Business layout method and device based on multi-modal data fusion and dynamic optimization

Abstract

The invention provides a business layout method and device based on multi-modal data fusion and dynamic optimization. The method comprises the following steps: acquiring overall economic situation data, platform big disk trend data and single item trend data currently corresponding to a target category; according to the overall economic situation data currently corresponding to the target category, determining a comprehensive feature Rwl of the target category; determining a feature value Rbt of the target category according to platform big disk trend data currently corresponding to the target category; according to target single item trend data currently corresponding to the target category, determining a feature value Anw of the target single item; and according to the comprehensive feature Rw1, the feature value Rbt and the feature value Anw, predicting the global sales volume and sales price of the target category, and performing international business layout on the target category. According to the method, comprehensive capture of market dynamics is realized, and the prediction accuracy is improved, so that powerful data support is provided for scientific and accurate international business layout.

Description

Technical Field

[0001] This invention relates to the field of market trend forecasting and price analysis technology, specifically to a business layout method and apparatus based on multimodal data fusion and dynamic optimization. Background Technology

[0002] In the complex ecosystem of international business activities, market trend assessment for specific product categories, precise pricing, and reasonable inventory layout in various regions have become key elements for enterprises to enhance competitiveness and achieve sustainable development.

[0003] Traditional business strategy approaches based on multimodal data fusion and dynamic optimization have long relied on single data sources and simple statistical analysis methods, making it difficult to accurately capture complex and ever-changing market dynamics. Taking the stock market as an example, while it can reflect market expectations to some extent, it suffers from significant information inconsistencies and is highly susceptible to manipulation by large investors and other human factors, greatly diminishing the reference value of the data it provides. Furthermore, there is currently a lack of effective integration and analysis methods for platform data and public product evaluations, making it impossible to fully explore the commercial value hidden behind this data and form comprehensive and accurate market insights.

[0004] Therefore, accurately predicting sales volume and prices in various regions, and thus formulating the optimal international business layout strategy for enterprises, has become a problem that needs to be solved. Summary of the Invention

[0005] This invention provides a business layout method and apparatus based on multimodal data fusion and dynamic optimization, which addresses the shortcomings of existing technologies in the inaccurate and unscientific layout of target product categories in international business layout.

[0006] This invention provides a business layout method based on multimodal data fusion and dynamic optimization, comprising: Obtain current overall economic situation data, platform market trend data, and individual product trend data for the target product category; Based on the current overall economic situation data corresponding to the target product category, the comprehensive characteristic Rwl of the target product category is determined; the comprehensive characteristic Rwl is used to characterize the overall environmental trend of the target product category. Based on the current overall market trend data of the target product category, the characteristic value Rbt of the target product category is determined; the characteristic value Rbt is used to characterize the supply and demand status of the target product category on the e-commerce platform. Based on the current trend data of the target product in the target category, the feature value Anw of the target product is determined; the feature value Anw of the target product is used to characterize the user evaluation of the target product. Based on the comprehensive characteristics Rwl, characteristic value Rbt, and characteristic value Anw of the target product category, predict the global sales volume and selling price of the target product category. Based on the global sales volume and selling price of the target product category, conduct international business planning for the target product category.

[0007] According to the business layout method based on multimodal data fusion and dynamic optimization provided by the present invention, the step of determining the comprehensive characteristics Rwl of the target product category based on the current overall economic situation data corresponding to the target product category includes: Obtain social media information, import / export information, and stock information for the target product category at the same time period; Based on the social media information, a feature value Rnw for the target product category is determined; the feature value Rnw is used to characterize the overall emotional inclination of the target product category in the future. Based on the import and export information, the characteristic value Rgv of the target product category is determined; the characteristic value Rgv is used to characterize the international market supply and demand trend of the target product category in the future period. Based on the stock information, a characteristic value Rst for the target category is determined; the characteristic value Rst is used to characterize the expected stock market performance of the target category in the future. Based on the characteristic values ​​Rnw, Rgv, and Rst of the target product category, the comprehensive characteristic Rwl of the target product category is determined.

[0008] According to the business layout method based on multimodal data fusion and dynamic optimization provided by the present invention, obtaining social media information of the target product category during the specified period includes: Identify the keywords corresponding to the target product category during this period; Select the top N most influential mainstream media outlets from the target major economic entities, and collect text data containing the keywords from the selected mainstream media outlets; From the text data containing the keywords, K items are randomly selected according to time sequence as initial social media information; The initial social media information is cleaned to obtain cleaned initial social media information; The top N mainstream media outlets are assigned corresponding weights, and the initial cleaned social media information corresponding to the top-ranked mainstream media outlets is subjected to multiplication learning to obtain multiplied social media information. The social media information after the increase and the initial social media information after the cleansing are used as the social media information of the target category during this period.

[0009] According to the business layout method based on multimodal data fusion and dynamic optimization provided by the present invention, determining the feature value Rnw of the target category based on the social media information includes: The social media information of the target product category during this period is partially labeled to obtain labeled historical social media information; the labeling is based on a score of the impact of social media information on the economy; Using labeled and unlabeled historical social media information, an initial overall market sentiment model is trained to obtain a trained overall market sentiment model; the input of the initial overall market sentiment model is social media information, and the output is the sentiment score corresponding to the social media information. The social media information corresponding to the target category is input into the trained overall market sentiment model. Based on the model, the sentiment score of the target category in the future is predicted. The feature value Rnw is determined based on the sentiment score over a certain future time period.

[0010] According to the business layout method based on multimodal data fusion and dynamic optimization provided by the present invention, the import and export information includes: GDP, CPI, PPI, Gini coefficient, Engel coefficient, foreign exchange reserves, exchange rate, interest rate, historical year-on-year growth rate of imports and exports, month-on-month growth rate, and actual transaction volume collected in chronological order; The step of determining the characteristic value Rgv of the target product category based on the import and export information includes: The historical import and export information of the target product category in the target trading country is preprocessed to obtain preprocessed historical import and export information. An initial market supply and demand trend prediction model is trained using preprocessed historical import and export information to obtain a trained market supply and demand trend prediction model; the market supply and demand trend prediction model is constructed using a Bi-LSTM model. Input the current import and export information of the target product category into the trained market supply and demand trend prediction model, and based on the model, predict the import and export information of the target product category in the future. The import and export information for the future period is converted into a feature value Rgv.

[0011] According to the business layout method based on multimodal data fusion and dynamic optimization provided by the present invention, the stock information includes: daily opening price, closing price, lowest price, highest price, and daily trading volume collected in chronological order; The step of determining the characteristic value Rst of the target category based on the stock information includes: The historical stock information is preprocessed to obtain the preprocessed historical stock information. The preprocessed historical stock information is used to train an initial stock expectation model to obtain a trained stock expectation model; the initial stock expectation model is constructed using a Bi-LSTM model. Input the current stock information of the target category into the trained stock expectation model, and based on the model, predict the daily opening price, closing price, lowest price, highest price and daily trading volume of the target category in the future period. The daily opening price, closing price, lowest price, highest price, and daily trading volume of the target product category over a future period are converted into the feature value Rst.

[0012] According to the business layout method based on multimodal data fusion and dynamic optimization provided by the present invention, the platform trend data includes: pricing, actual sales, exposure, conversion rate of target product categories collected in chronological order, and pricing, sales, exposure, and conversion rate of other merchants' products in the same product category; The step of determining the characteristic value Rbt of the target product category based on the current market trend data of the platform corresponding to the target product category includes: The historical platform market trend data is preprocessed to obtain the preprocessed historical platform market trend data; An initial supply and demand trend model is trained using preprocessed historical market trend data to obtain a trained supply and demand trend model; the initial supply and demand trend model is constructed using a Bi-LSTM model. Input the current platform trend data of the target product category into the trained supply and demand trend model, and based on the model, predict the sales volume and conversion rate of the target product category in the future. The sales volume and conversion rate of the target product category in the future period are converted into the feature value Rbt of the target product category.

