Intelligent cross-border commodity price comparison method and system based on deep learning

Abstract

This invention relates to the field of commodity transaction processing technology, specifically to a deep learning-based intelligent price comparison method and system for cross-border commodities. It includes constructing a multi-source, multi-modal feature index for cross-border commodities, fusing semantic and visual features to achieve accurate cross-language SKU alignment, applying a model to structurally decompose price elements, calculating a unified final price based on real-time exchange rates and tax rules, generating intelligent price comparison data, and achieving dynamic monitoring and early warning through prediction algorithms. This invention constructs a multi-modal extraction and cross-language mapping architecture, fusing multi-source features to overcome semantic differences and improve SKU alignment accuracy. It introduces a deep learning model to finely decompose price, tax, and logistics elements, utilizes dynamic exchange rates and tax topology to establish a normalization model, unifies pricing standards and eliminates biases, and combines time series analysis to achieve real-time monitoring and early warning, enhancing market responsiveness and providing high-precision, automated intelligent price comparison support for cross-border trade.

Description

Technical Field

[0001] This invention relates to the field of commodity transaction processing technology, and in particular to a cross-border commodity intelligent price comparison method and system based on deep learning. Background Technology

[0002] The field of commodity transaction processing technology involves the organization of commodity information, identification of trading partners, collection of price formation elements, and unification of settlement standards. Its core content revolves around the processing chain of commodities from display to transaction, unifying the representation and rule constraints of commodity names, brands, specifications, models, packaging quantities, applicable regions, pricing currencies, tax items, logistics costs, promotional conditions, and inventory status. Among these, intelligent price comparison for cross-border commodities based on deep learning utilizes deep learning models to identify identical products and unify price standards across different platforms, languages, and pricing rules. However, existing cross-border e-commerce processing systems still have certain limitations in multi-source data alignment, refined price breakdown, and cross-language identical product determination. A search revealed a patent with publication number CN121563650A that discloses an artificial intelligence-based cross-border e-commerce product recommendation and matching system. This system constructs user profiles by analyzing user behavioral data and emotional feedback, and intelligently recommends products and logistics solutions based on factors such as logistics, price, and culture. However, at the price comparison level, this solution focuses on personalized recommendation logic for users, limiting its ability to deeply integrate and identify the underlying attributes of multi-source commodities. When matching a large number of identical products across different platforms, the system has deviations in the accurate alignment at the SKU level. Furthermore, the refined breakdown and unified modeling mechanism for the complex pricing structure in cross-border scenarios is not yet perfect, resulting in the need to improve the granularity of the price comparison results.

[0003] On the other hand, patent CN120430749A discloses an intelligent collaborative enterprise rapid access and integrated product selection and fulfillment method and system. This system uses a multi-dimensional analysis model to screen high-potential products, utilizes intelligent price analysis to complete price negotiations, and automatically matches products and SKUs for intelligent listing and fulfillment. However, the application scenarios of this solution are mainly concentrated on enterprise-level product selection and supply chain flow, with its price analysis serving more commercial negotiation than multi-caliber price comparison. Regarding deep learning applications, the system's mechanisms for cross-language attribute extraction, sales unit conversion, and automatic restoration of preferential conditions still need improvement. There is room for improvement in addressing the issue of normalizing price comparison criteria under different pricing currencies and differentiated promotional rules.

[0004] In summary, existing cross-border e-commerce systems still have room for improvement in areas such as cross-language product recognition accuracy, reconstruction of complex price components, and construction of multi-caliber price comparison models. Therefore, this invention provides a deep learning-based intelligent price comparison method and system for cross-border goods. It aims to utilize deep learning technology to optimize cross-modal feature extraction and matching accuracy, and to automate the conversion and aggregation of currencies, taxes, shipping costs, and preferential conditions, thereby enhancing the ability of cross-border trade to support high-precision, real-time intelligent price comparison. Summary of the Invention

[0005] To address the technical problems existing in the prior art, embodiments of the present invention provide a method and system for intelligent price comparison of cross-border commodities based on deep learning. The technical solution is as follows: A deep learning-based intelligent price comparison method for cross-border goods includes the following steps: S1: Obtain the original product dataset from multiple cross-border platforms, extract multimodal feature information of products, perform semantic vector conversion on product titles in different languages, encode visual features of product images, establish a basic product feature index library, and output multimodal matching sequences; S2: Call the semantic vectors and visual codes in the multimodal sequence to be matched, perform cross-language semantic mapping and feature fusion operations, calculate the feature similarity between products on different platforms, filter product items with similarity exceeding the preset alignment threshold, perform precise alignment at the SKU level, and form a set of related products of the same type; S3: Call the pricing page data of each product in the same product association set, apply a deep learning sequence labeling model to identify the basic price, tax rules, shipping template and promotion weight in the price string, perform structured decomposition of complex price components, and output a refined price element list; S4: Call the refined price element list, obtain the real-time exchange rate matrix and the preset tax calculation topology, perform currency conversion, unit dimension alignment and preferential strategy restoration operations, calculate the final price of each aligned product under the unified standard, and generate intelligent price comparison result data.

