Enterprise ai case intelligent recommendation method and system

Abstract

The application relates to the technical field of enterprise big data processing, and provides an enterprise AI case intelligent recommendation method and system, which realizes semantic level matching through mapping enterprise products and case descriptions to a unified semantic vector space and using cosine similarity calculation, and eliminates the lexical gap from the root by combining the differentiated weight assignment of main and non-main products, thereby improving the recommendation matching precision. The enterprise is mapped to an industry chain knowledge graph, the accuracy and coverage of the matching are accurately balanced, and the discrimination granularity of the business overlap degree is enhanced. A case migratability evaluation channel is additionally added, and the gap between the recommendation and landing is bridged. The multiplication constraint synergy mechanism of the product matching, industry chain matching and migratability evaluation three channels makes each case finally recommended business-matched and technically feasible, thereby significantly improving the landing feasibility of the recommended cases.

Description

Technical Field

[0001] This invention belongs to the field of enterprise big data processing technology, and relates to an intelligent recommendation method and system for enterprise AI cases. Background Technology

[0002] As artificial intelligence (AI) technology penetrates deeper into various industries, many companies face the challenge of choosing and learning from AI application cases that suit their own business characteristics. Companies need to refer to the successful experiences of other companies in the same or similar industries in implementing AI applications to reduce decision-making risks and accelerate their intelligent transformation process.

[0003] Existing enterprise case recommendation technologies, such as those based on product string matching, match enterprise product names with case descriptions using keywords. A successful match is assigned a fixed score, and the recommendations are then sorted by score. However, these methods fail to recognize semantically similar but differently described products. For example, if an enterprise's product is injection molding processing while the case description is "plastic parts injection molding processing," the string matching result will be a mismatch, causing many semantically relevant cases to be missed. Furthermore, these methods use fixed weights and cannot dynamically adjust the matching contribution of different products based on the enterprise's core business focus.

[0004] Alternatively, recommendation methods based on a single industry directory hierarchy might categorize companies into a broad industry class and use only the consistency of industry directory node names as the matching criterion. This method is too crude, failing to distinguish between upstream and downstream business differences within the industry chain and ignoring the distance relationships and graph structure information between industry chain nodes. It suffers from a significant boundary cliff problem: two companies in adjacent industry chain nodes with different node names might score zero, while two companies in the same broad node with significantly different businesses might receive high scores. Another example is recommendation methods based on generalized hierarchical weights. While this method discloses the idea of ​​hierarchical weights, it doesn't address how weights dynamically relate to the actual importance of a company's business and the relevance of case studies. Essentially, it remains a static and fixed scoring system. None of these existing solutions assess whether a company possesses the basic conditions for implementing a case study, resulting in highly matched cases potentially being completely worthless, and a significant gap between recommendation results and actual implementation.

[0005] Therefore, how to eliminate the vocabulary gap caused by heterogeneous product names, the boundary cliff caused by fixed-level matching, and the migration gap between recommendation results and implementation feasibility has become a core technical problem that needs to be solved in enterprise AI case recommendation technology. Summary of the Invention

[0006] To address the problems existing in the aforementioned traditional methods, this invention proposes an intelligent recommendation method and a system for enterprise AI case studies. This system can eliminate the vocabulary gap caused by heterogeneous product names, the boundary cliff caused by fixed-level matching, and the migration gap between recommendation results and implementation feasibility. It ensures that the recommendation results are relevant to the business while also guaranteeing their technical feasibility.

[0007] To achieve the above objectives, the embodiments of the present invention adopt the following technical solutions: On the one hand, it provides a method for intelligent recommendation of enterprise AI cases, including: Collect enterprise AI application cases and perform structured analysis. Calculate the semantic vectors of the cases and store them as an AI case library. The target company's products are semantically standardized and sorted by business importance. The target company is mapped to the industry chain knowledge graph to obtain the industry chain coordinates and graph embedding vectors of the corresponding nodes. The company's publicly disclosed multi-dimensional weak signals are collected and the data readiness level is divided. The product matching score is calculated based on the cosine similarity between the semantic vector of the target company's product and the semantic vector of the case in the AI ​​case library, combined with the importance ranking value of the target company's product and the weight of the product attribute. The industry chain score is calculated based on the matching relationship between case companies and target companies in the AI ​​case library at the nodes of the industry chain, and the shortest path distance in the industry chain knowledge graph; the shortest path distance is calculated based on the graph embedding vector. Based on the implementation prerequisites of the case and the data readiness level of the target enterprise, logical rule reasoning and fuzzy matching are used to determine the transferability score. The product matching score and the industry chain score are combined and then multiplied by the transferability score to obtain the comprehensive score. Candidate cases are sorted in descending order of their overall scores, and the top cases with the highest overall scores are output. K One recommended case; K It is a positive integer.

[0008] In one embodiment, the steps of collecting enterprise AI application cases, performing structured parsing, calculating case semantic vectors, and storing them as an AI case library include: The crawling scope is limited by a pre-defined whitelist of authoritative websites. Semantic filtering is performed on candidate pages, and pages that simultaneously match AI technology keywords and business scenario keywords are identified as candidate case pages and downloaded. Extract the main text from the downloaded page, perform named entity recognition and structured information extraction on the extracted main text, and fill the extracted case entity information into the preset template. Using fragmented fields from multiple sources as fact anchors, a large language model is used to standardize and complete the aggregated fragmented information and generate case summaries. The generated content and the original fragment input are subjected to illusion detection based on a text entailment recognition verification model. The completed content that passes the verification is stored in the AI ​​case library; the completed content that fails the verification is discarded. The semantic vectors of cases stored in the AI ​​case library are calculated using a text embedding model.

[0009] In one embodiment, the process of calculating the cosine similarity between the product semantic vector of the target enterprise's product and the case semantic vectors of cases in the AI ​​case library includes: After calculating the product semantic vector of each product of the target company using a text embedding model, the cosine similarity is calculated with the case semantic vectors of each case in the AI ​​case library to obtain the product semantic similarity score; where the product semantic vector is the vector representation of the target company's product mapped to the semantic space by the text embedding model, and the case semantic vector is the vector representation of the case mapped to the semantic space by the same text embedding model.

[0010] In one embodiment, the process of mapping the target enterprise to a supply chain knowledge graph to obtain supply chain coordinates and graph embedding vectors of corresponding nodes includes: Map the target company to a pre-defined industry chain knowledge graph and record the target company's complete industry chain path; Extract the penultimate and tertiary level directory nodes and obtain the graph embedding vectors of the corresponding nodes; where the penultimate level directory node is the macro-level business sub-domain node and the tertiary level directory node is the subdivided link node.

[0011] In one embodiment, the process of calculating a product matching score by combining the importance ranking value of the target company's products and the weight of product attributes includes: Each product of the target company is assigned an importance ranking value based on its revenue share and business coreness, and corresponding product attribute weights are assigned according to whether the product is a main product or a non-main product; the product attribute weight of the main product is 1000, and the product attribute weight of the non-main product is 999. The product matching score is calculated based on the product semantic similarity score, product attribute weight, and importance ranking value.

[0012] In one embodiment, the process of calculating the industry chain score based on the matching relationship between case companies and target companies in the AI ​​case library at the nodes of the industry chain, and the shortest path distance in the industry chain knowledge graph, includes: If the node in the industry chain where the case company is located is exactly the same as the penultimate level directory node of the target company, then the basic weight of the case corresponding to the case company will be set to 1.0. If the case company's industry chain node is exactly the same as the target company's third-to-last directory node, but different from the second-to-last directory node, then the case's basic weight will be set to 0.0. In the industry chain knowledge graph, the shortest path distance between the target enterprise node and the case enterprise node is calculated based on the graph embedding vector. The basic weight of the case is continuously adjusted based on the shortest path distance; when the shortest path distance is 0, the basic weight of the case is multiplied by a coefficient of 1.5; when the shortest path distance is 1, the basic weight of the case is multiplied by a coefficient of 1.2; when the shortest path distance is 2, the basic weight of the case is multiplied by a coefficient of 0.8; when the shortest path distance is greater than 2, the corresponding path signal is ignored. Based on the case-based weights and distance-based functions f ( d Calculate the industry chain score; where the function is... f ( d ) is a decay function or gain function based on the shortest path distance.