[0013] According to the business layout method based on multimodal data fusion and dynamic optimization provided by the present invention, the step of obtaining single-product trend data of the target product category includes: Historical user review data for the target product is collected from relevant e-commerce platforms; the historical user review data covers feedback from users at different purchase times and in different regions; the historical user review data includes: text review content, rating data, review time, and whether the user is a repeat customer. The historical user review data of the target product is cleaned to obtain the cleaned historical user review data of the target product. The historical user review data of the target product after cleaning is labeled as positive, negative and neutral, and the positive / negative review ratio of the target product on the relevant e-commerce platform is calculated. Based on the positive / negative review ratio, the historical user review data marked as negative reviews are multiplied to obtain the multiplied historical user review data with negative reviews. The historical user review data of the target product marked as positive, the historical user review data of the target product marked as neutral, and the historical user review data of negative after doubling are used as the product trend data of the target category. The step of determining the feature value Anw of the target product based on the current trend data of the target product corresponding to the target category includes: Partial scoring and labeling are performed on the individual product trend data of the target product category to obtain labeled individual product trend data; Using labeled and unlabeled product trend data, an initial product evaluation model is trained to obtain a trained product evaluation model; the input of the initial product evaluation model is product trend data, and the output is the evaluation score of the target product. Input the current trend data of the target product corresponding to the target category into the trained product evaluation model, and predict the evaluation score of the target product in the future based on the model. The evaluation score of the target product over a future period is converted into the feature value Anw of the target product.

[0014] According to the business layout method based on multimodal data fusion and dynamic optimization provided by the present invention, the step of predicting the global sales volume and selling price of the target product category based on the comprehensive features Rwl, feature value Rbt, and feature value Anw of the target product category includes: Collect historical comprehensive features Rwl, historical feature value Rbt, historical feature value Anw, and corresponding historical sales volume and historical selling price aligned with the time dimension; The historical comprehensive feature Rwl, historical feature value Rbt, and historical feature value Anw are concatenated to obtain the comprehensive feature vector; Using the comprehensive feature vector as the input to the initial international business layout model, and the historical sales volume and historical selling price as the model output, the initial international business layout model is trained to obtain the trained international business layout model. The comprehensive features Rwl, feature value Rbt, and feature value Anw of the target product category are input into the trained international business layout model. Based on this model, the global sales volume and selling price of the target product category in the future are predicted.

[0015] The present invention also provides an international business layout forecasting device, comprising: The acquisition unit is used to acquire data on the overall economic situation, platform trends, and individual product trends corresponding to the target product category. The determining unit is used to determine the comprehensive characteristics Rwl of the target category based on the current overall economic situation data corresponding to the target category; the comprehensive characteristics Rwl are used to characterize the overall environmental trend of the target category. The determining unit is further configured to determine the feature value Rbt of the target product category based on the current trend data of the platform corresponding to the target product category; the feature value Rbt is used to characterize the supply and demand status of the target product category on the e-commerce platform. The determining unit is further configured to determine the feature value Anw of the target product based on the trend data of the target product corresponding to the current target category; the feature value Anw of the target product is used to characterize the user evaluation of the target product. The prediction unit allows users to predict the global sales volume and selling price of a target product category based on its comprehensive characteristics Rwl, feature value Rbt, and feature value Anw. The layout unit is used to conduct international business layout for the target product category based on its global sales volume and selling price.

[0016] The business layout method and apparatus based on multimodal data fusion and dynamic optimization provided by this invention achieves comprehensive capture of market dynamics by integrating overall economic situation data, platform market trend data, and individual product trend data, thereby improving the accuracy of prediction and providing strong data support for scientific and accurate international business layout. Attached Figure Description

[0017] Figure 1 This is one of the flowcharts for the business layout method based on multimodal data fusion and dynamic optimization provided by the present invention; Figure 2 The second flowchart of the business layout method based on multimodal data fusion and dynamic optimization provided by this invention; Figure 3 The structural block diagram of the international business layout prediction device provided by the present invention. Detailed Implementation

[0018] To make the objectives, technical solutions, and advantages of this invention clearer, the technical solutions of this 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 this invention. All other embodiments obtained by those skilled in the art based on the embodiments of this invention without creative effort are within the scope of protection of this invention.

[0019] Current international business layout planning suffers from problems such as limited data dimensions, lagging market response, and static risk assessment. Traditional methods rely on expert experience and struggle to integrate dynamic factors such as geopolitics, supply chain fluctuations, and exchange rate changes in real time. Existing patents lack technological capabilities in real-time data processing and multimodal feature fusion, specifically as follows: Patent application CN202411430751.7, entitled "Method and System for Data Management of Warehousing Process in Cross-border E-commerce," focuses solely on its own logistics channels. It integrates order, inventory, and logistics information from cross-border e-commerce platforms and logistics service providers to generate solutions for cargo allocation, packaging material recommendations, and logistics channel matching, but does not involve multi-dimensional data fusion. Patent application CN202411402522.4, entitled "Method and System for Price Analysis of Cross-border E-commerce Commodities Based on Big Data," determines the price of target commodities using only a small range of reference data (information and prices of the commodity to be analyzed and similar commodities), resulting in limited data coverage. Patent application CN202210434966.0, entitled "Method, Device, and Equipment for Intelligent Processing of Big Data in Cross-border E-commerce Based on Cloud Services," relies solely on user and product profiles from a single platform, constructing a sales prediction model and allocating warehouse space through a linear regression algorithm, lacking the ability to integrate cross-platform and multi-modal data. Therefore, there is an urgent need for a predictive method that integrates multi-dimensional data and advanced artificial intelligence models to improve the accuracy and scientific nature of international business planning. This application constructs a progressively interconnected predictive system from three levels: "macro-level, meso-level, and micro-level." The macro-level (overall economic situation) focuses on the macro level, integrating global market data (such as major stock market trends, GDP / CPI / PPI of various countries, exchange rate fluctuations), social media (policy texts, international sanctions lists, and overall market sentiment), and customs trade data (year-on-year and month-on-month changes). Cross-modal correlations are established through the splicing of feature parameters to capture the underlying impact of the global economic environment on target product categories. The meso-level (platform market trends) focuses on the e-commerce platform level, analyzing dynamic data from major international platforms (such as Amazon, Alibaba, and Temu), including competitor prices, sales volume, exposure, and conversion rates, to uncover market demand patterns and competitive dynamics within the platform, forming platform-level trend characteristics. Small-Scale Analysis (Single Product Trends): Focusing on specific products, this approach delves into micro-data such as user reviews, repurchase rates, and core function feedback across various platforms. Weights are adjusted based on the ratio of positive to negative reviews to accurately capture users' overall evaluation of each product. Each node across the three levels has an independent task, allowing for separate model training (e.g., using Bi-LSTM networks, large language models, etc.) and extraction of corresponding feature values ​​(e.g., Rnw reflecting market sentiment, Rbt reflecting platform trends, and Anw reflecting product evaluation). These feature values ​​are then integrated and input into a fully connected network. Through multi-dimensional data fusion analysis, sales volume and optimal selling prices in various regions are predicted, generating an optimal international business layout strategy that balances working capital and warehousing costs. This effectively overcomes the shortcomings of traditional methods, such as limited data and delayed response.

[0020] Figure 1This is one of the flowcharts for the business layout method based on multimodal data fusion and dynamic optimization provided by the present invention. Figure 2 This is the second flowchart of the business layout method based on multimodal data fusion and dynamic optimization provided by the present invention, as follows: Figure 1 , Figure 2 As shown, the method includes the following steps: Step 101: Obtain the current overall economic situation data, platform market trend data, and individual product trend data corresponding to the target product category.

