Cross-border e-commerce commodity recommendation system and method based on multi-source data fusion

Abstract

The invention relates to the technical field of data processing, and discloses a cross-border e-commerce commodity recommendation system and method based on multi-source data fusion. The system comprises an acquisition module for performing data acquisition to obtain a cross-border unified data warehouse and federal learning cooperation data; the classification module performs text classification processing to obtain a user preference analysis result and an interpretable attention mark; a quantization module carries out quantization processing to obtain a cross-border selection feature matrix; the fusion module carries out weighted fusion processing to obtain a basic comprehensive score and a weight convergence detection result; the evaluation module performs risk evaluation processing to obtain risk probability distribution and cross-border compliance evaluation results; and the sorting module carries out real-time processing through lightweight preprocessing and a flow-type calculation pipeline to obtain a comprehensive score sorting list of cross-border selected products. The problem that a traditional cross-border e-commerce product selection method cannot effectively integrate multi-source heterogeneous data and cannot analyze user feedback text semantic information is solved.

Description

Technical Field

[0001] The present application relates to the field of data processing technology, and in particular to a cross-border e-commerce product recommendation system and method based on multi-source data fusion. Background Art

[0002] With the acceleration of globalization and the rapid development of internet technology, cross-border e-commerce has become a significant growth driver for international trade. Cross-border e-commerce platforms offer consumers a rich and diverse selection of products while also providing sellers with convenient channels for expanding into overseas markets. Traditional cross-border e-commerce product selection methods rely primarily on manual judgment, simple market research, and basic statistical analysis. Sellers make product selection decisions by observing competitor product performance, analyzing historical sales data, and consulting market reports. Existing product selection analysis techniques typically utilize a single data source, focusing primarily on basic product attributes such as price and sales volume. Some solutions incorporate simple user behavior data for analysis.

[0003] However, traditional product selection methods have significant technical limitations. First, existing technologies have difficulty effectively integrating multi-source heterogeneous data and are unable to fully utilize multi-dimensional information resources such as internal data of cross-border e-commerce platforms, external market data, and social media data. They also lack federated learning collaboration mechanisms and differential privacy protection technologies, resulting in a lack of comprehensiveness and accuracy in product selection decisions. Second, traditional methods are technologically backward in user feedback text analysis. They lack multilingual BERT variants and cross-language semantic alignment technologies, and are unable to deeply analyze the complex sentiment tendencies and multi-level preference information in cross-border user reviews. They lack interpretable attention mechanisms and dynamic sub-classifier expansion capabilities, and are only capable of simple sentiment polarity judgments, making it difficult to capture users' detailed preference expressions for different product attributes. In addition, the existing product selection scoring mechanism uses a fixed-weight linear weighting method, lacks a reinforcement learning framework and Q-learning optimization algorithm, and is unable to dynamically adjust the scoring strategy based on changes in the market environment and historical decision-making results. It also lacks a Pareto optimal solution search and weight convergence detection mechanism, resulting in insufficient adaptive capabilities.

[0004] How to build an effective multi-source heterogeneous data collection and fusion mechanism, break through the limitations of traditional single data sources, and achieve distributed collector deployment and federated learning collaborative data generation has become a core technical issue that needs to be addressed urgently. How to design advanced user feedback text intelligent parsing technology that integrates multilingual processing capabilities and explainable attention annotation functions to deeply mine the hierarchical semantic information and fine-grained sentiment preferences in user comments, surpassing the shallow text analysis capabilities of existing technologies. At the same time, how to establish a dynamic weighted adaptive scoring optimization mechanism based on reinforcement learning, adjust scoring weights in real time based on market changes and decision feedback, and overcome the lack of adaptability of fixed weight methods; how to accurately quantify cross-border compliance risks through real-time event data fusion and Monte Carlo risk simulation, and achieve real-time sorting processing through lightweight preprocessing and streaming computing pipelines are all problems that traditional technical solutions cannot effectively solve. Summary of the Invention

[0005] This application provides a cross-border e-commerce product recommendation system and method based on multi-source data fusion, which solves the technical problems that traditional cross-border e-commerce product selection methods are unable to effectively integrate multi-source heterogeneous data, deeply analyze user feedback text semantic information, dynamically and adaptively adjust scoring weights, and accurately quantify cross-border trade risks.

[0006] In a first aspect, the present application provides a cross-border e-commerce product recommendation system based on multi-source data fusion, the cross-border e-commerce product recommendation system based on multi-source data fusion comprising: The acquisition module is used to collect and process multi-source heterogeneous data from cross-border e-commerce platform data, user behavior data, and external market data through distributed collectors to obtain cross-border raw data sets, clean and standardize the cross-border raw data sets, and obtain cross-border unified data warehouses and federated learning collaboration data; A classification module is used to perform hierarchical text classification processing based on the user feedback text data in the cross-border unified data warehouse through an improved BERT model to obtain user preference analysis results and explainable attention annotations; A quantification module is used to quantify the cross-border product selection features using a multi-dimensional feature extraction algorithm based on the cross-border unified data warehouse and the user preference analysis results to obtain a cross-border product selection feature matrix; A fusion module is used to perform weighted fusion processing on the product selection scores based on the cross-border product selection feature matrix through a dynamic weight adaptive algorithm to obtain a basic comprehensive score and weight convergence detection result; An assessment module, configured to perform risk assessment processing based on the basic comprehensive score through real-time event data fusion and Monte Carlo risk simulation to obtain risk probability distribution and cross-border compliance assessment results; The ranking module is used to perform real-time processing based on the federated learning collaborative data and the explainable attention annotation through lightweight preprocessing and streaming computing pipeline to obtain a comprehensive score ranking list of cross-border product selections.

[0007] In a second aspect, the present application provides a cross-border e-commerce product recommendation method based on multi-source data fusion, the cross-border e-commerce product recommendation method based on multi-source data fusion comprising: Through distributed collectors, multi-source heterogeneous data collection and processing is performed on cross-border e-commerce platform data, user behavior data, and external market data to obtain cross-border original data sets, which are then cleaned and standardized to obtain a cross-border unified data warehouse and federated learning collaboration data; Based on the user feedback text data in the cross-border unified data warehouse, hierarchical text classification processing is performed through an improved BERT model to obtain user preference analysis results and explainable attention annotations; Based on the cross-border unified data warehouse and the user preference analysis results, cross-border product selection features are quantified using a multi-dimensional feature extraction algorithm to obtain a cross-border product selection feature matrix; Based on the cross-border product selection feature matrix, the product selection scores are weighted and fused using a dynamic weight adaptive algorithm to obtain a basic comprehensive score and weight convergence detection result; Based on the basic comprehensive score, risk assessment is performed through real-time event data fusion and Monte Carlo risk simulation to obtain risk probability distribution and cross-border compliance assessment results; Based on the federated learning collaborative data and the explainable attention annotation, real-time processing is performed through lightweight preprocessing and streaming computing pipelines to obtain a comprehensive score ranking list of cross-border product selections.

[0008] In a third aspect, a cross-border e-commerce product recommendation device based on multi-source data fusion is provided, comprising: a memory and at least one processor, wherein the memory stores instructions; the at least one processor calls the instructions in the memory so that the cross-border e-commerce product recommendation device based on multi-source data fusion executes the above-mentioned cross-border e-commerce product recommendation method based on multi-source data fusion.

[0009] In a fourth aspect, a computer-readable storage medium is provided, wherein instructions are stored in the computer-readable storage medium, which, when executed on a computer, enables the computer to execute the above-mentioned cross-border e-commerce product recommendation method based on multi-source data fusion.

