A multi-source information verification commodity recommendation method, system, device and medium

Abstract

This application provides a product recommendation method, system, device, and medium based on multi-source information verification, relating to the field of information processing technology. The method includes: verifying the integrity of basic and related information of target products sent by each merchant; generating a first and second score for each target product based on the verification results; verifying the credibility of the basic and related information and generating a third score; weighted summing of the three scores to obtain a credibility score for each target product; adjusting the credibility scores based on the credit index of each merchant to generate a target credibility score; selecting candidate target products corresponding to the user's query request based on the user's request; sending the initial information of candidate target products with target credibility scores higher than a preset score to a large model; and enabling the large model to generate recommended content based on the initial information and push it to the user's end. The technical effect of this application is to ensure that users obtain authentic and valid product information and make correct purchasing decisions.

Description

Technical Field

[0001] This application relates to the field of information processing technology, specifically to a product recommendation method, system, device, and medium for multi-source information verification. Background Technology

[0002] With the rapid development of e-commerce and the increasing popularity of online shopping, finding products that meet users' needs quickly and accurately from a vast amount of product information has become a significant challenge. Traditional product recommendation systems typically rely on users' historical behavior data and collaborative filtering algorithms for recommendations. However, this approach often overlooks the authenticity and accuracy of the product information itself. This can lead to users receiving recommendations based on false parameters, exaggerated claims, or outdated data, severely impacting the quality of their purchasing decisions and shopping experience. Especially in the highly competitive e-commerce environment, some merchants may engage in dishonest practices such as tampering with product parameters, fabricating sales figures and reviews, and using invalid qualification documents to increase product exposure. This results in recommendation systems facing the dilemma of inconsistent product information quality. How to effectively verify the authenticity and reliability of product information during the recommendation process, filter out low-quality and false information, and recommend truly trustworthy products to users has become an urgent technical problem to be solved.

[0003] Currently, large language models are typically introduced to generate more natural, fluent, and personalized product recommendation copy, improving the readability and appeal of the recommended content. While these methods can improve the expression quality of the recommended content and the user reading experience, the lack of a systematic evaluation of the quality of the input information source may result in the direct use of product information of varying quality or even containing false information to generate recommended content. This leads to the generation of recommended content containing inaccurate or unreliable product parameters and attribute descriptions, affecting users' ability to obtain accurate and valid product information and make correct purchasing decisions. Summary of the Invention

[0004] This application provides a product recommendation method, system, device, and medium for multi-source information verification, which is used to ensure that users obtain true and valid product information and make correct purchasing decisions.

[0005] In a first aspect, this application provides a product recommendation method based on multi-source information verification. The method includes: obtaining initial information of target products sent by various merchants, the initial information including basic information and related information, wherein the related information is information generated by association retrieval based on the basic information; performing integrity verification on the basic information and the related information respectively, and generating a first score and a second score for each target product based on the integrity verification result; performing credibility verification on the basic information and the related information, and generating a third score for each target product based on the credibility verification result; performing a weighted summation of the first score, the second score, and the third score to generate a credibility score for each target product; adjusting the credibility score based on the credit index of each merchant to generate a target credibility score; receiving a query request for target products sent by a user terminal, selecting candidate target products corresponding to the query request from among the target products, sending the initial information of candidate target products with target credibility scores higher than a preset score to a large model, so that the large model generates recommended content based on the initial information, and pushing the recommended content to the user terminal.

[0006] By adopting the above technical solution, initial information containing basic and related information of the target product is obtained. Completeness verification of the basic and related information is performed to generate a first and second score, while credibility verification generates a third score. The three scores are then weighted and summed to obtain a credibility score, which is then adjusted based on the merchant's credit index to generate the target credibility score. This allows for a comprehensive evaluation of product information quality from multiple dimensions, including information completeness and data credibility. It effectively identifies and filters product information containing false parameters, incomplete descriptions, or low-credibility content. Upon receiving a user query request, only the initial information of candidate target products with a target credibility score higher than the preset score is sent to the large model to generate recommendation content. This ensures that the recommendation content is generated based on high-quality and reliable product information, thereby providing users with accurate and reliable product recommendations, improving the quality of users' purchase decisions and shopping experience. This solves the technical problem in existing technologies where insufficient assurance of product information quality leads to potentially inaccurate or unreliable descriptions in the recommendation content, ensuring that users obtain authentic and valid product information and make correct purchase decisions.

[0007] Secondly, this application provides a product recommendation system with multi-source information verification, the system comprising: an acquisition module, a first verification module, a second verification module, a generation module, and an output module; wherein, The acquisition module is used to acquire initial information of target products sent by each merchant. The initial information includes basic information and related information, and the related information is generated by related retrieval based on the basic information. The first verification module is used to perform integrity verification on the basic information and the related information respectively, and generate a first score and a second score for each target product based on the integrity verification result. The second verification module is used to perform credibility verification on the basic information and the related information, and generate a third score for each target product based on the credibility verification result. The generation module is used to perform a weighted sum of the first score, the second score, and the third score to generate a credibility score for each target product. The credibility score is adjusted according to the credit index of each merchant to generate a target credibility score. The output module is used to receive query requests for target products sent by the user terminal, select candidate target products corresponding to the query request from the target products, and send the initial information of candidate target products with a target credibility score higher than a preset score to the large model so that the large model can generate recommended content based on the initial information and push the recommended content to the user terminal.

[0008] Thirdly, this application provides an electronic device that adopts the following technical solution: it includes a processor, a memory, a user interface, and a network interface. The memory is used to store instructions, the user interface and the network interface are used to communicate with other devices, and the processor is used to execute the instructions stored in the memory to enable the electronic device to execute a computer program of any of the above-described multi-source information verification product recommendation methods.

[0009] Fourthly, this application provides a computer-readable storage medium that employs the following technical solution: storing a computer program capable of being loaded by a processor and executing any of the above-mentioned multi-source information verification product recommendation methods.

[0010] In summary, this application includes at least one of the following beneficial technical effects: By predicting the power generation of the generator within a preset time period and combining it with the remaining power to generate the total power, proactive management of power resources for IoT devices is achieved, avoiding unexpected equipment downtime due to inaccurate power estimation. Through a hierarchical and progressive strategy optimization mechanism, a first operating strategy is generated based on the total power constraint to ensure basic equipment operation. Then, a second operating strategy is generated using historical monitoring data to improve the targeting of monitoring. Finally, the target operating strategy is generated by combining current crop growth information and environmental information to predict state characteristics, achieving a comprehensive balance and optimization from power constraints to monitoring needs. This dynamic and adaptive strategy adjustment mechanism enables IoT devices to intelligently allocate monitoring resources according to the actual crop growth status and environmental changes, avoiding spatiotemporal mismatches between configuration and actual monitoring needs, and improving overall monitoring effectiveness and equipment operating efficiency. Attached Figure Description

[0011] Figure 1 This is a flowchart illustrating a product recommendation method based on multi-source information verification provided in an embodiment of this application. Figure 2 This is a complete technical architecture diagram of a product information credibility assessment and intelligent recommendation system based on a large model, provided in an embodiment of this application; Figure 3 This is a schematic diagram of a layered architecture of a multimedia data access network hardware stack provided in an embodiment of this application; Figure 4 This is a schematic diagram of the structure of a product recommendation system with multi-source information verification provided in an embodiment of this application; Figure 5 This is a schematic diagram of the structure of an electronic device provided in an embodiment of this application.

[0012] Explanation of reference numerals in the attached figures: 1000, electronic device; 1001, processor; 1002, communication bus; 1003, user interface; 1004, network interface; 1005, memory. Detailed Implementation

[0013] To enable those skilled in the art to better understand the technical solutions in this specification, the technical solutions in the embodiments of this specification will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, and not all embodiments.

[0014] In the description of the embodiments in this application, words such as "illustrative," "for example," or "for example" are used to indicate examples, illustrations, or explanations. Any embodiment or design described as "illustrative," "for example," or "for example" in the embodiments of this application should not be construed as being more preferred or advantageous than other embodiments or designs. Rather, the use of words such as "illustrative," "for example," or "for example" is intended to present the relevant concepts in a specific manner.

[0015] Figure 1 This is a flowchart illustrating a product recommendation method based on multi-source information verification provided in an embodiment of this application. Figure 1 As shown, the method includes S101-S105: S101, Obtain the initial information of the target products sent by each merchant. The initial information includes basic information and related information. The related information is generated by performing related searches based on the basic information.

[0016] First, a high-concurrency streaming heterogeneous data ingestion gateway deployed on the access side receives initial information about target products reported by various merchants through multiple terminal devices in real time and in parallel. The target products referred to here are specific physical goods that merchants wish to sell and promote on e-commerce or supply chain platforms, such as agricultural products, industrial parts, or consumer goods from specific production areas. The reason for building such a data ingestion mechanism is that in actual business scenarios, merchants are distributed in different geographical locations and use diverse types of data reporting devices, including PC-based web form filling systems, mobile smart terminal barcode scanning applications, and IoT sensor devices deployed at production sites. When these heterogeneous data sources simultaneously send data to the platform at high frequencies, the lack of a unified access gateway for buffering and initial data processing can lead to backend servers experiencing processing blockages or even system crashes when faced with sudden data surges. Therefore, the system architecture is configured with a streaming data ingestion gateway based on message queue middleware (such as Apache Kafka or RocketMQ). This gateway can continuously receive raw data streams from multiple ends in an asynchronous, non-blocking manner and perform peak smoothing and valley filling in an in-memory queue to ensure high availability of data collection and overall system stability.

[0017] Once the data stream enters the ingestion gateway, the system begins structured parsing and semantic alignment of the raw data, breaking it down into two core components: basic information and related information. The basic information refers to first-hand data actively submitted or uploaded by merchants, directly describing the attributes of the target product. This includes the product's name, category, brand, specifications, production date, declaring entity (i.e., the merchant's corporate identity), production area name (i.e., the product's geographical origin), and key physicochemical parameters described by the merchant (e.g., the effective ingredient content of agricultural products, material tolerances of industrial products, etc.). In addition, basic information also includes multimodal files uploaded by merchants, such as unstructured data attachments like product photos, PDF scans of quality inspection reports, and images of production licenses. The system uniformly labels this raw data input by merchants as basic information because it constitutes the most fundamental level of the product data profile, serving as the starting point and basis for all subsequent verification and analysis work.

[0018] However, relying solely on basic information provided by merchants has significant technical limitations. In actual business operations, due to market competition or marketing purposes, merchants often tend to exaggerate product quality parameters, fabricate origins, or selectively disclose core indicators when filling out forms, resulting in the inability to guarantee the authenticity and completeness of basic information. More importantly, if the large language model builds its corpus based solely on this thin and unverified basic information during product recommendation pre-training or retrieval enhancement generation, it will lack rich contextual background knowledge and multi-dimensional objective factual support, failing to fully understand the deep industrial logic and regional cultural characteristics behind the products. This will ultimately lead to misunderstandings when faced with complex user queries, and even generate false recommendations that seriously deviate from objective facts. To fundamentally solve this technical bottleneck, the system, upon receiving basic information, does not directly input it as final data into the database, but immediately initiates an automated retrieval and injection mechanism for related information.

[0019] Specifically, the system internally deploys a two-key adaptive routing engine. This engine automatically extracts the applicant unit and production area name—two key identifiers—from the structured fields of the basic information as composite foreign keys. The reason for choosing the applicant unit and production area name as composite foreign keys is that the applicant unit uniquely identifies the product's production entity, while the production area name pinpoints the product's geographical origin. Their combination can accurately locate the specific link in the industrial chain and geographical niche to which the product belongs. The system then automatically drills down to access a distributed static benchmark knowledge base. This knowledge base is a pre-built and independently running high-performance data storage cluster, containing authoritative benchmark data released by official industry institutions, publicly available standard geographic data from the National Geographic Information System, and industry encyclopedic knowledge reviewed by experts. Through a two-key joint query, the system can automatically retrieve and extract macro-environmental parameters and cultural background data related to the product from the knowledge base. This data is then marked as associated information by the system.