[0013] In one embodiment, the step of determining a transferability score by performing logical rule reasoning and fuzzy matching based on the implementation preconditions of the case and the data readiness level of the target enterprise, and then multiplying the product matching score and the industry chain score by the transferability score to obtain a comprehensive score, includes: Extract the required conditions from the implementation prerequisites of the case's structured fields, and perform logical rule reasoning and fuzzy matching with the target enterprise's data readiness level to obtain a transferability score; The business relevance score, obtained by combining the product matching score and the industry chain score, is then multiplied by the transferability score to obtain the comprehensive score.

[0014] In one embodiment, when sorting candidate cases in descending order of comprehensive score, the method further includes calculating a time-sensitivity decay factor for each case and incorporating the time-sensitivity decay factor into the comprehensive score calculation; wherein, the time-sensitivity decay factor is γ(t) = e -λt , t λ represents the number of months since the case was published, and λ is the decay coefficient.

[0015] In one embodiment, before sorting candidate cases in descending order of comprehensive score, the method further includes: Perform semantic deduplication on candidate cases.

[0016] On the other hand, it also provides an intelligent recommendation system for enterprise AI cases, including: The case library construction module is used to collect enterprise AI application cases, perform structured analysis, calculate case semantic vectors, and store them as an AI case library; The enterprise profiling module is used to semantically standardize the products of the target enterprise and sort them according to business importance. It maps the target enterprise to the industry chain knowledge graph to obtain the industry chain coordinates and graph embedding vectors of the corresponding nodes. It also collects multi-dimensional weak signals publicly disclosed by the enterprise and classifies the data readiness level. The product matching module is used to calculate the product matching score based on the cosine similarity between the product semantic vector of the target company's product and the case semantic vector of the cases in the AI ​​case library, combined with the importance ranking value of the target company's products and the product attribute weights. The supply chain matching module is used to calculate the supply chain score based on the matching relationship between case companies and target companies in the AI ​​case library at the nodes of the supply chain, as well as the shortest path distance in the supply chain knowledge graph; the shortest path distance is calculated based on the graph embedding vector. The transferability assessment module is used to determine the transferability score by logical rule reasoning and fuzzy matching based on the implementation prerequisites of the case and the data readiness level of the target enterprise. The product matching score and the industry chain score are combined and then multiplied with the transferability score to obtain the comprehensive score. The case output module sorts candidate cases in descending order of their overall score and outputs the top cases with the highest overall score. K One recommended case; K It is a positive integer.

[0017] One of the above technical solutions has the following advantages and beneficial effects: The aforementioned AI-powered intelligent recommendation method and system for enterprise cases effectively solves the challenges of identifying heterogeneous product names, synonyms, and related upstream and downstream products by mapping enterprise products and case descriptions to a unified semantic vector space and using cosine similarity calculation to achieve semantic-level matching. Simultaneously, by combining differentiated weighting of core and non-core products, it accurately focuses on the enterprise's core business needs, avoiding interference from irrelevant products. The synergistic effect of semantic vector matching and product importance ranking ensures that the semantic relevance of core products achieves a significantly higher score than that of peripheral products with perfect string matching, fundamentally eliminating the lexical gap and improving recommendation matching accuracy.

[0018] By mapping enterprises to the industry chain knowledge graph, not only is precise matching based on product-level nodes achieved, but the combination of dual-scale industry chain coordinates (i.e., product-level combinations) and the graph distance continuous perception function precisely balances the accuracy and coverage of matching, enhancing the granularity of business overlap judgment. Furthermore, a case transferability assessment pathway is added, using logical rule reasoning and fuzzy matching to examine the implementation prerequisites of the case and the target enterprise's digital transformation maturity profile. This effectively eliminates invalid cases that, while business-matched, lack the necessary digital infrastructure to support them, bridging the gap between recommendation and implementation. The multiplicative constraint synergy mechanism of the three pathways—product matching, industry chain matching, and transferability assessment—ensures that each recommended case is both business-relevant and technically feasible, significantly improving the implementability of the recommended cases. Attached Figure Description

[0019] To more clearly illustrate the technical solutions in the embodiments of the present invention or the conventional technology, the drawings used in the description of the embodiments or the conventional technology will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0020] Figure 1 This is a flowchart illustrating an intelligent recommendation method for enterprise AI cases in one embodiment; Figure 2 This is a schematic diagram of the module architecture of an enterprise AI case intelligent recommendation system in one embodiment. Detailed Implementation

[0021] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the invention. Unless otherwise defined, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this invention pertains. The terminology used in this specification is for the purpose of describing particular embodiments only and is not intended to limit the invention.

[0022] It should be noted that, in this document, the reference to "embodiment" means that a particular feature, structure, or characteristic described in connection with an embodiment may be included in at least one embodiment of the invention. The presentation of this phrase in various places throughout the specification does not necessarily refer to the same embodiment, nor is it a separate or alternative embodiment mutually exclusive with other embodiments. Those skilled in the art will understand that the embodiments described herein can be combined with other embodiments. The term "and / or" as used herein refers to any combination of one or more of the associated listed items, and all possible combinations, including such combinations.

[0023] The embodiments of the present invention will now be described in detail with reference to the accompanying drawings.

[0024] In one embodiment, such as Figure 1 As shown, an intelligent recommendation method for enterprise AI cases is provided, which may include the following processing steps S10 to S20: S10 collects enterprise AI application cases and performs structured analysis, calculates the semantic vectors of the cases, and stores them as an AI case library.

[0025] Specifically, the system can collect unstructured text (such as reports, white papers, and solution descriptions) from enterprises that have actually integrated AI into their R&D, production, supply chain, and marketing processes through a combination of targeted internet crawling and AI-assisted generation. The collected data covers all sub-fields of AI applications, such as computer vision, knowledge graphs, robotic process automation, and AI-enabled industry sub-scenarios. The system performs structured parsing of the crawled content, extracting case entity information, which may include the applying enterprise, its industry, its supply chain node, the specific AI technology used, the business problem solved, quantifiable results (such as percentage efficiency improvement), and implementation prerequisites (such as required data infrastructure and IT infrastructure). Each case is stored as a semantic vector calculated by a text embedding model (such as BERT-Industry). V ( c ), where the case semantic vector V ( c () indicates a case c The vector representation mapped to the semantic space by this model.

[0026] This step achieves a balance between high quality and large scale for the first time in the vertical field of AI application cases. It solves the long-standing paradox of scale versus quality in case library construction and provides a data foundation with both quantity and quality for upper-level recommendation algorithms.

[0027] S12 performs semantic standardization on the target company's products and sorts them according to business importance. The target company is mapped to the industry chain knowledge graph to obtain the industry chain coordinates and graph embedding vectors of the corresponding nodes. The multi-dimensional weak signals disclosed by the company are collected and the data readiness level is divided.

[0028] Specifically, in terms of product profiling, all products and services of the target company (also known as the search company) in its business registration information and operational data are obtained. After semantic standardization, these are divided into a main product set and a non-main product set. Each product can be assigned an importance ranking value based on its revenue share and business coreness. Ri Regarding the industry chain coordinates, the enterprise can be mapped to a pre-defined industry chain knowledge graph, recording its complete industry chain path. Product-level nodes are extracted, and the graph embedding vectors of the corresponding nodes in the industry chain graph are obtained. This is used to transform the implicit structural information between industry chain nodes into computable numerical features. In terms of digital transformation maturity profiling, weak signals such as the enterprise's publicly available IT system construction, data platform adoption, and AI talent recruitment are captured. These signals are then classified into data readiness levels (e.g., high, medium, and low) using a classification model. This level is used for subsequent migration assessment weights.

[0029] This step utilizes a dual modeling approach: semantic standardization of product profiles combined with quantifiable importance. This upgrades the recommendation logic from simply listing a company's products to identifying which products best represent the company. By combining product-level anchors based on the company's position within the industry chain, it precisely balances matching accuracy and coverage. Furthermore, by aggregating and inferring weak signals indicating digital transformation maturity, it addresses the challenge of obtaining fundamental digital information. The collaborative construction of this 3D profile provides a reliable foundation for subsequent accurate matching and transferability assessment, enabling the recommendation system to possess enterprise-adaptive capabilities.