[0021] Specifically, the overall economic situation data encompasses global market data (major stocks, GDP / PPI / CPI of major countries, exchange rate fluctuations, etc.), social media (policy texts, international sanctions lists, overall market sentiment, etc.), and customs trade data (year-on-year, month-on-month, etc.). This data reflects the overall environment and trends of the target product category in the international market from a macro perspective, thus overcoming the problem of single data dimensions in traditional methods. The platform's overall trend data comes from major international platforms such as Amazon, Alibaba, and TikTok, including information on its own and competitors' prices, sales volume, exposure, and conversion rates. This reflects the overall performance and competitive landscape of the target product category on e-commerce platforms, making up for the shortcomings of traditional methods in integrating platform data. Individual product trend data mainly consists of historical evaluation data for individual products on relevant platforms, reflecting users' comprehensive judgment of the products. This helps to understand the market feedback of the products themselves and avoids the problem of relying solely on expert experience while ignoring detailed product evaluations.

[0022] Step 102: Based on the current overall economic situation data corresponding to the target category, determine the comprehensive characteristic Rwl of the target category; the comprehensive characteristic Rwl is used to characterize the overall environmental trend of the target category.

[0023] Specifically, the comprehensive feature Rwl integrates social media, stock market data, and economic data from the main operating region, and uses deep learning models (such as Deepseek) to perform sentiment analysis and trend prediction to generate comprehensive features that reflect the overall environmental trends of the target category.

[0024] Step 103: Determine the feature value Rbt of the target product category based on the current trend data of the platform corresponding to the target product category; the feature value Rbt is used to characterize the supply and demand status of the target product category on the e-commerce platform.

[0025] Specifically, Rbt is a feature value extracted from the overall trend analysis of the platform, reflecting the supply and demand trend of this category on e-commerce platforms (such as Amazon, Alibaba International, TikTok, etc.). It is obtained by collecting historical data such as price, sales volume, exposure, and conversion rate of its own products and those of its competitors on the platform. After preprocessing, the data is input into a machine learning model to predict future sales volume (Dsm) and conversion data (Cvr), and then extracted.

[0026] Step 104: Determine the feature value Anw of the target product based on the current trend data of the target product corresponding to the target category; the feature value Anw of the target product is used to characterize the user evaluation of the target product.

[0027] Specifically, a target category refers to a collection of products with similar attributes or uses; it represents a meso- or macro-level classification of products. For example, "electronic products," "clothing," and "home furnishings" are all categories, encompassing multiple specific products. A single product, on the other hand, refers to a specific item within a category, a sub-category. For example, "a certain brand of smartphone" or "a certain model of laptop" within the "electronic products" category are both single products. The feature value Anw is extracted from single-product trend analysis, reflecting users' overall evaluation of the single product. This evaluation influences repurchase rates and other users' purchasing mentality. It is based on historical evaluation data of the single product on relevant platforms, using a processing method similar to social media analysis (such as multiplying negative reviews by sales volume R in the data sample), and is extracted after analysis.

[0028] Step 105: Based on the comprehensive characteristics Rwl, characteristic value Rbt, and characteristic value Anw of the target product category, predict the global sales volume and selling price of the target product category.

[0029] Step 106: Based on the global sales volume and selling price of the target product category, conduct international business planning for the target product category.

[0030] Specifically, this invention utilizes the extracted comprehensive features Rwl, feature value Rbt, and target product feature value Anw to predict the global sales volume and selling price of the target product category through a fully connected network or other machine learning models.

[0031] Based on the forecast results, enterprises can formulate optimal international business layout strategies, including pricing and inventory allocation. Specifically, this embodiment of the invention combines analysis of the overall economic situation to output forecasts of future economic indicators such as GDP and CPI (e.g., G0, G1...G10). Based on the platform's overall market trend, it outputs forecasts of sales volume (Ds0, Ds1...Ds10) and price (Pr0, Pr1...Pr10). Based on stock information analysis, it outputs forecasts of stock opening price, closing price, lowest price, and highest price (e.g., S0, S1, S2...S10). Finally, by combining feature values ​​such as Rwl, Anw, and Rbt, it outputs forecasts of its own product sales volume (MDsm) and conversion data (MCvr) that conform to the merchant's operating style. These forecasts can be used to evaluate the sales volume of its own products at a set price and balance working capital and inventory levels to achieve the optimal solution.

[0032] The business layout method based on multimodal data fusion and dynamic optimization provided by this invention achieves comprehensive capture of market dynamics by integrating overall economic situation data, platform market trend data, and individual product trend data, thereby improving the accuracy of prediction and providing strong data support for scientific and accurate international business layout.

[0033] Preferably, the following is a detailed explanation of how to determine the comprehensive characteristic Rwl of the target product category based on the current overall economic situation data corresponding to the target product category. Specifically, the following solutions are provided: The system acquires social media information, import / export information, and stock information for the target product category at the same time period. Based on the social media information, it determines the characteristic value Rnw for the target product category, which represents the overall sentiment of the target product category in the future. Based on the import / export information, it determines the characteristic value Rgv for the target product category, which represents the international market supply and demand trend of the target product category in the future. Based on the stock information, it determines the characteristic value Rst for the target product category, which represents the stock market expectations for the target product category in the future. Based on the characteristic values ​​Rnw, Rgv, and Rst, it determines the comprehensive characteristic Rwl for the target product category.

[0034] Specifically, in this embodiment of the invention, the comprehensive feature Rwl is a composite feature value composed of three feature values: feature value Rnw, feature value Rgv, and feature value Rst. Feature value Rnw is obtained through sentiment analysis of mainstream media text, reflecting the overall sentiment trend of the market. Feature value Rgv is extracted by using a machine learning model to predict time-series data such as GDP, CPI, and PPI of the trading country, reflecting the supply and demand trend of this category in the international market. Feature value Rst is generated by analyzing the historical stock data of companies related to the category using a machine learning model, reflecting the stock market's expectations for this category. The comprehensive feature Rnw, feature value Rgv, and feature value Rst (the three sub-feature values) are combined and integrated to form the comprehensive feature Rwl, which characterizes the overall environmental trend of the target category, thereby achieving the fusion of multi-dimensional information at the macro level.

[0035] The method provided by this invention decomposes "overall economic situation data" into three sub-dimensions: "social media, import and export information, and stock information." Each sub-dimension corresponds to a core influencing factor (public opinion, objective supply and demand, and market expectations), avoiding the generalized processing of macroeconomic data. By quantifying the sub-feature values ​​Rnw, Rgv, and Rst, the description of "overall environmental trends" becomes more specific and interpretable, providing more accurate underlying features for subsequent forecasts. Furthermore, the three types of sub-data belong to different modalities (text, structured economic data, and financial time-series data), corresponding to the three core influencing factors of international business: "market sentiment," "objective supply and demand," and "capital expectations." By integrating these data, the comprehensive feature Rwl can fully capture the complexity of the macroeconomic environment (such as the combined effect of "policy benefits (positive Rnw) + import and export growth (rising Rgv) + rising stock prices (optimistic Rst)," compensating for the shortcomings of traditional methods that neglect certain factors (such as capital expectations).