[0006] As a further aspect of the present invention, the multimodal sequence to be matched includes cross-language semantic vectors, visual feature encoding sets, and platform attribute tags; the same product association set includes feature alignment indexes, SKU mapping relationship tables, and cross-platform product clusters; the refined price element list includes basic price components, tax components, logistics cost components, and promotional deduction items; and the intelligent price comparison result data includes normalized price sequences, price difference analysis reports, and optimal procurement paths.

[0007] As a further aspect of the present invention, the step of obtaining S1 is as follows: S101: Obtain raw web page data from multiple preset cross-border e-commerce platforms through distributed data acquisition nodes, and use preset parsing rules to separate the product title string, product main image, detailed description text and real-time price area, and store each modal data in the preset cache area; S102: Call the preset cross-language pre-trained language model to perform word segmentation and word embedding operations on the product title string to obtain a title feature vector containing contextual semantic information. At the same time, call the convolutional neural network to perform feature extraction on the product main image to obtain an image feature vector containing texture, color and composition information. S103: The title feature vector and the image feature vector are concatenated, and position encoding and platform attribute weights are introduced. The dimensionality is transformed through a fully connected layer, and the transformed multimodal feature data is classified and numbered according to the platform source to obtain the multimodal matching sequence.

[0008] As a further aspect of the present invention, the step of obtaining S2 is as follows: S201: Call the first platform feature vector and the second platform feature vector in the multimodal sequence to be matched, use the preset shared semantic space mapping matrix to map the vectors of different languages ​​to a unified coordinate system, and perform vector dot product operation to obtain the initial similarity score of the candidate matching pair; S202: Based on the initial similarity score, filter product pairs whose scores are within the preset candidate interval, extract their corresponding SKU attribute list, including color, specifications, material and packaging quantity, and use a combination of string fuzzy matching algorithm and deep learning semantic similarity calculation to determine the attribute consistency status at the SKU level, retain the product index with completely consistent attributes, and obtain the SKU alignment index set; S203: Call the SKU alignment index set, retrieve the original global unique identifier of each product corresponding to each index, execute the data aggregation process, and group and arrange according to product category to generate a set of related products of the same type.

[0009] As a further aspect of the present invention, the step of obtaining S3 is as follows: S301: Call the real-time price area text of each product in the same product association set, apply the preset named entity recognition model to identify the numerical components, currency symbols and price prefixes and suffixes, and identify the original price, discounted price, member price and limited-time flash sale price according to the preset labeling rules to obtain the basic price component. S302: Based on the origin and destination information of the goods, retrieve the preset cross-border tax and fee rule library, extract the corresponding import and export tax rates, value-added tax and consumption tax calculation parameters, and at the same time analyze the first weight, additional weight and last delivery fees in the logistics strategy. Then, associate the extracted parameters with the preset fee calculation formula to obtain the tax and fee component and the logistics cost component. S303: Call the promotional activity description in the product details page, use a natural language processing model to identify discount conditions such as full reduction, discount coupons, multi-item discounts and points deduction, convert each discount condition into a preset numerical deduction operator, and logically arrange them according to the preset redemption priority to obtain a refined price element list.

[0010] As a further aspect of the present invention, the step of obtaining S4 is as follows: S401: Call the currency symbol in the refined price element list, obtain the real-time exchange rate conversion coefficient from the preset financial interface, perform linear multiplication on all basic price components, tax components and logistics cost components, convert them to the preset base currency, and obtain a unified currency price series. S402: Based on the packaging quantity and weight information in the product attributes, perform standardized conversion of unit dimensions, calculate the price benchmark value per unit weight or unit quantity, apply the promotional discount to the unified currency price series using a preset price restoration algorithm, identify the actual payment amount under different promotional combinations, and obtain the normalized price series. S403: Call the normalized price sequence, perform numerical sorting and extreme value detection operations, compare the price difference rate between different platforms, and combine the preset logistics timeliness weight and platform credit evaluation index to perform weighted scoring logic and obtain intelligent price comparison result data.

[0011] As a further aspect of the present invention, the method further includes: S5: Call the historical price fluctuation data in the intelligent price comparison results, combine the future price trend generated by the prediction algorithm, insert it into the monitoring task queue, and when the real-time price triggers the preset warning threshold, execute the abnormal signal sending process to generate the intelligent price comparison dynamic monitoring results of cross-border commodities. The results of the intelligent price comparison and dynamic monitoring of cross-border commodities include real-time price dashboards, threshold trigger records, and fluctuation trend prediction curves.

[0012] As a further aspect of the present invention, the step of obtaining S5 is as follows: S501: Call the historical price sampling points in the intelligent price comparison result data, perform time series analysis, extract the periodic features and random disturbance components of price fluctuations, input the extracted feature vector into the preset long short-term memory network, and obtain the price prediction trend vector; S502: Based on the price prediction trend vector, set dynamic early warning water level lines for different commodities, perform difference calculation between the real-time acquired price components and the dynamic early warning water level lines, record the time when the difference exceeds the preset fluctuation deviation limit and the corresponding price value, and generate a threshold trigger record. S503: Call the threshold trigger record and price prediction trend vector, perform visualization rendering processing, integrate the real-time price, predicted price and early warning status of each platform into a unified display framework, perform data refresh operation according to the preset update frequency, and obtain the dynamic monitoring results of intelligent price comparison of cross-border commodities.