[0030] S14. Calculate the product matching score based on the cosine similarity between the target company's product semantic vector and the case semantic vectors of cases in the AI ​​case library, combined with the target company's product importance ranking value and product attribute weight.

[0031] S16. Calculate the industry chain score based on the matching relationship between the case companies and the target company in the AI ​​case library at the nodes of the industry chain, and the shortest path distance in the industry chain knowledge graph; the shortest path distance is calculated based on the graph embedding vector.

[0032] S18. Based on the implementation prerequisites of the case and the data readiness level of the target enterprise, logical rule reasoning and fuzzy matching are used to determine the transferability score. The product matching score and the industry chain score are combined and then multiplied with the transferability score to obtain the comprehensive score.

[0033] Specifically, the above steps construct a collaborative scoring system for the three pathways.

[0034] Product semantic matching pathway, for each product of the target company p i The product semantic vector is calculated using the same text embedding model. V(p i ) The semantic vectors of each case in the AI ​​case library V ( c Cosine similarity calculation is performed to obtain the semantic similarity score of the products. Sim prod Combined with the product attribute weights of this product.W type For example, but not limited to, if the product is the main product, then W type The value is 1000; if the product is not a core product, then... W type The score is 999. The product matching score for each product is determined by the product's... Sim prod , W type and R i The calculation yields the highest matching signal from the most relevant products, retaining the maximum value to avoid dilution from multiple products. This approach upgrades product matching from string equality to semantic space-weighted search, resolving the lexical gap problem in string matching.

[0035] The supply chain path mapping matching pathway is divided into precise path matching and extended path matching. In precise path matching, if the node in the supply chain where the company in the case is located is related to the target company... C If the values ​​are completely identical (-2), it is considered a high overlap in business operations, and the case's base weight will be adjusted accordingly. BaseWeight For example, set it to 1.0. In extended path matching, if the node in the industry chain where the company in the case is located is the same as the target company... C -3 is completely identical but with C If -2 is different, then this case will be... BaseWeight For example, set it to 0.0. To provide more refined differentiation, the target enterprise node is further calculated in the industry chain knowledge graph. Ntarget Case study enterprise nodes Ncase Shortest path distance between d For example, if d A value of 0 indicates that the node is the same. BaseWeight Multiply by a coefficient of 1.5; if d Equals 1, BaseWeight Multiply by a coefficient of 1.2; if d Equals 2, BaseWeight Multiply by a coefficient of 0.8; if d If the value is greater than 2, the signal for that path is ignored. The supply chain score in this case is based on... BaseWeight and distance-based attenuation or gain function f(d) This approach upgrades supply chain matching from fixed-level comparison to continuous perception of distance within a graph, by introducing continuous distance coefficients. f(d) This transforms the matching process from a binary judgment to a continuous metric, effectively smoothing out boundary cliffs.

[0036] The case transferability assessment pathway extracts the necessary conditions from the implementation prerequisites of the case's structured fields, such as the need for an ERP system, structured data accumulation, and GPU computing power clusters. These conditions are then logically inferred and fuzzy matched against the target company's digital transformation maturity profile to obtain a transferability score. Score transfer ( c For example, if a case study requires the deployment of an MES system and the target enterprise's digital transformation maturity profile indicates a level of medium or higher in production informatization, the transferability score is 0.9; if the case study does not meet the required conditions at all, the transferability score is 0.1. Score transfer ( c The score is multiplied by the business relevance score, which combines product matching score and industry chain score, as a multiplicative factor. This results in a reduction in weight for companies with high matching scores but lacking basic implementation cases, making the recommendations more practical.

[0037] The three pathways form a close functional synergy: product matching and industry chain matching constitute a complementary dual pathway of business relevance, capturing the degree of correlation between enterprises and cases from semantic and structural dimensions, respectively; transferability assessment serves as a global constraint on implementation feasibility, intervening in the comprehensive score in a multiplicative manner. This architecture upgrades the recommendation system from interest matching to feasibility matching, ensuring that the final recommended cases are both business-relevant and implementable.

[0038] S20: Sort candidate cases in descending order of comprehensive score, and output the top cases with the highest comprehensive score. K One recommended case; K It is a positive integer.

[0039] Specifically, candidate cases will be ranked according to a comprehensive score (i.e., a comprehensive recommendation score). Score final Sort in descending order and take the highest score first. K The recommended cases are presented to enterprise users in the form of a list and radar chart comparison. The radar chart comparison covers three dimensions: matching degree, transferability, and timeliness, which intuitively presents the characteristics of the cases and helps enterprises quickly evaluate the AI ​​application path, effectively improving the success rate of intelligent transformation.

[0040] This embodiment introduces a feasibility assessment dimension independent of matching scores into the recommendation system through the above steps, ensuring that the recommendation results are not only business-relevant but also feasible. The collaborative relationship among the four pathways has a clear hierarchical structure: product matching and industry chain matching constitute a complementary dual pathway of business relevance; transferability assessment serves as a global constraint on implementation feasibility; and timeliness and diversity optimization serve as post-processing guarantees for result quality, forming a complete and hierarchical recommendation scheme.

[0041] The aforementioned AI-powered intelligent recommendation method for enterprise cases maps enterprise products and case descriptions to a unified semantic vector space, using cosine similarity calculation to achieve semantic-level matching. This effectively solves the challenges of identifying heterogeneous product names, synonyms, and related upstream and downstream products. Simultaneously, by combining differentiated weighting of core and non-core products, it accurately focuses on the enterprise's core business needs, avoiding interference from irrelevant products. The synergistic effect of semantic vector matching and product importance ranking ensures that the semantic relevance of core products achieves a significantly higher score than that of peripheral products with perfect string matching, fundamentally eliminating the lexical gap and improving recommendation matching accuracy.

[0042] By mapping enterprises to the industry chain knowledge graph, not only is precise matching based on product-level nodes achieved, but the combination of dual-scale industry chain coordinates (i.e., product-level combinations) and the graph distance continuous perception function precisely balances the accuracy and coverage of matching, enhancing the granularity of business overlap judgment. Furthermore, a case transferability assessment pathway is added, using logical rule reasoning and fuzzy matching to examine the implementation prerequisites of the case and the target enterprise's digital transformation maturity profile. This effectively eliminates invalid cases that, while business-matched, cannot be supported by the enterprise's digital foundation, bridging the gap between recommendation and implementation. The multiplicative constraint synergy mechanism of the three pathways—product matching, industry chain matching, and transferability assessment—ensures that each ultimately recommended case is both business-relevant and technically feasible, significantly improving the implementability of the recommended cases.

[0043] In one embodiment, step S10 described above may specifically include the following processing: The crawling scope is limited by a pre-defined whitelist of authoritative websites. Semantic filtering is performed on candidate pages, and pages that simultaneously match AI technology keywords and business scenario keywords are identified as candidate case pages and downloaded. Extract the main text from the downloaded page, perform named entity recognition and structured information extraction on the extracted main text, and fill the extracted case entity information into the preset template. Using fragmented fields from multiple sources as fact anchors, a large language model is used to standardize and complete the aggregated fragmented information and generate case summaries. The generated content and the original fragment input are subjected to illusion detection based on a text entailment recognition verification model. The completed content that passes the verification is stored in the AI ​​case library; the completed content that fails the verification is discarded. The semantic vectors of cases stored in the AI ​​case library are calculated using a text embedding model.

[0044] It is understandable that the intelligent construction of the AI ​​case library adopts a closed-loop pipeline of focused collection and intelligent completion, rather than a simple data crawling and stacking. The overall design idea of ​​this processing method is to redefine the construction of the AI ​​case library as a targeted collection strategy driven by the value density of cases and a human-machine collaborative construction mode with crawling as the skeleton and generation as the flesh and blood.