[0020] The specific content of the associated information includes two main dimensions. The first dimension is the macro-environmental parameter scalar field, which refers to objective natural environmental characteristic data associated with the geographical location of the commodity's production area. Examples include static environmental parameters such as the topographic elevation distribution of the production area, the content range of specific mineral components in the soil, the annual average precipitation, and accumulated temperature (i.e., the cumulative value of daily average temperature above biological zero during the growing season). The system needs to incorporate these macro-environmental parameters because, for categories such as agricultural products and geographical indication products, their quality characteristics are strongly correlated with the natural environmental conditions of the production area. For example, crops grown under specific altitude and accumulated temperature conditions will form unique patterns of effective component accumulation. This environmental background information is precisely the key contextual knowledge for the large model to understand "why the commodity has such physicochemical indicators." The second dimension is cultural background data, including narrative knowledge such as the traditional processing methods, historical and cultural lineage, and intangible cultural heritage certification information of the production area. For example, for traditional fermented tea products, the traditional techniques in their production process, such as fermentation time and temperature control, significantly affect the flavor and composition of the final product. This cultural and technological background information can help the large model more accurately describe the unique value and differentiated advantages of the product when generating recommendation content.

[0021] After automatically completing the associated retrieval through the routing engine, the system seamlessly injects the extracted macro-environmental parameters and cultural background data into the basic information data structure, performing feature cascading fusion to ultimately form complete initial information containing both basic and associated information. This automated cascading process is entirely driven by computer programs, requiring no manual intervention. Through standardized database JOIN operations and feature concatenation algorithms, the system integrates heterogeneous information originally scattered across different data sources into a unified high-dimensional feature matrix. Compared to the original basic information, the initial information after the associated information injection expands its data dimension from a single product self-description parameter to a three-dimensional knowledge graph encompassing spatiotemporal environmental background and cultural and technological context, significantly improving the semantic density and information completeness of the corpus.

[0022] S102, perform integrity verification on the basic information and related information respectively, and generate the first score and the second score for each target product based on the integrity verification results.

[0023] Specifically, the system first initiates a three-dimensional integrity check process for the basic information, which includes three progressively layered verification operators. The first-level verification operator performs parameter field integrity checks. Its core working principle is to match and compare the basic information data structure of the target product with a predefined product category specification template field by field. The product category specification template mentioned here refers to a structured data dictionary pre-built according to national standards, industry specifications, and platform business rules. This dictionary specifies in detail the list of required fields, the list of optional fields, and the data type definition of each field for each product category. For example, for agricultural products, the specification template requires that required fields must include core descriptive parameters such as product name, place of origin, production date, shelf life, and net content. For electronic products, it requires that key technical indicators such as brand and model, manufacturer, power parameters, and safety certification marks must be included. The system uses a field traversal algorithm to check each field in the basic information for null or missing fields. For each detected missing field, the system accumulates points based on the field's importance level in the specified template. The absence of critical, mandatory fields triggers a significant point deduction, while the absence of less important, optional fields only results in a minor point deduction. This design is based on the fact that different fields contribute significantly differently to product identification and quality judgment. For example, the absence of the production date field makes it impossible to determine whether the product is within its expiration date, constituting a serious integrity defect. In contrast, while the absence of the packaging material field affects information richness, it does not fundamentally hinder basic product identification and recommendation decisions.

[0024] The second-layer verification operator performs numerical range validity checks. This operator specifically verifies the reasonableness of value boundaries for numerical parameter fields in the basic information. The system extracts the physically valid or statistically reasonable ranges corresponding to each numerical field from the standard template. For example, the physically valid range for the moisture content parameter of agricultural products is usually between 5% and 95%. Values ​​outside this range are physically impossible. The industry statistically reasonable range for a specific agricultural product may be further narrowed to 10% to 20%. Although physically possible, values ​​outside this range are statistically extreme outliers. The system uses a threshold comparison algorithm to check whether the value of each numerical field in the basic information falls within the valid range. For values ​​outside the physically valid range, the system determines it as a hard violation, directly marks the field with zero points and triggers a data anomaly label. For values ​​inside the physically valid range but outside the statistically reasonable range, the system determines it as a suspected anomaly, deducts points appropriately, and adds a warning label for subsequent manual review. The necessity of this layer of verification lies in the fact that even if merchants fill in all the required fields, if the values ​​they fill in clearly violate physical laws or industry common sense, such as claiming that a certain agricultural product has a protein content as high as 95%, such data, although complete in form, is in fact invalid or even misleading junk data. If it is directly input into the large language model without legality filtering, it will pollute the model's training corpus or retrieval knowledge base, causing the model to learn incorrect parameter distribution patterns, and thus produce factual errors when generating recommended content.

[0025] The third-layer verification operator performs multimodal file metadata fingerprint verification. This operator is specifically designed for in-depth technical verification of unstructured multimodal attachments such as images and PDF documents contained in the basic information. The system first extracts the embedded metadata area of ​​each multimodal file using a file parsing library. For image files, the system focuses on extracting device fingerprint information from their EXIF ​​metadata segments, including hardware features such as the manufacturer's identifier, model code, and lens parameters of the shooting device, as well as environmental parameters such as the shooting timestamp, GPS geographic coordinates, and image resolution. EXIF, short for Exchangeable Image File Format, is a standardized metadata format automatically written to image files by digital cameras and smartphones when taking photos, recording the technical parameters and environmental context information of the shooting process. The system needs to verify this metadata because, in real-world business scenarios, some dishonest merchants, in order to enhance the image of their products, will download high-quality images taken by others from the internet or use image editing software to significantly modify or even fabricate original images. These operations often lead to abnormal characteristics in the image's metadata, such as the complete deletion of EXIF ​​data, timestamps that are significantly inconsistent with the shooting time claimed by the merchant, and GPS coordinates showing a shooting location thousands of kilometers away from the claimed production area—clear contradictions. The system uses a specially designed metadata consistency verification algorithm to compare the extracted device fingerprint with the device fingerprints of images uploaded by the merchant in the past. If multiple product images uploaded by the same merchant are found to be from completely different shooting devices with an abnormally high frequency of device changes, the system determines that there is a risk of suspicious image source and deducts points accordingly. Simultaneously, the system performs geographic matching verification between the GPS coordinates of the image and the production area name filled in the basic information. By calling the reverse geocoding interface of the geographic information system, the GPS latitude and longitude coordinates are converted into the corresponding administrative division name, and then compared with the production area name. If the geographical difference between the two exceeds the preset spatial tolerance threshold, such as a distance of more than 50 kilometers, it is judged as an abnormal geographical inconsistency, triggering a deduction mechanism. For PDF format quality inspection reports or certificate files, the system extracts fields such as creation time, modification time, and author information from its document attribute metadata to check whether the document's creation time is earlier than the product's production date. If the quality inspection report's creation time is later than the product's claimed production date by months or even years, it clearly violates the conventional logic that quality inspection is carried out during or immediately after production. Based on this, the system judges that the quality inspection report may be a forged report or a report from another batch of products, and imposes severe deductions.

[0026] Through the cascading execution of the three verification operators described above, the system conducts a comprehensive technical review of the completeness, legality, and authenticity of the basic information. Based on the verification results at each layer, cumulative deductions are calculated to ultimately generate the first score for each target product. The specific scoring algorithm employs an initial full-score deduction model. The system presets an initial first score value for each target product, such as 100 points. Then, it deducts points item by item based on the number and importance level of missing fields found in the field completeness check. It also accumulates deductions based on the number and severity of out-of-bounds parameters found in the numerical range legality check. Additional deductions are made based on the number and type of abnormal features found in the multimodal file metadata fingerprint verification. All deductions are summed and subtracted from the initial value to obtain the final first score. The physical meaning of this score reflects the comprehensive quality level of the basic information in three dimensions: structural completeness, numerical reasonableness, and source authenticity. A higher first score indicates that the basic information provided by the merchant is more complete and standardized, and there is no obvious suspicion of fraud. A lower first score indicates that the basic information has serious quality defects or may even contain malicious fraudulent elements, requiring close attention or direct removal of the product in subsequent processes.

[0027] After verifying the basic information and generating the first score, the system immediately initiates an independent verification process for the related information. Unlike the three-dimensional verification strategy for basic information, the verification of related information focuses on the data coverage dimension. This is because related information is objective background data extracted from a static knowledge base through automated retrieval. Its numerical accuracy and source authenticity have already been guaranteed during the knowledge base construction phase through expert review and certification by authoritative channels. Therefore, it does not require repeated verification of its legality and authenticity. However, the quality issues of related information mainly manifest in the imbalance of coverage dimensions and knowledge blind spots in specific production areas or product categories. Specifically, the system predefines a standard list of related dimensions that should be included in the related information. This list covers multiple sub-dimensions of macro-environmental parameters and cultural background data. For example, under the macro-environmental parameter dimension, the system requires that the related information should at least include three basic environmental sub-dimensions: topographic elevation data, soil composition data, and climate statistics data. Under the cultural background data dimension, the system requires at least two core cultural sub-dimensions: traditional process data and historical cultural lineage data. The system uses a dimension traversal algorithm to check which sub-dimensions are actually included in the associated information of the current target product. For each sub-dimension that is required in the standard list but is actually missing, the system counts it as a dimension coverage missing event.

[0028] The system further calculates the data coverage rate for each related dimension. This indicator is calculated by dividing the actual number of included sub-dimensions by the total number of sub-dimensions required by the standard list, resulting in a coverage rate value between zero and one. For example, if topographic elevation data and soil composition data are successfully retrieved in the related information of a target product, but climate statistics are missing, and only traditional process data is retrieved under the cultural dimension while historical cultural lineage data is missing, then the data coverage rate for this product in the macro-environmental parameters dimension is two-thirds (approximately 0.67%), and the data coverage rate in the cultural background data dimension is one-half (approximately 0.5%). After calculating the data coverage rate for each related dimension separately, the system performs a weighted average based on preset dimension weights to obtain the overall related information coverage rate comprehensive index for the target product. The reason for setting differentiated weights for different related dimensions is that different types of products have significantly different degrees of dependence on background information in each dimension. For example, for geographical indication agricultural products, their quality is highly correlated with natural environmental conditions; therefore, the macro-environmental parameters dimension should be given a higher weight. For industrial standardized products, the importance of cultural background data is relatively low, and the corresponding dimension's weight can be appropriately reduced.

[0029] The system then compares the calculated data coverage rate with a preset coverage threshold, which is typically set at a reasonable, slightly above-average level based on business experience and model training needs, such as 0.7 or 0.8. If the coverage rate of the target product's related information reaches or exceeds this threshold, it indicates that the product's background knowledge injection is relatively complete and comprehensive, providing sufficient contextual understanding material for the large language model. Based on this, the system assigns the product a higher second score, such as a score of 80 or higher on a 100-point scale. If the coverage rate is below the threshold but above the minimum acceptable level, such as between 0.5 and 0.7, it indicates that although some dimensions of the related information are missing, the main background knowledge framework is still preserved. The system assigns a moderate second score, such as between 50 and 80, with the specific score showing a direct linear correlation with the coverage rate. If the coverage rate is below the minimum acceptable threshold, such as below 0.5, it indicates that the associated information is seriously incomplete. The product is missing more than half of the background knowledge dimensions. The system determines that the product is not suitable as a high-quality recommendation material and gives it a lower second score, such as a score below 50. In extreme cases, it may even be given a punitive score of zero or close to zero.

[0030] By quantitatively detecting the coverage of associated information data and comparing thresholds, the system generates a second score for each target product. The physical meaning of this score is to reflect the completeness of the background information injection for the product at the knowledge base level. The higher the second score, the more abundant and comprehensive the environmental and cultural background knowledge that the system has successfully associated with the product. The large language model can obtain more sufficient semantic context support when processing the product. The lower the second score, the more data blind spots there are in the knowledge base for the production area or category corresponding to the product. There are obvious shortcomings in the injection of associated information. When the model understands and recommends the product, it may produce a one-sided understanding or logical gaps due to the lack of necessary background knowledge.

[0031] It is worth emphasizing that the independent generation mechanism of the first and second ratings has significant diagnostic and managerial value. When a target product has a low first rating but a high second rating, it indicates that the product's quality issue primarily stems from inaccurate or falsified data entry by the merchant, while the system's knowledge base is sound. In this case, a data rectification notice should be sent to the merchant, requiring them to supplement and improve basic information or provide more authentic supporting materials. Simultaneously, the merchant's credit rating can be downgraded based on their historical first rating accumulation. Conversely, when a target product has a high first rating but a low second rating, it indicates that the basic information provided by the merchant is complete and standardized, but the problem lies in the system's knowledge base's insufficient background knowledge of the product's production area or category. In this case, a knowledge base expansion task should be triggered, assigning data engineers to specifically supplement the environmental parameters and cultural data for that production area or category, without imposing any penalties or negative evaluations on the merchant. This hierarchical responsibility scoring mechanism effectively avoids attributing the system's own shortcomings to the merchant, ensuring the fairness and interpretability of the scoring results.