[0013] A deep learning-based intelligent price comparison system for cross-border commodities, comprising: The data acquisition module obtains the original product data stream from the cross-border e-commerce platform through a distributed network interface, extracts the multimodal feature data of the products, and uses a deep learning model to perform cross-language semantic transformation and image encoding processing to construct a feature vector set and generate a multimodal sequence to be matched. The feature alignment module calls the multimodal sequence to be matched, performs feature comparison operation through preset semantic space mapping logic, calculates the degree of association between products based on cosine similarity algorithm, performs fine-grained verification of SKU attributes, indexes and associates products that meet the alignment conditions, and obtains a set of related products of the same type. The element decomposition module calls the price description text in the same product association set, applies the sequence labeling model to extract the basic component, tax component and shipping component in the price composition, and uses the logistic regression model to identify the deduction weight of promotion conditions, transforming unstructured price information into structured data and generating a refined price element list. The standardization module calls the refined price element list, connects to the external real-time exchange rate data source, performs currency conversion and unit alignment operations, restores the real net price according to the preset tax calculation logic and preferential verification rules, performs horizontal comparison through numerical sorting algorithm, and generates intelligent price comparison result data. The monitoring and early warning module calls the intelligent price comparison result data, combines it with the time series prediction model to analyze the price evolution pattern, performs real-time monitoring and threshold judgment operations, executes the signal triggering process when the price deviates by a predetermined amount, updates the price comparison status, and generates dynamic monitoring results of intelligent price comparison for cross-border commodities.

[0014] The beneficial effects of the technical solutions provided in the embodiments of the present invention include at least the following: This invention addresses the issue of insufficient accuracy in identifying identical products in cross-border scenarios due to language differences and varying descriptive habits by constructing a multimodal feature extraction architecture and a cross-language semantic mapping mechanism, thereby significantly improving SKU-level alignment accuracy. By introducing a deep learning sequence labeling model to perform refined decomposition of price components, the basic price, taxes, logistics costs, and promotional weights are structurally separated, overcoming the limitations of existing technologies with coarse granularity when handling complex pricing structures. A normalized price comparison model is established using a dynamic exchange rate matrix and a tax calculation topology graph, achieving uniformity across different currencies and pricing rules, effectively eliminating price deviations caused by inaccurate tax calculations or ineffective promotional restoration. Furthermore, by combining time series analysis with real-time monitoring and early warning triggering, the price comparison system's responsiveness to market fluctuations is enhanced, providing high-precision, real-time, and decision-support-enabled intelligent price comparison assurance for cross-border trade, significantly improving the automation level and data consistency of commodity transaction processing. Attached Figure Description

[0015] Figure 1 This is a flowchart of the method of the present invention; Figure 2 This is a system flowchart of the present invention. Detailed Implementation

[0016] The technical solution of the present invention will now be described with reference to the accompanying drawings.

[0017] In embodiments of the present invention, words such as "exemplarily," "for example," etc., are used to indicate that something is an example, illustration, or description. Any embodiment or design described as "exemplary" in the present invention should not be construed as being more preferred or advantageous than other embodiments or designs. Specifically, the use of the word "exemplary" is intended to present the concept in a concrete manner. Furthermore, in embodiments of the present invention, the meaning expressed by "and / or" can be both, or either one.

[0018] In the embodiments of this invention, the terms "image" and "picture" may sometimes be used interchangeably. It should be noted that, without emphasizing the distinction between them, they convey the same meaning. Similarly, the terms "of," "corresponding (relevant)," and "corresponding" may sometimes be used interchangeably. It should be noted that, without emphasizing the distinction between them, they convey the same meaning.

[0019] In this embodiment of the invention, sometimes a subscript such as W1 may be written in a non-subscript form such as W1. When the difference is not emphasized, the meaning they express is the same.

[0020] To make the technical problems, technical solutions and advantages of the present invention clearer, a detailed description will be given below in conjunction with the accompanying drawings and specific embodiments.

[0021] Please see Figure 1 This invention provides a technical solution: a deep learning-based intelligent price comparison method for cross-border commodities, comprising the following steps: S1: Obtain the original product dataset from multiple cross-border platforms, extract multimodal feature information of products, perform semantic vector conversion on product titles in different languages, encode visual features of product images, establish a basic product feature index library, and output multimodal matching sequences; S2: Call the semantic vectors and visual encodings in the multimodal sequence to be matched, perform cross-language semantic mapping and feature fusion operations, calculate the feature similarity between products on different platforms, filter product items with similarity exceeding the preset alignment threshold, perform precise alignment at the SKU level, and form a set of related products of the same type; S3: Call the pricing page data of each product in the same product association set, apply a deep learning sequence labeling model to identify the basic price, tax rules, shipping template and promotion weight in the price string, perform structured decomposition of complex price components, and output a refined price element list; S4: Call the refined price element list, obtain the real-time exchange rate matrix and the preset tax calculation topology, perform currency conversion, unit alignment and discount strategy restoration, calculate the final price of each aligned product under the unified standard, and generate intelligent price comparison result data. S5: Call the historical price fluctuation data in the intelligent price comparison results, combine the future price trend generated by the prediction algorithm, insert it into the monitoring task queue, and when the real-time price triggers the preset warning threshold, execute the abnormal signal sending process to generate the intelligent price comparison dynamic monitoring results for cross-border commodities.

