implemented as a separate chip.

[0076] The memory 305 may include random access memory (RAM) or read-only memory. Optionally, the memory 305 may include a non-transitory computer-readable storage medium. The memory 305 may be used to store instructions, programs, code, code sets, or instruction sets. The memory 305 may include a program storage area and a data storage area, wherein the program storage area may store instructions for implementing an operating system, instructions for at least one function (such as touch function, sound playback function, image playback function, etc.), instructions for implementing the above-described method embodiments, etc.; the data storage area may store data involved in the above-described method embodiments, etc. The memory 305 may also be at least one storage device located remotely from the aforementioned processor 301. (Refer to...) Figure 3 The memory 305, which serves as a computer storage medium, may include an operating system, a network communication module, a user interface module, and an application program for generating cross-border product selection strategies based on dynamic knowledge graphs.

[0077] exist Figure 3In the illustrated electronic device 300, the user interface 303 is mainly used to provide an input interface for the user and to acquire user input data; while the processor 301 can be used to call an application program stored in the memory 305 for generating cross-border product selection strategies based on dynamic knowledge graphs. When executed by one or more processors 301, the electronic device 300 performs one or more methods as described in the above embodiments. It should be noted that, for the foregoing method embodiments, for the sake of simplicity, they are all described as a series of actions. However, those skilled in the art should understand that this application is not limited to the described order of actions, because according to this application, some steps can be performed in other orders or simultaneously. Secondly, those skilled in the art should also understand that the embodiments described in the specification are all preferred embodiments, and the actions and modules involved are not necessarily essential to this application.

[0078] In the above embodiments, the descriptions of each embodiment have different focuses. For parts not described in detail in a certain embodiment, please refer to the relevant descriptions in other embodiments.

[0079] This application also provides a computer-readable storage medium storing instructions. When executed by one or more processors, these instructions cause an electronic device to perform one or more of the methods described in the above embodiments.

[0080] In the various embodiments provided in this application, it should be understood that the disclosed apparatus can be implemented in other ways. For example, the apparatus embodiments described above are merely illustrative; for instance, the division of units is only a logical functional division, and in actual implementation, there may be other division methods. For example, multiple units or components may be combined or integrated into another system, or some features may be ignored or not executed. Furthermore, the coupling or direct coupling or communication connection shown or discussed may be through some service interface; the indirect coupling or communication connection between apparatuses or units may be electrical or other forms.

[0081] The units described as separate components may or may not be physically separate. The components shown as units may or may not be physical units; that is, they may be located in one place or distributed across multiple network units. Some or all of the units can be selected to achieve the purpose of this embodiment according to actual needs.

[0082] Furthermore, the functional units in the various embodiments of this application can be integrated into one processing unit, or each unit can exist physically separately, or two or more units can be integrated into one unit. The integrated unit can be implemented in hardware or as a software functional unit.

[0083] If the integrated unit is implemented as a software functional unit and sold or used as an independent product, it can be stored in a computer-readable storage device (CMD). Based on this understanding, the technical solution of this application, in essence, or the part that contributes to the prior art, or all or part of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a memory and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods of the various embodiments of this application. The aforementioned memory includes various media capable of storing program code, such as USB flash drives, portable hard drives, magnetic disks, or optical disks.

[0084] The above description is merely an exemplary embodiment of this disclosure and should not be construed as limiting the scope of this disclosure. Any equivalent changes and modifications made in accordance with the teachings of this disclosure shall still fall within the scope of this disclosure.

[0085] This application is intended to cover any variations, uses, or adaptations of this disclosure that follow the general principles of this disclosure and include common knowledge or customary techniques in the art not described in this disclosure. The specification and embodiments are to be considered exemplary only, and the scope of this disclosure is defined by the claims.

