User recommendation method and system based on enterprise product portrait analysis

Abstract

The invention relates to the technical field of portrait construction, in particular to a user recommendation method and system based on enterprise product portrait analysis. The method comprises the following steps: obtaining enterprise product information data; performing supply chain tracing based on the enterprise product information data, and generating enterprise product supply chain information data; performing associated enterprise screening on the enterprise product supply chain information data to generate product associated enterprise information data; performing product part production proportion analysis on the enterprise product supply chain information data according to the product associated enterprise information data to generate a product part production proportion value; and performing product part input label quantification on the enterprise product information data based on the product part production ratio to generate product label data. According to the method, product supply chain tracing, part proportion quantification, user deep behavior analysis and scene perception recommendation are fused, so that the accuracy and comprehensiveness of user recommendation are improved.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of portrait construction, and in particular to a user recommendation method and system based on enterprise product portrait analysis. BACKGROUND

[0002] Traditional user recommendation methods mainly rely on collaborative filtering, content recommendation and other technologies. These methods have achieved certain results in the early stage, but in the scene where user portraits are not perfect and product information is low in structuralization, they are prone to problems such as cold start, low recommendation accuracy, etc. The wide application of artificial intelligence, especially deep learning technology, has injected new vitality into the recommendation system. Enterprises have begun to build refined product portrait models, structurally modeling products from multiple dimensions such as product attributes, functional characteristics, market positioning, user evaluation, etc., to form enterprise product portraits. On this basis, combined with user behavior data and preference information, multi-dimensional matching analysis is carried out to achieve more personalized and accurate user recommendation. However, the current traditional recommendation system mainly focuses on user behavior, ignoring the deep understanding of the structure and source of enterprise products, resulting in low matching degree of recommendation, and it is difficult to perceive the user scene in real time, affecting personalization and practicality. SUMMARY

[0003] Therefore, it is necessary to provide a user recommendation method and system based on enterprise product portrait analysis to solve at least one of the above technical problems.

[0004] To achieve the above purpose, a user recommendation method based on enterprise product portrait analysis, the method comprising the following steps: Step S1: obtaining enterprise product information data; performing supply chain traceability based on the enterprise product information data to generate enterprise product supply chain information data; performing associated enterprise screening on the enterprise product supply chain information data to generate product associated enterprise information data; Step S2: performing product component production proportion analysis on the enterprise product supply chain information data according to the product associated enterprise information data to generate product component production proportion value; performing product component input label quantification on the enterprise product information data based on the product component production proportion value to generate product label data; Step S3: constructing an enterprise product portrait based on the product label data; collecting purchase user information data by using the constructed enterprise product portrait to obtain purchase user information data; analyzing user interaction behaviors of the purchase user information data to generate purchase user interaction behavior data, wherein the user interaction behaviors include time sequence offset browsing scanning and collaborative path reverse behavior tracking; Step S4: positioning the user scene based on the low-power Bluetooth beacon through the purchase user interaction behavior data, and dynamically generating a personalized user recommendation list according to the edge server to perform the user recommendation job based on the enterprise product portrait analysis.

[0005] The present application can comprehensively master the component source, production link and key node of the product by acquiring enterprise product information data and performing supply chain traceability analysis, improve the supply chain transparency and controllability, and facilitate accountability and optimization. Using product component production proportion analysis, the input proportion of each component in the product can be quantitatively described, which is helpful for product cost optimization, key component identification and fine management. The enterprise product portrait constructed based on the label quantitative data can multi-dimensionally display product attributes, market positioning and technical features, forming core data support for user behavior analysis and recommendation. Collecting purchase user information and mining interactive behaviors such as time sequence offset browsing and collaborative path reverse tracking, the deep insight of user intention is realized, and the behavior basis for personalized service is provided. With the help of low-power Bluetooth beacon technology, user scene perception is realized, and dynamic recommendation calculation combined with the edge server is realized, which effectively improves the real-time, accuracy and personalization level of user recommendation. The enterprise product portrait and user behavior data are deeply integrated to realize the closed-loop data flow and intelligent linkage between the enterprise end and the user end, and the intelligent level of product marketing strategy is improved. Through dynamic personalized recommendation and interactive behavior feedback mechanism, the enterprise can timely adjust the product strategy, better meet the personalized needs of users, and enhance the user stickiness and brand loyalty. Therefore, the present application improves the accuracy and comprehensiveness of user recommendation by integrating product supply chain traceability, component proportion quantization, user deep behavior analysis and scene perception recommendation.

[0006] Preferably, step S1 comprises the following steps: Step S11: acquiring enterprise product information data; Step S12: performing field standardization processing on the acquired enterprise product information data to generate standardized product information data; based on the standardized product information data, the raw material source, processing node and circulation record of each batch of products are tracked by calling the enterprise resource planning system and product barcode database, and the supply chain data link integration is performed to generate enterprise product supply chain initial path data; Step S13: classifying the enterprise product supply chain initial path data, identifying the key path constituent elements of raw material suppliers, processing manufacturing enterprises and warehousing logistics nodes, and generating supply chain node information data; performing link topology arrangement on the supply chain node information data, constructing a multi-level supply chain link graph, eliminating redundant or broken link nodes, and generating enterprise product supply chain information data; Step S14: Calculate the supply strength value of each level node based on the enterprise product supply chain information data, and use the supply strength value for associated enterprise screening to generate product associated enterprise candidate data set; cross compare the risk sources of the product associated enterprise candidate data set, screen out high stability upstream and downstream enterprises, and finally generate product associated enterprise information data.

[0007] The present application effectively solves the problems of inconsistent formats and field ambiguity of different data sources through field standardization processing, lays a solid foundation for subsequent data processing and supply chain analysis. With the help of enterprise resource planning (ERP) system and product barcode database, the source of raw materials, processing nodes and flow path are tracked, and the multi-dimensional supply chain initial path of enterprise products is constructed, which improves the data coverage and traceability depth. Through node classification, raw material suppliers, processing manufacturers, warehousing logistics and other core roles are identified, and the supply chain division structure is clear, which provides data basis for risk control and node optimization. Adopting link topology arrangement technology, redundant and broken link nodes are removed, forming a visual enterprise product supply chain information map, which improves the transparency and management efficiency of the supply chain structure. By calculating the supply strength value (such as supply frequency, performance rate, stability, etc.) of each level node, the quantitative evaluation of the supply chain node is realized, and the scientificity and objectivity of associated enterprise screening are improved. Combined with risk source cross comparison, potential high-risk enterprises are further screened out, only stable and reliable upstream and downstream partners are retained, which significantly enhances the resilience and risk resistance ability of the supply chain. The finally generated product associated enterprise information data has high quality and high reliability, which lays a solid data foundation for subsequent component input analysis, portrait construction and personalized recommendation, realizes the systematic optimization from product to enterprise ecology.

[0008] Preferably, the step S14 of calculating the supply strength value of each level node based on the enterprise product supply chain information data, and using the supply strength value for associated enterprise screening includes: Structurally layering the enterprise product supply chain information data to identify each level of supply node, including raw material supply node, production processing node, assembly integration node and distribution node, to generate multi-level supply chain node structure data; Constructing a supply record statistical matrix for each node in the multi-level supply chain node structure data to generate node supply statistical data, wherein the supply record statistical matrix includes historical supply quantity, annual performance times, and cooperation cycle length; Quantitative index standardization processing on the node supply statistical data to calculate the supply frequency, supply stability and delivery proportion index of each node, construct a multi-dimensional supply capacity parameter table, and generate supply capacity feature data; Defining a node supply strength formula based on the supply capacity feature data, wherein the node supply strength formula is as follows: ; In the formula, is the supply strength value of the first node, is the supply frequency of the first node, is the fulfillment reliability of the first node, is the delivery proportion of the first node, , , is a custom weight coefficient; The supply strength value of each node is calculated by a node supply strength formula to generate supply strength index data; the supply strength index data is classified and divided, a supply strength boundary is set and strong, medium and weak classification is performed, and associated nodes with high and medium supply strength are screened out to generate supply strong node screening data; The supply strong node screening data is merged for enterprise uniqueness and supply chain role recognition, and redundant or coincident enterprises are removed to finally generate product associated enterprise candidate data set.

[0009] The present application can effectively improve the understanding and control ability of the supply chain hierarchical structure by structurally layering the enterprise product supply chain information data, accurately identifying multiple types of nodes such as raw materials, production and processing, assembly integration and distribution, etc. The construction of supply record statistical matrix and the extraction of historical supply quantity, fulfillment frequency and cooperation time length and other indexes can systematically describe the actual supply performance of each node, lay a data foundation for scientific evaluation of suppliers. The supply behavior is quantified through three index dimensions of supply frequency, fulfillment reliability and delivery proportion, a complete supply capacity parameter table is formed, and the supply capacity data is more comparable and operable. Using the node supply strength formula , the index weight can be flexibly adjusted according to specific business needs to realize a node scoring model more in line with the actual situation of the enterprise. Based on the calculated supply strength value, strong, medium and weak classification is performed to accurately screen out key nodes with high and medium levels, effectively improving the stability and quality assurance level of the core links of the supply chain. Through enterprise uniqueness merging and role recognition, redundant or duplicate enterprise information is removed to ensure the accuracy and business correspondence of the final screening results, laying a clear enterprise boundary for subsequent portrait modeling and user recommendation. By building a standardized, structured and quantifiable supply strength evaluation mechanism, enterprises can dynamically monitor and adjust cooperation nodes, reduce supply risks and improve the ability to respond to unexpected events of the supply chain.