[0022] The multimodal matching sequence includes cross-language semantic vectors, visual feature encoding sets, and platform attribute tags. The same product association set includes feature alignment index, SKU mapping relationship table, and cross-platform product clusters. The refined price element list includes basic price component, tax component, logistics cost component, and promotional deduction items. The intelligent price comparison result data includes normalized price series, price difference analysis report, and optimal procurement path. The cross-border commodity intelligent price comparison dynamic monitoring results include real-time price dashboard, threshold trigger records, and fluctuation trend prediction curve.

[0023] The steps to obtain S1 are as follows: S101: Obtain raw web page data from multiple preset cross-border e-commerce platforms through distributed data acquisition nodes, and use preset parsing rules to separate the product title string, product main image, detailed description text and real-time price area, and store each modal data in the preset cache area; To obtain raw product data from multiple cross-border e-commerce platforms, the first step is to retrieve the webpage source code data of the target platforms one by one from a pre-set distributed crawler cluster or API interface. This source code data contains all the key information of the product display layer. In practical applications, if the current page being collected is a major overseas e-commerce platform, pre-set XPath or regular expression parsing rules are used to extract the product title string "Wireless Noise Cancelling Headphones", the main image URL, and the price tag area. The extracted text and image data are then preliminarily cleaned to remove HTML tags and redundant noise, and a mapping relationship is established between the product title and the corresponding image binary data. After performing the above operations on multiple platforms, a raw product feature data pool is formed. Each record contains a platform source identifier, a globally unique product identifier, and raw feature fragments for each modality. Finally, the processed modal data is stored in a pre-set high-speed cache area in a structured format.

[0024] S102: Call the preset cross-language pre-trained language model to perform word segmentation and word embedding operations on the product title string, obtain the title feature vector containing contextual semantic information, and call the convolutional neural network to perform feature extraction on the product main image to obtain the image feature vector containing texture, color and composition information; Based on the raw data in the preset cache area, the title string of each product is called sequentially during the feature extraction process. Let the title sequence be... Then, a pre-trained model (such as multilingual BERT) is used to perform word segmentation, converting the word into high-dimensional word embedding vectors. For example, mapping an English title to... The pre-defined residual neural network is synchronously invoked to perform visual feature encoding on the main product image, extracting deep features from the image and obtaining an image feature vector containing the product's appearance attributes. During the feature extraction process, each set of vectors can be normalized to ensure that the vector magnitudes within the same modality are within a consistent range, avoiding model distortion caused by differences in feature scale. After extraction, the feature components of each dimension are recorded in order of product ID, forming a modal feature dataset containing both semantic and visual information, ensuring that each product has a unique digital representation in the feature space.

[0025] S103: The title feature vector and the image feature vector are concatenated, and position encoding and platform attribute weights are introduced. The dimensionality is transformed through a fully connected layer, and the transformed multimodal feature data is classified and numbered according to the platform source to obtain the multimodal matching sequence. Based on the two sets of vector information generated in the feature extraction step, the title feature vector and the image feature vector are retrieved sequentially. When performing the feature fusion action, the vectors of different dimensions need to be concatenated first to form a composite feature vector. Subsequently, a platform attribute weighting factor was introduced. This is used to characterize the differences in description habits across different platforms, and positional encoding is superimposed to preserve the spatial correlation between features. A fully connected neural network layer is used to perform a linear transformation on the composite vector, uniformly compressing its dimensions to a preset matching dimension. According to the above rules, each fused feature vector is matched with a platform source ID and a product index ID. These data are then arranged according to a preset timestamp or category order to form a feature sequence. This sequence is used to represent the distribution of each product in a multidimensional feature space. Finally, all structured feature data are integrated to obtain a multimodal sequence to be matched.

[0026] The steps to obtain S2 are as follows: S201: Call the first platform feature vector and the second platform feature vector in the multimodal sequence to be matched, use the preset shared semantic space mapping matrix to map the vectors of different languages ​​to a unified coordinate system, and perform vector dot product operation to obtain the initial similarity score of the candidate matching pair; Multimodal sequences to be matched are extracted to ensure that product feature vectors from different language platforms are comparable within the same semantic space. The extraction process uses the primary and secondary features of the products as a basis, and calls a pre-defined linear mapping matrix. A spatial transformation is performed on the vectors to map the title vectors in non-native languages ​​such as German and Japanese to the baseline semantic space. Cosine similarity or dot product operations are then performed on the two extracted platform vector pairs. The relationship between the vector dot product and the product of their magnitudes is used to derive the scores, and the results are then processed. With the preset initial similarity threshold Compare them. If satisfied... If a pair of products is found to have the potential to be the same product, its index identifier is recorded. In actual price comparison, if the dot product of the headphone feature vector of platform A and the headphone feature vector of platform B is 0.85, which meets the initial matching condition, it is recorded as a candidate pair. This comparison process is then continuously executed, and finally, a candidate index set of the same product is formed.