[0032] Based on the above embodiments, as an optional implementation, in S102, the integrity verification of basic information and related information is performed respectively, and the first score and second score of each target product are generated according to the integrity verification results, specifically including S21-S22: S21, perform integrity checks and numerical range validity checks on each parameter field in the basic information, and perform metadata fingerprint verification on the multimodal files in the basic information. Based on the results of integrity checks, validity checks, and metadata fingerprint verification, generate the first score for each target product.

[0033] The system first performs integrity checks and numerical range validity checks on each parameter field in the basic information. The parameter fields referred to here are various product attribute data stored in structured form in the basic information, including standardized descriptive fields such as product name, price, inventory quantity, size specifications, weight, color, and material. Integrity checks involve the system traversing the data structure of the basic information, checking each required parameter field for anomalies such as null values, missing values, or placeholders. For example, it checks whether the price field is empty or zero, or whether the product name only contains the default text "Unnamed Product." Fields with integrity issues are marked and recorded. Numerical range validity checks involve the system determining whether the values ​​of numeric parameter fields are within a reasonable range based on business rules and common-sense constraints. For example, it checks whether the product price is negative or exceeds the normal market price limit, whether the product weight exceeds logistics carrying capacity limits, or whether the inventory quantity is negative—all clearly illogical anomalies. These anomalies usually indicate that errors occurred during data entry or transmission, or that the data was tampered with.

[0034] After completing the parameter field detection, the system further performs metadata fingerprint verification on the multimodal files in the basic information. The multimodal files referred to here are unstructured media data contained in the basic information, mainly including multimedia files such as product display images, video introductions, and audio descriptions. Compared to structured parameters, these files are more easily maliciously modified or forged, posing various quality risks such as content tampering, format spoofing, and source falsification. Metadata fingerprint verification involves the system extracting metadata information from the multimodal files and calculating the file's digital fingerprint characteristics, then comparing and verifying to determine the file's authenticity and integrity. Specifically, metadata refers to descriptive information embedded in the media file, including technical parameters such as the file's creation time, modification time, shooting device model, geographical coordinates, and file format version. This metadata can reveal the file's generation history and processing trajectory. The digital fingerprint is a unique identifier calculated using a hash algorithm on the file content; any minor modification to the file content will cause a significant change in the fingerprint value. The system extracts the EXIF ​​metadata of product images to check whether they have been processed by post-editing software. It also calculates the hash fingerprint of the file and compares it with historical records or trusted sources to verify whether the file has been tampered with, thereby identifying multimodal files that may be at risk of being forged or counterfeited.

[0035] After completing the three types of checks and verifications mentioned above, the system generates a first score for each target product based on the results of integrity checks, legality checks, and metadata fingerprint verification. The generation logic involves the system statistically analyzing the number and severity of issues found in each check for each target product. For integrity checks, the system calculates the proportion of missing required fields; for legality checks, the system counts the number of fields with abnormal values ​​and assigns different deduction weights based on the degree of abnormality; for metadata fingerprint verification, the system assigns a risk rating based on the suspiciousness of the document. The system then weights the results of these three dimensions according to preset weighting coefficients to obtain a first score that comprehensively reflects the quality of the basic information. A higher score indicates better completeness, accuracy, and authenticity of the basic information, and more reliable data quality at the core attribute level for the product.

[0036] S22, detect the data coverage of each associated dimension in the associated information, determine whether the data coverage meets the preset coverage threshold, and generate a second score for each target product based on the coverage detection results.

[0037] The system checks the data coverage of each related dimension in the associated information. Here, "related dimension" refers to the different types of data categories covered by the associated information, including multiple information levels such as user reviews, logistics and delivery, after-sales service, supplier qualifications, and historical sales. Each dimension represents extended information about a product in a specific aspect. Data coverage refers to the proportion of actual data items existing under a certain related dimension to the total number of data items that should exist in that dimension. For example, the coverage of the user reviews dimension can be calculated by counting the number of reviews for the product and comparing it with the average number of reviews for similar products. The coverage of the logistics and delivery dimension can be calculated by checking whether necessary information items such as delivery time, shipping costs, and delivery range are included.

[0038] After calculating the data coverage rate for each related dimension, the system determines whether the data coverage rate meets a preset coverage threshold. This coverage threshold is a minimum coverage standard pre-set by the system based on business experience and user needs analysis. For example, the system might specify a coverage threshold of 60% for the user review dimension, requiring the product to have at least 60% of the average number of reviews for similar products; and an 80% coverage threshold for the logistics and delivery dimension, requiring at least 80% of the standard information items under that dimension to be included. The system compares the actual coverage rate of each related dimension with the corresponding threshold requirements, calculating how many related dimensions meet the standard, how many do not, and the degree of gap in the non-compliant dimensions.

[0039] Based on the coverage detection results, the system generates a second score for each target product. The generation logic involves the system quantifying the coverage compliance of each related dimension, assigning positive bonuses to dimensions that meet the standards, and deducting points for dimensions that do not meet the standards based on the degree of difference between their coverage and the threshold. Simultaneously, different weighting coefficients are set to consider the varying importance of different related dimensions to user decision-making; for example, user reviews may be given higher weight because users rely more on the purchasing experiences of others when making decisions. The system calculates a weighted aggregate score reflecting the richness and completeness of the related information. A higher score indicates more comprehensive coverage of the product's related information, allowing users to obtain sufficient auxiliary decision-making information from multiple dimensions.

[0040] S103, verify the credibility of basic information and related information, and generate a third score for each target product based on the credibility verification results.

[0041] Specifically, the system first extracts the physicochemical parameter fields from the initial information and the macroscopic environmental parameter fields from the associated information, constructing a dual-source data mapping matrix. The physicochemical parameter fields refer to the specific numerical values ​​in the basic information, filled in by the merchants, describing the microscopic physicochemical properties of the target product. Examples include quantifiable quality indicators such as the effective ingredient content, moisture content, sugar content, and acidity of agricultural products, or material performance parameters such as hardness, tensile strength, and conductivity of industrial products. The macroscopic environmental parameter fields, on the other hand, refer to the numerical values ​​of the natural environmental conditions of the product's production area recorded in the associated information, such as the region's average annual temperature, precipitation, soil pH, altitude, and accumulated temperature. The system pairs these two types of parameters because, in categories such as agricultural products and geographical indication products, there is a clear causal relationship and physical constraint between the product's physicochemical parameters and the environmental parameters of the production area. This relationship has been scientifically verified and empirically summarized in numerous agricultural studies, geographical literature, and industrial practices. For example, the sugar accumulation of a certain fruit is positively correlated with the accumulated temperature during its growth period and the duration of light exposure. Under specific temperature and light conditions, the sugar content of this fruit has a theoretical maximum value. If the sugar content claimed by the merchant significantly exceeds this theoretical upper limit, it is physically impossible to achieve and must be considered false advertising or a measurement error.

[0042] The system then invokes a pre-built domain knowledge inference engine, an expert system based on a rule base and constraint solver. Its core consists of a vast collection of parameter association rules and physical constraint equations compiled by agricultural scientists, geographers, and industry experts. Specifically, the inference engine's rule base stores hundreds, even thousands, of conditional constraint rules, such as "If the accumulated temperature of the production area is within the range of X to Y, then the theoretical upper limit of the effective component content of crop A grown in that area is Z." These rules establish a clear inference chain between environmental parameters and physicochemical parameters through an "if-then" logical structure. The system substitutes the macroscopic environmental parameters of the current target product into the inference engine, which automatically derives and calculates the reasonable range or theoretical upper limit of the various physicochemical parameters of the target product under those environmental conditions based on the matched rules. For example, the system discovers that the production area of ​​a certain tea product is at an altitude of 1,500 meters and has an average annual temperature of 15 degrees Celsius. Based on the biochemical accumulation patterns of tea in the rule base, the inference engine calculates that the theoretical upper limit of tea polyphenol content under these environmental conditions is about 35%, while the reasonable range of theanine content should be between 3% and 6%.

[0043] The system then executes a parameter cross-consistency comparison algorithm, comparing the actual values ​​of the physicochemical parameters reported by the merchant in the basic information with the theoretical range or upper limit calculated by the inference engine. For each physicochemical parameter, the system determines whether the value reported by the merchant falls within the reasonable range given by the inference engine. If the value reported by the merchant significantly exceeds the theoretical upper limit, for example, the merchant of the aforementioned tea product claims that the tea polyphenol content is as high as 50%, far exceeding the theoretical upper limit of 35% calculated by the inference engine, the system determines that the parameter has a serious logical credibility defect, identifies it as high-risk false data, heavily deducts points from the parameter, and adds a "parameter abnormally exceeded" risk label to the product's data file. If the value reported by the merchant does not exceed the theoretical upper limit, but is on the edge of the reasonable range or near the extreme value, for example, the theanine content is reported as 5.9%, very close to the theoretical upper limit of 6%, the system determines that although the parameter may be physically achievable, it is a low-probability event statistically, and there is a certain degree of exaggeration, so the parameter is lightly deducted points from the parameter, and a "parameter suspected to be too high" warning label is added for subsequent manual sampling. If the value reported by the merchant consistently falls within the middle of a reasonable range, the system determines that the parameter has good logical credibility, does not deduct points, and assigns a positive label of "parameter credibility" to the parameter.

[0044] In addition to cross-validation of physicochemical and environmental parameters, the system also performs semantic consistency checks on the process flow description in the basic information and the cultural background data in the associated information. The process flow description refers to the textual description or key steps of the production process of the target product filled in by the merchant in the basic information. For example, it could describe the fermentation temperature control range, fermentation time, and types of added bacteria for a fermented food, or the material selection, processing sequence, and surface treatment methods for a handicraft. The cultural background data includes authoritative process knowledge retrieved from the knowledge base, such as traditional process specifications, intangible cultural heritage certification standards, and industry-recognized standard operating procedures for the product category or production area. The system calls the natural language processing module to parse the text and extract key information from the process flow description filled in by the merchant. Using a named entity recognition algorithm, it extracts key operational verbs, numerical parameters such as temperature and time, and noun entities such as materials and bacteria from the process description, forming a structured process feature vector. Simultaneously, the system performs the same parsing process on the traditional process specification text in the cultural background data, generating a standard process feature vector.

[0045] The system then calculates the semantic similarity between the feature vector of the process described by the merchant and the feature vector of the standard process. The techniques used include cosine similarity calculation based on word embeddings or sentence-level semantic matching based on a pre-trained language model. If the similarity score is high, for example, exceeding 0.8, it indicates that the process described by the merchant is highly consistent with the traditional standard process, and the system determines that the process description has good credibility and does not deduct points. If the similarity score is at a moderate level, for example, between 0.5 and 0.8, it indicates that the process described by the merchant differs somewhat from the traditional process, possibly due to some innovation or inaccurate description by the merchant. The system determines that the process description has a moderate degree of credibility doubt, deducts points appropriately, and marks it as "process description deviates from the standard." If the similarity score is very low, such as below 0.5, or if the key process steps described by the merchant completely contradict traditional standards, such as a traditional food that requires low-temperature fermentation, but the merchant claims to use a high-temperature rapid fermentation process, the system will determine that the process description is seriously unreliable. It is very likely that the merchant violated traditional process standards in order to shorten the production cycle and reduce costs, or that the process description is completely fictitious. The system will deduct a heavy score from the process description and mark it as "seriously untrue process description".

[0046] After completing cross-validation of physicochemical and environmental parameters and semantic consistency checks between process descriptions and cultural background, the system further performs causal consistency checks on the time logic chain. Specifically, the system extracts time-related fields from the basic information, including timestamp information such as production date, quality inspection date, and certificate issuance date. It also extracts seasonal production data for the production area from the associated information, such as the standard harvest season and harvest window for a particular agricultural product. The system uses a chronological reasoning algorithm to check for logical contradictions between these time-related information. For example, a fruit product may claim a production date of January, but according to agricultural timing knowledge in the associated information, the natural ripening period for this fruit variety is from July to September each year. It is impossible for a naturally ripened product to be available in January unless off-season greenhouse cultivation technology is used. The system further checks whether there is any explanation of greenhouse cultivation or off-season cultivation in the basic information. If there is no relevant explanation, the system determines that the production date seriously conflicts with natural laws, deducts points from the time information, and marks it as "abnormal production time." For example, if the system detects that the quality inspection date of a product is earlier than the production date, this is logically impossible because quality inspection must occur after the product is manufactured. Based on this, the system determines that there is a clear error in the reversal of time causality, severely deducts points from this time chain, and marks it as a "time logic error".

