Method and device for screening cooperative KOL for cross-border e-commerce, computer equipment and storage medium

Abstract

The invention relates to the technical field of data processing, and discloses a method and device for screening cooperation KOLs for cross-border e-commerce, computer equipment and a storage medium, and the method comprises the steps: determining a plurality of alternative KOLs of a target cross-border region; obtaining a user content data set of each KOL and generating a content feature vector; collecting fan interaction behavior time sequence data and calculating a fan authenticity index; extracting brand introduction data to generate brand attribute vectors; calling regional culture rule data to generate culture rule vectors; calculating the compatibility of the three-direction vector to obtain a comprehensive adaptation degree vector; fusing the vectors to form a comprehensive feature vector; the conversion rate is predicted through a pre-trained Transform model; and the cooperative KOL is selected according to a prediction result, so that the problems of low efficiency and insufficient matching precision existing in traditional manual screening are effectively solved.

Description

Technical Field

[0001] This invention relates to the field of data processing technology, and in particular to a method, apparatus, computer equipment, and storage medium for screening cooperative KOLs for cross-border e-commerce.

[0002] Background Techniques When brands enter cross-border markets, they often seek out KOLs (Key Opinion Leaders) in those regions to endorse or market their products in order to increase sales. Currently, the method of manually searching for KOLs in these regions is inefficient and has poor matching accuracy.

[0003] When brands enter cross-border markets, they typically leverage local KOLs (Key Opinion Leaders) for product endorsements or content marketing. Currently, this primarily relies on manual methods to find suitable KOLs in target cross-border regions. This manual screening method has significant limitations: 1. Low screening efficiency: Evaluating dozens of candidates in a single region takes an average of 3 to 4 weeks, making it difficult to adapt to rapidly changing market demands. 2. Poor matching accuracy: Due to the lack of a scientifically quantifiable evaluation system, it is difficult to find suitable KOLs. Summary of the Invention

[0004] Based on this, it is necessary to address the technical problems of low efficiency and poor accuracy in the existing technology for selecting KOLs for brands in cross-border regions, and to propose a method, device, computer equipment and storage medium for selecting cooperative KOLs for cross-border e-commerce.

[0005] Firstly, a method for selecting KOLs (Key Opinion Leaders) for cross-border e-commerce is provided, the method comprising: Identify multiple potential KOLs located in the target cross-border region; Obtain the user content dataset published on the account of the candidate KOL, and generate content feature vectors based on the user content dataset; the user content dataset includes: multiple user content data and metadata of each user content data, and the user content data is text data, video data or image data; Obtain the time-series dataset of fan interaction behavior of the candidate KOLs, and calculate the fan authenticity index based on the time-series dataset of fan interaction behavior; Obtain introductory data for the target brand, and generate a brand attribute vector based on the introductory data; The system retrieves cultural rule data corresponding to the target cross-border region from a pre-defined regional cultural rule base, and generates a cultural rule vector based on the cultural rule data. Calculate the compatibility between the content feature vector, the brand attribute vector, and the cultural rule vector to obtain a comprehensive fit vector; The content feature vector, the brand attribute vector, and the comprehensive adaptability vector are fused to obtain the comprehensive feature vector; The comprehensive feature vector is input into the pre-trained Transformer prediction model to obtain the conversion rate prediction value; Based on the conversion rate prediction, select one KOL from the multiple candidate KOLs to collaborate with.

[0006] Secondly, a device is provided for screening KOLs (Key Opinion Leaders) for cross-border e-commerce, the device comprising: The identification module is used to identify multiple candidate KOLs located in the target cross-border region; The acquisition module is used to acquire the user content dataset published by the candidate KOL on their account, and to generate content feature vectors based on the user content dataset; the user content dataset includes: multiple user content data and metadata of each user content data, and the user content data is text data, video data or image data; The acquisition module is also used to acquire the time-series dataset of fan interaction behavior of the candidate KOL, and to calculate the fan authenticity index based on the time-series dataset of fan interaction behavior. The acquisition module is also used to acquire introductory data of the target brand and generate a brand attribute vector based on the introductory data; The retrieval module is used to retrieve cultural rule data corresponding to the target cross-border region from a preset regional cultural rule base, and to generate a cultural rule vector based on the cultural rule data. The calculation module is used to calculate the compatibility between the content feature vector, the brand attribute vector, and the cultural rule vector to obtain a comprehensive fit vector; The fusion module is used to fuse the content feature vector, the brand attribute vector, and the comprehensive adaptability vector to obtain a comprehensive feature vector; The prediction module is used to input the comprehensive feature vector into the pre-trained Transformer prediction model to obtain the conversion rate prediction value; The selection module is used to select a cooperative KOL from the multiple candidate KOLs based on the conversion rate prediction value.

[0007] Thirdly, a computer device is provided, including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the computer program to implement the steps of the method for screening cooperative KOLs in cross-border e-commerce described above.

[0008] Fourthly, a computer-readable storage medium is provided, which stores a computer program that, when executed by a processor, implements the steps of the method for screening cooperative KOLs in cross-border e-commerce.

[0009] The beneficial effects of this application are: By automating the collection and processing of multimodal content data from key opinion leaders, the system achieves unified feature extraction from heterogeneous data such as text, images, and videos, replacing the traditional method of manual content review. Simultaneously, leveraging machine learning models to fuse and analyze multidimensional features, the system can process a large number of candidates in parallel, compressing the screening process, which originally took weeks, into a highly efficient automated cycle. A multimodal feature engine enables in-depth content understanding, temporal behavior analysis ensures accurate fan evaluation, and a cultural adaptation mechanism achieves precise cross-border matching. These three technological approaches together construct a complete automated screening system, significantly improving processing efficiency and eliminating the randomness of subjective human judgment through algorithmic models, achieving a technological leap from experience-based decision-making to data-driven decision-making.

[0010] In summary, by systematically processing data and using intelligent analysis to establish an objective and quantitative evaluation system, the selection of key opinion leaders has been transformed from subjective experience-based judgment to data-driven decision-making, which significantly improves the accuracy and reliability of the selection results while increasing operational efficiency. Attached Figure Description

[0011] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0012] in: Figure 1 This is an application environment diagram of a method for selecting cooperative KOLs for cross-border e-commerce in one embodiment; Figure 2 A flowchart illustrating a method for selecting KOLs for cross-border e-commerce in one embodiment; Figure 3 This is a structural block diagram of a device for screening KOLs for cross-border e-commerce in one embodiment; Figure 4 This is a structural block diagram of a computer device in one embodiment. Detailed Implementation