[0027] S202: Based on the initial similarity score, filter product pairs whose scores are within the preset candidate range, extract their corresponding SKU attribute list, including color, specifications, material and packaging quantity, and use a combination of string fuzzy matching algorithm and deep learning semantic similarity calculation to determine the attribute consistency status at the SKU level, retain the product index with completely consistent attributes, and obtain the SKU alignment index set; The SKU attribute information of candidate pairs is retrieved, and for each product pair pointed to by an index, its specification parameter list is extracted from the original webpage data. During SKU-level consistency checks, individual attribute items (such as "Color:Midnight Blue" and "Color: Midnight Blue") are used as the judgment unit, and their attribute values ​​are extracted for cross-language alignment. The judgment criterion is to set an attribute consistency score. If the weighted sum of the semantic similarity of attribute names and the literal similarity of attribute values ​​satisfies If both pairs of headphones have the same storage capacity of "256GB" and the same color code, then the SKU pair number is retained. If the packaging quantity differs (e.g., "1 pack" vs. "2 pack"), it is considered inconsistent, and the index is not retained. Following this logic, all candidate pairs are traversed, and records are executed for all product numbers with completely aligned attributes, ultimately resulting in the SKU-aligned index set.

[0028] S203: Call the SKU alignment index set, retrieve the original global unique identifier of the product corresponding to each index, execute the data aggregation process and group and arrange according to product category to generate a set of related products of the same type; Before extracting the original metadata for each product from the SKU alignment index set, it's crucial to ensure that each index points to a valid SKU record in the database. The extracted content includes the product title, SKU ID, affiliated store, and original price link. Based on the index mapping, data aggregation is performed on the extracted identical products, grouping the same product across different platforms into the same product cluster. The product cluster number is then bound to its corresponding platform data and output; for example, the index pair... Aggregate into cluster numbers Repeat the aggregation process to reorganize all SKU aligned data, re-sort it by category level, and output it as a set of related products of the same type, providing an accurate comparison benchmark for subsequent price element breakdown.

[0029] The steps to obtain S3 are as follows: S301: Call the real-time price area text of each product in the same product association set, apply the preset named entity recognition model to identify the numerical components, currency symbols and price prefixes and suffixes, and identify the original price, discounted price, member price and limited-time flash sale price according to the preset labeling rules to obtain the basic price component. The system retrieves real-time price text for each product in the associated set, extracting numerical fragments and currency identifiers using named entity recognition (NID) technology. A pre-defined price tag classifier is then invoked to define attributes for the values ​​based on surrounding modifiers (such as "MSRP," "Sale," and "Member Price"). The extracted values ​​are then bound to currency symbols; for example, "$299.00" is identified as the discounted price, and "$349.00" as the original price. If multiple price components exist, they are filtered according to a pre-defined priority rule (flash sale price > member price > discounted price) to determine the lowest base price available to the current user. This process is repeated for all associated products, extracting basic pricing information from each platform, and finally outputting a list of base price components arranged by product ID.

[0030] S302: Based on the origin and destination information of the goods, retrieve the preset cross-border tax and fee rule library, extract the corresponding import and export tax rates, value-added tax and consumption tax calculation parameters, and at the same time analyze the first weight, additional weight and last delivery fees in the logistics strategy. Then, associate the extracted parameters with the preset fee calculation formula to obtain the tax and fee component and the logistics cost component. Extract the shipping and logistics policies of each platform to determine the origin and destination coordinates and transportation methods. Search the pre-defined cross-border trade rule database and match the corresponding tariff rates based on the product category code. Synchronously parse the tiered pricing logic in the freight template to obtain the initial weight unit price. Additional weight unit price and last-mile delivery surcharge When calculating execution costs, the weight of the goods should be included. Substitute the volume into the formula to calculate the estimated total logistics cost. At the same time, estimated taxes and fees are calculated based on the base price. In practice, for example, if a product is shipped to China with a tax rate of 13% and a weight of 2kg, the corresponding tax and shipping costs are calculated and linked to the product ID to obtain the tax component and the logistics cost component.

[0031] S303: Call the promotional activity description in the product details page, use the natural language processing model to identify discount conditions such as full reduction, discount coupons, multi-item discounts and points deduction, convert each discount condition into a preset numerical deduction operator, and logically arrange them according to the preset redemption priority to obtain a refined price element list; This program retrieves unstructured text descriptions from product detail pages and promotional pages, using an NLP semantic extraction model to identify promotional logic. Descriptions such as "Buy 2 Get 10% Off" or "$20 Coupon" are converted into mathematical operators, such as multiplication operator 0.9 or subtraction operator -20. A discount redemption tree is built, logically arranging the operators according to a preset redemption order (store coupons > platform coupons > payment discounts). The maximum discount for a specific purchase combination is calculated, and the specific amounts of each deduction are recorded. After combining items, all price components (base, tax, fees, deductions) are summarized to ensure that every price comparison dimension is verifiable. Finally, a refined list of price elements arranged by structured fields is output.