[0047] Through the aforementioned multi-layered credibility verification process, the system comprehensively reviews the logical consistency between the basic and related information of each target product, accumulating and recording deductions for various credibility defects. The system then generates a third score based on the cumulative value of all deductions, using an initial full-score deduction model similar to the first score. Specifically, the system presets an initial value of one hundred points for the third score for each target product. Then, it deducts points based on the number and severity of parameters exceeding limits found in the cross-validation of physicochemical parameters, maps deductions based on the similarity score calculated in the semantic consistency test of the process description, and additional deductions based on the number and severity of contradictions found in the causal verification of the temporal logical chain. All deduction values ​​are then summed and subtracted from the initial value to obtain the final third score. It is important to note that the design of the deduction weights for the third rating should reflect the different degrees of harm caused by different types of credibility defects. For example, errors that seriously violate physical laws, such as physicochemical parameters that seriously exceed the theoretical upper limit, should have a much larger deduction than soft differences such as slight deviations in process description from traditional specifications. This is because the former clearly points to data falsification or serious measurement errors, while the latter may only reflect the merchant's innovative attempts or inaccurate descriptions.

[0048] Based on the above embodiments, as an optional implementation, in S103, the credibility verification of basic information and related information is performed, and a third score for each target product is generated based on the credibility verification result, specifically including S31-S33: S31, perform layout analysis and text recognition on the unstructured qualification documents attached to the basic information, and extract the target parameter values ​​from the unstructured qualification documents; The system performs layout analysis and text recognition processing on the unstructured qualification documents attached to the basic information. Unstructured qualification documents refer to various certification materials stored in image or document form, lacking a unified format standard. Qualification documents issued by different institutions vary greatly in layout, font style, and seal position, making it impossible to directly read information using standardized field names like structured data. Layout analysis refers to the system using computer vision technology to analyze the layout structure of the qualification document's image or page, identifying the position and type of different content blocks such as text areas, table areas, seal areas, and image areas, and understanding the document's logical organization structure. For example, it identifies layout patterns such as certificate numbers usually being located at the top right of the document, test result data usually being presented in table form in the middle of the document, and the institution's official seal usually being located at the bottom of the document. This structured understanding provides location navigation for subsequent information extraction. Text recognition refers to the system using optical character recognition technology to parse the text areas identified by layout analysis, converting the image-based text content into editable and searchable text data. This process needs to overcome common recognition challenges in qualification documents, such as blurry images, diverse fonts, and complex backgrounds.

[0049] After completing layout analysis and text recognition, the system further extracts target parameter values ​​from the unstructured qualification documents. These target parameter values ​​refer to key parameter data in the qualification documents that correspond to the basic or related information of the product. Examples include technical parameters such as net weight, dimensions, material composition, and performance indicators listed in quality inspection reports; traceability information such as production location, production date, and production batch listed in certificates of origin; and qualification information such as authorization period and scope listed in brand authorization letters. The system uses natural language processing technologies such as keyword matching, regular expression extraction, and semantic understanding to locate and extract these target parameter values ​​from the identified text content. For example, by searching for keywords such as "net content" and "weight," the system locates the text segment containing the weight parameter, and then uses regular expressions to extract the subsequent numerical value and unit, thereby obtaining the authoritative parameter data declared in the qualification documents.

[0050] S32, compare the target parameter value with the corresponding parameter value in the basic information and / or associated information, and calculate the deviation between the target parameter value and the corresponding parameter value; The system will perform a consistency comparison between the target parameter values ​​extracted from the qualification documents and the corresponding parameter values ​​in the basic information and / or associated information. The corresponding parameter values ​​referred to here are the parameter data entered by the merchant for the same product attribute dimension when filling in the basic information or associated information. For example, the product weight claimed by the merchant in the basic information, or the place of origin information marked by the merchant in the associated information. These values, entered by the merchant themselves, should theoretically be completely consistent with the values ​​certified by the third-party organization in the qualification documents, or within a reasonable margin of error, because the qualification documents are official certification records of the product's true attributes. The system establishes the correspondence between the parameters in the qualification documents and the product information parameters through parameter semantic matching, and then compares the values ​​of the same parameter in the two data sources.

[0051] During the comparison process, the system calculates the deviation between the target parameter value and the corresponding parameter value. The deviation, as described here, is a quantitative measure of the degree of difference between the numerical values ​​of the same parameter in two data sources. For numerical parameters such as weight and size, the deviation can be measured by calculating relative or absolute error. For example, if the qualification document indicates a weight of 500 grams while the basic information states 600 grams, the deviation is 20%. For categorical parameters such as origin and material, the deviation can be measured by determining whether they match perfectly; if they don't match, the deviation is 100%. The calculated deviation directly reflects the level of consistency between the information submitted by the merchant and the authoritative certification information. A larger deviation indicates lower credibility of the merchant's information and a higher possibility of false advertising or data errors.

[0052] S33. Based on the comparison results between the deviation degree and the preset deviation degree threshold, a third score is generated for each target product.

[0053] The system generates a third score for each target product based on the comparison between the deviation and a preset deviation threshold. The deviation threshold is an acceptable upper limit for deviation pre-set by the system based on the measurement accuracy and business tolerance of different parameter types. For example, for weight parameters, a deviation threshold of 5% might be set to allow a measurement error within 5%, while for origin parameters, a deviation threshold of zero might be set to require a perfect match. The system iterates through all comparable parameters, counting how many parameters have deviations within the threshold range, how many parameters exceed the threshold, and the severity of the deviation for parameters exceeding the threshold. Differential weights are assigned to different parameters based on their importance in judging product quality. The system generates a third score through comprehensive calculation. This score reflects the consistency between product information and qualification certification information; a higher score indicates more truthful and credible information provided by the merchant, and a higher level of product credibility.

[0054] Based on the above embodiments, as an optional implementation, in S33, generating a third score for each target product according to the comparison result between the deviation degree and the preset deviation degree threshold specifically includes S331-S332: S331, if the deviation is greater than the preset deviation threshold, the third score of the target product will be reset to zero, and the corresponding conflict information will be output to the diagnostic module so that the diagnostic module can trigger an abnormal alarm and send a data rectification notice to the corresponding merchant.

[0055] When the system detects that the deviation of a target product exceeds a preset deviation threshold, it immediately sets the third rating of that target product to zero. This zeroing process is an extreme punitive scoring strategy. Its design intent is to directly strip the product of all scores in the credibility dimension by assigning a zero score, causing its overall credibility score to drop significantly in subsequent comprehensive score calculations, potentially falling below the minimum standard for recommended candidates, thus automatically filtering it out of the recommendation results. The rationale for this severe measure is that when there is a significant deviation between the qualification documents and the product information, it indicates that at least one data source has a serious error. Whether the merchant intentionally misrepresents parameters or the qualification documents are forged or altered, it means that the product's information has lost its basic credibility and should not be recommended to users.

[0056] While resetting the third score to zero, the system also outputs the corresponding conflict information to the diagnostic module. This conflict information refers to the specific details of data conflicts discovered by the system during the consistency comparison process. This includes detailed records such as the conflicting product identifier, the names of the conflicting parameters, the target parameter values ​​in the qualification documents, the corresponding parameter values ​​in the product information, the calculated deviation values, and the source and type of the qualification documents. This information comprehensively describes the location and manifestation of the data conflict, providing necessary evidence for subsequent manual review and problem localization. The diagnostic module is a functional component in the system specifically responsible for anomaly monitoring and problem handling. It receives anomaly reports from various data verification stages and executes corresponding warning and handling procedures.

[0057] Upon receiving conflict information, the diagnostic module triggers an anomaly alert mechanism. This mechanism includes a series of warning actions such as sending real-time alerts to the system administrator, displaying a list of abnormal products on the monitoring panel, and generating anomaly event logs. This ensures that platform operators can promptly identify and intervene in handling these high-risk products. Simultaneously, the diagnostic module sends a data rectification notification to the corresponding merchant via in-app messages, email, or SMS. The notification includes details of the detected data conflict, a deadline for the merchant to verify and correct the errors, and information on the suspension of product recommendations until rectification is completed. This urges merchants to quickly identify the root cause of the problem and submit correct product data or qualification certificates. This closed-loop anomaly handling mechanism protects users from being misled by false information and provides merchants with an opportunity to correct errors.

[0058] S332, if the deviation is not greater than the preset deviation threshold, the third score of each target product is generated by accumulating the number of qualification documents that pass the consistency comparison.

[0059] When the system detects that the deviation of a target product is not greater than a preset deviation threshold, it indicates that the product has passed consistency verification on the comparable parameters, and the product information maintains good consistency with the qualification certification information. At this point, the system generates a third score for each target product using an additive method based on the number of qualification documents that have passed consistency comparison. The qualification documents that have passed consistency comparison refer to those that, after being extracted and verified by the system, show deviations between the target parameter values ​​recorded in them and the corresponding parameter values ​​in the product information within the threshold range. Each verified qualification document proves that the authenticity of the product's information in a certain dimension has been endorsed and supported by a third-party authoritative institution.

[0060] The system's cumulative generation logic treats each verified qualification document as a credibility gain unit. The more valid qualification documents a product provides, the more authoritative certifications it has obtained across multiple dimensions, resulting in higher overall credibility and a higher third-party score. The system assigns a base score to each verified qualification document and sets different weighting coefficients based on the document's type and authority. For example, quality inspection reports issued by national-level testing institutions may be given higher weight, while company self-inspection reports have relatively lower weight. Different types of qualification documents, such as brand authorization letters, certificates of origin, and international certifications, also have differentiated weights based on their influence on consumer decisions. The system iterates through all qualification documents provided by the target product, accumulating the weighted scores of those verified documents to ultimately obtain the product's third-party score. This cumulative mechanism encourages merchants to provide more comprehensive and authoritative qualification certification materials, creating a positive incentive for data quality improvement.

[0061] S104: The first, second, and third scores are weighted and summed to generate a credibility score for each target product; the credibility scores are adjusted based on the credit index of each merchant to generate a target credibility score.

[0062] Specifically, the system first performs a weighted linear combination operation on the first, second, and third scores. The calculation formula is: the credibility score equals the first score multiplied by the first weight coefficient, plus the second score multiplied by the second weight coefficient, plus the third score multiplied by the third weight coefficient. The sum of the three weight coefficients must equal one to ensure the normalization of the score. The weight coefficients mentioned here are numerical parameters reflecting the relative importance of different scoring dimensions in the overall data quality assessment. The setting of the weight coefficients needs to comprehensively consider factors such as business scenario characteristics, product category attributes, and model application requirements. The system uses a weighted summation method instead of a simple arithmetic average because in actual data quality assessment systems, the influence weights of different quality dimensions on the final recommendation effect vary significantly, requiring differentiated weight configurations to reflect this hierarchical structure of importance.

[0063] In designing the specific values ​​of the weighting coefficients, the system follows a hierarchical weighting principle based on business logic. For the first score, representing the completeness of basic information, since this basic information is first-hand data actively submitted by merchants, it directly determines whether the target product possesses the most basic identifiability and describability. If the basic information is severely incomplete or obviously falsified, even if the associated information is complete and logically consistent, the product data loses its minimum value threshold as recommendation material. Therefore, the system typically assigns a high weighting coefficient to the first score, such as 0.4 or 0.5, reflecting the fundamental and crucial position of basic information in the data quality system. For the third score, representing cross-source data consistency, since this score directly reveals the degree of consistency between the data submitted by merchants and objective facts, it is a core technical means of identifying false and exaggerated claims. Given the increasingly serious problem of data fraud in the current business environment, the system needs to give sufficient attention and punishment to logical credibility defects. Therefore, the third score is usually assigned a second-highest weighting coefficient, such as 0.35 or 0.4, to ensure that data that is formally complete but substantively false can be effectively identified and downgraded. Regarding the coverage of related information represented by the second score, although this dimension is of great value to enriching the contextual understanding ability of the large language model, the lack of related information mainly reflects the shortcomings of the system's knowledge base construction rather than the data quality problem of the merchant. Moreover, the incompleteness of related information can be gradually improved to a certain extent through subsequent knowledge base expansion, and its impact on the current recommendation decision is relatively small. Therefore, the second score is usually assigned a relatively low weight coefficient, such as 0.15 or 0.2, to play an auxiliary corrective role in the comprehensive score.