[0010] Preferably, step S2 comprises the following steps: Step S21: mapping the product associated enterprise information data to the supply chain path, identifying the upstream supply source and actual production division ratio of each product component, and generating product component source path data; Step S22: Accumulate and count each type of component production record in the enterprise product supply chain information data, and combine the product component source path data to extract the actual output proportion of different enterprises by component granularity, and generate component production distribution matrix data; Step S23: Normalization and integration calculation are performed on the component production distribution matrix data to generate standardized product component production proportion value data; based on the product component production proportion value data, a component label input function is constructed: ; In the formula, is the label weight value of the product component , is the production proportion value of the component by the first enterprise, is the production proportion value of the component by the first enterprise, is the reputation coefficient of the first enterprise; the component label input function is used to calculate the label weight of the product component production proportion value data to generate component label quantization data. Step S24: Embedding each label in the product component list in the enterprise product information data, binding the component label quantization data with the specific component, and generating structured labeled product list data; performing unified product ID annotation, label weight aggregation and version timestamp synchronization processing on the structured labeled product list data, and finally generating product label data.

[0011] The present application can identify the upstream source and enterprise division ratio of each component by mapping the product associated enterprise information, realize the supply chain visualization at the component level, and provide data support for accurate supply management. Combined with product supply chain data and component source path, the enterprise output behavior is accumulated and counted and the proportion is extracted to form a real and quantitative component production distribution matrix, which effectively reflects the participation degree of enterprises in component manufacturing. The component label input function: The enterprise reputation coefficient The production proportion calculation is integrated to effectively avoid the allocation of labels only according to the quantity or proportion, so as to guide high-quality enterprises to obtain higher label weight and enhance the credibility and incentive effect of the label mechanism. The label quantitative data corrected by the component granularity and reputation makes the label not only reflect "whether to participate", but also "how to participate with what quality and contribution", which significantly enhances the discriminability and credibility of the label. The label quantitative data is embedded in the component structure of the product information to form a structured labeled product list, which provides standardized basic data for intelligent retrieval, intelligent matching and visual display. Through unified product ID labeling, weight aggregation and timestamp synchronization, the label data and product information are updated synchronously and the historical version is traceable, which enhances the maintainability and data integrity of the system. The product label data is the core input source of the subsequent product portrait, and its structured, standardized and dynamic updating characteristics can significantly improve the portrait accuracy, label interpretation and user matching degree.

[0012] Preferably, the product label data-based enterprise product portrait construction of the enterprise product information data in step S3 comprises: label type classification is performed on the product label data to generate label category aggregation data; the label density of the label category aggregation data is calculated; weight normalization is performed on the label density data to generate label weight distribution data; a product feature vector is constructed based on the label weight distribution data, and principal component fusion is performed on the product feature vector data to generate principal component fusion feature data; label graph construction processing is performed on the principal component fusion feature data to generate label graph structure data; the label graph structure data is embedded and mapped to generate enterprise product portrait vector data; the enterprise product portrait vector data is subjected to spatial dimension reduction to generate enterprise product portrait data.

[0013] The application classifies and aggregates the label data, structurally manages the label information, reduces data redundancy, and improves the efficiency and accuracy of subsequent analysis and processing. The density of the label category is calculated to objectively reflect the distribution of labels of each category in the product, providing a data basis for accurately capturing product features. The label density data is subjected to weight normalization processing to eliminate dimensional influence and reasonably allocate the contribution of each category label in portrait construction, improving the balance of the portrait expression. Based on the label weight distribution, a product feature vector is generated to comprehensively reflect the multi-level and multi-dimensional features of the product. Through principal component fusion, feature redundancy is removed, the main components are highlighted, the discrimination and stability of the portrait are enhanced, and the efficiency and accuracy of the subsequent model are improved. The construction of the label atlas reveals the internal correlation and hierarchical relationship between the labels, enriches the semantic information of the product features, and provides multi-dimensional structural support for the portrait. The embedding technology is used to convert the label atlas into an enterprise product portrait vector, taking into account the expression ability and calculation efficiency, facilitating the storage, retrieval and analysis of the portrait. Through spatial dimension reduction processing, the portrait vector data is simplified, facilitating fast response and efficient operation in various application scenarios, while maintaining the core information of the portrait. The generated enterprise product portrait data has high information fusion and structural features, providing accurate input for subsequent purchase user information collection and interaction behavior analysis, and improving the intelligent level of the recommendation system.

[0014] Preferably, the step S3 of analyzing the user interaction behavior of the purchase user information data comprises: extracting user access records and page interaction logs in the purchase user information data; reconstructing user behavior trajectories based on the user access records to generate behavior path data; calculating time sequence features of the user behavior path data to obtain time sequence offset features of the user; based on the time sequence offset features, identifying the browsing scan mode of the user within a specific time window to generate browsing scan behavior data; based on the behavior path data, constructing a collaborative path atlas between users to identify existing collaborative path behaviors and generate collaborative path behavior data; performing reverse behavior path tracking on the collaborative path atlas according to the collaborative path behavior data to identify path convergence nodes and behavior reverse paths, and generating collaborative path reverse tracking data; performing fusion processing based on the browsing scan behavior data and the collaborative path reverse tracking data to identify user interaction behavior mode features, and generating user interaction behavior mode data; based on the user interaction behavior mode data, constructing structured purchase user interaction behavior data.

[0015] The application restores the user's behavior path completely by reconstructing the user access record and page interaction log, guarantees the space-time continuity and integrity of the user behavior data, and lays a solid foundation for accurate analysis. The time sequence characteristics of the user behavior path are calculated to capture the time sequence offset and dynamic fluctuation of the user behavior, and the understanding ability of the user interest change and behavior rhythm is improved. By using the time sequence offset feature, the browsing and scanning behavior of the user in a specific time window is analyzed, the focus and interest switching of the user on the page are captured, and the content display and recommendation strategy is optimized. The user collaborative path graph is established by the behavior path data, the implicit collaborative behavior mode between users is found, which helps to mine the group behavior characteristics and social influence factors. The convergent nodes and behavior reverse path in the collaborative path are identified by reverse behavior path tracking, which helps to accurately analyze the key touch points and causal relationship in the user behavior chain, and improves the explanation of the behavior mode. The browsing and scanning behavior and the reverse tracking data of the collaborative path are fused, the multi-dimensional information is integrated, and more representative user interaction behavior mode is extracted, which improves the accuracy and richness of the user behavior model. The structured and high-dimensional user interaction behavior data is formed, which supports various intelligent application scenarios such as personalized recommendation, user portrait improvement and behavior prediction, and enhances the matching degree and interaction efficiency of enterprise products and users. Deeply understand the user interaction behavior and its space-time evolution, provide data-driven user insight for enterprises, optimize product display and recommendation process, and significantly improve user satisfaction and conversion rate.

[0016] Preferably, the reverse behavior path tracking of the collaborative path graph based on the collaborative path behavior data includes: A directed graph model of the collaborative path graph is constructed, nodes represent user behavior events, edges represent behavior conversion relationships, and collaborative path graph model data is generated; Based on the collaborative path behavior data, a target behavior event node is selected as a backtracking starting point, a tracking depth and a path weight threshold are set, and reverse tracking parameters are initialized; Based on the reverse tracking parameters, the reverse path search is performed on the collaborative path graph model, the upstream behavior path that can be traced back from the target behavior node is identified, and a candidate reverse path set is generated; The candidate reverse path set is subjected to behavior mode similarity analysis and path convergence degree calculation, the path nodes that are commonly converged by multiple users in the behavior path are identified, and path convergence node data is generated; The path convergence node and the reverse path behavior sequence are combined to cluster and induce the user behavior path, extract the behavior reverse path structure mode, and generate behavior reverse path data; The path convergence node data and the behavior reverse path data are integrated to generate collaborative path reverse tracking data.

[0017] The application realizes systematic and visual expression of complex user behavior sequence by nodalizing user behavior events and converting behavior into edges, constructing a directed graph model of collaborative path atlas, and facilitating in-depth analysis of behavior evolution process. Taking the target behavior event as the backtracking starting point, combining the tracking depth and path weight threshold, the range and accuracy of reverse path search can be flexibly adjusted to meet different analysis needs and improve the pertinence and effectiveness of tracking. Through the reverse path search algorithm, the upstream multiple behavior paths of the target behavior node can be quickly traced back from the target behavior node, the hidden user behavior precursor events are mined, and the understanding of user decision-making process is enriched. The candidate reverse path set is compared in terms of behavior pattern similarity and path convergence degree, the key path nodes of multiple user intersections are effectively identified, the commonality and key touch points of user behavior are reflected, and the discovery of behavior rules is promoted. Combined with the path convergence nodes and reverse path behavior sequence, the behavior paths are induced through clustering analysis, and the typical reverse path structure mode is abstracted, which helps to reveal the internal logic and evolution mechanism of user behavior. The path convergence nodes and behavior reverse path data are organically integrated to form a complete collaborative path reverse tracking data, which provides a solid foundation for subsequent user behavior prediction, anomaly detection and personalized recommendation. The reverse tracking technology digs deep into the user behavior chain, accurately locates the key nodes and paths, enhances the enterprise's ability to understand the user behavior evolution process, and provides data support for product optimization and user experience improvement. By identifying the path convergence and behavior reverse structure among users, the collaborative behavior characteristics and group influence mechanism are revealed, and more effective marketing strategies and user relationship management are supported.

[0018] Preferably, step S4 comprises the following steps: Step S41: performing low-power Bluetooth beacon sensing processing based on the purchase user interaction behavior data to generate user scene sensing data; extracting spatial position information of the user scene sensing data to generate user positioning coordinate data; Step S42: performing scene semantic matching analysis on the user positioning coordinate data and enterprise product portrait data to generate user behavior portrait data; extracting behavior interest factors of the user behavior portrait data to generate user preference factor data; Step S43: performing correlation measurement analysis on the user preference factor data and enterprise product portrait to generate portrait correlation degree data; performing edge server task dispatching processing on the portrait correlation degree data to generate edge-level recommendation request data; Step S44: performing time window adaptive rearrangement on the edge-level recommendation request data to generate dynamic priority ordering data; performing high correlation matching on the dynamic priority ordering data and product feature vector data to generate personalized user recommendation list data; Step S45: performing client-side visual processing on the personalized user recommendation list data to generate user recommendation interface view data.