[0013] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some, not all, of the embodiments of the present invention. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0014] The present invention provides a method for screening KOLs for cross-border e-commerce, which can be applied to, for example... Figure 1 In this application environment, the client communicates with the server via a network. The client retrieves a dataset of user content posted on the accounts of potential KOLs from the server, and generates content feature vectors based on this dataset. The user content dataset includes multiple user content data and metadata for each data point; the user content data can be text, video, or image data. The client also retrieves a time-series dataset of fan interaction behavior from the server, and calculates a fan authenticity index based on this dataset. Furthermore, the client retrieves introduction data for the target brand from the server, and generates a brand attribute vector based on this data. It then retrieves cultural rule data corresponding to the target cross-border region from a pre-defined regional cultural rule library, and generates a cultural rule vector based on this data. The client calculates the compatibility between the content feature vector, brand attribute vector, and cultural rule vector to obtain a comprehensive fit vector. Finally, it fuses the content feature vector, brand attribute vector, and comprehensive fit vector to obtain a comprehensive feature vector. This comprehensive feature vector is then input into a pre-trained Transformer prediction model to obtain a conversion rate prediction value. Finally, based on the conversion rate prediction value, the client selects a collaborating KOL from among the multiple potential KOLs. This application enables brands to automatically select suitable KOLs in cross-border regions, solving the technical problem of low efficiency in manual screening through multimodal feature extraction and machine learning models. The client can be, but is not limited to, various personal computers, laptops, smartphones, tablets, and portable wearable devices. The server can be implemented using a dedicated server or a server cluster consisting of multiple servers. The invention will be described in detail below through specific embodiments.

[0015] Please see Figure 2 As shown, Figure 2 A flowchart illustrating a method for selecting KOLs for cross-border e-commerce, as provided in this embodiment of the invention, includes the following steps: S1: Identify multiple potential KOLs located in the target cross-border region.

[0016] In this context, KOLs (Key Opinion Leaders) refer to individuals or organizations on social media platforms that possess professional knowledge, authority, and influence. They build trust and appeal among specific audiences by consistently creating and disseminating distinctive original content, significantly impacting their followers' perceptions, attitudes, and purchasing decisions. The target cross-border region is specified by the brand owner, referring to a specific geographical area outside their existing markets where they plan to conduct product sales or service promotions. Brand owners can also specify specific social media platforms. Based on the brand owner's specified target cross-border region and target social media platform, the computer device can filter out multiple candidate KOLs. For example, social media platforms may rank different categories of KOLs monthly, year-end, or quarterly, allowing for the selection of multiple candidate KOLs based on these rankings. Current computer devices can also utilize custom filtering rules to identify and detect multiple candidate KOLs on social media platforms; this application does not impose any restrictions on this.

[0017] In one possible embodiment, determining multiple candidate KOLs located in the target cross-border region includes: The top N KOLs on the target social media platform in the target cross-border region that meet the screening criteria are selected as multiple candidate KOLs, where N is an integer greater than 1. The screening criteria include account verification status, number of followers threshold, and content update frequency threshold.

[0018] Among these, account verification status indicates the social media platform's official verification result of the account holder's identity. The follower count threshold represents the preset minimum number of followers. Content update frequency indicates the minimum number of times content is posted within a specific time period, typically measured in daily, weekly, or monthly increments.

[0019] Computer devices can access creator account information located in the target cross-border region through the API (Application Programming Interface) provided by the target social media platform. The computer device first performs an initial screening based on account verification status, retaining only accounts verified by the platform's official channels to ensure authenticity. Then, the computer device further filters the verified accounts, eliminating those with too few followers based on a follower count threshold, and those with a low update frequency based on a content update frequency threshold, ultimately resulting in a set of accounts that meet the screening criteria. Within this set, the computer device then quantitatively scores each KOL's account based on follower count and update frequency, ranking them in descending order of score, and finally selecting the top N KOLs as multiple candidate KOLs, where N is an integer greater than 1.

[0020] For example, when a brand plans to promote its beauty products in the Southeast Asian market, it uses a data interface provided by social media platform A to access 800 creator accounts in the Bangkok area of ​​Thailand, categorized as beauty creators. The initial screening criteria set by the computer equipment require that the accounts be officially verified by the platform, have at least 150,000 followers, and maintain at least three content updates per week within the past 30 days.

[0021] From the initial 800 accounts obtained, the computer system first filtered out 620 verified accounts based on their authentication status. Then, 300 accounts that did not meet the follower count threshold were excluded, and finally, 190 accounts with insufficient activity were removed based on content update frequency. This resulted in a final set of 130 accounts that met the basic criteria.

[0022] Then, a quantitative evaluation system was constructed using computer equipment to score the shortlisted accounts across multiple dimensions. The baseline score for follower count was calculated using a logarithmic function, while content update frequency was weighted based on consistency and stability. One account with 280,000 followers and daily updates received 85 points for follower count and 92 points for update frequency; another account with 350,000 followers but inconsistent updates received 90 points for follower count and 70 points for update frequency. By calculating a weighted composite score and ranking all accounts in descending order, the top 60 high-quality accounts by composite score were ultimately selected as KOLs (Key Opinion Leaders).

[0023] S2: Obtain the user content dataset published by the candidate KOL on their account, and generate a content feature vector based on the user content dataset.

[0024] The user content dataset represents a systematic collection of data comprised of content posted by key opinion leaders on their social media accounts. The user content dataset contains core data: the first category is raw content data, including multimedia content such as text, images, and videos posted by creators; the second category is metadata of the raw content data, including but not limited to descriptive information such as content publication time, geolocation tags, hashtags, and language identifiers.

[0025] The computer equipment employs multimodal fusion processing technology to uniformly analyze heterogeneous content data. For text data, a multilingual semantic understanding model is used to extract topic distribution and sentiment features, while for image and video data, a deep convolutional network is used to capture scene composition and style features. After all features are standardized, a feature-level fusion module generates a high-dimensional vector representing the comprehensive characteristics of the content.

[0026] In one possible embodiment, S2: obtaining the user content dataset published by the candidate KOL on their account, and generating a content feature vector based on the user content dataset, includes: S21: Extract text features from the text data in the user content dataset to obtain text features, and extract visual feature vectors from the image data and video data in the user content dataset. S22: Map the text features and the visual features to a unified semantic space, and use a cross-modal attention mechanism to align and weightedly fuse various features to generate fused features; S23: The fusion features are reduced in dimensionality to obtain the fusion feature vector; S24: Extract auxiliary features from the metadata in the user content dataset. The metadata includes posting time distribution, geographic location information, and topic tags. Generate an auxiliary feature vector based on the auxiliary features. S25: The fused feature vector and the auxiliary feature vector are concatenated to form the final content feature vector.

[0027] Specifically, in S21, the computer device classifies the multimodal data in the user content dataset. For text data, a multilingual semantic understanding model is used for deep analysis. Word embedding technology is used to map discrete words to a continuous vector space, and a sequence learning model is used to capture contextual semantic relationships, ultimately generating text feature vectors that represent the distribution of text topics and sentiment tendencies.

[0028] For keyframe sequences in image and video data, computer devices use deep convolutional neural networks to extract features in layers, gradually abstracting from low-level pixel features to high-level semantic features, generating visual feature vectors that include visual element composition, color style, and scene semantics.