[0064] It is important to note that the system employs a dynamic weight adaptive adjustment mechanism in its actual deployment. This means that the three weight coefficients are flexibly configured based on the characteristics of different product categories and the focus of different business stages, rather than being assigned fixed, uniform values ​​globally. For example, for products like geographical indication agricultural products or products with certification of origin, which are highly dependent on the characteristics of the production environment, their quality is strongly correlated with macro-environmental parameters. The system will automatically increase the weight coefficient of the second score because, for these products, complete coverage of related information directly affects whether the model can accurately understand the formation mechanism of their unique quality. Conversely, for industrially standardized products or branded consumer goods, their quality depends more on the company's production processes and quality control systems than on the natural environment of the production location. The system will correspondingly decrease the weight of the second score while increasing the weight of the first and third scores, focusing the evaluation on the authenticity and reliability of the process parameters and quality inspection data reported by the merchant. The system pre-configures the recommendation weight combination schemes corresponding to each category in the product category metadata. When performing weighted summation calculation, it automatically selects the matching set of weight coefficients according to the category to which the target product belongs. This achieves intelligent and refined weight configuration and avoids the one-size-fits-all, coarse-grained scoring fusion strategy.

[0065] After adaptively selecting the weighting coefficients, the system performs a specific weighted summation operation for each target product. For example, if an agricultural product has a first score of 85, a second score of 70, and a third score of 90, and the system selects weighting coefficients of 0.4, 0.2, and 0.4 based on the product's category, the calculated credibility score is 85 multiplied by 0.4, plus 70 multiplied by 0.2, plus 90 multiplied by 0.4, which equals 34 plus 14 plus 36, resulting in a final score of 84. This credibility score comprehensively reflects the weighted average performance level of the target product across three dimensions: basic information standardization, richness of related information, and logical consistency. Compared to single-dimensional scores, the credibility score has stronger comprehensiveness and representativeness, providing a more robust and reliable quantitative basis for subsequent data screening decisions.

[0066] However, after generating the credibility score, the system did not directly use it as the final target credibility score. Instead, it further introduced a merchant credit index, a vertical historical dimension adjustment factor, to perform a secondary correction and adjustment of the credibility score. The merchant credit index mentioned here refers to a dynamically updated quantitative credit indicator maintained by the system for each merchant on the platform. This indicator is calculated based on a comprehensive analysis of multiple dimensions of information, including the quality performance of all product data submitted by the merchant throughout history, user complaint records, number of violations and penalties, and duration of cooperation. The numerical range is typically normalized to between zero and one, or between zero and one hundredth. A higher credit index indicates that the merchant's historical data quality performance is more stable and reliable, while a lower credit index indicates that the merchant has a history of data falsification or violations, and their creditworthiness is questionable. The reason the system needs to include the merchant credit index in the scoring adjustment mechanism is that a credibility score based solely on a single data report cannot fully depict the merchant's true credit level and potential risks. In actual business scenarios, there are two typical problems of scoring distortion.

[0067] The first type of problem is the excessive punishment of reputable, long-established merchants for occasional data quality flaws. Some high-quality merchants who consistently maintain high standards of data submission may, due to accidental errors by operators or system integration malfunctions, submit data with missing fields or incorrect parameters, resulting in a credibility score significantly lower than their historical average. If the system severely demotes or removes a merchant's products based solely on this one instance of a credibility score, it will seriously discourage honest merchants and violate the principles of objective and fair credit evaluation, as a single accidental error should not erase a merchant's long-standing good reputation. The second type of problem is problematic merchants with poor credit who occasionally submit high-quality data to cover up their long-term fraudulent activities. Some dishonest merchants with a history of data fraud may, after being penalized by the platform or placed under close monitoring, strategically submit carefully crafted, high-quality data during certain data submissions. They may attempt to increase the system's trust in them through these localized high scores, thereby continuing their fraudulent behavior in subsequent data submissions. If the system only focuses on the credibility score of the current data submission and ignores the merchant's past misconduct, it will be unable to effectively prevent such opportunistic behavior, leading to the failure of the system's risk control capabilities.

[0068] To address the two types of scoring distortion issues mentioned above, the system designs a dynamic scoring adjustment algorithm based on the merchant's credit index. The core idea of ​​this algorithm is to smoothly handle occasional fluctuations and fully reflect long-term trends by weighting and fusing the current data's credibility score with the merchant's historical credit index. The specific adjustment formula is: the target credibility score equals the credibility score multiplied by the current data weight factor plus the merchant's credit index multiplied by the historical credit weight factor, where the sum of the current data weight factor and the historical credit weight factor equals one. Here, the current data weight factor refers to the proportion of the current reported data's credibility score in the final target credibility score, while the historical credit weight factor refers to the proportion of the merchant's historical credit index in the final score. The relative magnitude of these two factors determines the strength and direction of the score adjustment.

[0069] In designing the weighting factors, the system employs an asymmetric gradual adjustment strategy. When the credibility score is higher than the merchant's credit index, it indicates that the quality of the data submitted by the merchant this time exceeds its historical average, showing positive improvement. In this case, the system will appropriately increase the current data weighting factor while decreasing the historical credit weighting factor. For example, setting the current data weighting factor to 0.7 and the historical credit weighting factor to 0.3. This weighting configuration allows the high-quality performance in the current instance to significantly improve the final target credibility score, promptly reflecting the merchant's improvement efforts. At the same time, the historical credit index retains a certain restraining effect, preventing a single abnormally high score from completely masking historical problems. Conversely, when the credibility score is lower than the merchant's credit index, it indicates a negative decline in the quality of the data submitted by the merchant this time. In this case, the system's weighting strategy needs to be differentiated based on the merchant's historical credit rating. For high-quality merchants with a historical credit score of 80 or above, the system judges that the low score is more likely an accidental error than malicious fraud. Therefore, it significantly increases the historical credit weight factor while decreasing the current data weight factor. For example, setting the historical credit weight factor to 0.7 and the current data weight factor to 0.3. This weighting configuration allows the merchant's good historical credit to largely offset the impact of the low score, ensuring that the final target credibility score does not drop significantly due to a single error, reflecting trust and protection for trustworthy merchants. For problematic merchants with a historical credit score below 50, the system judges that the low score likely reflects persistent data quality issues. Therefore, it increases the current data weight factor while decreasing the historical credit weight factor. For example, setting the current data weight factor to 0.8 and the historical credit weight factor to 0.2. This weighting configuration ensures that the low score is fully reflected in the final target credibility score, imposing sufficient penalty on the merchant and preventing them from using limited historical credit to dilute the impact of current poor data.

[0070] When performing specific score adjustment calculations, the system first queries the merchant credit management database to obtain the current credit index of the merchant corresponding to the target product. Then, based on the relationship between the credibility score and the credit index, as well as the segment interval of the credit index, it automatically selects a matching combination of weighting factors and finally substitutes them into the adjustment formula for weighted fusion calculation.

[0071] Based on the above embodiments, as an optional implementation method, in S104, adjusting the credibility score according to the credit index of each merchant to generate the target credibility score specifically includes S41-S45: S41, Calculate the percentage of initial information submitted by each merchant within a preset time window that passes integrity verification and credibility verification, and generate the historical verification pass rate for each merchant.

[0072] The system calculates the percentage of initial information submitted by each merchant within a preset time window that passes integrity and credibility verification, thereby generating the historical verification pass rate for each merchant. The preset time window refers to a historical statistical period set by the system, such as the past six months or one year. The length of this time window needs to balance the timeliness of the data and the stability of the statistics, reflecting the merchant's recent performance while including a sufficient sample size to ensure the reliability of the statistical results. The system retrieves all initial product information records submitted by each merchant within this time window from the database, calculates the pass rate of these products after integrity verification in step S102 and credibility verification in step S103, and calculates the percentage of products that passed verification out of the total number of submissions. This percentage is the historical verification pass rate. The historical verification pass rate reflects the merchant's long-term performance in data quality management. Merchants with high pass rates indicate good data standardization awareness and quality control capabilities, and their submitted product information has a higher overall credibility. Merchants with low pass rates reveal chaotic data management or a tendency to deliberately falsify data, requiring stricter credibility discounts to be applied to their products.

[0073] S42 collects negative behavior data from user feedback in real time. Negative behavior data includes at least one of user complaint data, return rate data, and information correction trigger data.

[0074] The system collects negative behavior data from user feedback in real time, serving as another important dimension for merchant credit assessment. Negative behavior data refers to various negative feedback information generated by users during the purchase and use of goods. This data directly reflects users' dissatisfaction with merchants and their products, and is an objective indicator for measuring the quality of merchant services and the authenticity of products. Negative behavior data includes at least one of the following: user complaint data, return rate data, and information correction trigger data. User complaint data refers to records of complaints submitted by users through the platform's customer service system and complaint channels regarding discrepancies between the product and its description, quality issues, and false advertising. Return rate data refers to the percentage of users who initiate returns and refunds after purchasing goods from merchants; a high return rate usually indicates a significant deviation between the actual product and user expectations. Information correction trigger data refers to information correction suggestions submitted by users by actively clicking on function buttons such as "Information Error" and "Correction Feedback" on the product details page. This behavior indicates that the product information contains errors or misleading descriptions that users can identify. The system continuously monitors these negative behavior events generated by users through a real-time data collection interface, linking them to the corresponding merchant accounts for cumulative recording.

[0075] S43, weights the negative behavior data according to the time decay function to generate the cumulative negative feedback value for each merchant.

[0076] The system weights the collected negative behavior data using a time decay function to generate a cumulative negative feedback value for each merchant. The time decay function is necessary because negative behaviors occurring at different times have varying values ​​reflecting a merchant's current creditworthiness. Negative behaviors more recent are more representative of a merchant's current situation and potential improvements, while older, historical negative behaviors may no longer be representative, as the merchant may have already rectified and improved their operations. The time decay function is a mathematical function that decreases over time, such as an exponential decay function. The system calculates a decay weight for each piece of negative behavior data based on its occurrence time; more recent data has a higher weight, and older data has a lower weight. The system multiplies all negative behavior data by their respective time decay weights and then sums them to obtain the cumulative negative feedback value. This value comprehensively reflects the total weighted amount of negative feedback for the merchant; a higher value indicates more severe and concentrated negative feedback, suggesting a more concerning credit situation.

[0077] S44 combines historical verification pass rates and cumulative negative feedback values ​​to generate a credit index for each merchant.

[0078] The system combines historical verification pass rate and cumulative negative feedback value to generate a comprehensive credit index for each merchant. The credit index is a standardized quantitative indicator used to comprehensively characterize a merchant's overall creditworthiness. The system's generation logic uses historical verification pass rate as a positive credit factor, reflecting the merchant's performance at the data quality source; a higher pass rate contributes more to the credit index. Simultaneously, it uses cumulative negative feedback value as a negative credit factor, reflecting the merchant's performance at the user experience level; a higher cumulative value severely reduces the credit index. By setting reasonable weighting coefficients and normalization processing, the system merges these two factors with different dimensions and value ranges into a unified credit index. Typically, the credit index is standardized to a range of zero to one; the closer the value is to one, the better the merchant's credit; the closer the value is to zero, the worse the merchant's credit. This dual-dimensional credit index examines both the merchant's data management capabilities and the merchant's actual service quality, providing a more comprehensive and accurate characterization of the merchant's trustworthiness.

[0079] S45 uses the credit index as an adjustment factor and multiplies it with the credibility score to generate the target credibility score for each target product.

[0080] The system uses the calculated credit index as an adjustment coefficient and multiplies it with the credibility score to generate the target credibility score for each product. Here, the credibility score refers to the product-level credibility evaluation result calculated by integrating the first, second, and third scores in the preceding steps. This score only reflects the data quality of the product itself and does not yet consider the impact of merchant credit. The system applies the merchant credit index to the product credibility score through multiplication, achieving a linked adjustment of scores from the product dimension to the merchant dimension. Specifically, for high-quality merchants with high credit indices (close to one), their product credibility scores remain essentially unchanged or slightly increase after multiplication, reflecting an incentive for trust in high-quality merchants. For low-quality merchants with low credit indices (significantly less than one), their product credibility scores are significantly reduced after multiplication. Even if the product's data detection results are acceptable, the merchant's poor credit record will result in a punitive adjustment. This mechanism effectively prevents unscrupulous merchants from circumventing credibility checks by carefully packaging individual product data.

[0081] S105: Receive the query request for the target product sent by the user terminal, select the candidate target product corresponding to the query request from the target products, send the initial information of the candidate target products with a target credibility score higher than the preset score to the big model, so that the big model can generate recommended content based on the initial information and push the recommended content to the user terminal.