[0036] The following describes how to integrate the three sub-features Rnw, Rgv, and Rst to form a comprehensive feature Rwl that represents the overall environmental trend of the target category. The specific methods include the following: Specifically, firstly, the comprehensive feature Rnw, eigenvalue Rgv, and eigenvalue Rst are standardized and mapped to the same dimension (e.g., the 0-1 interval). Then, the time granularity of the comprehensive feature Rnw, eigenvalue Rgv, and eigenvalue Rst is checked (e.g., Rnw is the daily analysis result, Rgv is the monthly import / export forecast, and Rst is the daily stock expectation), unifying them to the same time dimension (e.g., "monthly"). If Rnw is daily data, the monthly average is used as the monthly Rnw; if Rst is daily data, the predicted value of the last trading day of the month is used as the monthly Rst; if Rgv is already a monthly forecast, it is directly retained. This ensures that the three sub-feature values ​​correspond one-to-one in time (e.g., Rnw in October 2024, Rgv in October 2024, and Rst in October 2024), providing a time consistency basis for subsequent concatenation. Next, based on the industry characteristics of the target product category, dynamic weights (not fixed ratios) are assigned to the three sub-feature values ​​to reflect the actual impact of different factors. For example, for policy-sensitive categories (such as new energy equipment), social media (Rnw) has a greater impact on the market, and its weight can be set to 40%; for bulk commodities (such as steel), the weight of import and export supply and demand (Rgv) can be increased to 50%; for new technology products, the weight of stock market expectations (Rst) can be set to 40%. The standardized and time-aligned Rnw, Rgv, and Rst are concatenated into a comprehensive feature Rwl as follows: First, each sub-feature value is multiplied by its corresponding weight (e.g., Rnw × 0.4, Rgv × 0.3, Rst × 0.3) to obtain the weighted sub-feature. The weighted sub-feature is then "dimensionally concatenated": the three values ​​are combined sequentially into a three-dimensional vector (e.g., [weighted Rnw, weighted Rgv, weighted Rst]), which is the comprehensive feature Rwl.

[0037] The method provided by this invention transforms the comprehensive feature Rnw, feature value Rgv, and feature value Rst into Rwl, which can comprehensively reflect the trend of the macro environment. This method not only retains the core information of each dimension, but also achieves the organic integration of multimodal data through standardization and weight adjustment, providing a reliable macro feature basis for subsequent international business layout forecasting.

[0038] Preferably, the following describes how to obtain social media information about the target product category during the stated period, specifically including the following methods: First, determine the keywords corresponding to the target product category during this period; select the top N mainstream media outlets with the highest influence from the target main economic entity, and collect text data containing the keywords from the selected mainstream media; randomly select K pieces of text data containing the keywords according to the time sequence as initial social media information; perform text cleaning on the initial social media information to obtain cleaned initial social media information; assign corresponding weights to the top N mainstream media outlets, and perform multiplication learning on the cleaned initial social media information corresponding to the top-ranked mainstream media outlets to obtain multiplied social media information; use the multiplied social media information and the cleaned initial social media information as the social media information of the target product category during this period.

[0039] Specifically, firstly, based on the target product category (such as electronic devices, textiles, etc.), select similar keywords that accurately represent international business-related information for that category, such as "new energy vehicle export policy" and "cross-border e-commerce tariff adjustments," ensuring that the keywords cover core dimensions such as market dynamics, policy impact, and consumption trends of the category. Secondly, select the top N (e.g., 30th) most influential mainstream media outlets in the target major economies (such as North America, Europe, Southeast Asia, etc.), covering authoritative news agencies, financial media, and industry journals, ensuring the authority and regional representativeness of the data sources. Next, data collection: collect text data containing the above keywords from the selected mainstream media, randomly selecting K (e.g., 1 million) entries as analysis samples, covering different time periods (e.g., the past 2-3 years), ensuring the integrity of the data's time series. Finally, clean the text, removing duplicate content, irrelevant advertisements, garbled text, and other interfering information, retaining the core semantics.

[0040] In addition, to avoid overfitting of neural networks and to strengthen the weight of important media: for texts from influential media (such as The New York Times and Reuters), a networked large language module (such as DeepSeek) is used to polish, convert synonyms or adjust sentence structure without changing the original sentiment and facts, generating equivalent new samples (i.e., "multiplicative learning"), thereby increasing the weight of important media data in training.

[0041] The method provided by this invention determines the keywords of the target product category, filters initial social media information based on the keywords, and performs text cleaning and multiplication on the initial social media information to ensure the relevance, representativeness and effectiveness of the social media information, providing reliable support for macro-environmental analysis in international business layout forecasting.

[0042] Preferably, the following describes a method for determining the feature value Rnw of a target product category based on social media information, specifically including the following scheme: Partial labeling of social media information for the target product category during the specified period yields labeled historical social media information. This labeling is based on scoring the economic impact of the social media information. Using both labeled and unlabeled historical social media information, an initial overall market sentiment model is trained, resulting in a trained overall market sentiment model. The initial overall market sentiment model takes social media information as input and outputs the sentiment score corresponding to that information. The current social media information for the target product category is input into the trained overall market sentiment model. Based on this model, the sentiment score for the target product category over a future period is predicted. The feature value Rnw is determined based on the sentiment score over the future period.

[0043] Specifically, social media information related to the target product category during this period is partially labeled. This labeling method involves using the target product category's social media information during this period as input x, and then using the DeepSeek model to score sentiment. Specifically, the model is asked: "How positive is the attitude towards [the economic field related to the target product category, such as the 'Southeast Asian home appliance market'] in the [article title] published in [media name]? Please rate it from 0 to 10 (0 being extremely negative, 10 being extremely positive)." The model outputs a score y. To improve the accuracy of the scoring, manual verification can be used: 1% of the labeled data is randomly selected for manual review. If the manual score differs significantly from the model score (e.g., more than 2 points), the manual labeling result is used to correct the model's output y. During model training, the text input x and the corresponding score y are input into the DeepSeek model for training. The loss function L1=|Ny| is used as the optimization objective. By iteratively adjusting the model parameters, the absolute error between the predicted value and the actual label is reduced, thereby improving the accuracy of sentiment classification. Where N is the model's predicted score and y is the labeled score. Finally, the social media information corresponding to the target category is input into the trained model for sentiment analysis, outputting its sentiment score (0-10). Then, the sentiment scores of new texts within a certain time period are comprehensively analyzed to extract the feature value Rnw. This value integrates information such as the average score, score trend, and the proportion of high / low-scoring texts, thus reflecting the market's overall sentiment towards the target category.

[0044] This invention requires the integration of multimodal data, including global economic data, cross-border e-commerce platform data, social media, and stock information. Social media alone involves millions (e.g., 1 million) of text samples. Manually labeling all of this data (e.g., manually judging the sentiment of each text) would require a significant amount of manpower (e.g., a team of dozens working continuously for months), resulting in excessively long data processing cycles and an inability to respond in real-time to dynamic changes in the international market (e.g., sudden policy adjustments, exchange rate fluctuations). This invention, through a "partial labeling + semi-supervised training" model, requires only manual processing of about 1% of the samples (e.g., 10,000 text samples). The model then autonomously learns the patterns in the remaining data, reducing the labeling cycle from "months" to "days." This ensures that key aspects such as analysis and sales forecasting can quickly keep up with market changes, providing timely data support for international business strategies.

[0045] Furthermore, the method provided by this invention can directly analyze "current social media information" using a trained model, quickly outputting sentiment scores for a certain future time period, thus achieving a leap from "historical data learning" to "future trend prediction." This forward-looking prediction can timely capture dynamic changes (such as policy shifts and sentiment fluctuations caused by unexpected events), injecting timely information into the feature value Rnw, making it more accurately reflect the overall sentiment trend that the market is about to form. Determining the feature value Rnw based on the sentiment score output by the model can avoid the subjective bias of human judgment (such as different experts interpreting the same text differently). At the same time, the quantified sentiment trend (e.g., 0-10 points) gives the feature value Rnw a clear quantitative attribute, which can be directly combined and integrated with other quantitative features such as Rgv (supply and demand trends) and Rst (market expectations), laying the foundation for subsequent multi-dimensional data fusion and ultimately improving the scientific nature of international business layout prediction.

[0046] Preferably, the import and export information provided in this embodiment of the invention includes: GDP, CPI, PPI, Gini coefficient, Engel coefficient, foreign exchange reserves, exchange rate, interest rate, historical year-on-year growth rate of imports and exports, month-on-month growth rate, and actual transaction volume, collected in chronological order.