[0010] In the technical solution provided in this application, multi-source heterogeneous data collection and processing is carried out on cross-border e-commerce platform data, user behavior data, and external market data through distributed collectors, which effectively solves the technical problem of the single data source of traditional product selection methods. The distributed collection architecture can simultaneously obtain product information, user interaction data, and external market trend data from multiple cross-border e-commerce platforms, breaking through the limitation of existing technologies that rely only on data from a single platform. The generation of federated learning collaborative data protects data privacy through differential privacy algorithms while realizing multi-party collaboration. The construction of a cross-border unified data warehouse provides a comprehensive and secure data foundation for subsequent analysis. The improved BERT model's hierarchical text classification processing technology has played a key role in the field of cross-border e-commerce user feedback analysis. The multilingual BERT variant and cross-language semantic alignment layer can effectively process multilingual comments in cross-border scenarios. The explainable attention labeling function improves the model's transparency and credibility through keyword weight calculation. The online learning mechanism enables dynamically updated sub-classifiers to automatically adapt to the emergence of new product categories. Compared with traditional flat text classification methods, this hierarchical processing method can more accurately identify users' preference expressions for different product attributes. Especially in the cross-border e-commerce environment where user feedback content is diverse and expression is complex, the improved BERT model's multilingual semantic understanding ability is significantly better than traditional text analysis technology.

[0011] The multidimensional feature extraction algorithm targets the special needs of cross-border e-commerce product selection and performs feature quantification from multiple dimensions such as user comprehensive features, product comprehensive features, market trend features, and competition features. The cross-border product selection feature matrix integrates the key factors that affect the sales performance of cross-border products. Compared with the existing technology that only considers the basic attributes of products, the multidimensional feature extraction method of this application can comprehensively characterize the competitive position and development potential of products in the cross-border market.

[0012] The Q-learning optimization strategy achieves intelligent weight updates through Markov decision process modeling. The Pareto optimal solution search algorithm balances conflicting objectives such as profit maximization and risk minimization. The weight convergence detection mechanism effectively prevents overfitting, overcoming the technical limitations of traditional fixed-weight methods that often fail to adapt to market fluctuations. Real-time event data fusion and Monte Carlo risk simulation technology access geopolitical news and customs policy change events through APIs. The NLP risk event level extraction function dynamically adjusts the risk index. The risk probability distribution generated by Monte Carlo simulation replaces the traditional standard deviation calculation method. Cross-border compliance assessment results integrate the target country's product access regulations as an independent risk dimension, significantly improving the accuracy of risk quantification. Lightweight preprocessing and streaming computing pipeline technology enable real-time user profile updates through the Apache Flink framework. Combined with federated learning collaborative data and interpretable attention annotation, the generated cross-border product selection ranking list fully considers the complexity and dynamic nature of the cross-border trade environment, providing cross-border e-commerce sellers with scientific and accurate product selection guidance and significantly reducing the subjectivity and blindness of product selection decisions. BRIEF DESCRIPTION OF THE DRAWINGS

[0013] In order to more clearly illustrate the technical solutions of the embodiments of the present invention, the following briefly introduces the drawings required for use in the description of the embodiments. Obviously, the drawings described below are some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without creative work.

[0014] Figure 1 This is a schematic diagram of an embodiment of a cross-border e-commerce product recommendation system based on multi-source data fusion in an embodiment of the present application; Figure 2 This is a schematic diagram of an embodiment of a cross-border e-commerce product recommendation method based on multi-source data fusion in an embodiment of the present application; Figure 3 It is a schematic block diagram of the structure of a cross-border e-commerce product recommendation device based on multi-source data fusion in an embodiment of the present invention. DETAILED DESCRIPTION

[0015] The embodiments of the present application provide a cross-border e-commerce product recommendation system and method based on multi-source data fusion. The terms "first", "second", "third", "fourth", etc. (if any) in the specification and claims of this application and the above-mentioned drawings are used to distinguish similar objects and are not necessarily used to describe a specific order or sequence. It should be understood that the data used in this way can be interchangeable where appropriate, so that the embodiments described herein can be implemented in an order other than that illustrated or described herein. In addition, the terms "including" or "having" and any variations thereof are intended to cover non-exclusive inclusions, for example, a process, method, system, product or device that includes a series of steps or units is not necessarily limited to those steps or units clearly listed, but may include other steps or units that are not clearly listed or inherent to these processes, methods, products or devices.

[0016] For ease of understanding, the specific process of the embodiment of the present application is described below. Figure 1 In the embodiments of the present application, an embodiment of a cross-border e-commerce product recommendation system based on multi-source data fusion includes: The acquisition module 101 is used to collect and process multi-source heterogeneous data from cross-border e-commerce platform data, user behavior data, and external market data through distributed collectors to obtain cross-border raw data sets, clean and standardize the cross-border raw data sets, and obtain a cross-border unified data warehouse and federated learning collaboration data; Classification module 102, which is used to perform hierarchical text classification processing based on the user feedback text data in the cross-border unified data warehouse through the improved BERT model to obtain user preference analysis results and explainable attention annotations; Quantification module 103 is used to quantify cross-border product selection features using a multi-dimensional feature extraction algorithm based on the cross-border unified data warehouse and user preference analysis results to obtain a cross-border product selection feature matrix; Fusion module 104 is used to perform weighted fusion processing on the product selection scores based on the cross-border product selection feature matrix through a dynamic weight adaptive algorithm to obtain a basic comprehensive score and weight convergence detection result; Assessment module 105, configured to perform risk assessment based on the basic comprehensive score through real-time event data fusion and Monte Carlo risk simulation to obtain risk probability distribution and cross-border compliance assessment results; The ranking module 106 is used to perform real-time processing based on federated learning collaborative data and explainable attention annotations through lightweight preprocessing and streaming computing pipelines to obtain a comprehensive score ranking list for cross-border product selection.

[0017] It is understandable that the execution subject of this application can be a cross-border e-commerce product recommendation system based on multi-source data fusion, or a terminal or server, which is not limited here. The embodiment of this application is described by taking the server as the execution subject as an example.

[0018] Specifically, the collection module 101 implements multi-source heterogeneous data collection through distributed collectors. First, crawlers are deployed for each cross-border e-commerce platform based on a preset collection node configuration table. The configuration table contains parameters such as the platform domain name, data interface address, and crawling frequency. Each collection node independently captures user browsing history, search keywords, purchase history, and other behavioral data according to the configuration. Simultaneously, external market data is obtained through search trend interfaces and social media interfaces. Internal behavioral data and external market data are merged and processed to form a cross-border raw data set after data source identification. A rule engine and anomaly detection algorithm are then used to clean the raw data. The rule engine identifies duplicate records using predefined rules, while the anomaly detection algorithm uses the isolation forest method to detect outliers in numerical fields. After cleaning, product classification standards are unified and currency units are converted. Finally, a star data model is used to construct product fact tables and user dimension tables to form a cross-border unified data warehouse. A differential privacy algorithm adds noise to the standardized data set to generate a privacy-preserving data set. Federated learning nodes generate federated learning collaborative data through model gradient sharing. The classification module 102 receives user feedback text data from a cross-border unified data warehouse and performs cross-language word vector conversion using a multilingual BERT variant. The multilingual BERT variant is a pre-trained model that supports multiple languages ​​and can map vocabulary in different languages ​​to a unified vector space. The cross-language semantic alignment algorithm maps word embedding vectors in different languages ​​to the same semantic space through linear transformation, ensuring that semantically similar words have similar vector representations in different languages. The position encoding algorithm uses sine and cosine functions to generate a unique code for each word position, using sine values ​​for even dimensions and cosine values ​​for odd dimensions, which are added element-by-element to the word embedding matrix to form a position encoding text matrix. The multi-head self-attention mechanism calculates the association weights between words through 12 parallel attention heads, and each attention head independently calculates the attention scores of the query matrix, key matrix, and value matrix. The interpretable attention layer normalizes the attention weights into a probability distribution through the softmax function, marking keywords that affect the classification results. The parent classifier divides the text into six categories: product quality evaluation, price sensitivity expression, logistics experience feedback, after-sales service evaluation, purchase intention inquiry, and competitive product comparison opinions. The sub-class classifier subdivides each parent category. The online learning mechanism updates the classifier parameters through the gradient descent algorithm, and automatically expands the sub-classifier label system when a new product category is detected. The quantification module 103 performs clustering analysis on user geographical distribution data and purchasing power data based on the K-means clustering algorithm. The K-means algorithm classifies similar users into the same cluster through iterative optimization, and calculates the center point of each cluster as the user feature vector. User preference weight quantization converts emotional tendencies into numerical weights, with positive evaluations given high weights and negative evaluations given low weights. The basic features of the product are converted into binary vectors through unique hot encoding, and the price and sales data are mapped to the zero to one range using maximum and minimum value normalization.The seasonal fluctuation analysis algorithm uses time series decomposition to separate historical sales into trend, seasonal, and random components, and identifies cyclical patterns through Fourier transforms. Market trend features are analyzed using a moving average algorithm to calculate the sliding average of keyword search popularity. Competition density is calculated using the Herfindahl Index to measure market concentration. Vector concatenation and fusion connects features from different dimensions column by column to form a unified matrix representation.