[0029] For example, when processing 50 posts from a travel blogger, text feature vectors containing themes such as "island vacation" and "cultural heritage" can be extracted from 30 of the text descriptions, and visual feature vectors reflecting visual elements such as "natural scenery" and "architectural details" can be extracted from 25 landscape photos and 15 aerial videos.

[0030] In S22, the computer device establishes a unified semantic embedding space and projects text and visual feature vectors into this space through linear transformation. Then, a cross-modal attention mechanism is deployed to calculate the correlation weights between text and visual features, achieving dynamic alignment of features from different modalities. For example, a Transformer encoder is used to calculate the attention weights for text and visual features. Multi-head attention layers capture semantic associations between modalities from different representation subspaces, and a feedforward neural network is used to non-linearly fuse the weighted feature representations. The final fused feature vector retains the most significant semantic associations between the text and visual modalities while reducing information redundancy between modalities.

[0031] For example, when processing the text and image content of a food blogger, the text feature vector describing "chocolate lava cake" is mapped to a unified semantic space along with the visual feature vector of the corresponding dessert image. A cross-modal attention mechanism identifies a strong correlation between "lava" in the text and "flowing chocolate sauce" in the image, assigning this feature pair a higher weight while weakening background features irrelevant to the topic. The final result is a 512-dimensional fused feature vector, where the dimension related to the dessert's texture and ingredient description has a weight of 0.88, accurately reflecting the core semantics of the text and image content.

[0032] In step S23, the computer device performs dimensionality reduction on the high-dimensional features fused across modalities. Principal component analysis (PCA) can be used to transform correlated high-dimensional features into linearly uncorrelated low-dimensional features through orthogonal transformation. Dimensionality reduction involves first calculating the covariance matrix of the fused features, solving for its eigenvalues ​​and eigenvectors, and then selecting a specified number of principal components in descending order of eigenvalues. This ultimately generates a low-dimensional content feature vector that retains the main information of the original features. This dimensionality reduction step significantly improves the computational efficiency of subsequent models and avoids the curse of dimensionality while maintaining the feature expressiveness.

[0033] In S24, the posting time distribution represents the temporal regularity of content posted by key opinion leaders, including but not limited to time-dimensional characteristics such as the concentration of posting periods, date regularity, and stability of update frequency. Geographic location information represents spatial location data associated with user content, including the geographic coordinates of the content's origin, administrative division information, and location feature tags. Topic tags represent metadata identifiers used to label topic categories in user content, including the set of tags added by the creator and their frequency, relevance, and semantic network structure.

[0034] Computer equipment analyzes metadata in user content datasets to extract discriminative auxiliary features. For posting time distribution data, time series analysis is used to identify the temporal regularity and active period characteristics of content posting. For geographic location information, geocoding technology is used to convert it into standardized regional identifiers, and corresponding socio-economic characteristics are supplemented by combining regional attribute databases. Regarding topic tag data, semantic clustering analysis is used to identify the structured features of core topic areas and content tag systems. All extracted auxiliary features are standardized and then used for feature combination and dimensionality regularization to generate auxiliary feature vectors.

[0035] In S25, the computer device performs a feature vector concatenation operation, horizontally connecting the dimensionality-reduced fused feature vector with the auxiliary feature vector extracted from metadata, keeping the original dimensions of each vector unchanged. The auxiliary feature vectors are then appended to the end of the fused feature vector in a predetermined order, forming an expanded composite feature vector. The concatenated feature vector is then standardized to eliminate the influence of differences in feature scales, ensuring the stability of subsequent model processing.

[0036] S3: Obtain the time-series dataset of fan interaction behavior of the candidate KOLs, and calculate the fan authenticity index based on the time-series dataset of fan interaction behavior.

[0037] The fan interaction behavior time-series dataset represents a complete collection of data recording fan interactions with key opinion leaders in chronological order. This dataset includes timestamp sequences of interactive events such as likes, comments, and shares, along with the corresponding text information and basic user account attributes.

[0038] The Fan Authenticity Index is a quantitative evaluation indicator calculated by analyzing time-series data of fan interaction behavior. Based on the difference between the natural human and mechanical characteristics of fan behavior, this index comprehensively evaluates multiple dimensions, including the regularity of interaction time distribution, the originality of content, and the authenticity of account attributes, resulting in a continuous value between 0 and 1. A value closer to 1 indicates higher fan authenticity, while a value closer to 0 indicates the presence of abnormal interaction patterns.

[0039] Computer devices can use time-series pattern analysis algorithms to identify clusters of abnormal behaviors. By calculating the concentration of interaction time distribution and the frequency of content repetition, a fan authenticity assessment model can be established. The fan authenticity assessment model comprehensively considers multiple dimensions such as behavioral synchronicity, content originality, and consistency of account attributes, and ultimately outputs quantitative fan authenticity assessment indicators.

[0040] In one possible embodiment, S3: Calculating the fan authenticity index based on the fan interaction behavior time-series dataset includes: S31: In the time series dataset of fan interaction behavior, extract the time series features of interaction behavior for each fan account. The time series features of interaction behavior include: interaction time distribution, interaction frequency change pattern and interaction content similarity.

[0041] Specifically, the time-series characteristics of each fan account's interaction behavior refer to a set of quantitative features extracted from the historical interaction behavior data of a single fan account and organized chronologically. Interaction time distribution indicates the distribution pattern of fan account interactions such as liking, commenting, or forwarding across different time periods. Statistical analysis methods reveal the central tendency and dispersion characteristics of interaction behavior in hourly, daily, or weekly dimensions, used to identify behavioral patterns that violate normal human circadian rhythms. Interaction frequency variation indicates the dynamic changes in the number of interactions by a fan account per unit of time. Time-series analysis techniques capture the periodic fluctuations, trend stability, and anomalous abrupt changes in interaction behavior, used to detect mechanical interaction behavior at a preset frequency. Interaction content similarity indicates the degree of semantic similarity between comments posted by different fan accounts. This feature calculates the semantic embedding vector distance of comment texts using natural language processing techniques, used to identify batch-generated templated comment content.

[0042] S32: Construct a fan interaction relationship graph based on the time series characteristics of the interactive behavior, where nodes in the fan interaction relationship graph represent fan accounts and edges represent interaction relationships.

[0043] Specifically, the fan interaction relationship graph represents a fan community relationship model in graph structure. This graph abstracts each fan account as a node, the interaction relationships between accounts as edges, and the interaction strength is characterized by the weight of the edges. The graph construction process comprehensively considers multi-dimensional characteristics such as interaction frequency, diversity of interaction types, and temporal co-occurrence, forming a network topology that can visually represent the structural characteristics and propagation paths of the fan group.

[0044] Interactive behavior refers to the recordable actions that occur between fan accounts and key opinion leader content, including but not limited to liking, commenting, forwarding, and saving. These actions are recorded as timestamped serialized data through social platform interfaces.