[0045] Specifically, the crawling scope is limited by a pre-defined whitelist of authoritative websites. This whitelist includes AI industry media, enterprise cloud service AI case studies, and demonstration project publicity platforms, which narrows the crawler's initial search space from trillions of URLs to thousands of high-quality seeds, thus achieving source focus.

[0046] Semantic filtering is performed on candidate pages, identifying and downloading only those pages that simultaneously contain both AI technology keywords and business scenario keywords. AI technology keywords can include computer vision, natural language processing, knowledge graphs, robotic process automation, intelligent recommendation, predictive maintenance, digital twins, and large language models; business scenario keywords can include cost reduction and efficiency improvement, yield improvement, intelligent scheduling, automated quality inspection, intelligent customer service, supply chain optimization, precision marketing, and assisted diagnostics. This dual-dimensional overlapping filtering effectively eliminates text that only mentions AI concepts without actual business applications, achieving semantic focus.

[0047] The HTML structure of the download page is extracted using an algorithm based on the DOM (Document Object Model) tree density ratio. This algorithm automatically identifies and removes non-content areas such as navigation bars, advertisements, and sidebar recommendations, retaining only core content areas such as titles, body paragraphs, and lists. The output is the original article content in plain text format, achieving structural focus.

[0048] Named entity recognition and structured information extraction are performed on the extracted plain text. The pre-trained named entity recognition model is used to extract the company name, industry category, application scenario, name of AI technology used, quantitative effect indicators (such as numerical expression of efficiency improvement of 30%), and project implementation time, etc., and fill them into the preset case data template to form a structured case record. The case data template can be uniformly set according to the structured needs of the case record.

[0049] Using fragmented fields from multiple sources as fact anchors, a large language model is employed to standardize and complete the aggregated fragmented information and generate case summaries. Specifically, for example, the company name is used as the aggregation key to gather all relevant fragmented information about the company from structured records across different sources into the same aggregation group. The aggregated fragmented fields are then organized into prompt words and input into the large language model, instructing the model to perform information integration, reasonable inference, completion, and summary generation tasks. For instance, if a fragmented field lacks a description of a business pain point, the model is instructed to infer the business problem that the case might solve based on existing application scenarios and effect data. The generation of the above-mentioned completion content is strictly constrained: the model can only perform logical reasoning based on the input fragmented information and must not add factual content beyond the scope of the fragmented information.

[0050] The case summary generated by the large language model and the original fragmented information are input into a text-entailment-based verification model for illusion detection to determine whether the generated content contradicts or exceeds the scope supported by the original information. If the verification model determines that there is a conflict, the generated paragraph is discarded, and only the structured fields from the original data are retained; if the verification passes, the generated summary is officially added to the database as a case description, and the source identifier for AI-assisted completion is marked. For the cases added to the database, the semantic vector of the case is automatically calculated using a text embedding model. V ( c And store it in the AI ​​case library.

[0051] It's important to note the paradox of scale versus quality in traditional case study database construction. Traditional manual screening and compilation involves domain experts reading industry reports and news articles to manually extract case study information. While quality is controllable, this is extremely inefficient and cannot support the need for large-scale, multi-industry, and continuous updates. On the other hand, using general web crawlers to indiscriminately crawl pages containing AI-related terms from the internet is efficient but introduces a lot of noise. Many pages merely mention AI concepts rather than providing real-world application examples, and the completeness of case study information varies greatly across different sources, with many cases lacking crucial fields.

[0052] This embodiment achieves a balance between case authenticity and information completeness through a three-layer focusing strategy and a closed-loop design combining AI completion and illusion verification. Source focusing narrows the crawling scope from trillions of URLs to high-quality seed sites; semantic focusing filters out invalid pages through dual-dimensional overlapping filtering; structural focusing extracts and refines effective information from the text; AI completion performs standardized completion under the constraint of factual anchor points; and the verification model provides a fallback for filtering out incorrect completions.

[0053] The data collection process focuses on providing reliable but incomplete skeleton information, while AI generation performs standardized completion under skeleton constraints. The verification model filters out potential errors in AI-generated completions. This collaborative design of the three stages enables the system to achieve a commercially viable level of automation and a level of quality reliability that can replace manual review, thus overcoming the paradox of scale versus quality in case library construction.

[0054] Furthermore, the three-stage pipeline proposed in this embodiment, which focuses on data collection, intelligent completion, and illusion verification, is not only applicable to AI application cases, but can theoretically be transferred to any scenario that requires building a professional knowledge base from unstructured web pages, such as digital transformation cases, green manufacturing cases, and intelligent manufacturing solutions. It has strong technical universality and reusability.

[0055] In one embodiment, the process of calculating the cosine similarity between the product semantic vector of the target enterprise's product and the case semantic vectors of cases in the AI ​​case library in step S14 above may specifically include the following processing: After calculating the product semantic vector of each product of the target company using a text embedding model, the cosine similarity is calculated with the case semantic vectors of each case in the AI ​​case library to obtain the product semantic similarity score; where the product semantic vector is the vector representation of the target company's product mapped to the semantic space by the text embedding model, and the case semantic vector is the vector representation of the case mapped to the semantic space by the same text embedding model.

[0056] It is understandable that the core of the product semantic matching pathway lies in using a text embedding model to map enterprise products and case descriptions to the same semantic vector space, achieving semantic-level matching. Specifically, this can include: using the same text embedding model as used in building the AI ​​case library to calculate the product semantic vector for each product of the target enterprise. V(p i ) And compared with the case semantic vectors of each case in the AI ​​case library. V ( c Cosine similarity calculation is performed to obtain the semantic similarity score of the products. Sim prod .

[0057] in, V(p i ) For target company products p i The vector representation mapped to the semantic space by this text embedding model. V ( c (This is a case study) c Vector representations mapped to the semantic space via the same text embedding model. Sim prodThe product and the case study are scored based on their similarity in the semantic dimension of the product. Sim prod The higher the value, the closer the semantic connection between the target company's product and the product involved in the case.

[0058] To address the lexical gap caused by product string matching in traditional solutions, existing technologies rely solely on string equality when matching company product names with case product descriptions. When the same product appears in different forms from different sources—for example, injection molding processing versus plastic parts injection molding processing, or photovoltaic modules versus solar panel manufacturing—string matching is deemed a mismatch, leading to the incorrect exclusion of many semantically highly related cases. This is because general-purpose language models have limited semantic understanding of specialized terms. This embodiment, however, employs a language model fine-tuned with industry-specific corpora to map product names to a unified semantic space.

[0059] This embodiment uses cosine similarity calculation in semantic vector space to automatically capture synonymous expressions and implicit relationships between upstream and downstream products, greatly expanding the scope of effective matching. Sim prod As a quantitative indicator of the semantic dimension of a product, it provides a precise numerical basis for subsequent collaborative calculations of importance ranking values ​​and product attribute weights.

[0060] In one embodiment, the process of mapping the target enterprise to the industry chain knowledge graph in step S12 above to obtain the industry chain coordinates and the graph embedding vectors of the corresponding nodes may specifically include the following processing: Map the target company to a pre-defined industry chain knowledge graph and record the target company's complete industry chain path; Extract the penultimate and tertiary level directory nodes and obtain the graph embedding vectors of the corresponding nodes; where the penultimate level directory node is the macro-level business sub-domain node and the tertiary level directory node is the subdivided link node.

[0061] It's understandable that constructing a supply chain coordinate system is a crucial step in upgrading a company's positioning from a rough, broad industry classification to a precise, map-based positioning. Specifically, this involves mapping the target company to a pre-defined supply chain knowledge graph, recording the complete supply chain path the target company occupies, and extracting the penultimate level directory nodes. C -2 and the third-to-last directory node C -3, and obtain the graph embedding vector of the corresponding node in the industry chain graph. E ( C -2) and E ( C -3).

[0062] in, C-2 represents a more macro-level business sub-domain node, used to identify case companies with similar business categories; C -3 represents a more refined segment node, used to identify weakly related companies that may differ in segment. E ( C -2) and E ( C -3) These are the vectorized representations of the corresponding nodes. Each industrial chain node is encoded into a vector using graph embedding technology, which transforms the implicit structural information such as the industrial distance and upstream and downstream relationships between nodes into computable numerical features, providing deeper information beyond whether the names are consistent for subsequent matching.