[0019] Fusion module 104 uses an LSTM neural network as a time series prediction model. LSTM controls information flow through forget gates, input gates, and output gates, inputting historical sales data in time series to predict future sales trends. A collaborative filtering algorithm constructs a user-product rating matrix, using matrix decomposition techniques to decompose the sparse matrix into a user latent factor matrix and a product latent factor matrix. User matching scores are calculated using cosine similarity. The reinforcement learning framework models weight adjustment as a Markov decision process, where the state space is composed of historical decision effects. The Q-learning algorithm updates the optimization weight strategy through a value function. The Pareto optimal solution search algorithm searches for a set of non-dominated solutions, which are solutions that are not strictly dominated by other solutions on all objectives. Weight convergence detection determines whether convergence has been achieved by calculating the weight change rate and comparing it with a preset threshold.

[0020] Assessment module 105 accesses geopolitical news and customs policy changes in real time through an API. An NLP algorithm uses named entity recognition and sentiment analysis to extract risk event levels. The exchange rate risk index is calculated using the standard deviation and coefficient of variation of historical exchange rate data. The logistics risk index is assessed based on the magnitude of cost fluctuations and the range of time-efficiency fluctuations. The competition risk index is quantified by market competition density and the degree of price competition. A Monte Carlo simulation algorithm uses a large number of random samples of risk variables as random variables and establishes a risk distribution model using a probability density function. Cross-border compliance assessments detect compliance deficiencies by matching them to the target country's commodity access regulations. A compliance risk quantification algorithm converts the severity of the deficiencies into a risk value.

[0021] Sorting module 106 feeds federated learning collaborative data into a lightweight preprocessing model for feature extraction. Lightweight models are neural network models with smaller parameters, making them suitable for resource-constrained environments. Apache Flink is a distributed stream processing framework that supports real-time data stream computation and uses window functions to group and aggregate data streams. The streaming computing pipeline organizes feature data into a continuous data stream, updating user profile information in real time. The sorting algorithm sorts products in descending order based on their matching scores, generating a final ranked list of cross-border product selections with a comprehensive score.

[0022] In a specific embodiment, the acquisition module 101 is configured to: Based on the preset collection node configuration table, a distributed crawler is deployed on the cross-border e-commerce platform to obtain the platform data collection node. The user's browsing trajectory, search keywords, and purchase records are input into the platform data collection node for behavior data crawling and processing to obtain the original user behavior data. Perform external data acquisition and processing on search trend data and social mention data through the search trend interface and social media interface to obtain external market raw data. Then, merge the user behavior raw data and external market raw data with data source identification to obtain a cross-border raw data set. Based on the rule engine and anomaly detection algorithm, duplicate data is removed and outliers are filtered out of the cross-border raw data set to obtain a cleaned data set. The cleaned data set is then subjected to unified commodity classification standards and currency unit conversion to obtain a standardized data set. Based on the star data model, the standardized data set is processed to construct product fact tables and user dimension tables to obtain a cross-border unified data warehouse; Based on the differential privacy algorithm, noise is added to the standardized dataset to obtain a privacy-preserving dataset. The privacy-preserving dataset is then shared through the federated learning node for model gradient sharing to obtain federated learning collaborative data.

[0023] Specifically, the collection module 101 implements distributed crawler deployment through a preset collection node configuration table. The preset collection node configuration table contains configuration parameters such as the domain name, API endpoint, access frequency, and request header of the target platform. The distributed crawler deployment process creates independent data collection programs on multiple servers according to the configuration table parameters. Each program is responsible for the data capture task of a specific platform, forming a platform data collection node. The behavioral data capture process takes user browsing trajectories, search keywords, and purchase records as input, extracts structured user interaction information by parsing the HTML page structure, calling the platform API, analyzing user session logs, etc., and outputs the original data of user behavior. The external data acquisition process obtains the changes in the search volume of keywords in different time periods through the search trend interface. The search trend interface returns time series data in JSON format, including fields such as keywords, search volume, and timestamp. The social media interface provides users with social interaction data such as product discussion content, forwarding volume, and likes on social platforms, and also returns structured information in JSON format. The data source identification merging process adds a source tag and timestamp to each data record, marking the original user behavior data as an "internal" source and the original external market data as an "external" source. It then performs correlation matching based on the product ID and time dimensions, combining the relevant data into a unified record to form a cross-border original data set.

[0024] The deduplication process identifies duplicate entries by calculating the similarity between data records. The similarity calculation is based on the text similarity and numerical difference of key fields such as product titles, prices, and merchant information. When the similarity exceeds the threshold, it is determined to be duplicate data and deleted. The anomaly detection algorithm uses the isolation forest method, which segments data points by constructing multiple random decision trees. Normal data points are clustered in the shallow layers of the tree and require fewer segmentations, while abnormal data points are distributed in the deep layers of the tree and require more segmentations. The anomaly score is calculated based on the average path length. Data points that exceed the threshold are identified as outliers and filtered. The unified processing of product classification standards establishes a cross-platform category mapping table, mapping the classification labels of different platforms to a unified standard classification system. The currency unit conversion processing converts various currencies into US dollars based on the real-time exchange rate.

[0025] The star-shaped data model construction and processing adopts a centralized data organization structure. The product fact table serves as the core to store metric information such as product ID, name, price, and sales volume. The user dimension table stores attribute information such as user ID, geographic location, and purchasing power level. A reference relationship is established between the fact table and the dimension table through foreign key associations, forming a cross-border unified data warehouse. The differential privacy algorithm protects individual privacy by adding carefully designed random noise to the original data. The noise intensity is controlled by the privacy budget parameter. The smaller the privacy budget, the greater the noise and the higher the degree of privacy protection, generating a privacy-protected dataset. The federated learning node is the computing unit in the distributed machine learning system. Model gradient sharing processing refers to each node uploading the gradient parameters to the central server for aggregation after training the model locally. The server calculates the global model parameters and distributes them to each node. Through multiple rounds of iteration, model collaborative optimization is achieved.