[0045] Computer devices construct a fan interaction relationship graph based on extracted time-series features of interactive behaviors, abstracting each fan account as an independent node in the graph and establishing connections between accounts based on their interaction relationships. The weight of the edges is determined by the interaction frequency, the diversity of interaction types, and the time synchronization. A graph computing engine is used to analyze the graph structure, identifying highly connected subgraph structures by calculating graph theory metrics such as node degree centrality and clustering coefficients.

[0046] S33: Cluster the nodes in the interaction relationship graph to identify account clusters with similar abnormal behavior characteristics.

[0047] Specifically, the computer equipment employs a density-based spatial clustering algorithm to divide nodes in the interaction graph into communities. The clustering process includes: first, calculating the density distribution of nodes in the feature space, and then iteratively optimizing the grouping of tightly connected nodes into the same cluster. Taking into account multi-dimensional indicators such as connection strength, synchronicity of interactive behaviors, and similarity of content features, graph embedding techniques are used to project the nodes into a low-dimensional space before performing cluster analysis. The resulting clusters exhibit highly consistent interaction patterns within each account cluster, while different clusters show significant differences in behavioral characteristics.

[0048] S34: Calculate abnormal behavior indicators for each account cluster, including: interaction time synchronization rate, content duplication rate, and account attribute consistency. Specifically, the interaction time synchronization rate indicates the degree of consistency in the timing of interaction behaviors among accounts within a specific fan cluster. It is calculated by statistically analyzing the proportion of interaction behaviors generated by accounts within the cluster within the same time period, and is used to detect abnormal behaviors with mechanical synchronization characteristics.

[0049] Content duplication indicates the degree of semantic and word repetition in the comments posted by a fan group. It is calculated using natural language processing technology to identify templated comment content.

[0050] Account attribute consistency indicates the degree of similarity in the basic attributes of accounts within a fan cluster. This metric analyzes the distribution patterns of attributes such as account registration time, profile completeness, and nickname rules to detect abnormal groups of accounts registered in bulk.

[0051] For example, the interaction time synchronization rate can be obtained by calculating the proportion of accounts within a cluster that interact simultaneously within a specific time window. If a cluster has 100 accounts, and 95 accounts consistently interact within the same minute across 10 monitoring periods, then the synchronization rate is 95%.

[0052] Content duplication can be calculated using the cosine similarity of text embedding vectors. If, after vectorization, 50 comments from a certain cluster have an average similarity of 0.9, and 45 comments contain identical core phrases, then the content duplication rate is determined to be 90%.

[0053] Account attribute consistency can be calculated using information entropy to determine the dispersion of attribute distribution. If 80% of accounts in a cluster are registered in the same week, 90% use the same email domain, and 85% have data completeness differences of less than 5%, then the attribute consistency index reaches 0.85.

[0054] S35: The fan authenticity index is obtained by weighting each account cluster based on the abnormal behavior indicators.

[0055] Specifically, the computer equipment standardizes three core indicators—interaction time synchronization rate, content duplication, and account attribute consistency—eliminating differences in measurement units. Each indicator is weighted according to its impact on fan authenticity, with content duplication given the highest weight because it directly reflects the characteristics of mechanized operations, followed by interaction time synchronization rate, and account attribute consistency, as an auxiliary indicator, receiving the lowest weight. Anomaly comprehensive scores for each account cluster are calculated through linear weighted fusion, and these scores are then mapped to a numerical range of 0 to 1 to ultimately obtain a quantitative index representing the overall fan authenticity.

[0056] For example, three abnormal account clusters were detected for a certain blogger. Cluster A had an interaction time synchronization rate of 0.92, content duplication rate of 0.88, and account attribute consistency of 0.95. Cluster B had corresponding indicators of 0.85, 0.91, and 0.82. Cluster C had indicators of 0.78, 0.83, and 0.90. Using a weighted calculation model, with a weight of 0.5 for content duplication rate, 0.3 for interaction time synchronization rate, and 0.2 for account attribute consistency, the blogger's fan authenticity index was finally calculated to be 0.36. This value reflects that the account has a large proportion of abnormal fans.

[0057] S4: Obtain the introductory data of the target brand, and generate a brand attribute vector based on the introductory data.

[0058] This process employs natural language processing (NLP) technology to perform deep semantic analysis on textual data. A pre-trained language model is used to extract multi-dimensional semantic features, including core brand value, product feature positioning, and target customer profiles. Then, a feature weighting layer is used to rank the extracted semantic features by importance, highlighting the brand's most distinctive attributes. Finally, dimensionality reduction is used to generate a vector representing the brand's comprehensive attributes.

[0059] In one possible embodiment, S4: obtaining the introductory data of the target brand and generating a brand attribute vector based on the introductory data includes: S41: Obtain introductory data from multiple sources, including the target brand's official website, product manuals, and marketing materials.

[0060] Specifically, the computer equipment retrieves raw introductory data from multiple official information channels of the target brand. First, it accesses the brand's official website's public pages to extract text and images containing brand stories, corporate philosophies, and product introductions. Then, it analyzes the technical parameters and functional descriptions in product manuals to obtain structured product feature data. Simultaneously, it collects marketing materials released by the brand, including advertising copy, promotional videos, and social media content, forming a complete set of brand information. Data from different sources undergoes format standardization and content deduplication to ensure the comprehensiveness and accuracy of the dataset.

[0061] S42: Preprocess the introductory data to obtain the text to be processed.

[0062] Specifically, the computer equipment cleans the collected introductory data, removing HTML tags, special characters, and irrelevant formatting symbols, while retaining the core text content. Text normalization processing is then performed, including standardizing character encoding formats, converting full-width and half-width characters, and standardizing numerical expressions. Next, text restructuring and paragraph segmentation are carried out, dividing continuous text into text units with independent semantics based on semantic integrity, while removing duplicate paragraphs and irrelevant advertising text, ultimately yielding the text to be processed.

[0063] S43: The language model using the attention mechanism processes the text to be processed and outputs a brand attribute vector with fixed dimensions.

[0064] Specifically, the computer device converts the text to be processed into a sequence of word embeddings, and performs contextual semantic modeling through a multi-layer Transformer encoder. During the encoding process, a self-attention mechanism dynamically calculates the importance weights of different semantic units, highlighting the core concepts and key attributes in the brand description. Finally, the Transformer encoder aggregates the hidden states at all positions through a pooling layer, generating a fixed-dimensional brand attribute vector with rich semantic information.

[0065] S5: Retrieve cultural rule data corresponding to the target cross-border region from the preset regional cultural rule base, and generate a cultural rule vector based on the cultural rule data.