[0063] To address the issue of coarse positioning caused by single-level directory matching in traditional solutions, which rely solely on the consistency of industry directory node names for matching, the finest-grained nodes, such as the fourth or fifth-to-last level directory node, may result in high matching accuracy but a narrow case coverage, potentially matching only a very small number of cases within the same sub-sector, leading to sparse recommendation results. Conversely, using coarse-grained nodes, such as the first-to-last level category, may result in broad coverage but insufficient accuracy, potentially leading to incorrect matches that are both in the manufacturing sector but completely unrelated to the business.

[0064] This embodiment is illustrated by... C -2 and C The -3 dual-scale combination precisely balances matching accuracy and coverage: C -2 Ensure that cases with high business overlap are not overlooked. C -3. Supplementary cases identifying businesses that are related but not direct competitors. Simultaneously, assign graph embedding vectors to nodes. E ( C -2) and E ( C -3), which allows for the subsequent calculation of supply chain distances (such as shortest path distances). d This allows for a more refined differentiation of the closeness or distance between nodes in the industrial chain, a capability that traditional directory matching lacks.

[0065] In one embodiment, the process of calculating the product matching score by combining the importance ranking value of the target company's products and the weight of product attributes in step S14 above may specifically include the following processing: Each product of the target company is assigned an importance ranking value based on its revenue share and business coreness, and corresponding product attribute weights are assigned according to whether the product is a main product or a non-main product; preferably, the product attribute weight of the main product is 1000, and the product attribute weight of the non-main product is 999. The product matching score is calculated based on the product semantic similarity score, product attribute weight, and importance ranking value.

[0066] It is understandable that in calculating product semantic similarity... Sim prod Building upon this foundation, this embodiment further incorporates product business importance information. Specifically, each product of the target company is assigned an importance ranking value based on its revenue share and business coreness. R i And assign product attribute weights based on whether the product is the main product or a non-main product. W type Among them, the main products W type 1000, non-core products W type For 999:

[0067] For each product of the target company, there is a specific case in the AI ​​case library. c Product matching score Score prod ( c ), by the product Sim prod , W type and R i The calculation yielded:

[0068] in, R i For the first i The importance ranking value of a product indicates that the higher the value, the stronger the representativeness of the product to the target company's business. W type The setting achieves a significant weight difference between core and non-core products, resulting in the semantic relevance of core products receiving a much higher score than the string matching of peripheral products.

[0069] The traditional approach treats all products equally, resulting in recommendations that fail to reflect the differences in a company's business focus. This approach typically extracts products by segmenting them into strings based on their business scope and assigns the same weight to all products for matching. Even if a product accounts for only a very small share of a company's revenue, its matching score is no different from that of a core product, leading to the recommendation results being interfered with by peripheral products.

[0070] This embodiment uses importance ranking values. R iThe introduction of AI technology upgrades the recommendation logic from "What products does a company have?" to "Which products best represent the company?" This allows subsequent matching scores to reflect differences in the company's business focus. For example, a company that mainly produces photovoltaic modules and also produces mounting systems will receive significantly higher weighting for photovoltaic module-related cases when matching AI cases. R i and W type The product combination of these factors results in the semantic relevance of the main products receiving a much higher matching weight contribution than peripheral products, truly reflecting the dominant position of the enterprise's core business in case retrieval.

[0071] In one embodiment, step S16 above, which involves calculating the industry chain score based on the matching relationship between the case companies and the target company in the AI ​​case library at the nodes of the industry chain, and the shortest path distance in the industry chain knowledge graph, may specifically include the following processing: If the node in the industry chain where the case company is located is exactly the same as the penultimate level directory node of the target company, then the basic weight of the case corresponding to the case company will be set to 1.0. If the case company's industry chain node is exactly the same as the target company's third-to-last directory node, but different from the second-to-last directory node, then the case's basic weight will be set to 0.0. In the industry chain knowledge graph, the shortest path distance between the target enterprise node and the case enterprise node is calculated based on the graph embedding vector. The basic weight of the case is continuously adjusted based on the shortest path distance; when the shortest path distance is 0, the basic weight of the case is multiplied by a coefficient of 1.5; when the shortest path distance is 1, the basic weight of the case is multiplied by a coefficient of 1.2; when the shortest path distance is 2, the basic weight of the case is multiplied by a coefficient of 0.8; when the shortest path distance is greater than 2, the corresponding path signal is ignored. Based on the case-based weights and distance-based functions f ( d Calculate the industry chain score; where the function is... f ( d ) is a decay function or gain function based on the shortest path distance.

[0072] Specifically, the supply chain path mapping matching pathways employ a three-tiered progressive matching strategy: precise path matching, extended path matching, and distance-aware adjustment. This can include the following operations.

[0073] The first level is precise path matching: if the industry chain node where the enterprise in the case is located is the penultimate directory node of the target enterprise... C If the case is completely identical to -2, then the case's basic weight will be... BaseWeight Setting it to 1.0 indicates high business overlap.

[0074] The second level is extended path matching: if the industry chain node where the enterprise in the case is located is the third-to-last directory node of the target enterprise... C -3 is completely consistent, but with the target company C If -2 is different, then this case will be... BaseWeight Set it to 0.0.

[0075] The third level is distance-aware adjustment: based on the node matching mentioned above, the node where the target enterprise is located is further calculated in the industry chain knowledge graph. Ntarget The node where the case company is located Ncase Shortest path distance between d ,according to d right BaseWeight Continuous adjustment is performed. When d When the value is 0, it means that they are at the same node. BaseWeight Multiply by a factor of 1.5; when d When the distance is equal to 1, that is, when the distance is one jump, BaseWeight Multiply by a coefficient of 1.2; when d When the distance is equal to 2, that is, when the distance is two jumps, BaseWeight Multiply by a coefficient of 0.8; when d When the value is greater than 2, the signal for that path is ignored. Ultimately, based on... BaseWeight With distance-based attenuation or gain function f(d) Calculate the industry chain score Score ind ( c ):

[0076] For the boundary cliff problem in traditional fixed-level matching, traditional directory matching only makes a binary judgment based on whether the node names are the same: two companies are in adjacent industry chain nodes and their businesses are highly related, but they get a match score of zero because the node names are different; two other companies are in the same large node but their businesses are very different, but they get a high score because the node names are the same. This binary judgment cannot reflect the continuous relationship between the closeness or distance between industry chain nodes.

[0077] This embodiment introduces a function of continuous distance. f(d) This transforms the matching process from a binary judgment to a continuous metric, effectively smoothing out boundary cliffs. d A gain of 1.5 when the sum is 0 significantly improves the score for cases with high business overlap within the same node. d When the value equals 1, multiplying by 1.2 ensures that strongly related cases of one-hop neighbors still receive positive weights. d Multiplying by 0.8 when the value equals 2 moderately reduces the weight of weakly associated cases, but not to the point of zero. dWhen the score is greater than 2, ignoring the signal avoids erroneous high matching scores from distant nodes. This distance-aware weighting design allows the implicit structural information between nodes in the industry chain to be incorporated into the recommendation score in a continuous and computable manner, an effect that simple weighting cannot achieve.

[0078] In one embodiment, step S18 described above may specifically include the following processing: Extract the required conditions from the implementation prerequisites of the case's structured fields, and perform logical rule reasoning and fuzzy matching with the target enterprise's data readiness level to obtain a transferability score; The business relevance score, obtained by combining the product matching score and the industry chain score, is then multiplied by the transferability score to obtain the comprehensive score.

[0079] It is understandable that the case transferability assessment pathway is one of the core innovations of this invention, and its design stems from a profound understanding of the unique characteristics of AI case recommendation: the reference value of AI cases is not universal, and it highly depends on the recipient's digital foundation; an incorrect recommendation of an advanced case may mislead enterprises into blindly investing, producing negative effects. Therefore, the AI ​​case recommendation system must undertake the function of predicting transferability.