[0026] In one embodiment, the classification module 102 is configured to: The user feedback text data in the cross-border unified data warehouse is input into the multilingual BERT variant for cross-lingual word vector conversion to obtain a multilingual word embedding matrix. The multilingual word embedding matrix is ​​then mapped into a semantic space based on a cross-lingual semantic alignment algorithm to obtain an aligned word embedding matrix. According to the position encoding algorithm, the position information of the aligned word embedding matrix is ​​fused to obtain the position encoding text matrix, which is then input into the multi-head self-attention mechanism for semantic association calculation to obtain the semantic feature vector. Based on the explainable attention layer, the keyword weights of the semantic feature vectors are calculated to obtain the attention weight distribution. The keywords that affect the classification results are annotated according to the attention weight distribution to obtain explainable attention annotations. The semantic feature vector is input into the parent category classifier to perform classification processing on six parent categories: product quality evaluation, price sensitivity expression, logistics experience feedback, after-sales service evaluation, purchase intention inquiry, and competitive product comparison opinion. The parent category classification result is obtained. Based on the parent category classification result, the corresponding subcategory classifier is activated to perform sub-category identification processing to obtain the sub-category classification result. Based on the online learning mechanism, the sub-classifier is expanded for the new product category to obtain a dynamically updated sub-classifier. The parent category classification results and the sub-category classification results are subjected to hierarchical label fusion processing through the dynamically updated sub-classifier to obtain the user preference analysis results.

[0027] Specifically, the classification module 102 first inputs the user feedback text data in the cross-border unified data warehouse into a multilingual BERT variant for processing. The multilingual BERT variant is a pre-trained neural network model that supports multiple languages ​​and can convert texts in different languages ​​into a unified numerical vector representation. The cross-language word vector conversion process maps each word to a 768-dimensional dense vector through the embedding layer of the model to form a multilingual word embedding matrix. The number of rows in the matrix corresponds to the number of words in the text, and the number of columns is fixed at 768 dimensions. The cross-language semantic alignment algorithm maps the word embedding vectors of different languages ​​to the same semantic space through a linear transformation matrix to ensure that semantically similar words have similar vector representations in different languages. The semantic space mapping process calculates the product of the transformation matrix and the word embedding matrix to obtain the aligned word embedding matrix.

[0028] The position encoding algorithm generates a unique position vector for each word in the text. The algorithm uses sine and cosine functions to calculate the position encoding value, using sine for even dimensions and cosine for odd dimensions. Position information fusion involves element-by-element addition of the position encoding vector with the word vector at the corresponding position in the aligned word embedding matrix, resulting in a position-encoded text matrix containing the position information. The multi-head self-attention mechanism includes 12 parallel attention heads, each of which independently calculates the query matrix, key matrix, and value matrix. The semantic association calculation process first calculates the matrix product of the transposed query matrix and the key matrix, then divides it by a scaling factor for normalization. The softmax function converts the attention scores into a probability distribution. Finally, the attention weights are multiplied by the value matrix to obtain a weighted feature representation. The outputs of multiple attention heads are concatenated to form a semantic feature vector.

[0029] The interpretable attention layer further processes the semantic feature vectors. Keyword weight calculation uses a fully connected neural network to calculate each word's contribution to the final classification result, outputting an attention weight distribution that represents the importance of each word in the text to the classification decision. The annotation process uses this attention weight distribution to identify the words with the highest weights as keywords. These keywords and their weights are combined to form interpretable attention annotations, visually displaying the important words that influence the classification result.

[0030] The parent category classifier uses a fully connected neural network structure, taking a semantic feature vector as input. After processing it through linear transformation and activation functions, it outputs a probability distribution of six categories. The six parent category classification process determines the primary category to which the text belongs based on the principle of maximum probability. The subcategory classifier is activated based on the parent category classification results. Each parent category corresponds to a dedicated subcategory classifier. The subcategory identification process inputs the same semantic feature vector into the activated subcategory classifier to achieve a finer classification granularity. For example, the parent category of product quality evaluation can be subdivided into subcategories such as functionality evaluation, appearance evaluation, and durability evaluation.

[0031] The online learning mechanism continuously updates classifier parameters through an incremental learning algorithm. When the system detects the emergence of new product categories, the sub-classifier expansion process automatically adds new classification labels and adjusts the network structure. The classifier weights are retrained using the new sample data through a gradient descent algorithm, forming a dynamically updated sub-classifier. The hierarchical label fusion process uses the parent category classification result as the root node and the sub-category classification results as leaf nodes to construct a tree-like label structure. Each label includes a corresponding confidence score, and the final output is a user preference analysis result containing complete hierarchical information.

[0032] In a specific embodiment, the quantization module 103 is configured to: Based on the user geographical distribution data and purchasing power data in the cross-border unified data warehouse, user feature clustering analysis is performed to obtain user feature vectors. Based on the user preference analysis results, the user's preference weights for product attributes are quantified and calculated to obtain the user preference weight vector. The product categories, prices, and sales data in the cross-border unified data warehouse are standardized and coded to obtain the basic feature vectors of the products. The historical sales data is subjected to periodic feature extraction based on the seasonal fluctuation analysis algorithm to obtain the product time series feature vectors. Conduct time series analysis of changes in target market popularity based on exchange rate fluctuation data and consumer trend data to obtain a market trend feature vector. Quantify the competitive intensity of market share data for similar products using a competition density calculation algorithm to obtain a competition feature vector. Perform vector splicing and fusion processing on the user feature vector and the user preference weight vector to obtain the user comprehensive feature vector; perform vector splicing and fusion processing on the product basic feature vector and the product time series feature vector to obtain the product comprehensive feature vector; The user comprehensive feature vector, product comprehensive feature vector, market trend feature vector and competition feature vector are concatenated and fused to obtain the cross-border product selection feature matrix.

[0033] Specifically, the quantification module 103 performs user feature clustering analysis on user geographic distribution data and purchasing power data based on the K-means clustering algorithm. The K-means algorithm groups similar users into the same cluster through iterative calculations. The algorithm first randomly initializes K cluster centers, then calculates the Euclidean distance from each user to each cluster center, assigns the user to the closest cluster center, and then recalculates the location of the center point of each cluster. The iteration is repeated until the cluster center no longer changes significantly. User geographic distribution data includes geographical attributes such as the user's continent, economic development level, and consumer cultural background. Purchasing power data includes economic attributes such as average order amount, purchase frequency, and consumer category preferences. Cluster analysis groups users with similar geographic and purchasing power characteristics into the same group, and the center point vector of each cluster serves as the representative user feature vector of that group. The user preference weight quantification calculation process converts the emotional tendencies in the user preference analysis results into numerical weights. Positive emotional evaluations are assigned high weights of 0.8 to 1.0, neutral emotional evaluations are assigned medium weights of 0.4 to 0.6, and negative emotional evaluations are assigned low weights of 0.1 to 0.3. The weight calculation is based on the product operation of emotional intensity and mention frequency to obtain the user preference weight vector.

[0034] The standardized encoding of basic product features uses one-hot encoding to convert the text labels of product categories into binary vectors. Each category corresponds to a dimension in the vector, with the corresponding dimension value being 1 and the remaining dimensions being 0. Price data is mapped to a numerical range of 0 to 1 through maximum and minimum value normalization. The normalization formula is the current price minus the minimum price, divided by the price range. Sales data is logarithmically transformed to address long-tail distribution issues. The logarithmic transformation takes the natural logarithm of the original sales value to compress the numerical range. The seasonal fluctuation analysis algorithm uses a time series decomposition method to separate historical sales data into trend, seasonal, and random components. Periodic feature extraction uses Fourier transforms to identify periodic patterns in sales data. The Fourier transform converts the time domain signal into the frequency domain signal, extracting the amplitude and phase information of different frequency components to identify time patterns such as monthly, quarterly, and annual cycles. The intensity coefficient and phase offset of each periodic component form the product time series feature vector.