[0019] The application can efficiently obtain real-time spatial position information of users by sensing user interaction behavior through low-power Bluetooth beacons, support accurate perception of specific scenes where users are located, and improve timeliness and accuracy of recommendations. By combining spatial position information with enterprise product portraits, scene semantic matching analysis is completed, user behavior characteristics and interest preferences are accurately captured, and rich user behavior portraits and preference factor data are generated to provide accurate basis for personalized recommendations. Based on the correlation measurement of user preference factors and enterprise product portraits, deep matching of user demand and product features is realized to ensure that recommended content highly matches user interest and purchase behavior, thereby enhancing user satisfaction and conversion rate. The edge server is used for task assignment and time window adaptive rearrangement to effectively optimize the processing order of recommendation requests, improve the response speed and real-time performance of the recommendation system, reduce network delay, and improve the stability and concurrent processing capacity of the system. Through dynamic priority sorting and high correlation matching of product feature vectors, the recommendation strategy can be adjusted in real time, user behavior changes can be flexibly responded to, and dynamic updating and accurate delivery of personalized recommendation lists can be realized. The client-side visual processing of personalized recommendation results generates a clear, intuitive and user-friendly recommendation interface view, improves user interaction experience, and promotes user acceptance and participation of recommended content. By using edge computing and low-power beacon technology, distributed collaboration of computing and sensing is realized, the load of the central server is reduced, real-time recommendation for a large number of users is supported, and the overall resource utilization efficiency and scalability of the system are improved.

[0020] Preferably, the step S44 of performing time window adaptive rearrangement on the edge-level recommendation request data to generate dynamic priority sorting data comprises: extracting the timestamps, request frequencies, user identifiers and content types in the edge-level recommendation request data to construct a request event sequence set; based on the request event sequence set, setting initial sliding time window parameters including time window size, sliding step and trigger threshold to generate a basic time window structure; analyzing the request density change and request interval distribution within the time window and calculating the dynamic activity index of each time window segment; According to the activity index, the time window size and the sliding step are adaptively adjusted to form an adaptive time window sequence with local optimal response performance; In each adaptive time window, the importance weight of the request is calculated, wherein the weight factor of the importance weight includes: request source node level, content popularity, request frequency and response delay requirement; Based on the importance weight of the request, the recommendation requests within the time window are locally prioritized, and the cross-window optimization is combined with the global sliding window position to generate dynamic priority sorting data.

[0021] The application can respond to changes in request density more flexibly by dynamically adjusting the time window size and sliding step, avoid delays or resource waste caused by fixed time windows, and ensure that recommendation requests are processed more timely. The adaptive time window design can automatically adjust the processing rhythm according to the fluctuations in user activity in different time periods, so that the system can quickly respond during peak periods and save computing resources during low activity periods, achieving load balancing. Combining request source node level, content popularity, request frequency and response delay, etc. Multi-dimensional weight factors can accurately assess the importance of requests, ensure that critical requests are processed first, and improve the relevance of recommendations and user satisfaction. The combination of local priority sorting and global sliding window cross optimization makes the request sorting more reasonable, effectively reduces resource conflicts and bottlenecks, and improves the concurrent processing capacity of edge servers and system throughput. Dynamic priority sorting data support system adjusts itself according to real-time environmental changes, ensures that the recommendation task still maintains efficient operation in complex network environment, and adapts to diversified user demand and scene changes. The priority sorting mechanism can be flexibly adjusted according to different user and content characteristics, helping to achieve more accurate personalized recommendation, and improving user experience and product conversion rate.

[0022] In the present specification, a user recommendation system based on enterprise product portrait analysis is provided for executing the user recommendation method based on enterprise product portrait analysis described above, and the user recommendation system based on enterprise product portrait analysis comprises: An enterprise screening module is configured to acquire enterprise product information data, perform supply chain traceability based on the enterprise product information data, and generate enterprise product supply chain information data; and perform associated enterprise screening on the enterprise product supply chain information data, and generate product associated enterprise information data. A product label module is configured to perform product component production proportion analysis on the enterprise product supply chain information data based on the product associated enterprise information data, and generate product component production proportion values; and perform product component input label quantification on the enterprise product information data based on the product component production proportion values, and generate product label data. A user interaction analysis module is configured to construct an enterprise product portrait based on the product label data, and perform purchase user information data acquisition using the constructed enterprise product portrait, and obtain purchase user information data; and analyze user interaction behaviors of the purchase user information data, and generate purchase user interaction behavior data, wherein the user interaction behaviors include time sequence offset browsing scanning and collaborative path reverse behavior tracking. A personalized recommendation module is configured to position a user scene based on a Bluetooth Low Energy beacon through the purchase user interaction behavior data, and dynamically generate a personalized user recommendation list based on an edge server, so as to perform a user recommendation job based on enterprise product portrait analysis.

[0023] The beneficial effect of the present invention is that through the enterprise screening module, the system not only obtains enterprise product information data, but also conducts in-depth supply chain node analysis based on supply chain traceability, and screens out key related enterprises in combination with the supply strength assessment model to ensure that the upstream and downstream nodes of the product are transparent and traceable. The product labeling module calculates the proportion of component production based on the enterprise capacity allocation data and the enterprise reputation factor, thereby realizing quantitative input analysis of labels, making product labels traceable and quantifiable, and laying the foundation for subsequent personalized recommendations. The user interaction analysis module integrates time series behavior offset recognition, browsing scanning analysis and collaborative path reverse tracing technology to extract users' real interaction intentions and potential interest preferences from access data, and generate structured and high-quality user behavior portraits. The personalized recommendation module combines low-power Bluetooth beacons to achieve precise user positioning, conducts interest deduction based on the semantic matching of the user's actual physical scene and product portrait, and realizes dynamic sorting and rapid response of recommendation requests through edge computing, thereby improving the real-time and relevance of recommendations. The module collaborative design enables efficient coupling of product labels and user portraits. Combined with the task dispatching capabilities of the edge server, it can significantly reduce the central computing burden, improve response speed and system throughput, and adapt to the concurrent recommendation needs of multiple scenarios. The clear functional boundaries between modules and the standardized data interface design facilitate rapid access and expansion across different industry chains, product structures, and user behavior models, resulting in excellent system portability and adaptability. Therefore, by integrating product supply chain traceability, component share quantification, in-depth user behavior analysis, and scenario-aware recommendations, this invention improves the accuracy and comprehensiveness of user recommendations. BRIEF DESCRIPTION OF THE DRAWINGS

[0024] Figure 1 This is a flowchart of the steps of a user recommendation method based on enterprise product portrait analysis; Figure 2 for Figure 1 Detailed implementation steps of step S1 in FIG. Figure 3 for Figure 1 Detailed implementation steps of step S2 in FIG. The purpose, features and advantages of the present invention will be further described with reference to the accompanying drawings and in conjunction with the embodiments. DETAILED DESCRIPTION

[0025] The following is a clear and complete description of the technical method of the present invention in conjunction with the accompanying drawings. It is obvious that the embodiments described are part of the embodiments of the present invention, but not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without making any creative efforts are within the scope of protection of the present invention.

[0026] In addition, the accompanying drawings are included to provide a further understanding of the present application, and are incorporated in and constitute a part of this specification. The drawings illustrate embodiments of the present application and, together with the description, serve to explain the principles of the present application. In the drawings:

[0027] It should be understood that, although the terms "first", "second" and the like can be used herein to describe various elements, these elements should not be limited by these terms. These terms are only used to distinguish one element from another. For example, a first element could be termed a second element, and, similarly, a second element could be termed a first element without departing from the scope of the example embodiments. The term "and / or" as used herein includes any and all combinations of one or more of the associated listed items.

[0028] To achieve the above object, there is provided Figures 1 to 3 A user recommendation method based on enterprise product portrait analysis, the method comprising the following steps: Step S1: obtaining enterprise product information data; performing supply chain traceability based on the enterprise product information data to generate enterprise product supply chain information data; performing associated enterprise screening on the enterprise product supply chain information data to generate product associated enterprise information data; Step S2: performing product component production proportion analysis on the enterprise product supply chain information data according to the product associated enterprise information data to generate product component production proportion values; performing product component input label quantification on the enterprise product information data based on the product component production proportion values to generate product label data; Step S3: constructing an enterprise product portrait based on the product label data for the enterprise product information data; collecting purchase user information data using the constructed enterprise product portrait to obtain purchase user information data; analyzing user interaction behaviors of the purchase user information data to generate purchase user interaction behavior data, wherein the user interaction behaviors include time sequence offset browsing scanning and collaborative path reverse behavior tracking; Step S4: positioning a user scene based on a low-power Bluetooth beacon through the purchase user interaction behavior data, and dynamically generating a personalized user recommendation list according to an edge server to perform a user recommendation job based on enterprise product portrait analysis.

[0029] The application can comprehensively master the component source, production link and key node of the product, improve the transparency and controllability of the supply chain, and facilitate accountability and optimization by obtaining enterprise product information data and performing supply chain traceability analysis. The product component production proportion analysis can quantitatively describe the input proportion of each component in the product, which is helpful for product cost optimization, key component identification and fine management. The enterprise product portrait constructed based on the label quantitative data can multi-dimensionally display product attributes, market positioning and technical characteristics, forming core data support for user behavior analysis and recommendation. The user information is collected and the interactive behaviors such as time offset browsing and collaborative path reverse tracking are mined to realize deep insight into user intent and provide behavior basis for personalized service. With the help of low-power Bluetooth beacon technology, user scene perception is realized, and combined with edge server dynamic recommendation calculation, the real-time, accuracy and personalization level of user recommendation are effectively improved. The enterprise product portrait and user behavior data are deeply integrated to realize closed-loop data flow and intelligent linkage between enterprise and user ends, and the intelligent level of product marketing strategy is improved. Through dynamic personalized recommendation and interactive behavior feedback mechanism, the enterprise can timely adjust product strategy to better meet the personalized needs of users and enhance user stickiness and brand loyalty. Therefore, the application improves the accuracy and comprehensiveness of user recommendation by integrating product supply chain traceability, component proportion quantification, user deep behavior analysis and scene perception recommendation.