[0080] Specifically, the required conditions are extracted from the implementation prerequisites of the case's structured fields, and logical rule reasoning and fuzzy matching are performed with the target company's data readiness level to obtain a transferability score. Score transfer ( c The prerequisites for implementing this case include, but are not limited to, the deployment of an ERP system, the accumulation of structured data, the provision of a GPU computing cluster, and the presence of an AI technology team. These conditions are then automatically matched with the digital transformation maturity profile constructed by the target enterprise in step S12 above, using rule-based fuzzy reasoning. A high degree of condition fulfillment results in a migration score close to 1.0, while a significant lack of fulfillment results in a migration score of 0.1 or lower.

[0081] Will Score transfer ( c As a multiplicative factor, it is multiplied with the business relevance score obtained by combining the product matching score and the industry chain score, so that cases that are highly relevant to both product matching and industry chain matching but do not have the basic conditions for implementation are effectively reduced in weight in the comprehensive score.

[0082] The gap between recommendations and practical application is a common problem in existing recommendation systems. Current enterprise recommendation technologies almost never incorporate the recipient's acceptance capacity as a scoring factor into the recommendation algorithm, rendering highly relevant cases completely unusable. For example, an AI vision inspection case requiring GPU computing power clusters, even if highly relevant to a traditional manufacturing company's products, is meaningless to that company if it lacks any digital foundation.

[0083] This embodiment intervenes in the comprehensive scoring through a multiplicative constraint mechanism rather than an additive weighting method, making the transferability score a decisive constraint on the final result. The advantage of the multiplicative method is that even if both the product matching score and the supply chain score are high, if... Score transfer ( c If the score is close to zero, the overall score will be dragged down to an extremely low level, ensuring that such cases will not appear in the recommendation results. This design upgrades the recommendation system from maximizing matching to responsible recommendation. The system proactively helps companies filter out cases that cannot be implemented at the current stage, guiding them to focus on feasible AI application paths.

[0084] The product semantic matching pathway, the industry chain mapping matching pathway, and the transferability assessment pathway form a close technological dependency and functional synergy. Product matching solves the problem of identifying whether a business is relevant; industry chain matching further refines the degree of relevance; and transferability assessment ultimately solves the feasibility problem of relevance and implementability. The outputs of the three pathways are not simply added together, but rather form a progressive decision chain with layers of constraints through multiplication: only cases that pass semantic matching verification, industry chain matching verification, and transferability verification can obtain a high comprehensive score.

[0085] In one embodiment, step S20 above, when sorting candidate cases in descending order of comprehensive score, may further include calculating a time-sensitivity decay factor for the cases and incorporating the time-sensitivity decay factor into the comprehensive score calculation; wherein, the time-sensitivity decay factor is γ(t) = e -λt , t λ represents the number of months since the case was published, and λ is the decay coefficient.

[0086] It is understandable that this embodiment introduces a timeliness decay factor when calculating the comprehensive score of candidate cases. Specifically, by incorporating γ(t) into the calculation of the comprehensive score, the timeliness of the case can affect the comprehensive score in an exponentially smooth manner, ensuring that new cases have relative priority in the recommendation ranking.

[0087] Traditional recommendation systems often lack a dedicated time decay mechanism to address the imbalance in timeliness caused by older cases consistently ranking high due to accumulated historical data, making it difficult for newer cases to emerge. They may simply use a basic time threshold for binary selection, excluding cases before a certain time point while treating all cases afterward equally. This crude approach results in cases near the time threshold being unfairly treated.

[0088] This embodiment introduces a timeliness decay factor to ensure that case freshness influences the overall score in a smooth and continuous manner, rather than by abruptly setting a time threshold. The choice of an exponential function ensures that the timeliness weight is close to 1.0 when a case is first published, decreasing exponentially over time, avoiding the problems of insufficient or excessive decay that may occur with linear decay. The introduction of the λ parameter provides flexible control over the decay rate; different λ values ​​can be set for different application scenarios to adapt to different case update frequencies.

[0089] The timeliness decay factor forms a further synergy with the product matching score, supply chain score, and transferability score: timeliness decay ensures that novel cases have a reasonable competitive opportunity, transferability constraints ensure that these new cases are also feasible to implement, and product matching and supply chain matching ensure that new cases remain highly relevant to the target company's business.

[0090] In one embodiment, before sorting the candidate cases in descending order of comprehensive score in step S20 above, the following processing may also be included: Perform semantic deduplication on candidate cases.

[0091] It is understandable that before sorting candidate cases in descending order of comprehensive score, this embodiment may also include a process of semantic deduplication using case semantic vectors. Specifically, in the initially determined candidate case set, the cosine similarity between each pair of cases is calculated using the case semantic vectors. If the cosine similarity of the semantic vectors of two cases exceeds a preset threshold of 0.92, the two cases are determined to be content homogeneous, and only the one with the higher comprehensive score is retained, while the one with the lower score is removed.

[0092] The Matthew effect, a common problem in recommender systems, occurs when a few popular cases dominate recommendation positions due to their high scores, while less popular but differentiated high-quality cases struggle to gain visibility, and homogeneous cases repeatedly occupy recommendation positions, resulting in low information density. Traditional deduplication methods, typically based on precise text hashing, can only remove completely identical cases and cannot identify cases with highly similar semantic content but different expressions.

[0093] This embodiment uses cosine similarity calculation of case semantic vectors to identify homogeneous cases at the semantic content level, rather than simply using precise text matching for deduplication. A preset threshold of 0.92 ensures that deduplication is triggered only when the core content, technical solutions, and implementation paths of the cases are highly similar, avoiding the accidental deletion of cases that, while similar in some dimensions, actually provide differentiated reference information. A closed-loop synergy is formed between diverse semantic deduplication and semantic matching score calculation: both are based on the same set of case semantic vectors. V ( c This ensures that deduplication and matching score calculation are performed within a unified semantic space, guaranteeing consistency. After deduplication, each example in the recommendation list provides differentiated reference information, improving the overall information density of the recommendation list.

[0094] In some implementations, after completing all the above scoring calculations and optimization processes, this embodiment may also perform the following operation on the final output: sort the cases in the candidate case set according to the comprehensive recommendation score. Score final Sort in descending order, take Score final The top 20 cases with the highest scores are used as the final recommendation results. These final recommendation results are displayed to enterprise users in both list and radar chart comparison formats. The radar chart comparison simultaneously presents the matching degree, transferability, and timeliness dimensions.

[0095] in, Score final To comprehensively consider product matching score, supply chain score, Score transfer ( c )and γ ( t The final comprehensive recommendation score for the case is calculated afterward. Score final The value comprehensively reflects the case's overall performance across three core dimensions: business relevance, implementation feasibility, and timeliness.

[0096] The list format allows enterprise users to view detailed, structured information for each case study, including the applying company, industry, AI technology used, business problem solved, quantifiable results, and implementation prerequisites. The radar chart comparison format uses a matching degree dimension to correspond to a comprehensive product matching score and industry chain score, and a transferability dimension to correspond to... Score transfer ( c And corresponding to the timeliness dimension γ ( t This tool visually presents the differences in features of each recommended case across three core dimensions, helping companies quickly assess their AI application paths.

[0097] This embodiment utilizes a dual-modal output approach combining lists and radar charts to simultaneously meet enterprise users' needs for detailed case information access and rapid comparison and decision-making. The three dimensions of the radar chart correspond one-to-one with the three core pathways of the actual scoring calculation, ensuring consistency between the visual presentation and the underlying algorithm logic, and avoiding the problem of a disconnect between the displayed dimensions and the actual evaluation dimensions. The selection of the first 20 cases ensures both the richness of the recommended content and avoids information overload that could negatively impact user decision-making efficiency.

[0098] The aforementioned AI-powered intelligent recommendation method for enterprise cases maps enterprise products and case descriptions to a unified semantic vector space, using cosine similarity calculation to achieve semantic-level matching. This effectively solves the challenges of identifying heterogeneous product names, synonyms, and related upstream and downstream products. Simultaneously, by combining differentiated weighting of core and non-core products, it accurately focuses on the enterprise's core business needs, avoiding interference from irrelevant products. The synergistic effect of semantic vector matching and product importance ranking ensures that the semantic relevance of core products achieves a significantly higher score than that of peripheral products with perfect string matching, fundamentally eliminating the lexical gap and improving recommendation matching accuracy.