[0082] Specifically, the system first receives target product query requests from users through a real-time query interface deployed in the user access layer. The user terminal refers to the terminal device used by the user to access the recommendation service, including mobile applications on smartphones, web browsers on personal computers, voice-interactive devices such as smart speakers, and other IoT terminals that support network communication. The query request is the user's information need or purchase intention regarding the target product expressed through natural language input. For example, a user might enter a text query like "suitable high-mountain organic tea for gifting" in the search box, or use voice interaction to say something like "I want to buy some high-quality Pu'er tea from Yunnan," or trigger an implicit query by clicking the "Recommend similar products for me" button on a product details page. The system's query interface uses a standardized RESTful API or WebSocket long-connection protocol, capable of receiving and parsing query requests from different terminal types with millisecond-level response speeds. For text-based query requests, the system directly extracts the query string; for voice-based query requests, the system first calls an automatic speech recognition engine to convert the speech signal into a text string; for implicitly triggered query requests, the system automatically generates structured query parameters based on the user's click behavior and the attributes of the currently viewed product.

[0083] Upon receiving a query request, the system immediately initiates a query intent parsing process based on deep semantic understanding. The core task of this process is to accurately extract the core intent dimensions and constraints of the query from the user's input natural language query, including multiple query dimensions such as the target product's category, key attribute features, application scenario requirements, price range preferences, and origin requirements. The system calls a pre-trained natural language processing model to perform lexical and syntactic analysis on the query text. Using a named entity recognition algorithm, it automatically identifies key information entities in the query, such as product category words like "tea," quality adjectives like "organic," origin names like "Yunnan," and application scenarios like "gift-giving." Simultaneously, it uses a dependency parsing algorithm to understand the semantic relationships between these entities. For example, the modifier relationship between "high mountain" and "organic tea" indicates that the user wants tea produced in high-altitude mountainous areas and meeting organic certification standards. The system represents the parsed structured query intent as a multi-dimensional feature vector. Each dimension of the vector corresponds to a query constraint and its value or range. For example, the category dimension is "tea", the place of origin dimension is "Yunnan", the quality certification dimension is "organic certification", and the application scenario dimension is "gift".

[0084] The system then performs an efficient multi-condition joint search from a pre-built product index based on the parsed query intent vector, initially identifying a set of candidate target products that meet the query constraints. This product index is an optimized distributed inverted index structure. When receiving initial target product information from merchants, the system simultaneously extracts features and builds an index, mapping the category, brand, origin, and key attribute parameters of each target product to an index. This allows for quick location of matching product records when faced with user queries without requiring a complete traversal of all product data. The system's retrieval algorithm combines exact matching and fuzzy matching strategies. For discrete constraints such as category and origin, the system performs strict exact matching, only considering a match successful when the corresponding field value of the target product completely matches the query conditions. For textual constraints such as quality descriptions and application scenarios, the system performs fuzzy matching based on semantic similarity. By calculating the vector distance between the target product's attribute description text and the query condition text in the semantic space, a match is considered successful if the similarity exceeds a preset threshold. By combining precise matching with fuzzy matching, the system can ensure search accuracy while maintaining query recall. This avoids missing products that users truly need due to overly strict matching rules, and also avoids falsely recalling a large number of irrelevant products due to overly lenient matching rules.

[0085] After completing the initial query matching and retrieval, the system obtains a preliminary candidate set containing dozens or even hundreds of target products. These products semantically satisfy the user's query intent, but their data quality varies greatly. They include both rigorously verified, highly reliable products and low-reliability products with data defects or even false information. If the system inputs all the initial information of all products in this preliminary candidate set into the large language model without filtering, it will lead to two serious problems. The first problem is the redundancy overload of input information. Although large language models have powerful contextual understanding capabilities, their input window length is physically limited. Current mainstream large language models, such as the GPT series or Claude series, typically have context windows ranging from thousands to tens of thousands of tokens. If the number of candidate products is too large and the initial information of each product contains rich textual descriptions and parameter data, the total length of the input information can easily exceed the model's window limit, causing some product information to be truncated and lost, or requiring significant compression and simplification of the information for each product. Both of these approaches will lose key contextual details, reducing the quality and accuracy of the recommendations generated by the model. The second problem is contamination from low-quality data. If the initial candidate set contains a large number of products with low target credibility scores, the initial information of these products may contain incomplete parameters, logically contradictory descriptions, or even false and exaggerated advertising content. When this inferior information is mixed with information from high-quality products and input into the model, the model has difficulty automatically identifying and filtering out the false components. Instead, it may learn and reference this erroneous information as real knowledge, ultimately resulting in factual errors or logical inconsistencies when generating recommendations. For example, the model may recommend a product that claims to have exceptional quality parameters but whose actual data is fabricated, or it may cite a fictitious description of a product's manufacturing process in the recommendation reason. These errors not only mislead users into making incorrect purchasing decisions but also seriously damage the credibility of the recommendation system.

[0086] To fundamentally address the aforementioned issues of information overload and data pollution, the system must implement a rigorous quality threshold screening process before inputting candidate product information into the large language model. Only candidate products with a target credibility score higher than a preset score will be retained, while other low-scoring products will be removed from the candidate set. The preset score refers to a target credibility score threshold pre-set by the system based on business quality requirements and user experience standards. It is typically between 70 and 80 points, and the specific value needs to be optimized by comprehensively considering factors such as the platform's current data quality, users' tolerance for recommendation accuracy, and the robustness of the large language model to input noise. The physical meaning of the preset score is to define a clear data quality admission line. Only candidate products with a target credibility score reaching or exceeding this threshold are deemed by the system to possess sufficient authenticity, completeness, and logical consistency, making them suitable as high-quality input material for the large language model. Products with scores below the threshold are considered to have significant quality defects or credibility risks and should not enter the model's recommendation process.

[0087] When performing quality threshold screening, the system iterates through each target product in the initial candidate set, retrieves the target credibility score calculated in step S104 from the database, and compares the score with a preset score. If the target credibility score is greater than or equal to the preset score, the product is retained in the candidate set and marked as "quality qualified." If the score is less than the preset score, the product is removed from the candidate set, and the reason for removal is recorded as "credibility not met." After this screening process, the initial candidate set, which originally contained dozens or even hundreds of products, is streamlined into a carefully selected candidate set containing only a dozen to several dozen highly credible products. These products not only match the user's query intent at the semantic level but also pass the system's rigorous verification at the data quality level, laying a reliable data foundation for subsequent large language model recommendation generation.

[0088] It's important to note that after performing quality threshold screening, the system dynamically adjusts the size of the selected candidate set to adapt to the optimal input size of a large language model. If the number of candidate products after quality screening is still large, for example, exceeding thirty, the system will further initiate a secondary screening mechanism based on relevance ranking. By calculating the semantic similarity score between each candidate product and the user's query intent, the candidate products are sorted in descending order, and then the top-ranked products, such as the top twenty, are selected as the final candidate set for input into the model. This relevance ranking mechanism ensures that, while meeting quality requirements, the products ultimately entering the model are those that best match the user's needs, further improving the accuracy of recommendations. Conversely, if the number of candidate products after quality screening is too small, for example, less than five, the system will appropriately relax the preset scoring threshold, for example, lowering the threshold from 80 to 70, and re-perform the screening to expand the size of the candidate set. This avoids insufficient recommended products due to overly strict quality requirements, which would fail to meet the diverse selection needs of users. Through this dynamic threshold adjustment mechanism, the system can flexibly balance the accuracy and richness of recommendation results while ensuring data quality, achieving the optimal trade-off between quality control and user experience.

[0089] After finalizing the candidate target products, the system begins preparing the contextual information to be input into the large language model. Specifically, the system extracts complete initial information for each final candidate product from the database. This initial information includes two parts: basic information and related information obtained in step S101. The basic information provides a description of the product's micro-attributes, including first-hand data such as name, specifications, physicochemical parameters, process flow, and multimodal files. The related information provides macro-background knowledge of the product, including objective background data such as environmental parameters of the production area and cultural lineage. The system integrates these two parts of information in a structured manner, arranging them in a format easily understood by the large language model. For example, it uses key-value pairs or natural language paragraphs such as "Product Name: XX, Origin: XX, Key Parameters: XX, Environmental Background: XX, Process Characteristics: XX" to concatenate and combine the initial information of multiple candidate products into a complete contextual text.

[0090] When constructing the context text input to the model, the system also inserts the user's original query request and a clear task instruction prompt at the beginning of the text. This prompt clearly informs the large language model that its current task is to generate personalized product recommendations based on the user's query requirements and the provided candidate product information. The recommendations should include practical information such as comparative analysis of the advantages and disadvantages of candidate products, purchasing suggestions, and explanations of applicable scenarios. Furthermore, the model is required to strictly base its reasoning and expression on the provided initial information when generating the content, and must not fabricate or speculate on product attributes or parameter values ​​not included in the initial information, ensuring the factual accuracy of the recommendations. By explicitly embedding task instructions and factual constraints in the input text, the system can effectively guide the large language model to generate recommendations in the expected manner, reducing the risk of the model generating illusions or deviating from the facts.

[0091] The system then sends the constructed context text to the large language model server via a standard API interface. This server can be a locally deployed model inference cluster or a large model API service provided by a third-party cloud service provider such as OpenAI or Anthropic. Upon receiving the input context text, the large language model first encodes and understands the key information in the context through its internal attention mechanism, identifies the core needs dimension of the user query, analyzes the characteristic attributes and quality advantages of each candidate product, and understands the environmental background and cultural connotations provided by the associated information. Based on this deep semantic understanding, it then generates recommended content text word by word and sentence by sentence through an autoregressive generation mechanism. Because the initial information of the input candidate products is highly reliable data that has undergone rigorous quality verification, including physicochemical parameters, process descriptions, and environmental background, the large language model can reason and express itself based on accurate and reliable facts when generating recommended content, avoiding inference errors caused by false or contradictory input data. Meanwhile, because the associated information infuses each product with rich macro-background knowledge, the large language model can generate deeper explanations and more persuasive arguments in the recommended content. For example, the model can explain that the reason why a certain tea product has a unique aroma and taste is because it is produced in a high-altitude area where the large temperature difference between day and night leads to the accumulation of rich internal substances in the tea. Compared with the traditional recommendation method of simply listing product parameters, this recommendation reason based on objective environmental causal relationship has stronger credibility and persuasiveness, and can significantly improve users' acceptance and satisfaction with the recommended content.

[0092] After generating recommended content, the large language model returns the generated text to the system's recommendation service module via API. Upon receiving the recommended content, the system performs necessary post-processing and formatting, such as segmenting the text to improve reading experience, adding hyperlinks to product names mentioned in the recommendations for users to view product details, and highlighting key quality parameters or price information. For content containing multiple product recommendations, the system also adjusts the recommendation order based on the sufficiency of the recommendation reasons and the overall product rating, ensuring that the highest quality and most relevant product recommendation is placed first. The system also appends a disclaimer or data source explanation at the end of the recommended content, informing users that the recommended content is generated based on system-verified merchant data, but the final purchase decision still requires users to make a comprehensive judgment based on their own circumstances. This transparent information disclosure helps build user trust in the recommendation system.

[0093] After post-processing and formatting the recommended content, the system immediately pushes it to the user's device via a real-time communication channel. For users accessing the site through a web browser, the system renders the recommended content as an HTML page and returns it to the browser via an HTTP response for display. For users accessing the site through a mobile application, the system encapsulates the recommended content as structured data in JSON format and pushes it to the user's mobile device via a mobile push service or a WebSocket long connection, where it is rendered and displayed locally by the mobile application's front-end interface. For users accessing the site through a voice interaction device, the system also calls a text-to-speech engine to convert the recommended content into audio and broadcast it to the user. By supporting multi-terminal and multi-format recommended content push methods, the system ensures that users in different usage scenarios can receive and consume recommended content in the most suitable way, improving the accessibility and ease of use of the recommendation service.

[0094] Based on the above embodiments, as an optional implementation method, after pushing the recommended content to the user's terminal, it further includes: Acquire user interaction data on recommended content. If the interaction data does not meet the preset behavior threshold conditions, convert the interaction data into a weakly supervised feedback signal. Based on the weakly supervised feedback signal, perform online incremental updates on the weighting weights used to generate credibility scores, so as to update the target credibility scores of each target product.

[0095] Specifically, the system uses a behavior tracking module deployed on the user's device to collect various interactive behavior data in real time after the user receives recommended content. This interactive behavior data refers to all observable behavior records generated when a user interacts with recommended content. This includes explicit behaviors such as whether the user clicked on a product link in the recommended content, the duration of time spent on the product details page after clicking, whether the user added the product to their shopping cart or favorites, whether they ultimately completed the purchase, and whether they liked or commented on the recommended content. It also includes implicit behaviors such as the user's completion rate of reading the recommended content, scrolling speed, and mouse hover position. The system records this multi-dimensional behavioral data in a structured manner and associates it with the context information of the current recommendation session, including metadata such as all candidate product identifiers involved in the recommendation, the target credibility score of each product, the user's query request content, and the generation time of the recommended content.