[0030] In the embodiment of the application, as shown in the reference Figure 1 The user recommendation method based on enterprise product portrait analysis includes the following steps: Step S1: Obtain enterprise product information data; perform supply chain traceability based on the enterprise product information data to generate enterprise product supply chain information data; perform associated enterprise screening on the enterprise product supply chain information data to generate product associated enterprise information data; Step S2: Perform product component production proportion analysis on the enterprise product supply chain information data according to the product associated enterprise information data to generate product component production proportion value; perform product component input label quantification on the enterprise product information data based on the product component production proportion value to generate product label data; Step S3: Construct an enterprise product portrait based on the product label data; collect purchase user information data by using the constructed enterprise product portrait to obtain purchase user information data; analyze user interactive behavior of the purchase user information data to generate purchase user interactive behavior data, wherein the user interactive behavior includes time offset browsing scanning and collaborative path reverse behavior tracking; Step S4: Position the user scene based on the low-power Bluetooth beacon through the purchase user interaction behavior data, and dynamically generate a personalized user recommendation list according to the edge server to perform the user recommendation job based on the enterprise product portrait analysis.

[0031] In the embodiment of the present application, by acquiring enterprise product information data, the data includes product number, category, component information, production batch, raw material details, etc. Based on this information, the blockchain or supply chain management system interface is called to perform supply chain traceability operation, track the whole process from raw material procurement, component manufacturing, whole machine assembly to final factory delivery, and generate enterprise product supply chain information data. Then, all upstream and downstream enterprises in the supply chain data are analyzed, and the associated enterprise screening is performed by setting the screening conditions such as cooperation frequency, product component importance, transaction amount, etc. to extract the enterprise structure with key contribution relationship and generate product associated enterprise information data. According to the product associated enterprise information data, the function component responsibility of each enterprise in the whole product supply chain is identified, and the product component production proportion of each enterprise is weighted calculated by combining the bill of materials BOM and production records, forming the product component production proportion value (for example, weighted by key component number or value). Then, according to the production proportion value, the product component input label quantity of each component in the enterprise product information data is quantified, that is, a weighted label is attached to each component to identify its importance in the whole machine product, supply risk level, etc. Finally, structured product label data is generated. Taking the product label data as the core, the full product of the enterprise is feature coded to build an enterprise product portrait with hierarchical structure, which can include product composition distribution, supply chain risk level, component stability index, pricing elasticity index, etc. Based on the constructed enterprise product portrait, the purchase user information data collection is performed, and the collection methods include: code scanning activation, terminal login, after-sales registration, etc. to obtain the purchase user information data. Further analyze the user behavior path contained in these data to identify user interaction behavior, including but not limited to time offset browsing scanning (i.e. the user has a delayed attention behavior to a specific product or component), collaborative path reverse behavior tracking (i.e. the user has a front and back staggered comparison behavior to products of the same type, etc.). Finally, the purchase user interaction behavior data with behavior mode characteristics is generated. Based on the user interaction behavior data, when the user enters the offline store, exhibition hall or industrial exhibition scene, the user is positioned in real time through the low-power Bluetooth beacon (BLE Beacon), and the historical behavior data of the user, the current area product exhibition, the interactive hot area information, etc. are combined to dynamically generate a personalized user recommendation list through an edge computing server. The recommendation list will fully integrate the enterprise product portrait dimensions (such as high component weight products, high stability products, key enterprise supply products) and the user's personal preferences and interaction path, so as to accurately perform user recommendation work based on enterprise product portrait analysis, and improve product guiding efficiency and conversion rate.

[0032] As an example of the present application, reference is made to Figure 2 In this example, the step S1 includes: Step S11: acquiring enterprise product information data; Step S12: Field standardization processing is performed on the acquired enterprise product information data to generate standardized product information data; based on the standardized product information data, the original material source, processing node and circulation record of each batch of products are tracked by calling the enterprise resource planning system and the product barcode database, and the supply chain data link is integrated to generate enterprise product supply chain initial path data; Step S13: The enterprise product supply chain initial path data is subjected to node classification to identify the key path constituent elements of the raw material suppliers, processing and manufacturing enterprises and warehousing logistics nodes to generate supply chain node information data; the supply chain node information data is subjected to link topology arrangement to construct a multi-level supply chain link map, and redundant or broken link nodes are eliminated to generate enterprise product supply chain information data; Step S14: The supply strength value of each node is calculated based on the enterprise product supply chain information data, and the supply strength value is used for associated enterprise screening to generate a product associated enterprise candidate data set; the product associated enterprise candidate data set is subjected to risk source cross comparison to screen out high-stability upstream and downstream enterprises to finally generate product associated enterprise information data.

[0033] In the embodiments of the present application, the core information data related to the product is obtained by connecting the product information management system (PIMS), enterprise resource planning system (ERP) and historical order database of the enterprise. The data includes but is not limited to product code, product name, specification, production batch, packaging barcode, factory time, business unit, etc. The data format is a structured table, usually in the form of JSON, CSV or database table. The collection process supports API interface calling or database direct connection, ensuring real-time data synchronization, and updating records in batch processing mode by day or hour, generating original enterprise product information data, providing basic input for subsequent processing. The original enterprise product information data is standardized by field processing, unifying the naming rules, data types and coding formats of each field, such as mapping the synonymous fields of "productID", "product number" and "commodity barcode" to "ProductCode", unifying the date format to ISO 8601 (YYYY-MM-DD), and filling in the missing fields or filtering the records, to generate standardized product information data. After standardization, the system automatically calls the raw material record interface integrated with the ERP system and the traceability module associated with the product barcode database to perform traceability analysis on each standardized product one by one. According to the batch number and barcode information of the product, the system traces the raw material source enterprise, production and processing nodes, quality inspection and packaging links, warehouse logistics flow records and other key events in the supply chain. This process supports graph structure storage, and constructs a chain path structure between the starting point (raw material), transfer point (processing and manufacturing, quality inspection and packaging) and end point (finished product) for each product, and finally generates enterprise product supply chain initial path data. Based on the supply chain initial path data, structure analysis and semantic recognition algorithms are used to classify and process various nodes involved in the path. Through node attribute matching and role label recognition, key links such as raw material supplier nodes, processing and manufacturing enterprise nodes, and warehouse logistics nodes are distinguished, and corresponding node role data tables are established to generate supply chain node information data. Subsequently, based on this data, graph modeling algorithms such as NetworkX or Neo4j are used to construct a multi-level supply chain link graph, and the dependency relationship between the links is hierarchically sorted and visualized modeled, supporting hierarchical retrieval and interactive query. During the graph construction process, the system automatically identifies isolated nodes, broken link nodes and duplicate paths, and prunes and cleans redundant nodes by setting weights and connection rules, and finally generates optimized enterprise product supply chain information data. Based on the enterprise product supply chain information data, a supply strength value calculation model is used to quantitatively analyze the stability and relevance of each node. The supply strength value (Supply Strength Index) calculation model considers the supply frequency, historical delivery cycle, supply continuity, abnormal interruption probability and other indicators of the node, and the calculation formula is: wherein, is the supply strength value of the i th node. The supply intensity value of each node, For the The supply frequency of each node, For the The performance reliability of each node, For the The delivery ratio of nodes, , , The default values ​​are 0.4, 0.3, and 0.3, respectively. Based on the calculated strength ranking, the companies represented by each node are screened for correlation, forming a candidate data set of product-related companies. Next, by invoking the enterprise risk monitoring platform, we cross-reference the candidate data set with multi-dimensional risk source data, including financial risk, credit rating, legal records, and industry compliance. We identify upstream and downstream companies with low risk levels, stable historical collaborations, and strong supply reliability, ultimately generating product-related company information data.

[0034] Preferably, in step S14, calculating the supply strength values ​​of nodes at all levels based on the enterprise product supply chain information data, and using the supply strength values ​​to screen related enterprises includes: Structural layering of enterprise product supply chain information data, identifying supply nodes at all levels, including raw material supply nodes, production and processing nodes, assembly and integration nodes, and distribution nodes, to generate multi-level supply chain node structure data; Construct a supply record statistical matrix for each node in the multi-level supply chain node structure data to generate node supply statistical data, where the supply record statistical matrix includes historical supply quantity, annual fulfillment times, and cooperation cycle length; Standardize the node supply statistics through quantitative indicators, calculate the supply frequency, supply stability, and delivery ratio of each node, construct a multi-dimensional supply capacity parameter table, and generate supply capacity characteristic data; The node supply strength formula is defined based on the supply capacity characteristic data, where the node supply strength formula is as follows: ; Where, For the The supply intensity value of each node, For the The supply frequency of each node, For the The performance reliability of each node, For the The delivery ratio of nodes, , , is a custom weight coefficient; The supply strength value of each node is calculated by a node supply strength formula to generate supply strength index data; the supply strength index data is classified and divided, the supply strength boundary is set and strong, medium and weak classification is performed, the associated nodes with high and medium supply strength are screened out, and supply strength node screening data is generated; The supply strength node screening data is merged and the supply chain role is identified to eliminate redundant or coincident enterprises, and finally a product associated enterprise candidate data set is generated.