[0066] The regional cultural rule base represents a structured database storing the cultural characteristics of different geographical regions. This rule base integrates multi-dimensional characteristics such as religious customs, festival traditions, aesthetic preferences, social etiquette, language habits, and cultural taboos of various regions, and forms a cultural rule system that can be recognized and processed by computers through standardized coding. Each cultural rule in the rule base is associated with a specific geographical region identifier and includes metadata attributes such as rule type, scope of application, constraint strength, and update time.

[0067] In this application, the method for constructing a regional cultural rule base includes the following steps: 1. Obtain original cultural characteristic data by crawling publicly available information from government cultural departments, academic research literature, localized marketing cases, and regional tags from social media platforms.

[0068] 2. Employ natural language processing techniques to extract cultural rule elements from unstructured text, establishing a unified cultural feature classification system, including fifteen main categories and forty-three subcategories such as religious norms, festival customs, color symbols, and numerical taboos. For example, use named entity recognition (NER) and relation extraction techniques to extract cultural rules from text.

[0069] 3. Label each cultural rule with the applicable administrative region code and language region label, and have it reviewed by local cultural experts to ensure accuracy. For example, rules for the Middle East should include the specific applicable country code and differentiate between Arabic-speaking and Persian-speaking regions.

[0070] 4. Establish a rule validity verification process and regularly update cultural rules through social media trend analysis and news event monitoring to ensure that the rule base is in sync with the current cultural environment.

[0071] For example, when constructing a cultural rules database for Southeast Asia, announcements from religious authorities in countries such as Malaysia and Indonesia were collected, regulations governing commercial activities during important local festivals were analyzed, and color preference survey data were integrated. After structuring, several standardized rules were formed, including "Regulations for Commercial Promotions During Ramadan," "Green Symbolizes Islamic Culture," and "Taboos Regarding Handing Items with the Left Hand," with each rule associated with a specific regional code.

[0072] Computer equipment accesses a regional cultural rule base based on the identification information of the target cross-border region. By parsing the regional code and language tags, it extracts the corresponding cultural norm dataset from the rule base. This dataset covers multi-dimensional cultural elements such as local religious customs, festival traditions, aesthetic preferences, and social etiquette. Then, semantic encoding technology is used to convert the textual cultural rules into numerical representations. By using feature weighting, it highlights cultural taboos and preferences that have a significant impact on marketing activities, ultimately generating a standardized cultural rule vector.

[0073] S6: Calculate the compatibility between the content feature vector, the brand attribute vector, and the cultural rule vector to obtain a comprehensive fit vector.

[0074] Among them, compatibility represents the degree of matching and coordination between content feature vectors, brand attribute vectors, and cultural rule vectors in a unified semantic space. This indicator comprehensively evaluates the degree of matching between key opinion leader content characteristics, brand core values, and regional cultural characteristics by quantitatively calculating the semantic association strength, value alignment, and cultural norm conformity among the three vectors.

[0075] The computer system spatially aligns content feature vectors, brand attribute vectors, and cultural rule vectors to ensure they are within the same dimensional and metric system. A multilayer perceptron model is used to calculate the semantic fit between content and brand, the compliance degree between content and culture, and the value consistency between brand and culture. An attention mechanism is incorporated into the calculation process to dynamically adjust the weight allocation of different compatibility dimensions, emphasizing key factors that significantly impact marketing effectiveness. Finally, vector synthesis technology is used to fuse the multidimensional compatibility scores into a single comprehensive fit vector.

[0076] In one embodiment of this application, S6: Calculating the compatibility between the content feature vector, the brand attribute vector, and the cultural rule vector to obtain a comprehensive fit vector includes: S61: Align the content feature vector, the brand attribute vector, and the cultural rule vector in the vector space, and map the three to the same dimension in the same semantic space through linear transformation; S62: In the aligned vector space, calculate the semantic similarity between the content feature vector and the brand attribute vector, the compliance score between the content feature vector and the cultural rule vector, and the consistency measure between the brand attribute vector and the cultural rule vector, respectively. S63: After weighting the semantic similarity, the compliance score, and the consistency measure, output a comprehensive fit vector.

[0077] In S61, the original dimensional distribution and numerical range of three vectors are analyzed. Then, a linear transformation matrix is ​​designed to project feature vectors from different sources onto a shared semantic space of the same dimension. This transformation process preserves the core semantic features of each vector while eliminating structural inconsistencies caused by differences in feature extraction methods. The projected vectors are comparable within a unified space, and their relative positions directly reflect the intrinsic semantic relationships.

[0078] For example, a beauty blogger's content feature vector originally had 512 dimensions, a brand attribute vector of 256 dimensions, and a cultural rule vector of 128 dimensions. These were uniformly mapped to a shared 256-dimensional semantic space using three independent linear transformation matrices. Within this space, the cosine distance between the blogger's "ingredient analysis" feature and the brand's "technological ingredients" attribute increased from 0.35 before projection to 0.78, and the semantic association strength between the "halal certification" requirement and the brand's "vegetarianism" claim increased from 0.42 to 0.81, significantly enhancing the accuracy of subsequent compatibility calculations.

[0079] In S62, semantic similarity represents the strength of the semantic association between content feature vectors and brand attribute vectors in a unified vector space. This metric is obtained by calculating the cosine similarity between two vectors; a higher value indicates a greater semantic alignment between the content created by key opinion leaders and the brand's core values.

[0080] The compliance score represents the degree of conformity between the content feature vector and the cultural rule vector. This indicator is calculated by measuring the distance in the vector space and comprehensively considering the strength of cultural rule constraints, and is used to assess whether the created content conforms to the cultural norms and taboos of the target region.

[0081] The consistency metric represents the degree of compatibility between the brand attribute vector and the cultural rule vector. This metric is calculated through vector projection analysis and by combining the matching relationship between the brand value proposition and cultural characteristics, reflecting the level of adaptation between the brand's core values ​​and the regional cultural environment.

[0082] In S63, weight coefficients for each indicator are set according to the business needs of the target market, with the sum of the weight coefficients for all indicators being 1. The specific values ​​can be determined based on actual needs. Compliance scores are given higher weight in regions with high cultural sensitivity, while semantic similarity plays a major role in scenarios with strict brand tone requirements. A weighted summation algorithm is used to calculate a comprehensive fit scalar value, which is then concatenated with the original feature vector to generate a comprehensive fit vector that includes both quantitative scores and retains semantic features.

[0083] For example, in a case study of a beauty brand in the Southeast Asian market, the compliance score was weighted at 0.5, the semantic similarity at 0.3, and the consistency measure at 0.2. The calculated semantic similarity from a certain blogger was 0.88, the compliance score at 0.92, and the consistency measure at 0.85. After weighted fusion, the overall fit was 0.89.

[0084] S7: The content feature vector, the brand attribute vector, and the comprehensive adaptability vector are fused to obtain a comprehensive feature vector.

[0085] The computer equipment performs unified verification on the dimensions of the three vectors to ensure their compatibility in terms of numerical range and distribution. Then, vector concatenation technology is used to connect the three vectors end-to-end in a preset order to form a high-dimensional composite feature representation. This composite feature is then standardized to eliminate dimensional differences between different feature subspaces, ultimately generating a comprehensive feature vector that includes content characteristics, brand attributes, and cross-cultural adaptation information.