[0035] Time series analysis processes exchange rate fluctuation data and consumption trend data in chronological order. A moving average algorithm is used to calculate averages across different time windows to smooth short-term fluctuations. The moving average algorithm takes the arithmetic mean of values ​​from N consecutive time points as the representative value for that time period. The standard deviation of exchange rate fluctuation data measures exchange rate stability; a larger standard deviation indicates more significant fluctuations. The first-order difference of consumption trend data reflects the market's dynamic rate of change. A first-order difference is the difference between values ​​at adjacent time points; a positive value indicates an upward trend and a negative value indicates a downward trend. The competition density calculation algorithm counts the number of sellers of similar products and their respective market share. Market concentration is calculated using the Herfindahl index, which is the sum of the squares of each seller's market share. An index value close to 0 indicates intense market competition, while an index value close to 1 indicates a highly concentrated market. Competition intensity is quantified by calculating the intensity of competition based on the inverse of market concentration, forming a competition characteristic vector.

[0036] Vector concatenation and fusion processing concatenates the user feature vector and the user preference weight vector by column to form a user comprehensive feature vector. The concatenation operation sequentially arranges all elements of the two vectors to form a new, longer vector. Product basic feature vectors and product time series feature vectors are similarly concatenated by column to form a comprehensive product feature vector, ensuring that both the static attributes and dynamic time series information of the product are fully preserved. Matrix concatenation and fusion processing concatenates the user comprehensive feature vector, product comprehensive feature vector, market trend feature vector, and competition feature vector by column to form a unified feature matrix. The number of rows in the matrix equals the number of candidate products, and the number of columns equals the sum of the dimensions of all feature vectors. Each row represents a complete feature description of a product.

[0037] In a specific embodiment, the fusion module 104 is configured to: Based on the time series prediction model, the sales-related features in the cross-border product selection feature matrix are predicted and calculated to obtain a sales prediction score. The user preference features and product features are matched and calculated using the collaborative filtering algorithm to obtain a user matching score. The market trend features are input into the trend analysis model to calculate the trend fit and obtain a market trend score. Based on the reinforcement learning framework, the weight adjustment is modeled as a Markov decision process. The state space of historical product selection decision effects is constructed to obtain a decision state matrix. The weight update strategy is optimized using the Q-learning algorithm based on the decision state matrix to obtain a dynamic weight coefficient group. The sales forecast score, user matching score, and market trend score are used as objective function inputs. A multi-objective optimization process is performed based on the Pareto optimal solution search algorithm to obtain a non-dominated solution set. The non-dominated solution set is weighted and summed according to the dynamic weight coefficient group to obtain a basic comprehensive score. Perform convergence detection on the dynamic weight coefficient group based on the weight fluctuation threshold to obtain the weight change rate, compare the weight change rate with the preset convergence threshold to obtain the weight convergence detection result; The sales forecast score, user matching score and market trend score are finally weighted and fused through a dynamic weight adaptive algorithm to obtain a basic comprehensive score and weight convergence detection result.

[0038] Specifically, the fusion module 104 processes the sales-related features in the cross-border product selection feature matrix based on the LSTM time series prediction model. The LSTM neural network controls the flow of information through three gating mechanisms: the forget gate, the input gate, and the output gate. The forget gate determines which information is discarded from the cell state, the input gate controls the storage level of new information, and the output gate determines which parts of the cell state are used as output. The prediction calculation process inputs historical sales data into the model in time series, and uses a sliding window mechanism to use the sales of the past 30 days as the input sequence to predict the sales trend for the next 7 days. The LSTM learns time dependencies through the backpropagation algorithm, and the sales prediction score is calculated by calculating the ratio of the predicted sales to the historical average sales. The collaborative filtering algorithm constructs a user-product rating matrix, in which the rows of the matrix represent users and the columns represent products. The element values ​​represent the user's implicit rating of the product. The matching calculation process uses matrix decomposition technology to decompose the sparse rating matrix into a user potential factor matrix and a product potential factor matrix. The matrix decomposition minimizes the reconstruction error through the gradient descent algorithm, and the user matching score is obtained by calculating the cosine similarity between the user potential factor and the product potential factor. The trend analysis model uses the linear regression method to fit the time-varying pattern of market trend characteristics. The trend fit calculation process inputs characteristics such as market heat changes and consumer preference shifts into the model, and estimates the slope and intercept parameters of the trend line through the least squares method. The market trend score is calculated based on the positivity and significance level of the trend slope.

[0039] The reinforcement learning framework models weight adjustment as a Markov decision process. The state space is composed of factors such as the current weight configuration, historical decision-making results, and changes in the market environment. The state space construction process encodes these factors into a state vector. The effectiveness of historical product selection decisions is measured through indicators such as the error between actual and predicted sales, user satisfaction, and market response, forming a decision state matrix. The Q-learning algorithm is a model-free reinforcement learning method that optimizes decision strategies by learning state-action value functions. The weight update strategy optimization process uses the Bellman equation to update Q values. Q values ​​represent the expected long-term reward of taking a specific action under a specific state. The algorithm uses an exploration-exploitation balance mechanism to learn the optimal strategy while maintaining exploration of new strategies, resulting in a dynamic set of weight coefficients.

[0040] The Pareto optimal solution search algorithm seeks solutions that are not strictly dominated by other solutions on multiple objective functions. A non-dominated solution is one for which no other solution is superior or equal to it on all objectives and strictly superior on at least one objective. Multi-objective optimization uses a genetic algorithm or particle swarm algorithm to search the Pareto frontier. The algorithm maintains a set of solutions and generates new solutions through operations such as crossover and mutation. Non-dominated sorting stratifies solutions according to dominance relationships to obtain a non-dominated solution set. The weighted summation process multiplies the weights in the dynamic weight coefficient group by the sales forecast score, user matching score, and market trend score, and then sums them. The sum of the weights is normalized to 1 to ensure the rationality of the score, resulting in a basic comprehensive score.

[0041] The convergence detection process calculates the magnitude of change in the dynamic weight coefficient group within consecutive time steps. The weight change rate is calculated by dividing the difference between the current weight and the previous weight by the previous weight. The absolute value of the change rate reflects the severity of the weight adjustment. The comparison judgment process compares the weight change rate with a preset convergence threshold. When the change rate is less than the threshold, the weight is determined to have stabilized. The weight stability indicator records the convergence status, forming the weight convergence detection result.

[0042] The dynamic weight adaptive algorithm comprehensively considers factors such as the current market environment, historical decision-making effects, and weight convergence status. The final weighted fusion processing decides whether to continue adjusting the weight based on the weight convergence detection results. When the weight converges, stable weights are used for score calculation. When the weight has not converged, Q-learning is continued to be used to optimize the weight. The three scores are weighted and summed according to the final weight, and the basic comprehensive score and weight convergence detection results are output at the same time.

[0043] In a specific embodiment, the evaluation module 105 includes: Through the API interface, real-time data acquisition and processing of geopolitical news and customs policy change events are carried out to obtain the real-time event raw data. The real-time event raw data is then correlated and matched with the basic comprehensive score to obtain the event impact score matrix; Based on the NLP algorithm, the risk event level of the real-time event raw data is extracted and processed to obtain the risk event classification identification. According to the risk event classification identification, the event impact score matrix is ​​fused to obtain the fused event data. Exchange rate risk, logistics risk, and competition risk are input as random variables into the Monte Carlo simulation algorithm for random sampling to obtain a risk variable sample set. Based on the risk variable sample set, distribution modeling is performed through probability density function calculation to obtain the risk probability distribution. Conduct compliance testing on the integrated event data based on the target country's commodity access regulations to obtain compliance defect identification. Based on the compliance defect identification, conduct an assessment using a compliance risk quantification algorithm to obtain cross-border compliance assessment results. The fused event data is processed through Monte Carlo risk simulation for final risk assessment to obtain risk probability distribution and cross-border compliance assessment results.