[0035] In the embodiment of the application, by performing structural hierarchical processing on enterprise product supply chain information data, combining the path dependence relationship and node role label in the supply chain map, different levels of supply node types are identified, including raw material supply nodes, production processing nodes, assembly integration nodes and product distribution nodes. By classifying and organizing the node attribute fields (such as supplier code, processing plant name, warehouse center location, channel merchant identification, etc.), a multi-level chain structure data model is constructed, multi-level supply chain node structure data is generated, and subsequent vertical and horizontal feature calculation between nodes is facilitated. Then, on the basis of the multi-level supply chain node structure data, a set of supply record statistical matrix is constructed for each node. The matrix comprehensively collects key quantitative data of the node in previous supply behavior, including total historical supply quantity, annual performance frequency in the past three or five years, and cumulative period (in months or quarters) of cooperation with the target enterprise, and the data is automatically aggregated and filled into the node attribute table by the database to generate node supply statistical data as the original input data set of supply behavior stability. Then, the node supply statistical data is uniformly quantified and standardized. The Z-score or Min-Max normalization method is used to standardize the core indexes such as supply frequency (supply frequency / period length), supply stability (performance success rate), and delivery proportion (node supply amount / product total amount), construct a structured multi-dimensional supply capability parameter table, and generate supply capability feature data of each node according to the table, to ensure that each dimension can be comprehensively evaluated within the same order of magnitude. On this basis, the quantitative calculation formula of node supply strength is defined to measure the supply capability and reliability of each node in the supply chain. The node supply strength formula is as follows: In the formula, is the supply strength value of the i th node, is the supply frequency of the i th node, is the performance reliability of the i th node, is the delivery proportion of the i th node, , , , , , , The weight coefficient is a customizable one; the weights can be flexibly configured based on the company's priorities (e.g., 0.4:0.4:0.2 or other ratios). Based on this formula, the system automatically performs batch calculations on each node to generate standardized supply strength index data. Subsequently, the supply strength index data is categorized into three levels: high, medium, and low. Supply strength cutoffs are set based on historical distribution ranges, company-level experience thresholds, or quantile methods (e.g., upper and lower quartiles). The system prioritizes nodes with medium-to-high strength ratings (e.g., supply strength values ​​in the top 60%) as reliable supplier candidates, generating a screening dataset of strong supply nodes. Finally, the screening data is consolidated to ensure enterprise uniqueness. Duplicate records of the same enterprise across multiple nodes are identified using a unified enterprise identifier (e.g., unified social credit code, organizational code, etc.). These records are then merged and supply chain roles are identified to ensure that a single enterprise is not selected repeatedly due to diverse roles. The system further eliminates enterprises with logical redundancy or duplication of paths, ultimately outputting a deduplicated candidate dataset of product-linked enterprises for subsequent risk assessment, partnership recommendations, or strategic supplier management modules.

[0036] As an example of the present invention, refer to Figure 3 As shown, in this example, step S2 includes: Step S21: mapping the supply chain path of the product-related enterprise information data, identifying the upstream supply source and actual production division ratio of each product component, and generating product component source path data; Step S22: accumulating and counting the production records of various components in the enterprise product supply chain information data, and combining the product component source path data, extracting the actual output proportion of different enterprises by component granularity, and generating component production allocation matrix data; Step S23: Normalize and integrate the component production allocation matrix data to generate standardized product component production ratio data; based on the product component production ratio data, construct a component label input function: ; Where, For product components The label weight value of For the Business-to-Component The production share of For the The company's reputation coefficient; using the component label input function to calculate the label weight of the product component production ratio data, to generate component label quantitative data; Step S24: embed labels item by item in the product component list in the enterprise product information data, associate and bind the component label quantification data with the specific components, and generate structured labeled product list data; perform unified product ID labeling, label weight aggregation and version timestamp synchronization on the structured labeled product list data, and finally generate product label data.

[0037] In an embodiment of the present invention, supply chain path mapping is performed on product-related enterprise information data. The system calls the constructed multi-level supply chain map and, in combination with the product structure BOM (Bill of Materials) list, traces back the actual source path of each component in the product in the supply chain network, and identifies each component's upstream supplier, transmission node, and processing unit layer by layer. Combined with historical procurement records and order collaboration data, the output responsibility of the actual participating enterprises in the component source path is determined, and the actual supply ratio and production division role of each enterprise for the component are clarified, thereby generating product component source path data as the basis for subsequent weight allocation. Based on the historical production and delivery records in the product supply chain information data, the focus is on the production responsibility of various components, and cumulative statistics are performed by time period (such as year, quarter) to obtain enterprise-level output data for each component. Combined with the product component source path data generated in step S21, the system matches the enterprise output records item by item at the component granularity, calculates the output quantity and frequency of each enterprise on a specific component, and unifies the standard into a percentage form (for example, a certain component is produced by enterprise A for 60% and enterprise B for 40%), forming a two-dimensional mapping relationship. Finally, the component production allocation matrix data is constructed, with the matrix rows representing the component numbers, the columns representing the participating enterprises, and the percentage values ​​filled in as matrix elements. The component production allocation matrix data is normalized to eliminate abnormal missing items or duplicate data, and the percentage values ​​of each item are standardized to a unified scale between 0 and 1. Subsequently, based on the normalized data, the component label input function is constructed as follows: Where, For product components The label weight value of For the Business-to-Component The production share of For the The reputation coefficient of each enterprise is calculated; the label weight is calculated for the product component production ratio data using the component label input function to generate component label quantitative data; this function is used to multiply and accumulate each enterprise's participation ratio for components with its reputation to form a label weight index that reflects the comprehensive supply quality of components. The system traverses and calculates the , output the corresponding component label quantization data, which is used for subsequent product life cycle identification. The component list in the enterprise product information data is structured and embedded with the corresponding label weight value for each component number as the primary key. The system binds the label value with the component to generate structured product component list data with quantized label characteristics, and further performs unified product ID labeling (such as associating product serial number or SKU), label weight aggregation (such as average label score or weighted score of the whole product), and version timestamp processing (recording label generation time, version number, etc.). Finally, the system outputs standardized product label data that can be used for product traceability, quality evaluation, and risk judgment, which can be called and displayed in regulatory, supply chain evaluation, or consumer visualization scenarios.

[0038] Preferably, the enterprise product portrait construction based on product label data in step S3 includes: label type classification of product label data to generate label category aggregation data; calculate the label density of the label category aggregation data; weight normalization of label density data to generate label weight distribution data; construct a product feature vector based on the label weight distribution data, and fuse the product feature vector data to generate principal component fusion feature data; construct a label graph based on the principal component fusion feature data to generate label graph structure data; embed and map the label graph structure data to generate enterprise product portrait vector data; reduce the dimensionality of the enterprise product portrait vector data to generate enterprise product portrait data.

[0039] In the embodiments of the present application, the product label data is classified by label type. The system divides the product labels into multiple categories according to the source attributes, performance dimensions and evaluation targets of the labels, such as quality labels (such as supply strength, raw material grade), supply labels (such as compliance timeliness, delivery stability), structure labels (such as key component coverage, structure complexity), safety labels (such as traceability reliability, quality inspection compliance) and the like. Through clustering and classification operations, structured label category aggregation data is formed to support fine modeling of subsequent label analysis. The label density of each label in the label category aggregation data is calculated. The label density reflects the coverage and refinement of the label in all products, and the calculation method can include label quantity density, label dimension distribution frequency and the like. The label density value is used to measure the representativeness and expression ability of a certain label to the overall product information structure, thereby providing a data basis for modeling. After obtaining the label density, the system performs weight normalization processing to unify the influence of different label categories in the overall label system to the same dimension range, and generates label weight distribution data. The weight data can be used as a weighting factor when combining features, giving higher weights to important labels and reducing the interference of secondary labels. Based on the label weight distribution data, the system constructs a multi-dimensional product feature vector for the product, and combines all label values according to the label category structure to form a vector representation. For high-dimensional product feature vectors, the system further performs principal component fusion processing using principal component analysis (PCA) or independent component analysis (ICA) and the like to extract key combined components with the maximum explained variance from multiple label dimensions, and generates principal component fusion feature data. Subsequently, the system performs label graph construction processing based on the principal component fusion feature data. The label graph models the association, similarity or dependency between product labels as a graph structure representation, with nodes representing various label elements and edges representing the weight relationship between labels. The edge weight can be calculated by co-occurrence frequency, correlation coefficient or mutual information, and the like, to finally form label graph structure data. Then, the system performs embedding mapping processing on the label graph structure data, and uses graph neural networks (GNN), node embedding algorithms (such as Node2Vec, DeepWalk) or graph auto-encoding methods (such as Graph AutoEncoder) to convert the graph structure into a dense low-dimensional vector representation, generating enterprise product portrait vector data. The vector can comprehensively reflect the label features, semantic relationships between labels and overall label distribution patterns of the product. Finally, in order to facilitate visual display and multi-dimensional analysis, the system performs spatial dimension reduction processing on the enterprise product portrait vector data, and uses t-SNE, UMAP or LLE and the like nonlinear dimension reduction algorithms to compress the high-dimensional vector to two-dimensional or three-dimensional space, and finally outputs enterprise product portrait data for evaluation, retrieval, clustering or trend analysis.

[0040] Preferably, the step S3 of analyzing the user interaction behavior of the purchase user information data comprises: extracting user access records and page interaction logs in the purchase user information data; reconstructing user behavior trajectories based on the user access records to generate behavior path data; calculating time sequence features of the user behavior path data to obtain time sequence offset features of the user; based on the time sequence offset features, identifying the user's browsing scanning mode within a specific time window to generate browsing scanning behavior data; based on the behavior path data, constructing a collaborative path graph between users to identify existing collaborative path behaviors and generate collaborative path behavior data; According to the collaborative path behavior data, the reverse behavior path tracking of the collaborative path graph is performed to identify the path convergence nodes and the behavior reverse path, and the collaborative path reverse tracking data is generated; According to the browsing scanning behavior data and the collaborative path reverse tracking data, the user interaction behavior mode features are identified, and the user interaction behavior mode data is generated; Based on the user interaction behavior mode data, a structured purchase user interaction behavior data is constructed.