[0086] For example, by concatenating a tech blogger's 512-dimensional content feature vector, an electronics brand's 256-dimensional brand attribute vector, and a 128-dimensional comprehensive fit vector, an 896-dimensional comprehensive feature vector is generated. The first 512 dimensions of this vector fully retain the blogger's characteristics in product review expertise, the middle 256 dimensions carry the brand's attributes in innovation and reliability, and the last 128 dimensions record the degree of fit between this combination and the target market's cultural environment.

[0087] S8: Input the comprehensive feature vector into the pre-trained Transformer prediction model to obtain the conversion rate prediction value.

[0088] The computer system inputs the fused feature vector into a pre-trained Transformer prediction model. The model uses a multi-layered self-attention mechanism to deeply mine the internal correlations of the input vector, capturing the non-linear impact of different feature dimensions on conversion rates. The model preserves feature sequence information through positional encoding, followed by feature reconstruction through multiple Transformer blocks, each containing a multi-head self-attention layer and a feedforward neural network layer. Finally, a regression output layer maps the high-dimensional features to specific conversion rate predictions, which represent an estimate of the probability of a transaction being completed between the key opinion leader and the brand under a given market environment.

[0089] In this application, the training process of the Transformer prediction model includes the following steps: A training sample set is constructed by collecting historical collaboration activity data. Each sample contains a comprehensive feature vector of key opinion leaders and their corresponding actual conversion rate label. The training set covers collaboration cases from different regions, industries, and time periods to ensure the diversity and representativeness of the samples. The training samples are randomly shuffled and proportionally divided into training, validation, and test sets.

[0090] A Transformer prediction network based on an Encoder architecture is constructed, configured with 12 attention heads and 6 encoder layers. Each attention head is responsible for capturing feature interactions in different dimensions, and the feedforward neural network layer uses the GeLU activation function. The output layer is a fully connected structure, and the output value is mapped to a predicted conversion rate between 0 and 1 using the Sigmoid function.

[0091] The training process uses the mean squared error loss function to calculate the difference between the predicted and actual conversion rates, and iteratively updates the model parameters using a stochastic gradient descent algorithm. The optimizer sets an initial learning rate of 0.001 and dynamically adjusts the learning rate using a cosine annealing strategy. A gradient clipping mechanism is introduced to prevent gradient explosion during training.

[0092] Training data is input into the model in batches, and the model performance is evaluated using a validation set after each training epoch. If the validation set loss function fails to improve for several consecutive epochs, an early stopping mechanism is automatically triggered to terminate training. After training is complete, a test set is used for final model evaluation to ensure the model has good generalization ability.

[0093] Cross-validation was used to evaluate the stability and consistency of the model across different subsets of data, and the model parameters that performed best on the validation set were ultimately selected. Before deployment, interpretability analysis was performed on the model to verify the rationality of the feature weight allocation and ensure that the prediction results conformed to business logic.

[0094] S9: Select a cooperative KOL from the multiple candidate KOLs based on the conversion rate prediction value.

[0095] The process involves sorting the predicted conversion rates of all candidate KOLs in descending order to generate a candidate list. The system then implements a multi-round screening mechanism, comprehensively considering the confidence interval of the predicted conversion rates, cooperation costs, and risk aversion requirements, to determine the final KOLs for collaboration from the top-ranked candidates.

[0096] Please see Figure 3 As shown, in one embodiment, an apparatus is provided for screening KOLs (Key Opinion Leaders) for cross-border e-commerce, the apparatus comprising: Module 301 is used to identify multiple candidate KOLs located in the target cross-border region; The acquisition module 302 is used to acquire the user content dataset published by the candidate KOL on the account, and to generate a content feature vector based on the user content dataset; the user content dataset includes: multiple user content data and metadata of each user content data, and the user content data is text data, video data or image data; The acquisition module 302 is also used to acquire the time series dataset of fan interaction behavior of the candidate KOL, and to calculate the fan authenticity index based on the time series dataset of fan interaction behavior. The acquisition module 302 is also used to acquire the introductory data of the target brand and generate a brand attribute vector based on the introductory data; The retrieval module 303 is used to retrieve cultural rule data corresponding to the target cross-border region from a preset regional cultural rule base, and to generate a cultural rule vector based on the cultural rule data. Calculation module 304 is used to calculate the compatibility between the content feature vector, the brand attribute vector and the cultural rule vector to obtain a comprehensive fit vector; The fusion module 305 is used to fuse the content feature vector, the brand attribute vector, and the comprehensive adaptability vector to obtain a comprehensive feature vector; Prediction module 306 is used to input the comprehensive feature vector into a pre-trained Transformer prediction model to obtain a conversion rate prediction value; Selection module 307 is used to select a cooperative KOL from the multiple candidate KOLs based on the conversion rate prediction value.

[0097] In one possible embodiment, determining multiple candidate KOLs located in the target cross-border region includes: The top N KOLs on the target social media platform in the target cross-border region that meet the screening criteria are selected as multiple candidate KOLs, where N is an integer greater than 1. The screening criteria include account verification status, number of followers threshold, and content update frequency threshold.

[0098] In one possible embodiment, obtaining the user content dataset published by the candidate KOL on their account, and generating content feature vectors based on the user content dataset, includes: Text features are obtained by extracting features from the text data in the user content dataset, and visual feature vectors are obtained by extracting features from the image data and video data in the user content dataset. The text features and visual features are mapped to a unified semantic space, and a cross-modal attention mechanism is used to align and weightedly fuse various features to generate fused features. The fusion features are then subjected to dimensionality reduction to obtain the fusion feature vector. Auxiliary features are obtained by extracting features from the metadata in the user content dataset. The metadata includes the distribution of posting time, geographic location information and topic tags. An auxiliary feature vector is generated based on the auxiliary features. The fused feature vector and the auxiliary feature vector are concatenated to form the final content feature vector.

[0099] In one possible embodiment, calculating the fan authenticity index based on the time-series dataset of fan interaction behavior includes: In the time series dataset of fan interaction behavior, the time series features of interaction behavior for each fan account are extracted. The time series features of interaction behavior include: interaction time distribution, interaction frequency variation pattern and interaction content similarity. A fan interaction relationship graph is constructed based on the time series characteristics of the interactive behavior. In the fan interaction relationship graph, the nodes represent fan accounts and the edges represent interaction relationships. Cluster the nodes in the interaction relationship graph to identify account clusters with similar abnormal behavior characteristics; For each account cluster, calculate abnormal behavior indicators, including: interaction time synchronization rate, content duplication, and account attribute consistency. The fan authenticity index is obtained by weighting each account cluster based on the aforementioned abnormal behavior indicators.