[0032] The steps to obtain S4 are as follows: S401: Call the currency symbol in the refined price element list, obtain the real-time exchange rate conversion coefficient from the preset financial interface, perform linear multiplication on all basic price components, tax components and logistics cost components, convert them to the preset base currency, and obtain a unified currency price series. Iterate through each currency identifier in the element list, and for non-base currencies (such as USD, EUR, JPY), call the foreign exchange API in real time to obtain the current mid-rate exchange rate. According to the exchange rate conversion formula Perform linear conversion on each price component. During the calculation process, ensure the timeliness of the exchange rate; set the update frequency to minutes. In the example, if the price on a platform is 100 Euros, the real-time exchange rate... This is converted to a base value of 785 RMB. The same process is followed for taxes and freight, ensuring all amounts are added under the same unit of measurement. The converted values ​​for each component are then output, and finally, a unified currency price sequence is formed.

[0033] S402: Based on the packaging quantity and weight information in the product attributes, perform standardized conversion of unit dimensions, calculate the price benchmark value per unit weight or unit quantity, apply the promotional discount to the unified currency price series using a preset price restoration algorithm, identify the actual payment amount under different promotional combinations, and obtain the normalized price series. Determine the sales packaging specifications for each product and extract the unit quantity. Or net weight When performing dimensional standardization conversion, calculate the benchmark value for the unit price. This eliminates price incomparability caused by different packaging specifications. Then, a preset price restoration logic is applied, sequentially applying the previously identified promotional discount operators to the unified currency price. The detection method simulates a real order placement process to calculate the final payment amount. This is then combined with the unit of measurement. In practice, if one item is a two-pack costing 180 yuan and another is a single-pack costing 100 yuan, the calculated unit prices are 90 yuan and 100 yuan respectively, thus identifying the actual difference in the final price. After processing all product pairs, the final price across all platforms under a unified standard is selected to obtain a normalized price series.

[0034] S403: Call the normalized price series, perform numerical sorting and extreme value detection operations, compare the price difference rate between different platforms, and combine the preset logistics timeliness weight and platform credit evaluation indicators to perform weighted scoring logic and obtain intelligent price comparison result data. The system retrieves the final price score from each platform in the normalized price series and sorts the data in ascending order to identify the platform with the lowest price. Extreme value detection is performed to exclude abnormally low prices (such as incorrect prices or phishing links); if a price deviates from the mean by more than 50%, it is marked as potentially abnormal. Subsequently, a multi-dimensional weighted evaluation system is introduced, with logistics timeliness as a weight. With platform credit weight Calculate the overall score The results are split and sorted, and the optimal purchasing recommendation is extracted from the sequence. The remaining platform indices are then rearranged and displayed in a cluster. All price comparison requests are processed repeatedly, and the final output is intelligent price comparison results data that includes price ranking, detailed breakdown of the final price, and recommendation level.

[0035] The steps to obtain S5 are as follows: S501: Call the historical price sampling points in the intelligent price comparison result data, perform time series analysis, extract the periodic characteristics and random disturbance components of price fluctuations, input the extracted feature vector into the preset long short-term memory network, and obtain the price prediction trend vector; For each aligned commodity, historical price observations within a preset time window are extracted to construct a time series dataset. Sequence decomposition is performed to extract trend components, seasonality components, and residual perturbation terms. The processed sequence data is then input into a Long Short-Term Memory (LSTM) network model, utilizing the model's gating mechanism to capture long-term dependencies in price evolution. Through forward propagation, the model outputs predicted price trends for the next 7 or 30 days, denoted as the trend vector. In practice, if a product regularly raises its price before a promotional season, the Long Short-Term Memory (LSTM) network model will identify this feature and output a corresponding price increase warning signal. This prediction process is repeated to form a price prediction trend vector, providing data support for dynamic monitoring.

[0036] S502: Based on the price prediction trend vector, set dynamic early warning water level lines for different commodities, perform difference calculation between the real-time acquired price components and the dynamic early warning water level lines, record the time when the difference exceeds the preset fluctuation deviation limit and the corresponding price value, and generate threshold trigger records. Based on the mean and variance of the predicted trend vector, the warning threshold for each product is dynamically calculated. For each monitoring task, obtain the latest final price from each platform in real time. And perform a difference calculation between it and the dynamic early warning water level line. Replace the rigid logic of fixed thresholds with a dynamic water level to identify abnormal, sudden price drops or increases. If the deviation exceeds a preset limit (e.g., 15%), a recording action is triggered in real time, saving the current timestamp, platform name, and abnormal price. All price fluctuations are processed according to all monitoring indices, ultimately generating a threshold-triggered record.

[0037] S503: Calls threshold trigger records and price prediction trend vectors, performs visualization rendering processing, integrates real-time prices, predicted prices and alert status of various platforms into a unified display framework, performs data refresh operation according to preset update frequency, and obtains intelligent price comparison dynamic monitoring results of cross-border commodities. The system triggers abnormal data from records based on thresholds, summarizing and organizing them chronologically. For each record, the corresponding product information is visualized and mapped to ensure the predicted curve and real-time sampling points are displayed within a unified coordinate system. First, all price comparison dimensions are re-labeled with status tags (e.g., "stable price," "rapid price drop") to ensure the intuitiveness of the monitoring results. Then, for price comparison data from each platform, multi-dimensional dashboard information is output, including price trends, tax percentages, and shipping fluctuations. During the processing, all warning messages are sorted by severity, and data from each platform is integrated into the monitoring dashboard framework. If it's a price drop warning, a "buy the dip" flag is added; otherwise, it's marked "under monitoring." Finally, all integrated data is output as a dynamic monitoring result for intelligent price comparison of cross-border goods, enabling real-time control of price comparison status.