[0041] In the embodiments of the present application, the original user access records and page interaction logs are extracted from the purchase user information data. Such data includes user access timestamp, page click path, page dwell time, page scroll depth, interaction control click event, etc. By structurally analyzing these raw data, the system can obtain the complete interaction track of the user in the product platform or service page. Then, based on the user access records, the system reconstructs the user behavior track. By time sorting and event recognition method, the continuous access behavior sequence of the same user within a certain time window is serialized and arranged to form an identifiable user behavior path chain, generating behavior path data. The behavior path data is used to represent the specific browsing order, jump logic and dependency relationship between the pages of interest of the user on the platform. Subsequently, the system calculates the time series features of the behavior path data, including user click rhythm, page dwell time distribution, behavior periodicity, etc., further generating the time sequence offset features of the user. This feature is used to reveal the tendency and behavior activity law of the user behavior in different time periods. After obtaining the time sequence offset features, the system identifies the browsing scan mode of the user within a certain time window based on the features. Through clustering analysis or pattern matching, the user's behavior logic is identified, such as linear browsing, cross-jumping, backtracking, etc., and finally the browsing scan behavior data is generated, which is used to analyze the user's attention mode to a specific product or information board. At the same time, the system uses the behavior path data to construct the collaborative path graph of users. By comparing the coincidence degree and similarity of the behavior paths between different users, the collaborative behavior modes such as information co-reading and product co-searching are identified, and the collaborative path behavior data is generated. This data can be used to reveal the behavior commonality and potential correlation between different user groups. Subsequently, the system traces the reverse behavior path of the collaborative path graph according to the collaborative path behavior data. By analyzing the convergence points (i.e. multiple user behavior convergence nodes) and their upstream behavior paths on the behavior path, the system can identify the key influencing factors of the user in the decision path, generate collaborative path reverse tracking data, and analyze group behavior attribution and guide node identification. The browsing scan behavior data and the collaborative path reverse tracking data are fused and processed to extract the comprehensive mode features of individual user behavior and group collaborative behavior, identify the behavior decision logic, interest shift point and behavior trigger key node of different users, and finally generate user interaction behavior mode data. The user interaction behavior mode data is structured and processed, and is organized and arranged according to user ID, behavior type, behavior time sequence feature, collaborative path feature, key node label, etc. dimensions, to construct structured purchase user interaction behavior data.

[0042] Especially important is that, based on the time sequence offset features, identifying the browsing scan mode of the user within a certain time window further includes: Based on the time series offset feature, the user's browsing and scanning pattern within a specific time window is identified to generate browsing and scanning behavior data; the interaction behavior time series data is sliced ​​into time windows to generate multi-segment time window behavior fragment data; Perform offset difference calculation and rhythm density analysis on multi-segment time window behavior segment data to generate time series offset feature data; perform user stay period identification and scroll jump pattern analysis on time series offset feature data to generate browsing rhythm pattern data; Perform focus area recognition and page hotspot trajectory mapping on the browsing rhythm pattern data to generate page scanning path mark data; perform clustering extraction and feature map construction on the page scanning path mark data to generate user scanning pattern feature map data; The user scanning pattern feature map data is used to extract behavioral labels and classify user behaviors, and finally generate browsing scanning behavior data.

[0043] In the embodiment of the application, by using the interaction data of the user's page browsing behavior on the digital terminal device, core parameters such as behavior timestamp, stay duration, scrolling speed and jump frequency are extracted to generate interaction behavior time sequence data. Then, the interaction behavior time sequence data is processed by time window slicing according to a set time interval to construct multi-section time window behavior segment data with a unified time length or user active period length, and each section of behavior segment represents a continuous browsing activity sequence of the user in a time window. Then, offset difference calculation is performed on each section of time window behavior segment data to identify the time difference sequence between adjacent behavior events and calculate the rhythm density index (i.e. the number of behavior events per unit time) to generate time sequence offset feature data. Based on the feature data, the stay period features (e.g. fast browsing, slow reading and short stay) and scrolling jump patterns (e.g. continuous scrolling, intermittent jumping and regional stay) of the user are identified, and the output is browsing rhythm pattern data. Further, based on the browsing rhythm pattern data, the view position coordinates in the user behavior data are analyzed in combination with the page structure and region identification information to identify the focus area of the user in the page, the user browsing path is reconstructed through hotspot trajectory mapping technology, and page scanning path marking data is formed. The page scanning path marking data includes hotspot region number, scanning order and stay intensity. Then, the page scanning path marking data is extracted by multi-user clustering, the common mode of the user groups in the scanning behavior is identified through clustering algorithms (such as K-Means or DBSCAN), and a multi-dimensional feature map is constructed to generate user scanning mode feature map data. The map reflects the structured mode of the user's browsing method, attention distribution and behavior rhythm for the page content. Finally, the user scanning mode feature map data is extracted by behavior label extraction, including labels such as "fast skimmer", "deep focus type" and "regional interest concentration type", and the user is classified into the corresponding behavior feature group in combination with the label system to finally generate browsing scanning behavior data. The data can be used in application scenarios such as user intention prediction, content optimization recommendation or intelligent adjustment of page layout.

[0044] Preferably, the collaborative path graph is tracked in reverse according to the collaborative path behavior data to identify the path convergence node and the behavior backtracking path, which includes: A directed graph model of the collaborative path graph is constructed, nodes represent user behavior events, and edges represent behavior transition relationships to generate collaborative path graph model data; Based on the collaborative path behavior data, a target behavior event node is selected as a backtracking starting point, a tracking depth and a path weight threshold are set, and reverse tracking parameters are initialized; Based on the reverse tracking parameters, a reverse path search is performed on the collaborative path graph model to identify the upstream behavior path that can be backtracked from the target behavior node to generate a candidate reverse path set; The behavior pattern similarity analysis and path convergence degree calculation are performed on the candidate reverse path set, a path node converging multiple user behaviors is identified, and path convergence node data is generated; The user behavior path is clustered and induced in combination with the path convergence node and the reverse path behavior sequence, a reverse path behavior structure mode is extracted, and reverse path behavior data is generated. The path convergence node data and the reverse path behavior data are integrated to generate collaborative path reverse tracking data.

[0045] In the embodiment of the application, a directed graph model of a collaborative path graph is constructed. User behavior events are taken as nodes in the graph, behavior conversion relationships are taken as directed edges, and the time stamp, conversion frequency and other attributes of the behavior are combined to generate complete collaborative path graph model data, thereby providing a structured basis for subsequent path tracking. Secondly, based on the collaborative path behavior data, a target behavior event node is selected as the starting point of reverse tracking. According to the business requirements, the tracking depth (i.e. the maximum number of levels of reverse backtracking) and the path weight threshold (used to filter low-weight or low-frequency paths) are set to initialize the parameters required by the reverse tracking algorithm. Then, the reverse path search algorithm is executed on the collaborative path graph model using the above parameters. Through the depth-first or breadth-first strategy, all paths from the target behavior node to the upstream behavior node are identified and collected, forming a candidate reverse path set, covering the starting point of user behavior and intermediate key nodes. Then, the behavior pattern similarity analysis is performed on the candidate reverse path set, the convergence degree of the path is calculated by combining the node frequency, path length and path weight, and the key nodes converging multiple user behavior paths are identified, i.e. the path convergence nodes, thereby forming a path convergence node data set. Subsequently, the path convergence node and the reverse path behavior sequence are combined, and a clustering algorithm (such as spectral clustering, hierarchical clustering or density-based clustering method) is applied to induce and pattern the user behavior path, thereby extracting a typical reverse path behavior structure mode, generating reverse path behavior data. Finally, the path convergence node data and the reverse path behavior data are integrated to form systematic collaborative path reverse tracking data.

[0046] Especially important is that the behavior pattern similarity analysis and path convergence degree calculation on the candidate reverse path set to identify the path nodes converging multiple user behaviors also include: The user behavior sequence of the candidate reverse path set data is standardized to generate path behavior standard sequence data, and the path behavior standard sequence data is time-aligned and behavior action coded to generate uniform time-action matrix data. performing multi-path behavior similarity clustering analysis on the unified time-action matrix data to generate behavior pattern similarity clustering result data; extracting path intersection points of the behavior pattern similarity clustering result data, and performing path convergence frequency weighting calculation and node transition stability evaluation on the path intersection points to generate path convergence degree scoring data; performing threshold filtering and high-frequency node screening on the path convergence degree scoring data to generate path common high-frequency convergence node data; performing user source distribution verification and behavior direction consistency test on the path common high-frequency convergence node data to generate path convergence node data.