[0047] The following describes how to determine the characteristic value Rgv of a target product category based on import and export information, specifically including the following methods: The historical import and export information of the target product category in the target trading country is preprocessed to obtain preprocessed historical import and export information. An initial market supply and demand trend prediction model is trained using this preprocessed historical import and export information to obtain a trained market supply and demand trend prediction model. This market supply and demand trend prediction model is constructed using a Bi-LSTM model. The current import and export information corresponding to the target product category is input into the trained market supply and demand trend prediction model. Based on this model, the import and export information of the target product category in the future is predicted. This future import and export information is then converted into feature values ​​Rgv.

[0048] Specifically, the process begins with data collection, which involves gathering multi-dimensional historical economic data from the target trading country, including but not limited to: macroeconomic indicators such as GDP (Gross Domestic Product), CPI (Consumer Price Index), and PPI (Producer Price Index); socioeconomic indicators such as the Gini coefficient (income distribution equity) and Engel coefficient (household food expenditure as a percentage of GDP); financial indicators such as foreign exchange reserves, exchange rates (exchange rate between the local currency and the target country's currency), and interest rates; and trade data such as historical year-on-year growth rates, month-on-month growth rates, and actual transaction volumes for imports and exports. Next, data preprocessing is performed, which involves cleaning and standardizing the collected raw data to eliminate noise and adapt it to the model's input requirements. Specific measures include: smoothing: using methods such as moving averages to reduce short-term fluctuations and highlight long-term trends; outlier detection: identifying abnormal data caused by statistical errors or sudden events (such as extreme exchange rate fluctuations) and correcting them through interpolation or replacement; and missing value imputation: filling in missing data using methods such as the mean of preceding and following time series data and linear interpolation to ensure the continuity of the time series. Finally, model training (based on a Bi-LSTM network) is conducted. Then, a bidirectional long short-term memory network is trained using historical data to capture the temporal dependencies of the data. Specifically, preprocessed historical time-series data is used as input (such as GDP and CPI series for the past 36 months), and the corresponding actual values ​​at the next moment (such as GDP and CPI for the 37th month) are used as labels. For example, when inputting data such as "GDP, CPI, PPI, exchange rate, and import / export month-on-month changes for the past 12 months", the model predicts the corresponding data: GDP, CPI, PPI, exchange rate, and import / export month-on-month changes for the 13th month.

[0049] This invention employs a Bi-LSTM model to construct a market supply and demand trend prediction model because the core advantage of the Bi-LSTM model lies in its ability to handle long-term dependencies and seasonal characteristics in time-series data. Furthermore, the introduction of multi-dimensional data allows the model to more comprehensively understand economic patterns. For example, the model learns the potential month-on-month trend of import transaction volume when "PPI rises for three consecutive months + currency depreciation." By integrating socioeconomic indicators such as the Gini coefficient and Engel coefficient, the model can more accurately predict changes in import demand related to household consumption. These data collectively constitute the "contextual information" for model training, ensuring that the model can not only predict single indicators (such as GDP) but also improve overall prediction accuracy through the synergistic relationships of multiple indicators.

[0050] Furthermore, during training, the model employs a loss function L2=sum(|Ny|) to optimize the model, where "N" represents the model's predicted values ​​for all indicators, and "y" represents the actual values ​​of all indicators. By calculating the cumulative sum of prediction errors for all indicators, the model's ability to predict multi-dimensional data is simultaneously optimized. Iterative adjustments to model parameters minimize prediction bias and improve training accuracy. Then, the trained market supply and demand trend prediction model is used to predict economic indicators for the future: inputting the most recent time-series data (such as indicators for the current month), the model outputs predicted values ​​for GDP, CPI, and year-on-year / month-on-month import and export figures for the next period (e.g., next month). Based on these predictions, a feature value Rgv is extracted. This value comprehensively reflects the supply and demand trends of the target category in the international market (e.g., an increase in Rgv may indicate increased demand, while a decrease may indicate oversupply). Finally, the feature value Rgv will serve as one of the core inputs for international business layout prediction, merging with other dimensional data (such as social media and stock market trends) to determine the category's feature value Rwl.

[0051] The eigenvalue Rgv is a specific value extracted from the model's predicted future indicators using pre-defined rules. For example, it weights and integrates the direction and magnitude of changes in predicted indicators (such as GDP growth rate, import / export month-on-month changes, etc.) to transform them into a single eigenvalue Rgv. Rgv acts as a "quantitative representation" of the supply and demand trends of this product category in the international market; its magnitude and fluctuation direction directly reflect the trend characteristics. A higher Rgv indicates that the predicted results of indicators such as GDP and CPI collectively point to increased demand or tight supply for this product category (the supply and demand trend leans towards "supply shortage" or "demand expansion"); a lower Rgv indicates that the combination of indicators points to shrinking demand or oversupply (the supply and demand trend leans towards "oversupply" or "demand contraction"). A month-on-month increase in Rgv (e.g., from 0.5 to 0.7) reflects a "strengthening" supply and demand trend (e.g., accelerated demand growth); a month-on-month decrease reflects a "weakening" trend (e.g., slowing demand growth).

[0052] Preferably, the stock information provided by the present invention includes: daily opening price S0, closing price S1, lowest price Smin, highest price Smax, and daily trading volume, collected in chronological order.

[0053] The following describes how to determine the characteristic value Rst of a target category based on stock information, specifically including the following methods: Historical stock information is preprocessed to obtain preprocessed historical stock information; an initial stock expectation model is trained using the preprocessed historical stock information to obtain a trained stock expectation model; the initial stock expectation model is constructed using a Bi-LSTM model; the current stock information of the target category is input into the trained stock expectation model, and based on the model, the daily opening price, closing price, lowest price, highest price, and daily trading volume of the target category in the future period are predicted; the daily opening price, closing price, lowest price, highest price, and daily trading volume of the target category in the future period are converted into the feature value Rst.

[0054] Specifically, firstly, historical stock data of companies related to the target category is collected. This data is recorded continuously in chronological order (e.g., trading days) to form a complete stock time-series dataset. This dataset specifically includes: daily opening price (S0), closing price (S1), lowest price (Smin), highest price (Smax), and daily trading volume. Then, this data is recorded continuously in chronological order (e.g., trading days) to form a complete stock time-series dataset. Next, the historical stock data undergoes normalization processing to eliminate interfering factors and adapt it to model training. This includes: normalization processing: converting data of different magnitudes (e.g., opening price may be hundreds of yuan, trading volume may be in the millions) to a uniform numerical range (e.g., 0-1) to avoid the model's learning of key features being affected by differences in data magnitude. Missing value handling: for missing values ​​caused by holiday closures, data recording errors, etc., missing values ​​are filled by interpolation (e.g., linear interpolation) or mean averaging of data from preceding and following trading days to ensure the continuity of the time series. Extreme outliers (e.g., single-day surges or plunges caused by market manipulation) are identified and corrected to reduce their interference with model training. Then, an initial stock expectation model is trained using historical stock data to learn the time-dependent patterns in the data. Specifically, preprocessed historical stock time-series data (such as the opening price and trading volume sequence of the past 120 trading days) is used as input, and the model outputs the actual values ​​for a future trading day (such as the opening price and trading volume of the 121st trading day). Next, the target category's stock information (recent stock time-series data, such as data from the last 30 trading days) is input, and the model outputs predicted values ​​for the daily opening price, closing price, lowest price, and highest price for the next six months (approximately 120 trading days). Based on these predictions, a feature value Rst is extracted using preset rules (such as considering the direction of price trends, volatility, and the correlation between trading volume and price). For example, if the predicted closing price for the next six months shows a continuous upward trend, and the trading volume increases simultaneously, Rst may be assigned a higher value, reflecting the stock market's optimistic expectation for that category; if the price shows a downward trend, Rst may be a lower value, reflecting a more pessimistic market expectation.