[0100] In one possible embodiment, obtaining the introductory data of the target brand and generating a brand attribute vector based on the introductory data includes: Information was obtained from multiple sources, including the target brand's official website, product brochures, and marketing materials. The introductory data is preprocessed to obtain the text to be processed; The language model using the attention mechanism processes the text to be processed and outputs a fixed-dimensional brand attribute vector.

[0101] In one possible embodiment, the step of calling cultural rule data corresponding to the target cross-border region from a preset regional cultural rule base, and generating a cultural rule vector based on the cultural rule data, includes: Obtain the identification information of the target cross-border region; The corresponding structured cultural rule data is obtained by searching the regional cultural rule database based on the identification information. The cultural rule data includes: religious custom norms, festival traditions, aesthetic preferences, and social etiquette guidelines. Multimodal features are obtained by extracting features from the cultural rule data; The multimodal features are converted into cultural rule vectors.

[0102] In one possible embodiment, the step of calculating the compatibility between the content feature vector, the brand attribute vector, and the cultural rule vector to obtain a comprehensive fit vector includes: The content feature vector, the brand attribute vector, and the cultural rule vector are aligned in vector space, and linear transformation is used to map the three to the same dimension in the same semantic space. In the aligned vector space, the semantic similarity between the content feature vector and the brand attribute vector, the compliance score between the content feature vector and the cultural rule vector, and the consistency measure between the brand attribute vector and the cultural rule vector are calculated respectively. After weighting the semantic similarity, compliance score, and consistency metric, a comprehensive fit vector is output.

[0103] It should be noted that the device 3 provided in the above embodiments, when executing the method for screening cooperative KOLs for cross-border e-commerce, is only illustrated by the division of the above functional modules. In actual 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 above functions. In addition, the device for screening cooperative KOLs for cross-border e-commerce provided in the above embodiments and the method embodiments for screening cooperative KOLs for cross-border e-commerce belong to the same concept, and the implementation process is detailed in the method embodiments, which will not be repeated here.

[0104] The sequence numbers of the embodiments in this application are for descriptive purposes only and do not represent the superiority or inferiority of the embodiments.

[0105] In one embodiment, a computer device is provided, which may be a client, and its internal structure diagram may be as follows: Figure 4As shown, the computer device includes a processor, memory, network interface, display screen, and input devices connected via a system bus. The processor provides computing and control capabilities. The memory includes non-volatile storage media and internal memory. The non-volatile storage media stores the operating system and computer programs. The internal memory provides an environment for the operation of the operating system and computer programs stored in the non-volatile storage media. The network interface is used to communicate with an external server via a network connection. When executed by the processor, the computer program implements client-side functions or steps for a method of screening cooperative KOLs for cross-border e-commerce.

[0106] In one embodiment, a computer device is provided, including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor, when executing the computer program, performs the following steps: Identify multiple potential KOLs located in the target cross-border region; Obtain the user content dataset published on the account of the candidate KOL, and generate content feature vectors based on the user content dataset; the user content dataset includes: multiple user content data and metadata of each user content data, and the user content data is text data, video data or image data; Obtain the time-series dataset of fan interaction behavior of the candidate KOLs, and calculate the fan authenticity index based on the time-series dataset of fan interaction behavior; Obtain introductory data for the target brand, and generate a brand attribute vector based on the introductory data; The system retrieves cultural rule data corresponding to the target cross-border region from a pre-defined regional cultural rule base, and generates a cultural rule vector based on the cultural rule data. Calculate the compatibility between the content feature vector, the brand attribute vector, and the cultural rule vector to obtain a comprehensive fit vector; The content feature vector, the brand attribute vector, and the comprehensive adaptability vector are fused to obtain the comprehensive feature vector; The comprehensive feature vector is input into the pre-trained Transformer prediction model to obtain the conversion rate prediction value; Based on the conversion rate prediction, select one KOL from the multiple candidate KOLs to collaborate with.

[0107] This application establishes an objective and quantitative evaluation system through systematic data processing and intelligent analysis, transforming the selection of key opinion leaders from subjective experience-based judgment to data-driven decision-making. This significantly improves the accuracy and reliability of the selection results while enhancing operational efficiency.

[0108] In one embodiment, a computer-readable storage medium is provided that stores a computer program, which, when executed by a processor, performs the following steps: Identify multiple potential KOLs located in the target cross-border region; Obtain the user content dataset published on the account of the candidate KOL, and generate content feature vectors based on the user content dataset; the user content dataset includes: multiple user content data and metadata of each user content data, and the user content data is text data, video data or image data; Obtain the time-series dataset of fan interaction behavior of the candidate KOLs, and calculate the fan authenticity index based on the time-series dataset of fan interaction behavior; Obtain introductory data for the target brand, and generate a brand attribute vector based on the introductory data; The system retrieves cultural rule data corresponding to the target cross-border region from a pre-defined regional cultural rule base, and generates a cultural rule vector based on the cultural rule data. Calculate the compatibility between the content feature vector, the brand attribute vector, and the cultural rule vector to obtain a comprehensive fit vector; The content feature vector, the brand attribute vector, and the comprehensive adaptability vector are fused to obtain the comprehensive feature vector; The comprehensive feature vector is input into the pre-trained Transformer prediction model to obtain the conversion rate prediction value; Based on the conversion rate prediction, select one KOL from the multiple candidate KOLs to collaborate with.

[0109] This application establishes an objective and quantitative evaluation system through systematic data processing and intelligent analysis, transforming the selection of key opinion leaders from subjective experience-based judgment to data-driven decision-making. This significantly improves the accuracy and reliability of the selection results while enhancing operational efficiency.

[0110] It should be noted that the functions or steps that can be implemented by the computer-readable storage medium or computer device described above can be referred to the relevant descriptions on the server side and client side in the foregoing method embodiments. To avoid repetition, they will not be described one by one here.

[0111] 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 non-volatile computer-readable storage medium. When executed, the computer program can include the processes of the embodiments of the above methods. Any references to memory, storage, databases, or other media used in the embodiments provided in this application can include non-volatile and / or volatile memory. Non-volatile memory may include read-only memory (ROM), programmable ROM (PROM), electrically programmable ROM (EPROM), electrically erasable programmable ROM (EEPROM), or flash memory. Volatile memory may include random access memory (RAM) or external cache memory. By way of illustration and not limitation, RAM is available in a variety of forms, such as static RAM (SRAM), dynamic RAM (DRAM), synchronous DRAM (SDRAM), dual data rate SDRAM (DDRSDRAM), enhanced SDRAM (ESDRAM), synchronous link DRAM (SLDRAM), RAMbus direct RAM (RDRAM), direct memory bus dynamic RAM (DRDRAM), and memory bus dynamic RAM (RDRAM), etc.

[0112] Those skilled in the art will clearly understand that, for the sake of convenience and brevity, the above-described division of functional units and modules is used as an example. In practical applications, the above functions can be assigned to different functional units and modules as needed, that is, the internal structure of the device can be divided into different functional units or modules to complete all or part of the functions described above.