[0038] Please see Figure 2 A deep learning-based intelligent price comparison system for cross-border goods, comprising: The data acquisition module obtains the original product data stream from the cross-border e-commerce platform through a distributed network interface, extracts the multimodal feature data of the products, and uses a deep learning model to perform cross-language semantic transformation and image encoding processing to construct a feature vector set and generate a multimodal sequence to be matched. The feature alignment module calls the multimodal sequence to be matched, performs feature comparison operation through the preset semantic space mapping logic, calculates the degree of association between products based on the cosine similarity algorithm, performs fine-grained verification of SKU attributes, indexes and associates products that meet the alignment conditions, and obtains the set of related products of the same type. The element decomposition module calls the price description text in the associated set of the same product, applies the sequence labeling model to extract the basic components, tax components and shipping components in the price composition, and uses the logistic regression model to identify the deduction weight of promotion conditions, transforming unstructured price information into structured data and generating a refined price element list. The standardization module calls a refined price element list, connects to an external real-time exchange rate data source, performs currency conversion and unit alignment operations, restores the true net price according to the preset tax calculation logic and preferential verification rules, and performs horizontal comparison through a numerical sorting algorithm to generate intelligent price comparison result data. The monitoring and early warning module calls up intelligent price comparison results data, combines time series prediction models to analyze price evolution patterns, performs real-time monitoring and threshold judgment operations, and executes signal triggering processes when prices deviate by a predetermined range, updates the price comparison status, and generates dynamic monitoring results for intelligent price comparison of cross-border commodities.

[0039] The above description is merely a specific embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the technical scope disclosed in the present invention should be included within the scope of protection of the present invention. Therefore, the scope of protection of the present invention should be determined by the scope of the claims.

Claims

1. A deep learning-based intelligent price comparison method for cross-border commodities, characterized in that, Includes the following steps: S1: Obtain the original product dataset from multiple cross-border platforms, extract multimodal feature information of products, perform semantic vector conversion on product titles in different languages, encode visual features of product images, establish a basic product feature index library, and output multimodal matching sequences; S2: Call the semantic vectors and visual codes in the multimodal sequence to be matched, perform cross-language semantic mapping and feature fusion operations, calculate the feature similarity between products on different platforms, filter product items with similarity exceeding the preset alignment threshold, perform precise alignment at the SKU level, and form a set of related products of the same type; S3: Call the pricing page data of each product in the same product association set, apply a deep learning sequence labeling model to identify the basic price, tax rules, shipping template and promotion weight in the price string, perform structured decomposition of complex price components, and output a refined price element list; S4: Call the refined price element list, obtain the real-time exchange rate matrix and the preset tax calculation topology, perform currency conversion, unit dimension alignment and preferential strategy restoration operations, calculate the final price of each aligned product under the unified standard, and generate intelligent price comparison result data.

2. The intelligent price comparison method for cross-border commodities based on deep learning according to claim 1, characterized in that: The product feature description sequence includes visual representation vectors, text semantic encoding, attribute tag sets, and platform source identifiers; the same product association mapping matrix includes cross-language semantic mapping, SKU attribute alignment index, same product confidence score, and category affiliation identifier; the multi-dimensional price composition vector includes FOB benchmark price, import tariff amount, value-added tax amount, international logistics fee, last-mile delivery fee, and tiered discount amount; the product price comparison analysis report includes normalized price ranking, price fluctuation trend map, optimal procurement suggestions, and tax percentage analysis.

3. The intelligent cross-border commodity price comparison method based on deep learning according to claim 1, characterized in that, The steps for obtaining S1 are as follows: S101: Obtain raw web page data from multiple preset cross-border e-commerce platforms through distributed data acquisition nodes, and use preset parsing rules to separate the product title string, product main image, detailed description text and real-time price area, and store each modal data in the preset cache area; S102: Call the preset cross-language pre-trained language model to perform word segmentation and word embedding operations on the product title string to obtain a title feature vector containing contextual semantic information. At the same time, call the convolutional neural network to perform feature extraction on the product main image to obtain an image feature vector containing texture, color and composition information. S103: The title feature vector and the image feature vector are concatenated, and position encoding and platform attribute weights are introduced. The dimensionality is transformed through a fully connected layer, and the transformed multimodal feature data is classified and numbered according to the platform source to obtain the multimodal matching sequence.

4. The intelligent cross-border commodity price comparison method based on deep learning according to claim 1, characterized in that, The steps for obtaining S2 are as follows: S201: Call the first platform feature vector and the second platform feature vector in the multimodal sequence to be matched, use the preset shared semantic space mapping matrix to map the vectors of different languages ​​to a unified coordinate system, and perform vector dot product operation to obtain the initial similarity score of the candidate matching pair; S202: Based on the initial similarity score, filter product pairs whose scores are within the preset candidate interval, extract their corresponding SKU attribute list, including color, specifications, material and packaging quantity, and use a combination of string fuzzy matching algorithm and deep learning semantic similarity calculation to determine the attribute consistency status at the SKU level, retain the product index with completely consistent attributes, and obtain the SKU alignment index set; S203: Call the SKU alignment index set, retrieve the original global unique identifier of each product corresponding to each index, execute the data aggregation process, and group and arrange according to product category to generate a set of related products of the same type.