[0055] Finally, the extracted feature value Rst will be aligned with the feature value Rnw, which reflects market sentiment, and the feature value Rgv, which reflects supply and demand trends, in the time dimension (ensuring that the three are based on predictions within the same time interval), and then concatenated into a comprehensive feature Rwl, providing a decision-making basis for intelligent prediction of subsequent international business layout.

[0056] Furthermore, during model training, the L3 loss function (the sum of the absolute errors between the predicted and actual values) is used to measure model accuracy. By iteratively adjusting network parameters, the error is minimized, enabling the model to accurately predict key stock price indicators. Specifically, L3 = |S0 - S0| + |S1 - S1| + |Smin - Smin| + |Smax - Smax|.

[0057] Furthermore, the initial stock expectation model of this invention is constructed using a Bi-LSTM model. Bi-LSTM, through forward and backward memory units, can simultaneously capture short-term fluctuations (such as intraday price changes) and long-term trends (such as consecutive months of upward / downward cycles) in stock data, solving the long-term dependency problem that traditional models cannot handle, thereby making stock expectation prediction more accurate.

[0058] Preferably, the platform trend data provided by the present invention includes: historical pricing, actual sales volume, exposure, conversion rate of the target product category collected in chronological order, and pricing, sales volume, exposure, and conversion rate of other merchants' products in the same product category.

[0059] The following describes how to determine the characteristic value Rbt of a target product category based on the current market trend data of the corresponding platform. The specific methods include the following: Historical platform market trend data is preprocessed to obtain preprocessed historical platform market trend data. An initial supply and demand trend model is trained using this preprocessed historical platform market trend data to obtain a trained supply and demand trend model. This initial supply and demand trend model is constructed using a Bi-LSTM model. The current platform market trend data for the target product category is input into the trained supply and demand trend model. Based on this model, the sales volume and conversion rate of the target product category in the future are predicted. The sales volume and conversion rate of the target product category in the future are then converted into the feature value Rbt of the target product category.

[0060] Specifically, firstly, multi-dimensional historical data on the target product category is collected from mainstream e-commerce platforms (such as Amazon, Alibaba International, and Temu). This includes: data on the target product category itself, namely: historical pricing, actual sales volume, exposure (e.g., number of times users viewed the product), conversion rate (e.g., the proportion of views leading to orders), and pricing, sales volume, exposure, and conversion rate of other merchants in the same category. This data is recorded by time dimension (e.g., daily, weekly) to form a continuous platform operation time-series dataset. Then, the collected historical platform trend data is cleaned and standardized to ensure data quality and suitability for model requirements. This includes: data cleaning: removing outliers caused by system failures or statistical errors (e.g., negative daily sales) and correcting duplicate records. Data processing: converting heterogeneous data from different platforms (e.g., different statistical methods for exposure across platforms) into a unified format for cross-platform comparative analysis. For missing data (e.g., missing conversion rate on a certain date on a certain platform), it is filled using methods such as averaging adjacent time periods and interpolating data from similar products on the same platform to ensure time-series continuity. Then, an initial supply and demand trend model is trained using preprocessed historical platform market trend data to learn the potential patterns in the platform data, thereby obtaining a well-trained supply and demand trend model. Specifically, preprocessed historical time-series data (such as the pricing and exposure sequences of the past 60 days for both the model itself and its competitors) is used as input, along with the corresponding actual values ​​for the next moment (such as sales volume and conversion rate on day 61). During model training, the loss function L4 = |Dsm - yDsm| + |Cvr - yCvr| is used to optimize the model, where Dsm is the model's predicted sales volume, yDsm is the actual sales volume, Cvr is the model's predicted conversion rate, and yCvr represents the actual conversion rate. The model accuracy is measured by the loss function L4 (the absolute error between predicted and actual sales volume + the absolute error between predicted and actual conversion rate), and the network parameters are iteratively adjusted to minimize the error, enabling the model to accurately predict sales volume (Dsm) and conversion rate (Cvr). Next, using the trained model, the latest platform operation data (such as pricing and exposure over the past 30 days) is input, and the predicted sales volume (Dsm) and conversion rate (Cvr) for a future period (e.g., the next 30 days) are output. Then, based on the predicted sales volume and conversion rate trends, combined with the correlation patterns of "sales volume-conversion-supply and demand" in historical data (e.g., a consistently high conversion rate usually corresponds to strong demand), a feature value Rbt is extracted through preset rules (e.g., the overall sales growth rate, the direction of conversion rate fluctuations, etc.). For example, if the predicted sales volume continues to rise and the conversion rate increases simultaneously, Rbt may be high, reflecting that the supply and demand trend of this category on the platform is biased towards "strong demand"; if the predicted sales volume declines and the conversion rate decreases, Rbt may be low, reflecting "weak demand".This invention utilizes the eigenvalue Rbt as a core indicator reflecting the supply and demand status of a target product category on an e-commerce platform. It integrates this eigenvalue with other dimensional features (such as the overall economic situation feature Rwl and the single product evaluation feature Anw) to provide platform-side data analysis support for enterprises' pricing strategies, inventory planning, and other international business layout decisions.

[0061] The initial supply and demand trend model provided by this invention is constructed using a Bi-LSTM model. The Bi-LSTM network captures short-term fluctuations (such as a surge in daily sales due to promotional activities) and long-term trends (such as the seasonal peak consumption patterns) of data through bidirectional memory units, while learning the linkage between indicators such as "pricing and sales volume" and "exposure and conversion" (such as price reductions may increase sales volume and conversion rate).

[0062] Preferably, the following describes how to obtain the trend data of individual products in the target category, specifically including the following methods: Historical user review data for the target product is collected from relevant e-commerce platforms. This historical user review data covers feedback from users at different purchase times and in different regions. The historical user review data includes: text review content, rating data, review time, and whether the user is a repeat customer. The historical user review data for the target product is cleaned to obtain cleaned historical user review data. The cleaned historical user review data is labeled as positive, negative, or neutral, and the positive / negative review ratio for the target product on the relevant e-commerce platforms is calculated. Based on the positive / negative review ratio, the historical user review data marked as negative is multiplied to obtain multiplied negative historical user review data. The historical user review data marked as positive, the neutral historical user review data, and the multiplied negative historical user review data are used as the product trend data for the target category.

[0063] Specifically, firstly, historical user review data (including text reviews, ratings, review times, and whether the user is a repeat customer) for the target product is collected from relevant e-commerce platforms (such as Amazon, Alibaba, and Temu). This data ensures that the collected historical user review data covers feedback from users at different purchase times and in different regions to guarantee data diversity. Next, the historical user review data for the target product is cleaned, specifically by removing duplicate reviews, invalid spam content, and irrelevant chat information to retain text with substantive review value. Outliers in the rating data (such as malicious low / high scores) are also addressed to ensure data authenticity. Then, the cleaned historical user review data for the target product is labeled as positive, negative, or neutral, and the "ratio of positive to negative reviews" is calculated, i.e., R = total positive reviews for this category on the platform ÷ total negative reviews for this category on the platform. For example, if the total number of positive reviews for all products in this category on the platform is 10,000 and the total number of negative reviews is 2,000, then R = 5, meaning that the overall number of positive reviews for this category is 5 times the number of negative reviews. Then, based on the positive / negative review ratio, negative reviews undergo "multiplicative learning": using a large language model (such as DeepSeek), each negative review is paraphrased, its sentence structure adjusted, or details supplemented (e.g., rewriting "This product is terrible, it broke after two days" as "This product is of poor quality and malfunctioned after only two days of use") without altering its core sentiment and facts, generating equivalent samples R times the number of original negative reviews. For example, if a single product has 100 original negative reviews and R=5, then 500 equivalent negative review samples are generated through multiplication, forming the negative review portion of the training data together with the original negative reviews. Finally, the historical user review data of the target product marked as positive, the historical user review data of the target product marked as neutral, and the historical user review data of the multiplied negative reviews are used as the product trend data of the target category for model training.