[0096] After collecting user interaction data, the system needs to determine whether this data meets preset behavioral threshold conditions. These behavioral threshold conditions are a set of pre-defined behavioral standards used by the system to judge user satisfaction with the recommended content. They typically include multiple evaluation dimensions such as click-through rate threshold, conversion rate threshold, and dwell time threshold. For example, the system might set behavioral threshold conditions as "the user clicked on at least one recommended product and stayed on the product details page for more than 30 seconds" or "the user added the recommended product to their shopping cart or completed a purchase." The physical meaning of these threshold conditions is that they define a minimum behavioral standard for user satisfaction. Only when the user's actual interaction behavior reaches or exceeds these standards does the system consider the recommendation to have achieved a positive effect, indicating that the highly-rated products used in the recommendation have indeed been recognized by the user, and that the current rating model and weight configuration are reasonable and effective.

[0097] When the system detects that the interaction data of a recommendation session does not meet the preset behavior threshold conditions—for example, if a user closes the page without clicking any product links after viewing the recommended content, or clicks a product but only stays for a few seconds before returning—this negative user response indicates a potential problem with the recommendation. While the recommended products may perform well in terms of data quality scores, they fail to truly meet the user's actual needs or preferences, suggesting a possible weight configuration bias in the current scoring model, requiring adjustment and optimization. In this case, the system will convert this negative interaction data into a weakly supervised feedback signal. This weakly supervised feedback signal refers to supervisory information that does not directly provide explicit labels but can indirectly reflect the direction of model prediction bias. Unlike traditional supervised learning where each sample has a clear correct answer label, user interaction behavior can only provide a vague quality feedback. For example, a user not clicking on a recommended product only indicates that the product may not be attractive enough, but it cannot directly tell the system what the correct score for the product should be, nor can it clearly indicate which scoring dimension's weight configuration is problematic. This ambiguity makes this feedback signal usable only as a weakly supervised signal.

[0098] When the system transforms interactive behavior data into weakly supervised feedback signals, it constructs a feedback sample record. This record contains information such as all candidate products involved in the current recommendation, their current target credibility scores, the scores and weight configurations of each scoring dimension, and the user's actual interactive behavior. The sample is weakly labeled "negative feedback" to indicate that it reflects a mismatch between the predicted score and the user's preferences. The system then performs online incremental updates to the weighted weights used in generating the target credibility score based on the accumulated weakly supervised feedback signals. The weighted weights mentioned here refer to the weight coefficients used in step S104 when calculating the target credibility score by weighting and summing multiple scoring dimensions, including the weight values ​​corresponding to the base score, consistency score, traceability score, and other dimension scores. Online incremental updates mean that the system does not need to retrain on all historical data; instead, it makes small gradient adjustments to the existing weight parameters based on newly collected feedback signals. This update method has low computational overhead and can quickly respond to changes in user feedback trends.

[0099] The system employs a gradient descent-type optimization algorithm to update weights. By analyzing product samples with high ratings but lacking positive user responses, it identifies potentially overestimated rating dimensions and reduces their corresponding weights. Simultaneously, by comparing product samples with relatively low ratings but which received positive user responses in historical recommendations, it identifies potentially underestimated rating dimensions and increases their corresponding weights. Through continuous online incremental updates, the system's weighted weight configuration gradually converges towards a direction that better reflects users' true preferences. This allows the target credibility score to not only reflect the objective quality of the data but also, to some extent, predict users' subjective acceptance of products. After completing the weight update, the system recalculates the target credibility score for all target products using the updated weights, making the score distribution more aligned with users' actual needs. This enables the system to filter out more popular products in subsequent recommendation sessions, continuously improving the overall performance of the recommendation system and user satisfaction.

[0100] Based on the above embodiments, as an optional implementation, in S105, sending the initial information of candidate target products with target credibility scores higher than preset scores to the large model, so that the large model generates recommended content based on the initial information, specifically includes: The initial information of candidate target products is structurally assembled to generate prompt text containing basic information summaries and related information summaries. The prompt text is combined with the user intent tags carried in the query request to generate input prompt words for the large model. The input prompt words are sent to the large model so that the large model can generate recommendation content containing product comparison analysis and recommendation reasons based on the input prompt words. The recommendation content is checked for factual consistency to determine whether there is a discrepancy between the parameter descriptions in the recommendation content and the corresponding parameter values ​​in the initial information. If a discrepancy exists, the recommendation content is corrected and pushed to the user.

[0101] The system first performs structured concatenation on the initial information of candidate target products, generating a prompt text containing both basic and related information summaries. This structured concatenation refers to the process by which the system, following a predetermined text organization logic, orderly integrates and formats the various data fields scattered throughout the initial information. Since the initial information contains two main categories of data—basic and related information—the system extracts key fields from each to generate summaries. The basic information summary covers core attribute parameters such as product name, price, specifications, material, and origin, while the related information summary includes auxiliary decision-making information such as inventory status, sales data, user review statistics, and merchant qualifications. The system then concatenates these two types of summaries according to a fixed template format, forming a clearly structured and complete prompt text. This text retains the objectivity and accuracy of the product data while its logical organization facilitates understanding and extraction of key information by larger models, providing a reliable factual basis for subsequent content generation.

[0102] The system further combines the generated prompt text with the user intent tags carried in the query request to generate input prompt words for the large model. The user intent tags mentioned here refer to the user demand feature identifiers extracted by the system through natural language understanding technology when receiving user query requests. For example, the user may be looking for high-value products, valuing brand quality, preferring specific functions, or having budget constraints. These tags characterize the user's personalized preferences and decision-making focus. By combining objective product information prompt text with subjective user intent tags, the system constructs a complete input prompt word. This prompt word informs the large model which products are available for recommendation and their detailed parameters, and also clarifies which dimensions the recommendation content should emphasize to match user needs. This enables the large model to generate highly targeted and personalized recommendation copy, rather than general, generic descriptions.

[0103] The system sends carefully crafted input prompts to a large-scale model, which then generates recommended content including product comparison analysis and rationale. Upon receiving the prompts, the model leverages its language understanding and generation capabilities, trained on massive amounts of text data, to intelligently compare and analyze the parameters of multiple candidate products. It identifies differences and advantages / disadvantages in price, performance, quality, and other dimensions, and combines this with user intent tags to determine which products better meet the user's needs and preferences. Finally, it generates natural and fluent recommendation reasons explaining why these products are recommended and their suitable usage scenarios. The generated recommendations typically include a summary of the product's core selling points, a multi-product comparison table, a matching analysis based on user intent, and purchase suggestions—rich in decision support information. Presented in natural language that aligns with human reading habits, this significantly enhances the readability and persuasiveness of the recommendations.

[0104] However, the system does not directly push the raw content generated by the large model. Instead, it performs a rigorous factual consistency check on the recommended content, determining whether there are discrepancies between the parameter descriptions in the recommended content and the corresponding parameter values ​​in the initial information. This factual consistency check involves the system extracting all descriptive statements related to product parameters from the natural language recommended content generated by the large model using text parsing and data comparison techniques. These descriptive statements include specific parameter statements such as "This product weighs 500 grams," "Made in Germany," and "Battery capacity reaches 5000 mAh." These textual descriptions are then converted into structured parameter values ​​and compared one by one with the authoritative values ​​of the corresponding parameters stored in the initial information. The system checks for discrepancies such as inconsistent parameter values, misattributed parameter attributes, or fabricated parameters. This verification mechanism effectively identifies potential illusions or misunderstandings that may occur during the generation process of the large model, ensuring the factual accuracy of the recommended content.

[0105] When the system detects a discrepancy between the recommended content and the initial information, it triggers a content correction process. The system automatically replaces or deletes erroneous descriptions in the recommended content based on the accurate parameter values ​​in the initial information, correcting all inconsistent parameter expressions to perfectly match the initial information. The corrected recommended content retains the natural and fluent language style and personalized recommendation logic generated by the large model, while ensuring the absolute accuracy of all factual statements, eliminating the risk of information distortion caused by model illusions. The system finally pushes the corrected and verified recommended content to the user. The recommended information received by the user not only includes professional product comparison analysis and tailored recommendations, but more importantly, all product parameter descriptions have undergone end-to-end credibility assurance from the data source to the generated output, truly achieving a perfect combination of accurate and reliable recommended content with intelligent personalization.

[0106] like Figure 2 As shown, Figure 2This is a complete technical architecture diagram of a product information credibility assessment and intelligent recommendation system based on a large model, provided in this application embodiment. It shows the entire process chain from multi-source data access to recommended content output. The system uses the Kafka streaming data processing framework to access four key data sources in real time, including product description information from PG Mall, clipboard tag image data as extended qualification files, semantic similarity results detected by Image Search for originality verification, and text parsing results generated by the original image description system. These heterogeneous data are first aggregated into the Dataset_raw raw dataset for unified storage and preliminary cleaning. After preprocessing by the Kafka data quality verification framework, the system enters the core credibility assessment stage. Through traffic diversion and population processing, products are grouped and managed according to dimensions such as category and merchant. Then, three assessment sub-modules run in parallel to perform integrity verification, related information richness analysis, and qualification document cross-validation of product information. Integrity verification checks the compliance of basic fields using a central database product quality evaluator. Related information assessment relies on the exposure rate management module of the graph search engine to analyze data quality such as sales and reviews. Credibility verification utilizes the Caffeine local secondary cache manager in the operation report and downstream product information verification servers to identify data conflicts. The scores from the three dimensions are aggregated and adjusted based on the merchant's historical credit records to generate a target credibility score. The system uses threshold decision nodes set by a diamond-shaped judgment box to filter qualified products. Products that fail are imported into the risk point data filing and demonstration information review database. Qualified candidate products are sent to the large model context window problem database, i.e., the Ollam / FAISS vector retrieval framework. The system structurally concatenates product information and combines it with user intent tags to construct input prompts, which are then sent to a guiding model. This model is implemented using PythonApp LightGBM integration and utilizes a gradient boosting framework to assist in generating natural language content that includes product comparison analysis and recommendation reasons. The generated recommendation content then enters an automated online service evaluation module for factual consistency verification. The system compares the parameter descriptions in the recommendation text with the authoritative values ​​of the initial information, automatically correcting any discrepancies. Finally, through consumer service AI and predictive editing knowledge base reasoning, the verified content is pushed to the user's end via database access. The entire architecture ensures that the recommended content remains accurate and reliable while maintaining intelligent personalization through multi-dimensional evaluation and screening and dual verification mechanisms.

[0107] like Figure 3 As shown, Figure 3This is a schematic diagram of the layered architecture of a multimedia data acquisition network hardware stack provided in this application embodiment, illustrating the complete hardware processing link from data access to local storage. The top layer of the system is the multimedia data acquisition network hardware stack module, which is responsible for receiving raw signals from various data acquisition terminals such as PCI slot network cards, IoT smart chip MCUs, or field sensors. These signals are routed and transmitted to the second layer, the streaming heterogeneous data acquisition gateway server, through Ethernet or IoT bus manager. This server uses a multi-core CPU cluster or high-throughput small storage device array architecture to achieve high-speed data reception and preliminary processing. After transmission through the PCIe high-speed local external bus, the data stream enters the core processing unit of the system, namely the central processing server, which is the core computing power hub. This module is equipped with a tensor acceleration chip GPU or a multi-core main control CPU to provide powerful parallel computing capabilities, becoming the scheduling and computing center of the entire architecture. Multiple data paths extend outward from this core node. On the left, a distributed storage bus or RDMA network connects to a distributed disk storage cluster and a high-speed addressing module, using an NVMe SSD cluster or Milvus vector database storage array to achieve fast read and write of massive amounts of data. In the middle, a hardware encryption bus and a dedicated full-disk bus connect to a hardware security module and a dedicated cryptographic chip, using an HSM multi-signature node or a Secure Enclave isolation security chip to ensure the security of data processing. On the right, a high-speed network link or a streaming interface physical connector connects to a large-scale distributed computing cluster at the downstream consumer end. This module uses a large-scale context window co-processing storage or computing power cluster array architecture to support subsequent AI inference and data analysis tasks. All processed data is ultimately converged to the local secondary high-performance cache entity and status sensor module via a bidirectional feedback physical bus and a reverse feedback physical bus. This module uses DRAM cache space or an edge-side interactive behavior status descriptor to implement temporary data caching and status monitoring. The entire architecture achieves end-to-end hardware-level optimization from data acquisition, secure processing, high-speed storage to computation and distribution through multi-level bus interconnection and dedicated hardware acceleration units, ensuring that the system has comprehensive capabilities of high throughput, low latency and strong security when processing large-scale multimedia data.