[0044] Specifically, the evaluation module 105 obtains geopolitical news and customs policy change events in real time through the API interface. The API interface is an application programming interface that allows the system to access a standardized interface of external data sources. The real-time data acquisition process calls the news aggregation API and the government policy release API through HTTP requests to obtain structured information such as news headlines, release time, event type, and impact range in JSON format to form real-time event raw data. The association matching process matches real-time events with basic comprehensive scores based on dimensions such as product category, target market, and time window. The matching algorithm identifies the relevance between events and products through keyword extraction and semantic similarity calculation, quantifies the impact of related events into numerical values, and combines them with the basic scores of the corresponding products to form an event impact score matrix. The matrix rows represent products and the columns represent event types. The element values ​​represent the intensity of the impact of specific events on specific products.

[0045] The NLP algorithm uses named entity recognition and sentiment analysis techniques to process real-time event raw data. Named entity recognition uses a pre-trained BERT model to identify key entities in the text, such as country names, policy types, and product categories. Sentiment analysis uses a sentiment dictionary and deep learning model to determine the positive or negative impact of an event on the market. Risk event level extraction categorizes events into three levels: low, medium, and high risk, based on factors such as the scope of impact, duration, and severity. This grading standard is based on historical event data and an expert knowledge base, resulting in a risk event classification identifier. Data fusion weights the values ​​in the event impact score matrix based on the risk level. The impact coefficient for high-risk events is set to 1.5, for medium-risk events to 1.0, and for low-risk events to 0.5. This adjusted score matrix is ​​linearly combined with the basic product score to produce the fused event data.

[0046] The Monte Carlo simulation algorithm is a numerical calculation method based on random sampling. It treats exchange rate risk, logistics risk, and competition risk as random variables that follow specific probability distributions. Random sampling involves fitting probability distribution functions based on historical data for each risk variable. Exchange rate risk follows a normal distribution, logistics risk follows an exponential distribution, and competition risk follows a uniform distribution. A pseudo-random number generator is used to generate a large number of random samples that conform to each distribution, resulting in a sample set of risk variables. Probability density function calculation is based on kernel density estimation, smoothing sample points with a Gaussian kernel function. Distribution modeling represents the joint distribution of multiple risk variables as a multidimensional probability density function. The function value represents the probability density of a specific risk combination. Numerical integration is then used to calculate the cumulative probability of different risk levels, forming a risk probability distribution.

[0047] Compliance testing checks the integrated event data against the target country's product access regulations, which include safety standards, environmental requirements, and certification procedures. The detection algorithm uses rule matching and text analysis to identify the degree of compliance between product attributes and regulatory requirements. Compliance defect identification records product non-compliance with various regulatory requirements, including defect type, severity, and difficulty of correction. The compliance risk quantification algorithm converts defect information into a numerical risk score, assigning a high risk weight to severe defects and a low risk weight to minor defects. The algorithm considers the cost and time required to correct defects, calculating a comprehensive compliance risk value through a weighted summation to produce the cross-border compliance assessment results.

[0048] The final risk assessment process uses the fused event data as the input parameters of the Monte Carlo simulation. The simulation algorithm considers the impact of real-time events on various risks, reflects the event impact by adjusting the distribution parameters of risk variables, re-performs random sampling and probability calculations, generates an updated risk probability distribution, and outputs cross-border compliance assessment results that include the impact of events.

[0049] In one embodiment, the sorting module 106 is configured to: The federated learning collaboration data is input into a lightweight preprocessing model for feature extraction to obtain a collaboration feature vector. Based on interpretable attention annotation, the key product attributes are weighted to obtain attention-weighted features. Perform feature fusion processing based on the collaborative feature vector and the attention weighted feature to obtain a fused feature matrix, and input the fused feature matrix into the streaming computing pipeline for real-time data stream processing to obtain a real-time feature stream; Based on the Apache Flink framework, real-time feature streams are processed in a streaming manner to obtain dynamically updated user profiles. Product matching scores are then recalculated based on the dynamically updated user profiles to obtain real-time matching scores. Compare and analyze the real-time matching scores with historical score data to obtain the score change trend. Based on the score change trend, the products are prioritized using a sorting algorithm to obtain a preliminary ranking result. The final sorting optimization is performed based on the preliminary sorting results to obtain a comprehensive score ranking list for cross-border product selection.

[0050] Specifically, the sorting module 106 inputs the federated learning collaborative data into a lightweight preprocessing model for processing. The lightweight preprocessing model is a neural network model with a smaller parameter scale. It uses deep separable convolution and parameter quantization technology to reduce computational complexity and is suitable for operation in a resource-constrained environment. The feature extraction process maps high-dimensional data such as user behavior patterns, product attribute information, and market trend changes in the federated learning collaborative data into low-dimensional dense vector representations through the encoder part of the model. The encoder adopts a multi-layer perceptron structure. Each layer extracts features at different levels of abstraction through linear transformation and activation function, and finally outputs a collaborative feature vector of fixed dimension. The weight allocation process is based on the keyword weights marked in the interpretable attention annotation. The attention weight is used as an indicator of the importance of the product attributes. Different weight coefficients are assigned to the attributes of the product such as price, quality, brand, and function. The weight normalization ensures that the sum of all attribute weights is 1. The attention weighted feature is obtained by weighted average calculation.

[0051] Feature fusion processing fuses collaborative feature vectors and attention-weighted features at the element level using a weighted combination. The collaborative feature vector reflects the overall preference patterns of the user group, while the attention-weighted features highlight the key attributes that influence user decisions. The two are linearly combined according to preset weights to form a fused feature vector. The fused feature vectors of multiple products are arranged in rows to form a fused feature matrix. The streaming computing pipeline is a computational framework for processing continuous data streams. Real-time data stream processing converts the fused feature matrix into a chronological data stream. Each element in the data stream contains a timestamp and feature vector information. Streaming processing supports incremental updates and sliding window calculations to produce a real-time feature stream.

[0052] Apache Flink is a distributed stream processing framework that supports low-latency real-time data processing and complex event processing. Stream processing uses event-time semantics and a watermarking mechanism to handle out-of-order data. Dynamically updating user profiles aggregates users' recent behavior data through a sliding window, with the window size set to the past 24 hours. Aggregation operations include statistical functions such as sum, average, and maximum. These operations calculate profile features such as user preference strength for different product categories, changes in purchase frequency, and price sensitivity. Recalculation is performed based on the similarity between the updated user profile and product features. The similarity algorithm uses cosine similarity to measure the degree of match between the user preference vector and the product feature vector. A higher similarity value indicates that the product more closely matches the user's preferences, resulting in a real-time matching score.

[0053] The comparative analysis process compares the current real-time matching score with the score data within the historical time window. The historical score data is stored in a time series database. The comparison method includes calculating indicators such as score difference, rate of change, and trend slope. The score change trend is identified through time series analysis to identify trend patterns such as rising, falling, and stable. The trend detection algorithm uses moving average and linear regression methods to fit the change pattern of the score over time. The positive and negative trend slope and significance level reflect the dynamic changes in the popularity of the product. The sorting algorithm prioritizes the products according to the score change trend. The priority calculation comprehensively considers the current score value and the trend change amplitude. Products with an upward trend receive priority bonuses, while products with a downward trend receive lower priority. The sorting uses a quick sorting algorithm to sort by priority score from high to low to obtain a preliminary sorting result.

[0054] The final sorting optimization process makes fine adjustments to the preliminary sorting results. The optimization algorithm considers additional factors such as product inventory status, supply chain stability, and user diversity needs, and balances recommendation accuracy and diversity through multi-objective optimization. The optimization process uses a local search algorithm to make small adjustments in the sorting results to ensure that the recommendation list not only meets user preferences but also maintains a reasonable product distribution, and finally outputs a cross-border product selection comprehensive score ranking list.