5. The intelligent cross-border commodity price comparison method based on deep learning according to claim 1, characterized in that, The steps for obtaining S3 are as follows: S301: Call the real-time price area text of each product in the same product association set, apply the preset named entity recognition model to identify the numerical components, currency symbols and price prefixes and suffixes, and identify the original price, discounted price, member price and limited-time flash sale price according to the preset labeling rules to obtain the basic price component. S302: Based on the origin and destination information of the goods, retrieve the preset cross-border tax and fee rule library, extract the corresponding import and export tax rates, value-added tax and consumption tax calculation parameters, and at the same time analyze the first weight, additional weight and last delivery fees in the logistics strategy. Then, associate the extracted parameters with the preset fee calculation formula to obtain the tax and fee component and the logistics cost component. S303: Call the promotional activity description in the product details page, use a natural language processing model to identify discount conditions such as full reduction, discount coupons, multi-item discounts and points deduction, convert each discount condition into a preset numerical deduction operator, and logically arrange them according to the preset redemption priority to obtain a refined price element list.

6. The intelligent cross-border commodity price comparison method based on deep learning according to claim 1, characterized in that, The steps for obtaining S4 are as follows: S401: Call the currency symbol in the refined price element list, obtain the real-time exchange rate conversion coefficient from the preset financial interface, perform linear multiplication on all basic price components, tax components and logistics cost components, convert them to the preset base currency, and obtain a unified currency price series. S402: Based on the packaging quantity and weight information in the product attributes, perform standardized conversion of unit dimensions, calculate the price benchmark value per unit weight or unit quantity, apply the promotional discount to the unified currency price series using a preset price restoration algorithm, identify the actual payment amount under different promotional combinations, and obtain the normalized price series. S403: Call the normalized price sequence, perform numerical sorting and extreme value detection operations, compare the price difference rate between different platforms, and combine the preset logistics timeliness weight and platform credit evaluation index to perform weighted scoring logic and obtain intelligent price comparison result data.

7. The intelligent cross-border commodity price comparison method based on deep learning according to claim 1, characterized in that, The method further includes: S5: Call the historical price fluctuation data in the intelligent price comparison results, combine the future price trend generated by the prediction algorithm, insert it into the monitoring task queue, and when the real-time price triggers the preset warning threshold, execute the abnormal signal sending process to generate the intelligent price comparison dynamic monitoring results of cross-border commodities. The results of the intelligent price comparison and dynamic monitoring of cross-border commodities include real-time price dashboards, threshold trigger records, and fluctuation trend prediction curves.

8. The intelligent cross-border commodity price comparison method based on deep learning according to claim 7, characterized in that, The steps for obtaining S5 are as follows: S501: Call the historical price sampling points in the intelligent price comparison result data, perform time series analysis, extract the periodic features and random disturbance components of price fluctuations, input the extracted feature vector into the preset long short-term memory network, and obtain the price prediction trend vector; S502: Based on the price prediction trend vector, set dynamic early warning water level lines for different commodities, perform difference calculation between the real-time acquired price components and the dynamic early warning water level lines, record the time when the difference exceeds the preset fluctuation deviation limit and the corresponding price value, and generate a threshold trigger record. S503: Call the threshold trigger record and price prediction trend vector, perform visualization rendering processing, integrate the real-time price, predicted price and early warning status of each platform into a unified display framework, perform data refresh operation according to the preset update frequency, and obtain the dynamic monitoring results of intelligent price comparison of cross-border commodities.

9. A cross-border commodity intelligent price comparison system based on deep learning, characterized in that: The system is used in the deep learning-based intelligent price comparison method for cross-border commodities as described in any one of claims 1-8, and the system comprises: The data acquisition module is used to acquire the raw product data stream from the cross-border e-commerce platform, extract the multimodal feature data of the products, and use a deep learning model to perform cross-language semantic transformation and image encoding processing to construct a feature vector set and generate a multimodal sequence to be matched. The feature alignment module is used to call the multimodal sequence to be matched, perform feature comparison operation through preset semantic space mapping logic, calculate the feature similarity between products, perform fine-grained verification of SKU attributes, index and associate products that meet the alignment conditions, and obtain a set of related products of the same type. The element decomposition module is used to call the price description text in the associated set of the same product, apply the sequence labeling model to extract the basic components, tax components and shipping components in the price composition, identify the deduction weight of promotional conditions, transform unstructured price information into structured data, and generate a refined price element list. The standardization module is used to call the refined price element list, access the real-time exchange rate data source, perform currency conversion and unit alignment operations, restore the true final price, perform horizontal comparison through numerical sorting algorithm, and generate intelligent price comparison result data. The monitoring and early warning module is used to call the intelligent price comparison result data, combine it with the time series prediction model to analyze the price evolution pattern, perform real-time monitoring and threshold judgment operations, and execute the signal triggering process when the price deviates by a predetermined amount to generate dynamic monitoring results of intelligent price comparison for cross-border commodities.

Images