Claims

1. A method for generating cross-border product selection strategies based on dynamic knowledge graphs, characterized in that, The method includes: Consumer multimodal behavior data is collected in real time from multiple data sources, including cross-border e-commerce platforms, social media platforms, and search engines. The multimodal behavior data includes text data and image data. The semantic feature vector of the text data is extracted by a multilingual large language model, and the visual feature vector of the image data is extracted by a computer vision model. The semantic feature vector and the visual feature vector are mapped to a shared embedding space to obtain a unified intent feature vector. The structured consumer intent is determined based on the unified intent feature vector. Based on the consumer intent, consumer identifiers, demand descriptions, product attributes, and usage scenarios are extracted. The consumer identifiers are mapped to consumer group nodes, the demand descriptions are mapped to demand expression nodes, the product attributes are mapped to product attribute nodes, and the usage scenarios are mapped to usage scenario nodes. Edges are established between corresponding nodes according to the relationships in the consumer intent to construct a dynamic demand knowledge graph. The relationship weights of the edges are dynamically updated according to a time decay function. A time-series analysis is performed on the relational weights of the edges of the demand expression nodes in the dynamic demand knowledge graph. The popularity value of the demand expression nodes in different preset time windows is calculated. Based on the change of the popularity value over time, the popularity change trend of the demand expression nodes is identified. Based on the popularity change trend, newly emerging demand hotspots and demand migration paths are determined. Based on the existing relationship edges between the demand expression node and the product attribute node, the newly emerging demand hotspots are mapped to the candidate product attribute combinations associated with the demand expression node, generating the current hot product selection direction. Combined with the demand migration path, the candidate product attribute combinations that will become hotspots in the future within a preset time period are predicted, generating the predicted hot product selection direction. Obtain user profile data, best-selling category data, and recommendation algorithm parameters from different cross-border e-commerce platforms. Perform differentiated adaptation on the current hot product selection direction and the predicted hot product selection direction, and output product selection strategy suggestions for the different cross-border e-commerce platforms.

2. The method according to claim 1, characterized in that, The process involves extracting semantic feature vectors from the text data using a multilingual large language model and visual feature vectors from the image data using a computer vision model. The semantic feature vectors and visual feature vectors are then mapped to a shared embedding space to obtain a unified intent feature vector. Based on this unified intent feature vector, structured consumer intent is determined. Specifically, this includes: The text data is input into the multilingual large language model, and the cross-lingual semantic encoder in the multilingual large language model maps the texts of different languages ​​to the shared embedding space to obtain the semantic feature vector. The image data is input into the computer vision model, and the features of visual elements related to the product are extracted through the target detection and attribute recognition network to obtain the visual feature vector; The semantic feature vector and the visual feature vector are weighted and summed using a cross-modal attention mechanism to obtain the unified intent feature vector; The unified intent feature vector is input into the intent classifier, which outputs the consumer intent, which includes the consumer identifier, the demand description, the product attributes, and the usage scenario.

3. The method according to claim 1, characterized in that, The relation weights of the edges are dynamically updated according to a time decay function, specifically including: An initial weight and a decay coefficient are configured for each edge, and the decay coefficient is adaptively determined based on the demand cycle characteristics of the product attributes associated with the edge; Each time a new consumer intent is detected that corresponds to the edge, an increment is applied to the current weight of the edge, and the matching timestamp is recorded; The system iterates through all edges in the dynamic demand knowledge graph at a preset frequency. Based on the time interval between the current time and the matching timestamp, it calculates the current weight of each edge using an exponential decay function and updates it to the dynamic demand knowledge graph.

4. The method according to claim 1, characterized in that, The process of identifying newly emerging demand hotspots based on the trend of popularity changes specifically includes: Obtain the popularity value of the same demand expression node in multiple consecutive time-series snapshots to form a popularity value sequence; The heat value sequence is detected, and when the first difference of the heat value sequence changes from negative to positive and exceeds a preset difference threshold, the demand expression node is identified as the newly emerging demand hotspot.

5. The method according to claim 1, characterized in that, The process of acquiring user profile data, best-selling category data, and recommendation algorithm parameters from different cross-border e-commerce platforms, and then applying differentiated adaptations to the current and predicted hot product selection directions to output product selection strategy suggestions for different cross-border e-commerce platforms, specifically includes: For each of the aforementioned cross-border e-commerce platforms, the platform adaptation parameters of the target platform are obtained. The platform adaptation parameters include the user profile data, the platform's best-selling category data, and the platform's recommendation algorithm parameters. The user profile data includes price sensitivity, decision-making cycle, and content preferences. Extract the product attribute combination of the products to be matched from the current hot product selection direction and the predicted hot product selection direction; Based on the platform adaptation parameters, calculate the matching weight of each attribute in the product attribute combination with the target platform, remove attributes with matching weights lower than a preset cropping threshold, and obtain the product feature combination adapted to the target platform. The product selection feature combination is matched with a preset strategy template to generate product selection strategy suggestions for the target platform.