[0055] The above describes the cross-border e-commerce product recommendation system based on multi-source data fusion in the embodiment of the present application. The following describes the cross-border e-commerce product recommendation method based on multi-source data fusion in the embodiment of the present application. Figure 2 In the embodiment of the present application, an embodiment of a cross-border e-commerce product recommendation method based on multi-source data fusion includes: S201. Collect and process multi-source heterogeneous data on cross-border e-commerce platform data, user behavior data, and external market data through distributed collectors to obtain a cross-border raw data set. Clean and standardize the cross-border raw data set to obtain a cross-border unified data warehouse and federated learning collaboration data. S202. Based on the user feedback text data in the cross-border unified data warehouse, perform hierarchical text classification processing using an improved BERT model to obtain user preference analysis results and explainable attention annotations; S203. Based on the cross-border unified data warehouse and user preference analysis results, quantify the cross-border product selection features using a multi-dimensional feature extraction algorithm to obtain a cross-border product selection feature matrix; S204. Based on the cross-border product selection feature matrix, the product selection scores are weighted and integrated using a dynamic weight adaptive algorithm to obtain a basic comprehensive score and weight convergence detection result; S205. Based on the basic comprehensive score, conduct risk assessment through real-time event data fusion and Monte Carlo risk simulation to obtain risk probability distribution and cross-border compliance assessment results; S206. Based on the federated learning collaborative data and explainable attention annotation, lightweight preprocessing and streaming computing pipelines are used for real-time processing to obtain a comprehensive score ranking list of cross-border product selections.

[0056] above Figure 2 The cross-border e-commerce product recommendation method based on multi-source data fusion in the embodiment of the present invention is described in detail from the perspective of modular functional entities. The cross-border e-commerce product recommendation device based on multi-source data fusion in the embodiment of the present invention is described in detail from the perspective of hardware processing.

[0057] Reference Figure 3 In the embodiment of the present invention, there is also provided a method in which the cross-border e-commerce product recommendation device based on multi-source data fusion can be a server, and its internal structure can be as follows: Figure 3 As shown. The cross-border e-commerce product recommendation device based on multi-source data fusion includes a processor, a memory, a display screen, an input device, a network interface and a database connected via a system bus. Among them, the computer-designed processor is used to provide computing and control capabilities. The memory of the cross-border e-commerce product recommendation device based on multi-source data fusion includes a non-volatile storage medium and an internal memory. The non-volatile storage medium stores an operating system, a computer program and a database. The internal memory provides an environment for the operation of the operating system and the computer program in the non-volatile storage medium. The database of the cross-border e-commerce product recommendation device based on multi-source data fusion is used to store the corresponding data in this embodiment. The network interface of the cross-border e-commerce product recommendation device based on multi-source data fusion is used to communicate with an external terminal through a network connection. When the computer program is executed by the processor, the above method is implemented.

[0058] Those skilled in the art will understand that Figure 3 The structure shown in is merely a block diagram of a portion of the structure related to the solution of the present invention, and does not constitute a limitation on the cross-border e-commerce product recommendation device based on multi-source data fusion to which the solution of the present invention is applied.

[0059] The present invention also provides a computer-readable storage medium, which may be a non-volatile computer-readable storage medium or a volatile computer-readable storage medium. The computer-readable storage medium stores instructions, which, when executed on a computer, enable the computer to execute the steps of the cross-border e-commerce product recommendation system based on multi-source data fusion.

[0060] Those skilled in the art will clearly understand that, for the convenience and brevity of description, the specific working processes of the above-described systems, systems and units can refer to the corresponding processes in the aforementioned method embodiments and will not be repeated here.

[0061] If the integrated unit is implemented in the form of a software functional unit and sold or used as an independent product, it can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of the present invention is essentially 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. The computer software product is stored in a storage medium and includes several instructions for enabling a cross-border e-commerce product recommendation device based on multi-source data fusion (which can be a personal computer, server, or network device, etc.) to execute all or part of the steps of the method described in each embodiment of the present invention. The aforementioned storage medium includes: U disk, mobile hard disk, read-only memory (ROM), random access memory (RAM), disk or optical disk, and other media that can store program code.

[0062] The above embodiments are only used to illustrate the technical solutions of the present invention, rather than to limit the same. Although the present invention has been described in detail with reference to the aforementioned embodiments, those skilled in the art should understand that they can still modify the technical solutions described in the aforementioned embodiments, or make equivalent replacements for some of the technical features therein. However, these modifications or replacements do not deviate the essence of the corresponding technical solutions from the spirit and scope of the technical solutions of the various embodiments of the present invention.

Claims

1. A cross-border e-commerce product recommendation system based on multi-source data fusion, characterized by: The system includes: The acquisition module is used to collect and process multi-source heterogeneous data from cross-border e-commerce platform data, user behavior data, and external market data through distributed collectors to obtain cross-border raw data sets, clean and standardize the cross-border raw data sets, and obtain cross-border unified data warehouses and federated learning collaboration data; A classification module is used to perform hierarchical text classification processing based on the user feedback text data in the cross-border unified data warehouse through an improved BERT model to obtain user preference analysis results and explainable attention annotations; A quantification module is used to quantify the cross-border product selection features using a multi-dimensional feature extraction algorithm based on the cross-border unified data warehouse and the user preference analysis results to obtain a cross-border product selection feature matrix; A fusion module is used to perform weighted fusion processing on the product selection scores based on the cross-border product selection feature matrix through a dynamic weight adaptive algorithm to obtain a basic comprehensive score and weight convergence detection result; An assessment module, configured to perform risk assessment processing based on the basic comprehensive score through real-time event data fusion and Monte Carlo risk simulation to obtain risk probability distribution and cross-border compliance assessment results; The ranking module is used to perform real-time processing based on the federated learning collaborative data and the explainable attention annotation through lightweight preprocessing and streaming computing pipeline to obtain a comprehensive score ranking list of cross-border product selections.

2. The cross-border e-commerce product recommendation system based on multi-source data fusion according to claim 1 is characterized in that: The acquisition module is used to: Based on the preset collection node configuration table, a distributed crawler is deployed on the cross-border e-commerce platform to obtain the platform data collection node, and the user's browsing trajectory, search keywords, and purchase history are input into the platform data collection node to perform behavior data crawling and processing to obtain the original user behavior data; Performing external data acquisition processing on the search trend data and social mention data according to the search trend interface and the social media interface to obtain external market raw data, and merging the user behavior raw data and the external market raw data with data source identification to obtain the cross-border raw data set; Based on the rule engine and anomaly detection algorithm, duplicate data is removed and outliers are filtered out of the cross-border original dataset to obtain a cleaned dataset, and commodity classification standards and currency unit conversion are performed on the cleaned dataset to obtain a standardized dataset; Constructing a product fact table and a user dimension table for the standardized data set according to a star data model to obtain the cross-border unified data warehouse; The standardized dataset is subjected to noise addition processing based on a differential privacy algorithm to obtain a privacy-preserving dataset, and the privacy-preserving dataset is subjected to model gradient sharing processing through a federated learning node to obtain the federated learning collaborative data.