[0047] In the embodiment of the application, the behavior sequences of different users in the candidate reverse path set data are standardized to construct a behavior event unified coding system, the user behavior events are uniformly represented as behavior action sequences in time sequence, and the event granularity and marking method are unified to generate path behavior standard sequence data. Then, the path behavior standard sequence data is subjected to time alignment processing, the dynamic time warping algorithm (DTW) or the sliding window matching method is used to perform sequence length uniformization processing on behavior paths of different lengths, and a unified time-action matrix data is constructed based on the behavior action coding, the rows of the matrix represent different users, the columns represent time slices, and the cells are behavior action codes corresponding to the time slices. Next, multi-path behavior similarity clustering analysis is performed on the unified time-action matrix data, and the hierarchical clustering or density clustering (such as DBSCAN) method is used to cluster according to the edit distance or Jaccard similarity between behavior action sequences to obtain behavior pattern similarity clustering result data. According to the clustering result, the path intersection points (i.e. nodes repeatedly appearing in multiple user paths) with higher frequency in each class are extracted, and the path convergence frequency (i.e. the number of times the node appears in the cluster) and the node transition stability (i.e. the reciprocal of the variance of the behavior transition probability between the nodes) are calculated to form path convergence degree scoring data. Then, threshold filtering processing is performed on the path convergence degree scoring data, the minimum convergence frequency threshold and the minimum stability score threshold are set, the path nodes with significant behavior commonality are screened out, and path common high-frequency convergence node data is generated. Further, the user source distribution verification (judging whether the convergence node involves multiple source user groups) is performed on the path common high-frequency convergence node data, and the behavior direction consistency test (i.e. whether the subsequent behavior of the node tends to be consistent in multiple user paths) is combined to exclude isolated high-frequency nodes and path noise points. Finally, the qualified nodes are marked as path convergence nodes, and the output is path convergence node data, which can be used as an important basis for user path reverse modeling and collaborative behavior decision-making.

[0048] Preferably, step S4 comprises the following steps: Step S41: based on the purchase user interaction behavior data, a Bluetooth low energy beacon sensing processing is performed to generate user scene sensing data; spatial position information of the user scene sensing data is extracted to generate user positioning coordinate data; Step S42: scene semantic matching analysis is performed on the user positioning coordinate data and enterprise product portrait data to generate user behavior portrait data; behavior interest factors of the user behavior portrait data are extracted to generate user preference factor data; Step S43: correlation measurement analysis is performed on the user preference factor data and enterprise product portrait to generate portrait correlation degree data; edge server task distribution processing is performed on the portrait correlation degree data to generate edge-level recommendation request data; Step S44: time window adaptive rearrangement is performed on the edge-level recommendation request data to generate dynamic priority ordering data; high correlation matching is performed on the dynamic priority ordering data and product feature vector data to generate personalized user recommendation list data; Step S45: client visualization processing is performed on the personalized user recommendation list data to generate user recommendation interface view data.

[0049] In the embodiment of the application, by deploying BLE (Bluetooth Low Energy) beacon devices (such as iBeacon or Eddystone) in the shopping mall or exhibition hall, the broadcast power of each beacon is set to -59dBm, and the broadcast interval is 200ms; the user terminal integrates a BLE scanning module, and the scanning radius is set to 20m and the scanning period is 500ms. The obtained original RSSI value is smoothed by a multipath attenuation model (path loss index n=2.2) and a Kalman filtering algorithm, and is converted into a distance estimation value; combined with the distance data of at least three different beacons, the two-dimensional plane coordinates (X, Y) are calculated by a trilateration algorithm, the error is controlled within ±1.5m, and the user scene perception data and the corresponding user positioning coordinate data are generated. The system pre-maintains an enterprise product portrait library in the background, including the attribute label (such as “technology”, “home”, “sports”) and the scene label (such as “entrance”, “rest area”, “exhibition area”) of each product or service. The user positioning coordinates are mapped with the scene map (based on an indoor GIS platform) to identify the current scene of the user; at the same time, combined with the interactive logs such as terminal touch click and browsing time, the behavior events are extracted and mapped to the product portrait label, the TF-IDF weighted model is used to calculate the scene-label matching degree, and the user behavior portrait data is generated. Then, the weight of each label in the behavior portrait is normalized, and the behavior interest factor (0-1 interval) is output to form the user preference factor data. The cosine similarity calculation is performed between the user preference factor and the feature vector (128-dimensional vector obtained by pre-training the Word2Vec or Doc2Vec model) in the enterprise product portrait, and a list of portrait correlation scores of each product and the user is obtained. The products ranked in the top 20% of the correlation degree and their scores are packaged as edge-level recommendation requests, wherein the request message body includes the user ID, the preference label and the corresponding score threshold (such as ≥0.75), and is pushed to the edge computing server deployed in the same local area network through the MQTT protocol or HTTP / 2, to generate edge-level recommendation request data. After receiving the recommendation request, the edge server rearranges the request list according to the request timestamp and the user active period model (a pre-trained time series LSTM model predicts the time window in which the user is most likely to accept the push), and the priority of the product with a high predicted response rate is increased by 10-20%. Then, the rearranged list and the product feature vector library are matched again for high-precision matching (Annoy or FAISS high-performance vector retrieval library is used), and the top 10 most relevant results are retained to generate the final personalized user recommendation list data. In the user terminal application, the front-end framework (such as React Native or Flutter) is used to render the recommendation list, and the card layout is used to display the product thumbnail, name, similarity score and scene matching degree progress bar.The recommended card component supports user sliding, clicking and collecting interactions, and the component style is dynamically adjusted by CSS to adapt to different screen sizes and night mode; meanwhile, the recommended interface is connected with the background in real time through WebSocket, and is automatically refreshed and displayed when new preference factors are generated or the position is changed, so that a smooth user recommended interface view data is formed.

[0050] Preferably, the step S44 of performing time window self-adaptive rearrangement on the edge-level recommendation request data generates dynamic priority sorting data, including: The time stamp, request frequency, user identifier and content type in the edge-level recommendation request data are extracted to construct a request event sequence set; based on the request event sequence set, initial sliding time window parameters are set, including time window size, sliding step and trigger threshold, to generate a basic time window structure; The request density change and request interval distribution in the time window are analyzed, and a dynamic activity index of each time window segment is calculated; The time window size and sliding step are adaptively adjusted according to the activity index, to form an adaptive time window sequence with local optimal response performance; In each adaptive time window, the importance weight of the request is calculated, wherein the weight factor of the importance weight includes: request source node level, content heat, request frequency and response delay requirement; The recommendation requests in the time window are locally prioritized based on the importance weight of the request, and cross-window optimization is performed in combination with the global sliding window position to generate dynamic priority sorting data.

[0051] In the embodiment of the application, the time stamp (accurate to milliseconds), request frequency (the number of repetitions of the same content of the same user within a specified period), user identifier and content type of each request are extracted from the edge-level recommendation request data, and all requests are constructed into a request event sequence set in ascending order of time. Based on the event sequence, the sliding time window parameters are initialized: the default time window size is set to 300 seconds, the sliding step is 60 seconds, and the trigger threshold is that when the number of requests in the window exceeds 10 or the same user requests 3 times, the adaptive mechanism is started, and at this time, the basic time window structure is formed. Then, the request density (number of requests / window length) and request interval distribution (adjacent request time difference sequence) in each basic time window segment are counted, and the dynamic activity index is calculated where p is the request density of the current window, p0 is the global average density, σ _ Δt is the request interval standard deviation, a = 0.6, b = 0.4. If the activity A of a time window exceeds a preset high threshold (such as 1.2 times the average activity of all windows), the window is reduced to 200 seconds and the sliding step is adjusted to 40 seconds to improve the response to peak periods; if A is lower than the low threshold (such as 0.8 times the average value), the window is expanded to 400 seconds and the step is increased to 80 seconds to reduce the fragmentation of empty windows, thereby automatically forming a series of adaptive time window sequences. Then, in each adaptive time window, the importance weight W = ω1·L + ω2·H + ω3·F + ω4·(1 / D) is calculated for the recommendation requests in the window, where L is the request source node level (the regional priority of the node in the edge server cluster, which can be divided into 1-5 levels), H is the content popularity (the normalized score based on the number of recommendations in the last 24 hours), F is the user request frequency normalized score, and D is the response delay requirement (unit: seconds, the smaller the value, the higher the weight); the default weight coefficients ω1…ω4 are 0.25, 0.35, 0.25, and 0.15, respectively. The request list in each window is locally prioritized according to the W value from large to small, and then the ranking between adjacent windows is cross-corrected in combination with the global sliding window position index: when the same request appears in multiple windows, the final priority takes the highest local ranking position and migrates downward by no more than 5, to ensure the smoothness across windows. Finally, the system merges the local ranking results in all adaptive windows, outputs the complete dynamic priority ranking data arranged continuously in the time axis and dynamically changed in priority, for the push scheduling and resource allocation of the edge server.

[0052] Therefore, embodiments should be considered in all respects as illustrative and not restrictive, the scope of the application being defined by the appended claims rather than the description given above, and all changes which come within the meaning and range of equivalency of the claims are therefore intended to be embraced therein.

[0053] The foregoing is considered as illustrative only of the principles of the application. Numerous modifications and adaptations will be apparent to those skilled in the art in view of the above descriptions of the embodiments. This specification and the embodiments are not to be limited or confined to the specific embodiments disclosed and changes can be made by those skilled in the art without departing from the scope of the application as set forth in the appended claims and the legal equivalents thereof.

Claims

1. A user recommendation method based on enterprise product portrait analysis, characterized in that: The following steps are involved: Step S1: Obtain enterprise product information data; Conduct supply chain traceability based on enterprise product information data to generate enterprise product supply chain information data; Screen the related enterprises of the enterprise product supply chain information data to generate product related enterprise information data; Step S2: Analyze the production proportion of product components on the enterprise product supply chain information data based on the product-related enterprise information data to generate a product component production proportion value; Quantify product component input labels based on the production ratio of product components in the enterprise product information data to generate product label data; Step S3: Constructing an enterprise product profile for the enterprise product information data based on the product tag data; using the constructed enterprise product profile to collect purchasing user information data to obtain purchasing user information data; analyzing the user interaction behavior of the purchasing user information data to generate purchasing user interaction behavior data, where the user interaction behavior includes time-series offset browsing and scanning and collaborative path reverse behavior tracking; Step S4: Locate user scenarios based on low-power Bluetooth beacons by purchasing user interaction behavior data, and dynamically generate personalized user recommendation lists based on edge servers to perform user recommendation tasks based on enterprise product portrait analysis.