[0064] As the sample size of negative reviews increases by R times, the impact of negative review data on the loss function is amplified during model training: the model will pay more attention to the frequently occurring problems in negative reviews (such as "short battery life" and "low cost performance"), thereby being able to more accurately identify key negative factors affecting users' purchasing mentality and ultimately improve the accuracy of international business layout for target products.

[0065] The following describes how to determine the characteristic value Anw of a target product based on the current trend data of the target product in the target category. The specific methods include the following: Partial scoring and labeling are performed on the individual product trend data of the target category to obtain labeled individual product trend data. Using the labeled and unlabeled individual product trend data, an initial individual product evaluation model is trained to obtain a trained individual product evaluation model. The input to the initial individual product evaluation model is the individual product trend data, and the output is the evaluation score of the target individual product. The current target product trend data corresponding to the target category is input into the trained individual product evaluation model. Based on this model, the evaluation score of the target individual product in the future is predicted. The evaluation score of the target individual product in the future is converted into the feature value Anw of the target individual product.

[0066] Specifically, a subset of samples is selected from the product trend data of the target category for manual or model-assisted scoring and labeling. The labeling content is a quantitative score reflecting the sentiment of the evaluation (e.g., 0-10 points, where 0 represents extremely negative and 10 represents extremely positive), along with specific scores for key evaluation dimensions (such as "quality," "price," and "logistics"). Then, the labeled product trend data (with score labels) and unlabeled product trend data (without labels) are used together as training data to input into the initial product evaluation model. The model's input is product trend data (such as user review text, historical ratings, etc.), and its output is the product's evaluation score (consistent with the labeled score logic, 0-10 points). During training, the model learns the mapping pattern of "evaluation content → score" through labeled data, while simultaneously mining potential features in the unlabeled data (such as the correlation between the frequently occurring "poor quality" in negative reviews and low scores), continuously adjusting parameters to reduce prediction errors. After training, a "trained product evaluation model" that can stably output product evaluation scores is obtained. Then, the "product trend data" (such as recent user reviews, feedback related to short-term sales fluctuations, etc.) of the target product is input into the trained model. Based on the learned patterns, the model predicts the product's rating score over a future period (e.g., 1 month, 3 months). Finally, the model's predicted "future rating score" is converted into a feature value Anw. The feature value Anw ultimately becomes the core indicator reflecting the target product, providing data support for subsequent international business planning (such as inventory adjustments and pricing strategy optimization).

[0067] The following describes how to predict the global sales volume and selling price of a target product category based on its comprehensive characteristics Rwl, characteristic value Rbt, and characteristic value Anw. The specific methods include the following: Collect historical comprehensive features Rwl, historical feature values ​​Rbt, and historical feature values ​​Anw aligned to the time dimension, along with corresponding historical sales volume and historical selling price; concatenate the historical comprehensive features Rwl, historical feature values ​​Rbt, and historical feature values ​​Anw to obtain a comprehensive feature vector; use the comprehensive feature vector as input to an initial international business layout model, and use the historical sales volume and historical selling price as model output to train the initial international business layout model, resulting in a trained international business layout model; input the comprehensive features Rwl, feature values ​​Rbt, and feature values ​​Anw of the target product category into the trained international business layout model, and based on this model, predict the global sales volume and selling price of the target product category in the future.

[0068] Specifically, first, historical data is collected, including the historical comprehensive characteristics Rwl (macroeconomic trend), historical feature values ​​Rbt (platform supply and demand status), and historical feature values ​​Anw (single product user reviews) of the target product category, as well as the corresponding historical sales volume and historical selling price. All historical data are ensured to correspond one-to-one in time (e.g., Rwl, Rbt, and Anw in January 2023 correspond to the actual sales volume and selling price in January 2023), forming a time series sample of "feature-result". Then, the aligned comprehensive characteristics Rwl, feature values ​​Rbt, and feature values ​​Anw are concatenated dimensionally into a comprehensive feature vector (e.g., [Rwl=0.6, Rbt=0.7, Anw=0.8]), where each vector represents comprehensive information about the market environment, platform status, and user feedback at a specific point in time. Using a comprehensive feature vector (e.g., [Rwl, Rbt, Anw]) as input and corresponding historical sales volume and historical selling price as output labels, training samples are constructed (e.g., "Feature vector in January 2023 → Sales volume of 500 units and selling price of 80 yuan in January 2023"). An initial international business layout model (e.g., a fully connected neural network) is used, and the model is optimized using the loss function L5 = |MDsm - yMDsm| + |MCvr - yMCvr|, where MDsm is the predicted sales volume, yMDsm is the actual sales volume, MCvr is the predicted selling price, and yMCvr is the actual selling price. This allows the model to learn the mapping relationship between features and sales volume and selling price, thus obtaining a trained international business layout model. The latest comprehensive features Rwl, feature value Rbt, and feature value Anw are then input into the trained model. Based on the learned patterns, the model outputs a prediction of the global sales volume and selling price of the target product category over a future period. This prediction result can ultimately be used to guide international business layout (e.g., adjusting production plans and formulating regional pricing strategies).

[0069] The international business layout prediction device provided by the present invention is described below. The international business layout prediction device described below can be referred to in correspondence with the business layout method based on multimodal data fusion and dynamic optimization described above.

[0070] Figure 3 The structural block diagram of the international business layout prediction device provided by the present invention is as follows: Figure 3 As shown, the device includes: Acquisition unit 301 is used to acquire the overall economic situation data, platform market trend data, and individual product trend data corresponding to the target product category. The determining unit 302 is used to determine the comprehensive characteristic Rwl of the target product category based on the current overall economic situation data corresponding to the target product category; the comprehensive characteristic Rwl is used to characterize the overall environmental trend of the target product category. The determining unit 302 is further configured to determine the feature value Rbt of the target product category based on the current trend data of the platform corresponding to the target product category; the feature value Rbt is used to characterize the supply and demand status of the target product category on the e-commerce platform. The determining unit 302 is further configured to determine the feature value Anw of the target product based on the trend data of the target product corresponding to the current target category; the feature value Anw of the target product is used to characterize the user evaluation of the target product. Prediction unit 303: The user predicts the global sales volume and selling price of the target product category based on the comprehensive features Rwl, feature value Rbt, and feature value Anw of the target product category. The layout unit 304 is used to conduct international business layout for the target product category based on its global sales volume and selling price.

[0071] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, and not to limit them; 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; and these 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.

Claims

1. A business layout method based on multimodal data fusion and dynamic optimization, characterized in that, include: Obtain current overall economic situation data, platform market trend data, and individual product trend data for the target product category; Based on the current overall economic situation data corresponding to the target product category, the comprehensive characteristic Rwl of the target product category is determined; the comprehensive characteristic Rwl is used to characterize the overall environmental trend of the target product category. Based on the current overall market trend data of the target product category, the characteristic value Rbt of the target product category is determined; the characteristic value Rbt is used to characterize the supply and demand status of the target product category on the e-commerce platform. Based on the current trend data of the target product in the target category, the feature value Anw of the target product is determined; the feature value Anw of the target product is used to characterize the user evaluation of the target product. Based on the comprehensive characteristics Rwl, characteristic value Rbt, and characteristic value Anw of the target product category, predict the global sales volume and selling price of the target product category. Based on the global sales volume and selling price of the target product category, conduct international business planning for the target product category.