3. The cross-border e-commerce product recommendation system based on multi-source data fusion according to claim 1 is characterized in that: The classification module is used to: Inputting the user feedback text data in the cross-border unified data warehouse into a multilingual BERT variant for cross-lingual word vector conversion processing to obtain a multilingual word embedding matrix, and performing semantic space mapping processing on the multilingual word embedding matrix based on a cross-lingual semantic alignment algorithm to obtain an aligned word embedding matrix; Performing position information fusion processing on the aligned word embedding matrix according to the position encoding algorithm to obtain a position encoding text matrix, and inputting the position encoding text matrix into the multi-head self-attention mechanism to perform semantic association calculation processing to obtain a semantic feature vector; Performing keyword weight calculation processing on the semantic feature vector based on the explainable attention layer to obtain an attention weight distribution, and annotating the keywords that affect the classification result according to the attention weight distribution to obtain the explainable attention annotation; Input the semantic feature vector into the parent category classifier to perform six parent category classification processes: product quality evaluation, price sensitivity expression, logistics experience feedback, after-sales service evaluation, purchase intention inquiry, and competitive product comparison opinion, to obtain a parent category classification result; activate the corresponding child category classifier based on the parent category classification result to perform sub-category identification processing, to obtain a sub-category classification result; Based on the online learning mechanism, the new product category is subjected to sub-classifier expansion processing to obtain a dynamically updated sub-classifier, and the parent category classification result and the sub-category classification result are subjected to hierarchical label fusion processing through the dynamically updated sub-classifier to obtain the user preference analysis result.

4. The cross-border e-commerce product recommendation system based on multi-source data fusion according to claim 1 is characterized in that: The quantization module is used to: Performing user feature clustering analysis based on the user geographical distribution data and purchasing power data in the cross-border unified data warehouse to obtain a user feature vector, and performing quantitative calculation of the user's preference weights for product attributes based on the user preference analysis results to obtain a user preference weight vector; Standardize and encode the commodity categories, prices, and sales data in the cross-border unified data warehouse to obtain commodity basic feature vectors. Perform periodic feature extraction on the historical sales data based on a seasonal fluctuation analysis algorithm to obtain commodity time series feature vectors. Conduct time series analysis of changes in target market popularity based on exchange rate fluctuation data and consumer trend data to obtain a market trend feature vector. Quantify the competitive intensity of market share data for similar products using a competition density calculation algorithm to obtain a competition feature vector. Performing vector splicing and fusing processing on the user feature vector and the user preference weight vector to obtain a user comprehensive feature vector, and performing vector splicing and fusing processing on the product basic feature vector and the product time series feature vector to obtain a product comprehensive feature vector; The user comprehensive feature vector, the product comprehensive feature vector, the market trend feature vector and the competition feature vector are subjected to matrix splicing and fusion processing to obtain the cross-border product selection feature matrix.

5. The cross-border e-commerce product recommendation system based on multi-source data fusion according to claim 1 is characterized in that: The fusion module is used to: Based on the time series prediction model, the sales-related features in the cross-border product selection feature matrix are predicted and calculated to obtain a sales prediction score. The user preference features and product features are matched and calculated based on the collaborative filtering algorithm to obtain a user matching score. The market trend features are input into the trend analysis model to calculate the trend fit and obtain a market trend score. Based on the reinforcement learning framework, the weight adjustment is modeled as a Markov decision process. The state space of the historical product selection decision effects is constructed to obtain a decision state matrix. The weight update strategy is optimized using the Q-learning algorithm based on the decision state matrix to obtain a dynamic weight coefficient group. Using the sales forecast score, the user matching score, and the market trend score as objective function inputs, performing multi-objective optimization processing based on a Pareto optimal solution search algorithm to obtain a non-dominated solution set, and performing weighted summation processing on the non-dominated solution set according to the dynamic weight coefficient group to obtain the basic comprehensive score; Performing convergence detection processing on the dynamic weight coefficient group based on the weight fluctuation threshold to obtain a weight change rate, and comparing the weight change rate with a preset convergence threshold to obtain the weight convergence detection result; The sales forecast score, the user matching score and the market trend score are finally weighted and fused by a dynamic weight adaptive algorithm to obtain the basic comprehensive score and the weight convergence detection result.

6. The cross-border e-commerce product recommendation system based on multi-source data fusion according to claim 1 is characterized in that: The evaluation module is used to: Real-time data acquisition and processing of geopolitical news and customs policy change events are performed through the API interface to obtain real-time event raw data, and the real-time event raw data is correlated and matched with the basic comprehensive score to obtain an event impact score matrix; Perform risk event level extraction processing on the real-time event raw data based on the NLP algorithm to obtain a risk event classification identifier, and perform data fusion processing on the event impact scoring matrix according to the risk event classification identifier to obtain fused event data; Inputting exchange rate risk, logistics risk, and competition risk as random variables into a Monte Carlo simulation algorithm for random sampling to obtain a risk variable sample set, and performing distribution modeling based on the risk variable sample set by calculating a probability density function to obtain the risk probability distribution; Performing compliance testing on the fused event data according to the commodity access regulations of the target country to obtain compliance defect identifications, and performing assessment processing using a compliance risk quantification algorithm based on the compliance defect identifications to obtain the cross-border compliance assessment results; The fused event data is subjected to final risk assessment processing through Monte Carlo risk simulation to obtain the risk probability distribution and the cross-border compliance assessment result.

7. The cross-border e-commerce product recommendation system based on multi-source data fusion according to claim 1 is characterized in that: The sorting module is used to: Inputting the federated learning collaborative data into a lightweight preprocessing model for feature extraction to obtain a collaborative feature vector, and performing weight assignment processing on key product attributes based on the explainable attention annotation to obtain an attention-weighted feature; Performing feature fusion processing on the collaborative feature vector and the attention weighted feature to obtain a fused feature matrix, and inputting the fused feature matrix into a streaming computing pipeline to perform real-time data stream processing to obtain a real-time feature stream; Performing streaming computing on the real-time feature stream based on the Apache Flink framework to obtain a dynamically updated user profile, and recalculating the product matching degree based on the dynamically updated user profile to obtain a real-time matching score; Comparing and analyzing the real-time matching score with historical score data to obtain a score change trend, and prioritizing the products using a sorting algorithm based on the score change trend to obtain a preliminary sorting result; A final sorting optimization process is performed based on the preliminary sorting results to obtain a comprehensive score sorting list of cross-border product selections.

8. A cross-border e-commerce product recommendation method based on multi-source data fusion, characterized in that: Based on the cross-border e-commerce product recommendation system based on multi-source data fusion according to any one of claims 1 to 7, the cross-border e-commerce product recommendation method based on multi-source data fusion includes: Through distributed collectors, multi-source heterogeneous data collection and processing is performed on cross-border e-commerce platform data, user behavior data, and external market data to obtain cross-border original data sets, which are then cleaned and standardized to obtain a cross-border unified data warehouse and federated learning collaboration data; Based on the user feedback text data in the cross-border unified data warehouse, hierarchical text classification processing is performed through an improved BERT model to obtain user preference analysis results and explainable attention annotations; Based on the cross-border unified data warehouse and the user preference analysis results, cross-border product selection features are quantified using a multi-dimensional feature extraction algorithm to obtain a cross-border product selection feature matrix; Based on the cross-border product selection feature matrix, the product selection scores are weighted and fused using a dynamic weight adaptive algorithm to obtain a basic comprehensive score and weight convergence detection result; Based on the basic comprehensive score, risk assessment is performed through real-time event data fusion and Monte Carlo risk simulation to obtain risk probability distribution and cross-border compliance assessment results; Based on the federated learning collaborative data and the explainable attention annotation, real-time processing is performed through lightweight preprocessing and streaming computing pipelines to obtain a comprehensive score ranking list of cross-border product selections.

9. A cross-border e-commerce product recommendation device based on multi-source data fusion, characterized in that: It includes a memory and a processor, the memory stores a computer program that can be run on the processor, and when the processor executes the computer program, it implements the cross-border e-commerce product recommendation method based on multi-source data fusion as described in claim 8.

10. A computer-readable storage medium having a computer program stored thereon, characterized in that: When the computer program is executed by a processor, the processor executes the cross-border e-commerce product recommendation method based on multi-source data fusion as described in claim 8.

Images