6. The method according to claim 1, characterized in that, After generating the predicted hot product selection directions, the method further includes: The candidate product attribute combination corresponding to the current hot product selection direction and the predicted hot product selection direction is matched with the preset supplier database to obtain candidate factories that meet the production conditions of the target product, and the modification cost and production cycle are calculated. The preset supplier database includes factory process capacity, minimum order quantity and historical sampling cycle. The target product is the product corresponding to the candidate product attribute combination. Add the information of the candidate factories, the cost of the model change, and the production cycle to the product selection strategy recommendations.

7. The method according to claim 1, characterized in that, After generating the predicted hot product selection directions, the method further includes: The consumer group nodes are divided according to geographic market tags to obtain the consumer group set corresponding to each geographic market; Based on the different preset time windows, the heat values ​​of the demand expression nodes associated with the consumer group set are aggregated to obtain the demand heat evolution curve of each geographic market; The market where demand first rises is taken as the source market, and the market where demand rises later is taken as the target market. The time difference between the start of demand rise in the source market and the target market is calculated to obtain the demand transmission delay. When the demand transmission delay is lower than a preset delay threshold and the demand heat value of the source market is higher than a preset heat threshold, the demand hotspot in the source market is marked as a transmittable hotspot. The transmissible hotspots are incorporated into the predicted hotspot product selection direction of the target market.

8. A cross-border product selection strategy generation device based on dynamic knowledge graph, characterized in that, The device includes an acquisition unit (201) and a processing unit (202): The acquisition unit (201) is used to collect multimodal behavioral data of consumers in real time from multiple data sources, including cross-border e-commerce platforms, social media platforms and search engines, and the multimodal behavioral data includes text data and image data. The processing unit (202) is used to extract the semantic feature vector of the text data through a multilingual large language model, and extract the visual feature vector of the image data through a computer vision model, map the semantic feature vector and the visual feature vector to a shared embedding space to obtain a unified intent feature vector, and determine the structured consumer intent based on the unified intent feature vector. The processing unit (202) is further configured to extract consumer identifiers, demand descriptions, product attributes, and usage scenarios based on the consumer intent, map the consumer identifiers to consumer group nodes, map the demand descriptions to demand expression nodes, map the product attributes to product attribute nodes, map the usage scenarios to usage scenario nodes, and establish edges between corresponding nodes according to the association relationships in the consumer intent to construct a dynamic demand knowledge graph, wherein the relationship weights of the edges are dynamically updated according to a time decay function; The processing unit (202) is also used to perform time-series analysis on the relational weights of the edges of the demand expression nodes in the dynamic demand knowledge graph, calculate the popularity value of the demand expression nodes in different preset time windows, identify the popularity change trend of the demand expression nodes based on the change of the popularity value over time, and determine the newly emerging demand hotspots and demand migration paths according to the popularity change trend. The processing unit (202) is further configured to map the newly emerging demand hotspots to candidate product attribute combinations associated with the demand expression node based on the existing relationship edges between the demand expression node and the product attribute node, generate the current hotspot product selection direction, and predict the candidate product attribute combinations that will become hotspots within a preset time period in conjunction with the demand migration path, thereby generating the predicted hotspot product selection direction. The processing unit (202) is also used to acquire user profile data, platform best-selling category data and platform recommendation algorithm parameters of different cross-border e-commerce platforms, perform differentiated adaptation on the current hot product selection direction and the predicted hot product selection direction, and output product selection strategy suggestions for the different cross-border e-commerce platforms.

9. An electronic device, characterized in that, The device includes a processor (301), a memory (305), a user interface (303), and a network interface (304). The memory (305) is used to store instructions. The user interface (303) and the network interface (304) are used to communicate with other devices. The processor (301) is used to execute the instructions stored in the memory (305) to cause the electronic device (300) to perform the method as described in any one of claims 1 to 7.

10. A computer-readable storage medium storing a computer program, characterized in that, When the computer program is executed by a processor, it implements the method as described in any one of claims 1 to 7 above.

Images