2. The business layout method based on multimodal data fusion and dynamic optimization according to claim 1, characterized in that, The determination of the comprehensive characteristics Rwl of the target product category based on the current overall economic situation data includes: Obtain social media information, import / export information, and stock information for the target product category at the same time period; Based on the social media information, a feature value Rnw for the target product category is determined; the feature value Rnw is used to characterize the overall emotional inclination of the target product category in the future. Based on the import and export information, the characteristic value Rgv of the target product category is determined; the characteristic value Rgv is used to characterize the international market supply and demand trend of the target product category in the future period. Based on the stock information, a characteristic value Rst for the target category is determined; the characteristic value Rst is used to characterize the expected stock market performance of the target category in the future. Based on the characteristic values ​​Rnw, Rgv, and Rst of the target product category, the comprehensive characteristic Rwl of the target product category is determined.

3. The business layout method based on multimodal data fusion and dynamic optimization according to claim 2, characterized in that, Obtaining social media information for the target product category during the stated period includes: Identify the keywords corresponding to the target product category during this period; Select the top N most influential mainstream media outlets from the target major economic entities, and collect text data containing the keywords from the selected mainstream media outlets; From the text data containing the keywords, K items are randomly selected according to time sequence as initial social media information; The initial social media information is cleaned to obtain cleaned initial social media information; The top N mainstream media outlets are assigned corresponding weights, and the initial cleaned social media information corresponding to the top-ranked mainstream media outlets is subjected to multiplication learning to obtain multiplied social media information. The social media information after the increase and the initial social media information after the cleansing are used as the social media information of the target category during this period.

4. The business layout method based on multimodal data fusion and dynamic optimization according to claim 3, characterized in that, The step of determining the feature value Rnw of the target category based on the social media information includes: The social media information of the target product category during this period is partially labeled to obtain labeled historical social media information; the labeling is based on a score of the impact of social media information on the economy; Using labeled and unlabeled historical social media information, an initial overall market sentiment model is trained to obtain a trained overall market sentiment model; the input of the initial overall market sentiment model is social media information, and the output is the sentiment score corresponding to the social media information. The social media information corresponding to the target category is input into the trained overall market sentiment model. Based on the model, the sentiment score of the target category in the future is predicted. The feature value Rnw is determined based on the sentiment score over a certain future time period.

5. The business layout method based on multimodal data fusion and dynamic optimization according to claim 2, characterized in that, The import and export information includes: GDP, CPI, PPI, Gini coefficient, Engel coefficient, foreign exchange reserves, exchange rate, interest rate, historical year-on-year growth rate of imports and exports, month-on-month growth rate, and actual transaction volume, collected in chronological order. The step of determining the characteristic value Rgv of the target product category based on the import and export information includes: The historical import and export information of the target product category in the target trading country is preprocessed to obtain preprocessed historical import and export information. An initial market supply and demand trend prediction model is trained using preprocessed historical import and export information to obtain a trained market supply and demand trend prediction model; the market supply and demand trend prediction model is constructed using a Bi-LSTM model. Input the current import and export information of the target product category into the trained market supply and demand trend prediction model, and based on the model, predict the import and export information of the target product category in the future. The import and export information for the future period is converted into a feature value Rgv.

6. The business layout method based on multimodal data fusion and dynamic optimization according to claim 2, characterized in that, The stock information includes: daily opening price, closing price, lowest price, highest price, and daily trading volume, collected in chronological order. The step of determining the characteristic value Rst of the target category based on the stock information includes: The historical stock information is preprocessed to obtain the preprocessed historical stock information. The preprocessed historical stock information is used to train an initial stock expectation model to obtain a trained stock expectation model; the initial stock expectation model is constructed using a Bi-LSTM model. Input the current stock information of the target category into the trained stock expectation model, and based on the model, predict the daily opening price, closing price, lowest price, highest price and daily trading volume of the target category in the future period. The daily opening price, closing price, lowest price, highest price, and daily trading volume of the target product category over a future period are converted into the feature value Rst.

7. The business layout method based on multimodal data fusion and dynamic optimization according to any one of claims 1-6, characterized in that, The platform's overall trend data includes: pricing, actual sales volume, exposure, conversion rate of target product categories collected in chronological order; pricing, sales volume, exposure, and conversion rate of other merchants' products in the same category; The step of determining the characteristic value Rbt of the target product category based on the current market trend data of the platform corresponding to the target product category includes: The historical platform market trend data is preprocessed to obtain the preprocessed historical platform market trend data; An initial supply and demand trend model is trained using preprocessed historical market trend data to obtain a trained supply and demand trend model; the initial supply and demand trend model is constructed using a Bi-LSTM model. Input the current platform trend data of the target product category into the trained supply and demand trend model, and based on the model, predict the sales volume and conversion rate of the target product category in the future. The sales volume and conversion rate of the target product category in the future period are converted into the feature value Rbt of the target product category.

8. The business layout method based on multimodal data fusion and dynamic optimization according to claim 7, characterized in that, The acquisition of individual product trend data for the target product category includes: Historical user review data for the target product is collected from relevant e-commerce platforms; the historical user review data covers feedback from users at different purchase times and in different regions; the historical user review data includes: text review content, rating data, review time, and whether the user is a repeat customer. The historical user review data of the target product is cleaned to obtain the cleaned historical user review data of the target product. The historical user review data of the target product after cleaning is labeled as positive, negative and neutral, and the positive / negative review ratio of the target product on the relevant e-commerce platform is calculated. Based on the positive / negative review ratio, the historical user review data marked as negative reviews are multiplied to obtain the multiplied historical user review data with negative reviews. The historical user review data of the target product marked as positive, the historical user review data of the target product marked as neutral, and the historical user review data of negative after doubling are used as the product trend data of the target category. The step of determining the feature value Anw of the target product based on the current trend data of the target product corresponding to the target category includes: Partial scoring and labeling are performed on the individual product trend data of the target product category to obtain labeled individual product trend data; Using labeled and unlabeled product trend data, an initial product evaluation model is trained to obtain a trained product evaluation model; the input of the initial product evaluation model is product trend data, and the output is the evaluation score of the target product. Input the current trend data of the target product corresponding to the target category into the trained product evaluation model, and predict the evaluation score of the target product in the future based on the model. The evaluation score of the target product over a future period is converted into the feature value Anw of the target product.

9. The business layout method based on multimodal data fusion and dynamic optimization according to claim 7, characterized in that, The prediction of global sales volume and selling price of the target product category based on its comprehensive characteristics Rwl, characteristic value Rbt, and characteristic value Anw includes: Collect historical comprehensive features Rwl, historical feature value Rbt, historical feature value Anw, and corresponding historical sales volume and historical selling price aligned with the time dimension; The historical comprehensive feature Rwl, historical feature value Rbt, and historical feature value Anw are concatenated to obtain the comprehensive feature vector; Using the comprehensive feature vector as the input to the initial international business layout model, and the historical sales volume and historical selling price as the model output, the initial international business layout model is trained to obtain the trained international business layout model. The comprehensive features Rwl, feature value Rbt, and feature value Anw of the target product category are input into the trained international business layout model. Based on this model, the global sales volume and selling price of the target product category in the future are predicted.

10. An international business layout forecasting device, characterized in that, include: The acquisition unit is used to acquire data on the overall economic situation, platform trends, and individual product trends corresponding to the target product category. The determining unit is used to determine the comprehensive characteristics Rwl of the target category based on the current overall economic situation data corresponding to the target category; the comprehensive characteristics Rwl are used to characterize the overall environmental trend of the target category. The determining unit is further configured to determine the feature value Rbt of the target product category based on the current trend data of the platform corresponding to the target product category; the feature value Rbt is used to characterize the supply and demand status of the target product category on the e-commerce platform. The determining unit is further configured to determine the feature value Anw of the target product based on the trend data of the target product corresponding to the current target category; the feature value Anw of the target product is used to characterize the user evaluation of the target product. The prediction unit allows users to predict the global sales volume and selling price of a target product category based on its comprehensive characteristics Rwl, feature value Rbt, and feature value Anw. The layout unit is used to conduct international business layout for the target product category based on its global sales volume and selling price.

Images