[0108] Based on the above method, this application also discloses a product recommendation system with multi-source information verification, such as... Figure 4 As shown, Figure 4 This is a schematic diagram of the structure of a multi-source information verification product recommendation system provided in an embodiment of this application. The system includes: an acquisition module, a first verification module, a second verification module, a generation module, and an output module; wherein, The module consists of four parts: an acquisition module and an output module. The acquisition module retrieves initial information about the target products sent by each merchant. This initial information includes basic information and related information, which is generated by searching for related information based on the basic information. The first verification module verifies the completeness of the basic and related information, generating a first and second score for each target product based on the verification results. The second verification module verifies the credibility of the basic and related information, generating a third score for each target product based on the credibility verification results. The generation module performs a weighted sum of the first, second, and third scores to generate a credibility score for each target product. Based on the credit index of each merchant, the credibility scores are adjusted to generate a target credibility score. The output module receives query requests for target products from the user, selects candidate target products corresponding to the query request from among the target products, and sends the initial information of candidate target products with a credibility score higher than a preset score to the large model. This allows the large model to generate recommended content based on the initial information and push the recommended content to the user.

[0109] It should be noted that the system provided in the above embodiments is only illustrated by the division of the above functional modules. In actual applications, the above functions can be assigned to different functional modules as needed, that is, the internal structure of the device can be divided into different functional modules to complete all or part of the functions described above. In addition, the system and method embodiments provided in the above embodiments belong to the same concept, and the specific implementation process can be found in the method embodiments, which will not be repeated here.

[0110] Please see Figure 5 This is a schematic diagram of the structure of an electronic device provided in an embodiment of this application. Figure 5 As shown, the electronic device 1000 may include: at least one processor 1001, at least one network interface 1004, a user interface 1003, a memory 1005, and at least one communication bus 1002.

[0111] The communication bus 1002 is used to realize the connection and communication between these components.

[0112] The user interface 1003 may include a display screen and a camera. Optionally, the user interface 1003 may also include a standard wired interface and a wireless interface.

[0113] The network interface 1004 may optionally include a standard wired interface or a wireless interface (such as a Wi-Fi interface).

[0114] The processor 1001 may include one or more processing cores. The processor 1001 connects to various parts of the server using various interfaces and lines, and performs various server functions and processes data by running or executing instructions, programs, code sets, or instruction sets stored in the memory 1005, and by calling data stored in the memory 1005. Optionally, the processor 1001 may be implemented using at least one hardware form of Digital Signal Processing (DSP), Field-Programmable Gate Array (FPGA), or Programmable Logic Array (PLA). The processor 1001 may integrate one or a combination of several of the following: Central Processing Unit (CPU), Graphics Processing Unit (GPU), and modem. The CPU primarily handles the operating system, user interface, and applications; the GPU is responsible for rendering and drawing the content to be displayed on the screen; and the modem handles wireless communication. It is understood that the modem may also not be integrated into the processor 1001 and may be implemented as a separate chip.

[0115] The memory 1005 may include random access memory (RAM) or read-only memory. Optionally, the memory 1005 may include a non-transitory computer-readable storage medium. The memory 1005 can be used to store instructions, programs, code, code sets, or instruction sets. The memory 1005 may include a program storage area and a data storage area, wherein the program storage area may store instructions for implementing an operating system, instructions for at least one function (such as touch function, sound playback function, image playback function, etc.), instructions for implementing the above-described method embodiments, etc.; the data storage area may store data involved in the above-described method embodiments, etc. Optionally, the memory 1005 may also be at least one storage device located remotely from the aforementioned processor 1001. Figure 5 As shown, the memory 1005, which serves as a computer storage medium, may include an operating system, a network communication module, a user interface module, and an application program for a product recommendation method that verifies multi-source information.

[0116] exist Figure 5In the electronic device 1000 shown, the user interface 1003 is mainly used to provide an input interface for the user and to obtain the user input data; while the processor 1001 can be used to call an application program stored in the memory 1005 that performs a multi-source information verification method for product recommendation. When executed by one or more processors, the electronic device performs one or more of the methods described in the above embodiments.

[0117] An electronic device readable storage medium stores instructions that, when executed by one or more processors, cause the electronic device to perform one or more of the methods described in the above embodiments.

[0118] It should be noted that, for the sake of simplicity, the foregoing method embodiments are all described as a series of actions. However, those skilled in the art should understand that this application is not limited to the described order of actions, as some steps may be performed in other orders or simultaneously according to this application. Furthermore, those skilled in the art should also understand that the embodiments described in the specification are preferred embodiments, and the actions and modules involved are not necessarily essential to this application.

[0119] In the above embodiments, the descriptions of each embodiment have different focuses. For parts not described in detail in a certain embodiment, please refer to the relevant descriptions in other embodiments.

[0120] In the several embodiments provided in this application, it should be understood that the disclosed apparatus can be implemented in other ways. For example, the apparatus embodiments described above are merely illustrative; for instance, the division of units is only a logical functional division, and in actual implementation, there may be other division methods. For example, multiple units or components may be combined or integrated into another system, or some features may be ignored or not executed. Furthermore, the displayed or discussed mutual couplings, direct couplings, or communication connections may be through some service interfaces; indirect couplings or communication connections between devices or units may be electrical or other forms.

[0121] The units described as separate components may or may not be physically separate. The components shown as units may or may not be physical units; that is, they may be located in one place or distributed across multiple network units. Some or all of the units can be selected to achieve the purpose of this embodiment according to actual needs.

[0122] Furthermore, the functional units in the various embodiments of this application can be integrated into one processing unit, or each unit can exist physically separately, or two or more units can be integrated into one unit. The integrated unit can be implemented in hardware or as a software functional unit.

[0123] If the integrated unit is implemented as a software functional unit and sold or used as an independent product, it can be stored in a computer-readable storage device (CMD). Based on this understanding, the technical solution of this application, in essence, or the part that contributes to the prior art, or all or part of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a memory and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods described in the various embodiments of this application. The aforementioned memory includes various media capable of storing program code, such as USB flash drives, portable hard drives, magnetic disks, or optical disks.

[0124] The foregoing description is merely an exemplary embodiment of this disclosure and should not be construed as limiting the scope of this disclosure. Any equivalent changes and modifications made in accordance with the teachings of this disclosure shall still fall within the scope of this disclosure. Other embodiments of this disclosure will be readily apparent to those skilled in the art upon consideration of the specification and practice of the disclosure herein. This application is intended to cover any variations, uses, or adaptations of this disclosure that follow the general principles of this disclosure and include common knowledge or customary techniques in the art not described herein. The specification and embodiments are to be considered exemplary only, and the scope and spirit of this disclosure are defined by the claims.

Claims

1. A product recommendation method based on multi-source information verification, characterized in that, The method includes: Obtain initial information about the target products sent by each merchant. The initial information includes basic information and related information. The related information is generated by performing related searches based on the basic information. The basic information and the associated information are verified for completeness, and a first score and a second score are generated for each target product based on the completeness verification results. The credibility of the basic information and the associated information is verified, and a third score is generated for each of the target products based on the credibility verification results. The first score, the second score, and the third score are weighted and summed to generate a credibility score for each target product; the credibility scores are adjusted according to the credit index of each merchant to generate a target credibility score. The system receives a query request for a target product from a user, selects candidate target products corresponding to the query request from the target products, sends the initial information of the candidate target products with a target credibility score higher than a preset score to the big model, so that the big model can generate recommended content based on the initial information and push the recommended content to the user.

2. The product recommendation method based on multi-source information verification according to claim 1, characterized in that, After pushing the recommended content to the user's device, the process also includes: Acquire user interaction behavior data for the recommended content; if the interaction behavior data does not meet the preset behavior threshold conditions, convert the interaction behavior data into a weak supervision feedback signal. Based on the weak supervision feedback signal, the weighting weights used in generating the credibility score are updated online incrementally to update the target credibility score of each target product.

3. The product recommendation method based on multi-source information verification according to claim 1, characterized in that, The process of performing integrity verification on the basic information and the associated information, and generating a first score and a second score for each target product based on the integrity verification results, includes: The integrity and validity of the numerical range of each parameter field in the basic information are checked, and the metadata fingerprint of the multimodal file in the basic information is verified. Based on the results of the integrity check, the validity check, and the metadata fingerprint verification, a first score for each target product is generated. The data coverage of each associated dimension in the associated information is detected to determine whether the data coverage meets the preset coverage threshold. Based on the coverage detection results, a second score is generated for each of the target products.

4. The product recommendation method based on multi-source information verification according to claim 1, characterized in that, The step of verifying the credibility of the basic information and the associated information, and generating a third rating for each target product based on the credibility verification results, includes: The layout analysis and text recognition of the unstructured qualification documents attached to the basic information are performed to extract the target parameter values ​​from the unstructured qualification documents; The target parameter value is compared with the corresponding parameter value in the basic information and / or the associated information to calculate the deviation between the target parameter value and the corresponding parameter value. Based on the comparison between the deviation and the preset deviation threshold, a third score is generated for each of the target products.

5. The product recommendation method based on multi-source information verification according to claim 4, characterized in that, The step of generating a third rating for each target product based on the comparison result between the deviation and a preset deviation threshold includes: If the deviation is greater than the preset deviation threshold, the third score of the target product will be reset to zero, and the corresponding conflict information will be output to the diagnostic module so that the diagnostic module can trigger an abnormal alarm and send a data rectification notice to the corresponding merchant. If the deviation is not greater than a preset deviation threshold, a third score for each target product is generated by accumulating the number of qualification documents that pass the consistency comparison.

6. The product recommendation method based on multi-source information verification according to claim 1, characterized in that, The step of adjusting the credibility score based on each merchant's credit index to generate a target credibility score includes: The percentage of initial information submitted by each merchant within a preset time window that passes integrity verification and credibility verification is statistically analyzed to generate the historical verification pass rate for each merchant. Real-time collection of negative behavior data from user feedback, including at least one of user complaint data, return rate data, and information correction trigger data; The negative behavior data is weighted according to a time decay function to generate a cumulative negative feedback value for each merchant; By combining the historical verification pass rate with the cumulative value of negative feedback, a credit index is generated for each merchant; The credit index is used as an adjustment factor and multiplied with the credibility score to generate the target credibility score for each target product.

7. The product recommendation method based on multi-source information verification according to claim 1, characterized in that, The step of sending initial information of candidate target products with target credibility scores higher than preset scores to the large model, so that the large model can generate recommended content based on the initial information, includes: The initial information of the candidate target products is structurally assembled to generate a prompt text containing a basic information summary and a related information summary; The prompt text is combined with the user intent tags carried in the query request to generate input prompt words for the large model; The input prompts are sent to the large model so that the large model can generate recommendation content containing product comparison analysis and recommendation reasons based on the input prompts. The recommended content is subjected to factual consistency verification to determine whether there is a deviation between the parameter description in the recommended content and the corresponding parameter value in the initial information. If there is a deviation, the recommended content is corrected and then pushed to the user.

8. A product recommendation system with multi-source information verification, characterized in that, The system includes: an acquisition module, a first verification module, a second verification module, a generation module, and an output module; wherein, The acquisition module is used to acquire the initial information of the target products sent by each merchant. The initial information includes basic information and related information. The related information is generated by performing related searches based on the basic information. The first verification module is used to perform integrity verification on the basic information and the associated information respectively, and generate a first score and a second score for each target product based on the integrity verification results; The second verification module is used to verify the credibility of the basic information and the associated information, and generate a third score for each target product based on the credibility verification result; The generation module is used to perform a weighted summation of the first score, the second score, and the third score to generate a credibility score for each of the target products; and to adjust the credibility scores according to the credit index of each merchant to generate a target credibility score. The output module is used to receive a query request for a target product sent by the user terminal, select candidate target products corresponding to the query request from the target products, send the initial information of the candidate target products with a target credibility score higher than a preset score to the big model, so that the big model can generate recommended content based on the initial information and push the recommended content to the user terminal.

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

10. A computer-readable storage medium, characterized in that, The computer program is stored that can be loaded by a processor and executed as described in any one of claims 1-7.

Images