[0113] The above-described embodiments are only used to illustrate the technical solutions of the present invention, and are not intended to limit it. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features. Such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present invention, and should all be included within the protection scope of the present invention.

Claims

1. A method for selecting KOLs for cross-border e-commerce, characterized in that, The method includes: Identify multiple potential KOLs located in the target cross-border region; Obtain the user content dataset published on the account of the candidate KOL, and generate content feature vectors based on the user content dataset; the user content dataset includes: multiple user content data and metadata of each user content data, and the user content data is text data, video data or image data; Obtain the time-series dataset of fan interaction behavior of the candidate KOLs, and calculate the fan authenticity index based on the time-series dataset of fan interaction behavior; Obtain introductory data for the target brand, and generate a brand attribute vector based on the introductory data; The system retrieves cultural rule data corresponding to the target cross-border region from a pre-defined regional cultural rule base, and generates a cultural rule vector based on the cultural rule data. Calculate the compatibility between the content feature vector, the brand attribute vector, and the cultural rule vector to obtain a comprehensive fit vector; The content feature vector, the brand attribute vector, and the comprehensive adaptability vector are fused to obtain the comprehensive feature vector; The comprehensive feature vector is input into the pre-trained Transformer prediction model to obtain the conversion rate prediction value; Based on the conversion rate prediction, select one KOL from the multiple candidate KOLs to collaborate with.

2. The method for selecting KOLs for cross-border e-commerce as described in claim 1, characterized in that, The determination of multiple candidate KOLs located in the target cross-border region includes: The top N KOLs on the target social media platform in the target cross-border region that meet the screening criteria are selected as multiple candidate KOLs, where N is an integer greater than 1. The screening criteria include account verification status, number of followers threshold, and content update frequency threshold.

3. The method for selecting KOLs for cross-border e-commerce as described in claim 1, characterized in that, The step of obtaining the user content dataset published on the accounts of the candidate KOLs and generating content feature vectors based on the user content dataset includes: Text features are obtained by extracting features from the text data in the user content dataset, and visual feature vectors are obtained by extracting features from the image data and video data in the user content dataset. The text features and visual features are mapped to a unified semantic space, and a cross-modal attention mechanism is used to align and weightedly fuse various features to generate fused features. The fusion features are then subjected to dimensionality reduction to obtain the fusion feature vector. Auxiliary features are obtained by extracting features from the metadata in the user content dataset. The metadata includes the distribution of posting time, geographic location information and topic tags. An auxiliary feature vector is generated based on the auxiliary features. The fused feature vector and the auxiliary feature vector are concatenated to form the final content feature vector.

4. The method for selecting KOLs for cross-border e-commerce as described in claim 1, characterized in that, The calculation of the fan authenticity index based on the time-series dataset of fan interaction behavior includes: In the time series dataset of fan interaction behavior, the time series features of interaction behavior for each fan account are extracted. The time series features of interaction behavior include: interaction time distribution, interaction frequency variation pattern and interaction content similarity. A fan interaction relationship graph is constructed based on the time series characteristics of the interactive behavior. In the fan interaction relationship graph, the nodes represent fan accounts and the edges represent interaction relationships. Cluster the nodes in the interaction relationship graph to identify account clusters with similar abnormal behavior characteristics; For each account cluster, calculate abnormal behavior indicators, including: interaction time synchronization rate, content duplication, and account attribute consistency. The fan authenticity index is obtained by weighting each account cluster based on the aforementioned abnormal behavior indicators.

5. The method for selecting KOLs for cross-border e-commerce as described in claim 1, characterized in that, The steps of obtaining introductory data for the target brand and generating a brand attribute vector based on the introductory data include: Information was obtained from multiple sources, including the target brand's official website, product brochures, and marketing materials. The introductory data is preprocessed to obtain the text to be processed; The language model using the attention mechanism processes the text to be processed and outputs a fixed-dimensional brand attribute vector.

6. The method for selecting KOLs for cross-border e-commerce as described in claim 1, characterized in that, The step of retrieving cultural rule data corresponding to the target cross-border region from a preset regional cultural rule base, and generating a cultural rule vector based on the cultural rule data, includes: Obtain the identification information of the target cross-border region; The corresponding structured cultural rule data is obtained by searching the regional cultural rule database based on the identification information. The cultural rule data includes: religious custom norms, festival traditions, aesthetic preferences, and social etiquette guidelines. Multimodal features are obtained by extracting features from the cultural rule data; The multimodal features are converted into cultural rule vectors.

7. The method for selecting KOLs for cross-border e-commerce as described in claim 1, characterized in that, The calculation of the compatibility between the content feature vector, the brand attribute vector, and the cultural rule vector to obtain a comprehensive fit vector includes: The content feature vector, the brand attribute vector, and the cultural rule vector are aligned in vector space, and linear transformation is used to map the three to the same dimension in the same semantic space. In the aligned vector space, the semantic similarity between the content feature vector and the brand attribute vector, the compliance score between the content feature vector and the cultural rule vector, and the consistency measure between the brand attribute vector and the cultural rule vector are calculated respectively. After weighting the semantic similarity, compliance score, and consistency metric, a comprehensive fit vector is output.

8. A device for screening KOLs for cross-border e-commerce, characterized in that, The device includes: The identification module is used to identify multiple candidate KOLs located in the target cross-border region; The acquisition module is used to acquire the user content dataset published by the candidate KOL on their account, and to generate content feature vectors based on the user content dataset; the user content dataset includes: multiple user content data and metadata of each user content data, and the user content data is text data, video data or image data; The acquisition module is also used to acquire the time-series dataset of fan interaction behavior of the candidate KOL, and to calculate the fan authenticity index based on the time-series dataset of fan interaction behavior. The acquisition module is also used to acquire introductory data of the target brand and generate a brand attribute vector based on the introductory data; The retrieval module is used to retrieve cultural rule data corresponding to the target cross-border region from a preset regional cultural rule base, and to generate a cultural rule vector based on the cultural rule data. The calculation module is used to calculate the compatibility between the content feature vector, the brand attribute vector, and the cultural rule vector to obtain a comprehensive fit vector; The fusion module is used to fuse the content feature vector, the brand attribute vector, and the comprehensive adaptability vector to obtain a comprehensive feature vector; The prediction module is used to input the comprehensive feature vector into the pre-trained Transformer prediction model to obtain the conversion rate prediction value; The selection module is used to select a cooperative KOL from the multiple candidate KOLs based on the conversion rate prediction value.

9. A computer device, characterized in that, The computer device includes a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the computer program to implement the steps of the method for screening cooperative KOLs for cross-border e-commerce as described in any one of claims 1 to 7.

10. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores a computer program that, when executed by a processor, implements the steps of the method for screening cooperative KOLs for cross-border e-commerce as described in any one of claims 1 to 7.

Images