[0099] By mapping enterprises to a knowledge graph of the industry chain, not only is precise matching based on product-level nodes achieved, but also a decay or gain function based on the shortest path distance of the graph is introduced. This allows for fine-grained weight adjustment of node cases involving the same node (shortest path distance equals 0), one-hop neighbors (shortest path distance equals 1), and two-hop neighbors (shortest path distance equals 2), overcoming the boundary cliff effect caused by traditional fixed weights of 0 or 1. The combination of dual-scale industry chain coordinates—that is, product-level combinations—and the continuous perception function of graph distance precisely balances the accuracy and coverage of matching, enhancing the granularity of business overlap judgment.

[0100] Furthermore, a case transferability assessment pathway has been added. This pathway uses logical rules and fuzzy matching to logically deduce and match the case implementation prerequisites, such as ERP systems and data accumulation, with the target company's digital transformation maturity profile. The assessment results are then used as a multiplicative factor in the comprehensive scoring calculation. This effectively eliminates invalid cases that, while matching the business, cannot be supported by the company's digital infrastructure, bridging the gap between recommendation and implementation. The multiplicative constraint synergy mechanism of the three pathways—product matching, supply chain matching, and transferability assessment—ensures that each ultimately recommended case is both business-relevant and technically feasible, thereby significantly improving the implementability of the recommended cases.

[0101] Furthermore, by introducing a time-exponential decay factor, the weight of older cases is automatically reduced, ensuring the freshness of recommended content. Simultaneously, a similarity deduplication mechanism based on case semantic vectors identifies homogeneous cases within a preset threshold, avoiding redundant recommendations of similar content and enriching the recommendation scope. The synergistic post-processing of timeliness decay and diversity deduplication ensures the information density and timeliness of the recommendation list, optimizing the timeliness and diversity of recommended content. The final output can be compared with a list display and a radar chart. This radar chart covers three dimensions: matching degree, transferability, and timeliness, intuitively presenting case characteristics and assisting enterprises in quickly evaluating AI application paths. This effectively improves the success rate of intelligent transformation, achieves cost reduction and efficiency improvement, and enhances human-computer interaction experience and decision-making efficiency.

[0102] It should be understood that, although Figure 1 The steps are shown sequentially as indicated by the arrows, but these steps are not necessarily executed in the order indicated by the arrows. Unless otherwise specified in this document, there is no strict order in which these steps are executed; they can be performed in other orders. Figure 1 At least some of the steps may include multiple sub-steps or multiple stages. These sub-steps or stages are not necessarily executed at the same time, but can be executed at different times. The execution order of these sub-steps or stages is not necessarily sequential, but can be executed in turn or alternately with other steps or at least some of the sub-steps or stages of other steps.

[0103] In one embodiment, such as Figure 2 As shown, an intelligent recommendation system 100 for enterprise AI cases is also provided, including a case library construction module 11, an enterprise profile module 13, a product matching module 15, an industry chain matching module 17, a transferable evaluation module 19, and a case output module 21. The case library construction module 11 is used to collect enterprise AI application cases and perform structured parsing, calculate case semantic vectors, and store them as an AI case library. The enterprise profile module 13 is used to semantically standardize the products of target enterprises and sort them according to business importance, map target enterprises to an industry chain knowledge graph to obtain industry chain coordinates and graph embedding vectors of corresponding nodes, collect publicly available multi-dimensional weak signals from enterprises, and classify data readiness levels.

[0104] Product matching module 15 calculates a product matching score based on the cosine similarity between the target company's product semantic vector and the case semantic vectors of cases in the AI ​​case library, combined with the target company's product importance ranking and product attribute weights. Supply chain matching module 17 calculates a supply chain score based on the matching relationship between case companies and the target company at supply chain nodes in the AI ​​case library, and the shortest path distance in the supply chain knowledge graph; the shortest path distance is calculated using graph embedding vectors. Transferability assessment module 19 determines a transferability score by performing logical rule reasoning and fuzzy matching based on the case's implementation prerequisites and the target company's data readiness level. It then combines the product matching score and the supply chain score, multiplies this by the transferability score to obtain a comprehensive score. Case output module 21 sorts candidate cases in descending order of comprehensive score and outputs the top cases with the highest comprehensive scores. K One recommended case; K It is a positive integer.

[0105] The aforementioned AI-powered intelligent recommendation system 100 for enterprise cases maps enterprise products and case descriptions to a unified semantic vector space, using cosine similarity calculation to achieve semantic-level matching. This effectively solves the challenges of identifying heterogeneous product names, synonyms, and related upstream and downstream products. Simultaneously, by combining differentiated weighting of core and non-core products, it accurately focuses on the enterprise's core business needs, avoiding interference from irrelevant products. The synergistic effect of semantic vector matching and product importance ranking ensures that the semantic relevance of core products achieves a significantly higher score than that of peripheral products with perfect string matching, fundamentally eliminating the lexical gap and improving recommendation matching accuracy.

[0106] By mapping enterprises to the industry chain knowledge graph, not only is precise matching based on product-level nodes achieved, but the combination of dual-scale industry chain coordinates (i.e., product-level combinations) and the graph distance continuous perception function precisely balances the accuracy and coverage of matching, enhancing the granularity of business overlap judgment. Furthermore, a case transferability assessment pathway is added, using logical rule reasoning and fuzzy matching to examine the implementation prerequisites of the case and the target enterprise's digital transformation maturity profile. This effectively eliminates invalid cases that, while business-matched, lack the necessary digital infrastructure to support them, bridging the gap between recommendation and implementation. The multiplicative constraint synergy mechanism of the three pathways—product matching, industry chain matching, and transferability assessment—ensures that each recommended case is both business-relevant and technically feasible, significantly improving the implementability of the recommended cases.

[0107] Each module in the aforementioned enterprise AI case intelligent recommendation system 100 can be implemented entirely or partially through software, hardware, or a combination thereof. These modules can be embedded in hardware or independently of a device with data processing capabilities, or stored in software within the memory of the aforementioned device, so that the processor can call and execute the corresponding operations of each module. The aforementioned device can be, but is not limited to, various types of computer devices already existing in the art.

[0108] It is understood that the specific limitations of the above-mentioned enterprise AI case intelligent recommendation system 100 can be understood by referring to the corresponding explanations and limitations of the various embodiments of the enterprise AI case intelligent recommendation method above, and will not be repeated here.

[0109] Those skilled in the art will understand that all or part of the processes in the methods of the above embodiments can be implemented by a computer program instructing related hardware. The computer program can be stored in a non-volatile computer-readable storage medium. When executed, the computer program can include the processes of the embodiments of the above methods. Any references to memory, storage, databases, or other media used in the embodiments provided by this invention can include non-volatile and / or volatile memory. Non-volatile memory can include read-only memory (ROM), programmable ROM (PROM), electrically programmable ROM (EPROM), electrically erasable programmable ROM (EEPROM), or flash memory. Volatile memory can include random access memory (RAM) or external cache memory. By way of illustration and not limitation, RAM is available in various forms, such as static RAM (SRAM), dynamic RAM (DRAM), synchronous DRAM (SDRAM), dual data rate SDRAM (DDRSDRAM), enhanced SDRAM (ESDRAM), synchronous link DRAM (SLDRAM), memory bus DRAM (RDRAM), and interface DRAM (DRDRAM), etc.

[0110] The technical features of the above embodiments can be combined in any way. For the sake of brevity, not all possible combinations of the technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this specification.

[0111] The above embodiments merely illustrate several implementation methods of the present invention, and their descriptions are relatively specific and detailed, but they should not be construed as limiting the scope of protection of the invention. It should be noted that those skilled in the art can make various modifications and improvements without departing from the concept of the present invention, and all such modifications and improvements fall within the scope of protection of the present invention.