2. The user recommendation method based on enterprise product portrait analysis according to claim 1 is characterized in that: Step S1 includes the following steps: Step S11: Obtain enterprise product information data; Step S12: Field standardization is performed on the acquired enterprise product information data to generate standardized product information data. Based on the standardized product information data, the enterprise resource planning system and product barcode database are called to track the raw material sources, processing nodes, and flow records of each batch of products, and the supply chain data links are integrated to generate the initial path data of the enterprise product supply chain. Step S13: Node classification is performed on the initial path data of the enterprise product supply chain to identify the key path components of raw material suppliers, processing and manufacturing enterprises, and warehousing and logistics nodes, and generate supply chain sub-node information data; link topology is organized on the supply chain sub-node information data to construct a multi-level supply chain link map, redundant or broken nodes are eliminated, and enterprise product supply chain information data is generated; Step S14: Calculate the supply strength values ​​of nodes at all levels based on the enterprise product supply chain information data, and use the supply strength values ​​to screen associated enterprises to generate a candidate data set of product-associated enterprises; perform a risk source cross-comparison on the candidate data set of product-associated enterprises to screen out upstream and downstream enterprises with high stability, and finally generate product-associated enterprise information data.

3. The user recommendation method based on enterprise product portrait analysis according to claim 2 is characterized in that: In step S14, the supply strength values ​​of nodes at all levels are calculated based on the enterprise product supply chain information data, and the supply strength values ​​are used to screen related enterprises, including: Structural layering of enterprise product supply chain information data, identifying supply nodes at all levels, including raw material supply nodes, production and processing nodes, assembly and integration nodes, and distribution nodes, to generate multi-level supply chain node structure data; Construct a supply record statistical matrix for each node in the multi-level supply chain node structure data to generate node supply statistical data, where the supply record statistical matrix includes historical supply quantity, annual fulfillment times, and cooperation cycle length; Standardize the node supply statistics through quantitative indicators, calculate the supply frequency, supply stability, and delivery ratio of each node, construct a multi-dimensional supply capacity parameter table, and generate supply capacity characteristic data; The node supply strength formula is defined based on the supply capacity characteristic data, where the node supply strength formula is as follows: ； Where, For the The supply intensity value of each node, For the The supply frequency of each node, For the The performance reliability of each node, For the The delivery ratio of nodes, , , is a custom weight coefficient; Calculate the supply strength value of each node using the node supply strength formula to generate supply strength index data; classify the supply strength index data into different levels, set the supply strength dividing line and classify them into strong, medium and weak, and screen out related nodes with medium and high supply strength to generate strong supply node screening data; The data of strong supply nodes are screened to merge enterprises’ uniqueness and identify supply chain roles, eliminate redundant or overlapping enterprises, and finally generate a candidate data set of product-related enterprises.

4. The user recommendation method based on enterprise product portrait analysis according to claim 1 is characterized in that: Step S2 includes the following steps: Step S21: mapping the supply chain path of the product-related enterprise information data, identifying the upstream supply source and actual production division ratio of each product component, and generating product component source path data; Step S22: accumulating and counting the production records of various components in the enterprise product supply chain information data, and combining the product component source path data, extracting the actual output proportion of different enterprises by component granularity, and generating component production allocation matrix data; Step S23: Normalize and integrate the component production allocation matrix data to generate standardized product component production ratio data; based on the product component production ratio data, construct a component label input function: ； Where, For product components The label weight value of For the Business-to-Component The production share of For the The company's reputation coefficient; using the component label input function to calculate the label weight of the product component production ratio data, to generate component label quantitative data; Step S24: embed labels item by item in the product component list in the enterprise product information data, associate and bind the component label quantification data with the specific components, and generate structured labeled product list data; perform unified product ID labeling, label weight aggregation and version timestamp synchronization on the structured labeled product list data, and finally generate product label data.

5. The user recommendation method based on enterprise product portrait analysis according to claim 1 is characterized in that: In step S3, the enterprise product profile is constructed based on the product tag data, including: Classify product tag data into tag types and generate tag category aggregate data; Calculate the label density of the label category aggregated data; Normalize the label density data to generate label weight distribution data; Construct product feature vectors based on label weight distribution data, and perform principal component fusion on product feature vector data to generate principal component fusion feature data; Perform label graph construction processing on the principal component fusion feature data to generate label graph structure data; Embed and map the label graph structure data to generate enterprise product portrait vector data; Perform spatial dimensionality reduction on enterprise product portrait vector data to generate enterprise product portrait data.

6. The user recommendation method based on enterprise product portrait analysis according to claim 1 is characterized in that: The analysis of user interaction behaviors of purchasing user information data in step S3 includes: Extract user access records and page interaction logs from purchasing user information data; Reconstruct user behavior trajectories based on user access records and generate behavior path data; Calculate the time series characteristics of user behavior path data and obtain the user's time series offset characteristics; Based on the time series offset feature, the user's browsing and scanning patterns within a specific time window are identified to generate browsing and scanning behavior data; Based on the behavioral path data, a collaborative path map between users is constructed, the existing collaborative path behaviors are identified, and collaborative path behavior data is generated; Based on the collaborative path behavior data, reverse behavior path tracing is performed on the collaborative path graph, path convergence nodes and behavior reverse path are identified, and collaborative path reverse tracing data is generated; Based on the fusion processing of browsing and scanning behavior data and collaborative path reverse tracking data, the user interaction behavior pattern characteristics are identified and the user interaction behavior pattern data is generated; Based on user interaction behavior pattern data, construct structured purchasing user interaction behavior data.

7. The user recommendation method based on enterprise product portrait analysis according to claim 6 is characterized in that: Based on the collaborative path behavior data, reverse behavior path tracing is performed on the collaborative path graph to identify the path convergence nodes and behavior reverse path, including: Construct a directed graph model of the collaborative path graph, where nodes represent user behavior events and edges represent behavior conversion relationships, and generate collaborative path graph model data; Based on the collaborative path behavior data, the target behavior event node is selected as the backtracking starting point, the tracing depth and path weight threshold are set, and the reverse tracing parameters are initialized; Perform reverse path search on the collaborative path graph model based on the reverse tracing parameters, identify upstream behavior paths that can be traced back from the target behavior node, and generate a set of candidate reverse paths; Perform behavioral pattern similarity analysis and path convergence calculation on the candidate reverse path set, identify the path nodes where multiple users converge in the behavioral path, and generate path convergence node data; Combining the path convergence nodes and reverse path behavior sequences, clustering and summarizing the user behavior paths, extracting the behavior reverse path structure pattern, and generating behavior reverse path data; The path convergence node data and the behavior reverse path data are integrated to generate collaborative path reverse tracking data.

8. The user recommendation method based on enterprise product portrait analysis according to claim 1 is characterized in that: Step S4 includes the following steps: Step S41: performing low-power Bluetooth beacon sensing processing based on the purchase user interaction behavior data to generate user scene perception data; extracting spatial location information of the user scene perception data to generate user positioning coordinate data; Step S42: Perform scene semantic matching analysis on the user location coordinate data and the enterprise product portrait data to generate user behavior portrait data; extract the behavior interest factor of the user behavior portrait data to generate user preference factor data; Step S43: performing correlation measurement analysis on the user preference factor data and the enterprise product profile to generate profile correlation data; performing edge server task dispatching processing on the profile correlation data to generate edge-level recommendation request data; Step S44: adaptively rearrange the edge-level recommendation request data in a time window to generate dynamic priority ranking data; perform high-correlation matching on the dynamic priority ranking data and the product feature vector data to generate personalized user recommendation list data; Step S45: Perform client-side visualization processing on the personalized user recommendation list data to generate user recommendation interface view data.

9. The user recommendation method based on enterprise product portrait analysis according to claim 8, characterized in that: In step S44, the time window adaptive rearrangement of the edge-level recommendation request data to generate dynamic priority ranking data includes: Extract the timestamp, request frequency, user ID, and content type from edge-level recommendation request data to construct a request event sequence set. Based on the request event sequence set, set the initial sliding time window parameters, including the time window size, sliding step, and trigger threshold, to generate a basic time window structure. Analyze the change in request density and the distribution of request intervals within the time window, and calculate the dynamic activity index for each time window segment; Adaptively adjust the time window size and sliding step size according to the activity index to form an adaptive time window sequence with local optimal response performance; In each adaptive time window, the importance weight of the request is calculated, where the weighting factors of the importance weight include: the level of the request source node, content popularity, request frequency and response delay requirements; Based on the importance weight of the requests, the recommended requests within the time window are locally prioritized, and cross-window tuning is performed in combination with the global sliding window position to generate dynamic priority ranking data.

10. A user recommendation system based on enterprise product portrait analysis, characterized in that: The method for recommending users based on enterprise product portrait analysis according to claim 1 is configured to include: The enterprise screening module is used to obtain enterprise product information data; conduct supply chain traceability based on enterprise product information data to generate enterprise product supply chain information data; screen related enterprises based on enterprise product supply chain information data to generate product related enterprise information data; The product labeling module is used to analyze the production ratio of product components in the enterprise product supply chain information data based on the product-related enterprise information data, and generate the product component production ratio value; based on the product component production ratio value, the enterprise product information data is quantified by product component input labels to generate product label data; The user interaction analysis module is used to construct an enterprise product profile based on product tag data; use the constructed enterprise product profile to collect purchasing user information data to obtain purchasing user information data; analyze the user interaction behavior of the purchasing user information data to generate purchasing user interaction behavior data, where the user interaction behavior includes time-series offset browsing and scanning and collaborative path reverse behavior tracking; The personalized recommendation module is used to locate user scenarios based on low-power Bluetooth beacons by purchasing user interaction behavior data, and dynamically generate personalized user recommendation lists based on edge servers to perform user recommendation tasks based on enterprise product portrait analysis.

Images