Claims

1. A method for intelligent recommendation of enterprise AI cases, characterized in that, include: Collect enterprise AI application cases and perform structured analysis. Calculate the semantic vectors of the cases and store them as an AI case library. The target company's products are semantically standardized and sorted by business importance. The target company is mapped to the industry chain knowledge graph to obtain the industry chain coordinates and graph embedding vectors of the corresponding nodes. The company's publicly disclosed multi-dimensional weak signals are collected and the data readiness level is divided. The product matching score is calculated based on the cosine similarity between the semantic vector of the target company's product and the semantic vector of the case in the AI ​​case library, combined with the importance ranking value of the target company's product and the weight of the product attribute. The industry chain score is calculated based on the matching relationship between case companies and target companies in the AI ​​case library at the nodes of the industry chain, and the shortest path distance in the industry chain knowledge graph; the shortest path distance is calculated based on the graph embedding vector. Based on the implementation prerequisites of the case and the data readiness level of the target enterprise, logical rule reasoning and fuzzy matching are used to determine the transferability score. The product matching score and the industry chain score are combined and multiplied by the transferability score to obtain the comprehensive score. Among them, the required conditions are extracted from the implementation prerequisites of the case's structured fields, and the required conditions are used to obtain the transferability score by logical rule reasoning and fuzzy matching with the data readiness level of the target enterprise. The implementation prerequisites include at least one of the following: deploying an ERP system, having accumulated structured data, having a GPU computing power cluster, and having an AI technology team. The transferability score is used as a multiplication factor. Candidate cases are sorted in descending order of their overall scores, and the top cases with the highest overall scores are output. K One recommended case; K It is a positive integer.

2. The enterprise AI case intelligent recommendation method according to claim 1, characterized in that, The steps involved in collecting enterprise AI application cases, performing structured analysis, calculating semantic vectors for the cases, and storing them as an AI case library include: The crawling scope is limited by a pre-defined whitelist of authoritative websites. Semantic filtering is performed on candidate pages, and pages that simultaneously match AI technology keywords and business scenario keywords are identified as candidate case pages and downloaded. Extract the main text from the downloaded page, perform named entity recognition and structured information extraction on the extracted main text, and fill the extracted case entity information into the preset template. Using fragmented fields from multiple sources as fact anchors, a large language model is used to standardize and complete the aggregated fragmented information and generate case summaries. The generated content and the original fragment input are subjected to illusion detection based on a text entailment recognition verification model. The completed content that passes the verification is stored in the AI ​​case library; the completed content that fails the verification is discarded. The semantic vectors of cases stored in the AI ​​case library are calculated using a text embedding model.

3. The enterprise AI case intelligent recommendation method according to claim 1, characterized in that, The process of calculating the cosine similarity between the target company's product semantic vector and the case semantic vectors of cases in the AI ​​case library includes: After calculating the product semantic vector of each product of the target company using a text embedding model, the cosine similarity is calculated with the case semantic vectors of each case in the AI ​​case library to obtain the product semantic similarity score. The product semantic vector is the vector representation of the target company's product mapped to the semantic space by the text embedding model, and the case semantic vector is the vector representation of the case mapped to the semantic space by the same text embedding model.

4. The enterprise AI case intelligent recommendation method according to any one of claims 1 to 3, characterized in that, The process of mapping target enterprises to a supply chain knowledge graph to obtain supply chain coordinates and graph embedding vectors of corresponding nodes includes: Map the target company to a pre-defined industry chain knowledge graph and record the target company's complete industry chain path; Extract the penultimate and tertiary level directory nodes and obtain the graph embedding vectors of the corresponding nodes; where the penultimate level directory node is the macro-level business sub-domain node and the tertiary level directory node is the subdivided link node.

5. The enterprise AI case intelligent recommendation method according to claim 3, characterized in that, The process of calculating a product matching score by combining the importance ranking of the target company's products and the weighting of product attributes includes: Each product of the target company is assigned an importance ranking value based on its revenue share and business coreness, and corresponding product attribute weights are assigned according to whether the product is a main product or a non-main product; the product attribute weight of the main product is 1000, and the product attribute weight of the non-main product is 999. The product matching score is calculated based on the product semantic similarity score, product attribute weight, and importance ranking value.

6. The enterprise AI case intelligent recommendation method according to claim 4, characterized in that, Based on the matching relationship between case companies and target companies in the AI ​​case library at the nodes of the industrial chain, and the process of calculating the industrial chain score based on the shortest path distance in the industrial chain knowledge graph, the process includes: If the node in the industry chain where the case company is located is exactly the same as the penultimate level directory node of the target company, then the basic weight of the case corresponding to the case company will be set to 1.

0. If the case company's industry chain node is exactly the same as the target company's third-to-last directory node, but different from the second-to-last directory node, then the case's basic weight will be set to 0.

0. In the industry chain knowledge graph, the shortest path distance between the target enterprise node and the case enterprise node is calculated based on the graph embedding vector. The basic weight of the case is continuously adjusted based on the shortest path distance; when the shortest path distance is 0, the basic weight of the case is multiplied by a coefficient of 1.5; when the shortest path distance is 1, the basic weight of the case is multiplied by a coefficient of 1.2; when the shortest path distance is 2, the basic weight of the case is multiplied by a coefficient of 0.8; when the shortest path distance is greater than 2, the corresponding path signal is ignored. Based on the case-based weights and distance-based functions f ( d Calculate the industry chain score; where the function is... f ( d ) is a decay function or gain function based on the shortest path distance.

7. The enterprise AI case intelligent recommendation method according to claim 1, characterized in that, Based on the implementation prerequisites of the case and the target company's data readiness level, logical rule reasoning and fuzzy matching are used to determine the transferability score. The steps include combining the product matching score and the industry chain score, and then multiplying this combined score by the transferability score to obtain the overall score. The business relevance score, obtained by combining the product matching score and the industry chain score, is then multiplied by the transferability score to obtain the comprehensive score.

8. The enterprise AI case intelligent recommendation method according to claim 1, characterized in that, When ranking candidate cases in descending order of comprehensive score, the calculation of a time-sensitivity decay factor is also included, and this factor is incorporated into the comprehensive score calculation; whereby the time-sensitivity decay factor is γ(t) = e -λt , t λ represents the number of months since the case was published, and λ is the decay coefficient.

9. The enterprise AI case intelligent recommendation method according to claim 1, characterized in that, Before sorting candidate cases in descending order of comprehensive score, the following are also included: Perform semantic deduplication on candidate cases.

10. An intelligent recommendation system for enterprise AI case studies, characterized in that, Including the following steps: The case library construction module is used to collect enterprise AI application cases, perform structured analysis, calculate case semantic vectors, and store them as an AI case library; The enterprise profiling module is used to semantically standardize the products of the target enterprise and sort them according to business importance. It maps the target enterprise to the industry chain knowledge graph to obtain the industry chain coordinates and graph embedding vectors of the corresponding nodes. It also collects multi-dimensional weak signals publicly disclosed by the enterprise and classifies the data readiness level. The product matching module is used to calculate the product matching score based on the cosine similarity between the product semantic vector of the target company's product and the case semantic vector of the cases in the AI ​​case library, combined with the importance ranking value of the target company's products and the product attribute weights. The supply chain matching module is used to calculate the supply chain score based on the matching relationship between case companies and target companies in the AI ​​case library at the nodes of the supply chain, as well as the shortest path distance in the supply chain knowledge graph; the shortest path distance is calculated based on the graph embedding vector. The transferability assessment module is used to determine the transferability score by logical rule reasoning and fuzzy matching based on the implementation prerequisites of the case and the data readiness level of the target enterprise. The product matching score and the industry chain score are combined and multiplied by the transferability score to obtain the comprehensive score. Specifically, the required conditions are extracted from the implementation prerequisites of the case's structured fields, and the required conditions are logical rule reasoning and fuzzy matching with the data readiness level of the target enterprise to obtain the transferability score. The implementation prerequisites include at least one of the following: deploying an ERP system, having accumulated structured data, having a GPU computing power cluster, and having an AI technology team. The transferability score is used as a multiplication factor. The case output module sorts candidate cases in descending order of their overall score and outputs the top cases with the highest overall score. K One recommended case; K It is a positive integer.

Images