Enterprise recommendation information system and enterprise recommendation method

Abstract

The invention relates to the technical field of information recommendation, and provides an enterprise recommendation information system and an enterprise recommendation method. The system comprises a data acquisition module, an atlas storage library, a feature generation module, a sorting model module and an intervention engine. According to the sequencing model module, by combining a wide model and a deep model and introducing a TW-GAT technology, effective fusion of multi-modal features is achieved, the dynamic timeliness of an enterprise relationship is accurately captured, and therefore the problems that feature utilization is insufficient and dynamic updating lags are effectively solved. The intervention engine can carry out precise processing on territorial sensitive and long-tail query scenes, so that the solution efficiency of the bad case is remarkably improved, and the user experience is optimized. Through the improvement, the sorting effect, the maintenance cost and the generalization ability of the system are remarkably improved.

Description

Technical Field

[0001] The present application relates to the field of information recommendation technology, and in particular to an enterprise recommendation information system and an enterprise recommendation method. Background Art

[0002] Enterprise recommendation information systems play an important role in the field of enterprise information service technology, especially in enhancing the accuracy of enterprise search rankings, broadening generalization capabilities, and improving maintenance efficiency. As enterprise information services continue to develop, companies have an increasing demand for accurate and efficient recommendations. These systems can help companies quickly match ideal partners, improve supply chain management efficiency, and optimize resource allocation on bidding platforms, profoundly impacting the overall development and operation of the enterprise ecosystem. A well-designed enterprise recommendation information system can quickly locate the most relevant companies within a vast amount of enterprise information, saving companies time and costs, promoting cooperation and development between companies, and driving overall industry progress. Both large and small enterprises can gain strong support from these systems.

[0003] Traditional enterprise recommendation information systems are primarily based on static rule-based recommendation engines. These systems rely on over ten simple features, such as registered capital and geographic distance, to rank and recommend enterprises. While these systems meet basic recommendation needs to a certain extent, their rigid rules and limited feature dimensions lead to poor recommendation results. For example, the average DCG@5 (Discounted Cumulative Gain at 5) is only 2.3.

[0004] Some enterprise recommendation information systems have introduced basic graph technology (such as invention patent application publication number CN115422459A), and embedded learning of enterprise graphs through technologies such as graph convolutional networks (GCN), expanding the feature dimension to more than 30 items. This system can capture more complex relationships between enterprises, thereby improving the recommendation effect. According to the nDCG evaluation standard, the DCG@5 index reached 2.7, indicating that the ranking performance of the recommendation system has reached a high level after considering the degree of relevance and ranking correctness. Unfortunately, this type of system still has significant shortcomings in its ability to dynamically capture the complex relationship between query terms and enterprise characteristics, making it difficult to effectively solve the timeliness problem of dynamic relationships.

[0005] In addition, some enterprise recommendation information systems have introduced integrated dynamic learning mechanisms (such as invention patent application publication number CN119417250A) to enhance the system's adaptability. However, it still has significant defects, mainly manifested in insufficient feature utilization (on average, less than 40% of potential association paths are effectively utilized) and dynamic feature update delays (usually greater than 24 hours), making it difficult to adapt to rapidly changing business environments and user needs. For example, in real-time recommendation systems, dynamic feature updates are an important means to improve recommendation accuracy and practicality, but traditional recommendation systems often fail to capture changes in user interests and needs in a timely manner, resulting in update delays. In addition, research on user portrait update strategies and dynamic recommendation methods in e-commerce personalized recommendation systems has shown that untimely updates of user portraits will lead to inaccurate recommendation results, which further emphasizes the importance of dynamic feature updates in improving the performance of recommendation systems.

[0006] However, the enterprise recommendation information systems in the aforementioned related technologies still suffer from the following issues: First, there are deficiencies in feature engineering. In current systems, textual, graph, and business features are isolated and lack effective integration methods. This results in a loss rate of over 65% in cross-modal correlation information, severely reducing recommendation accuracy. Second, the related technologies lack dynamic adaptability. According to industry analysis reports, traditional rule bases are frequently updated, averaging up to three times per month, resulting in high maintenance costs and requiring two maintenance personnel per month. For example, updates to KPI indicator libraries have traditionally relied on regular meetings and expert reviews. While these ensure a certain level of update frequency and accuracy, they are inefficient and prevent real-time data analysis and prediction. This rigid maintenance model makes it difficult to quickly adapt to market changes and meet user needs. Furthermore, the related technologies lack privacy protection. The lack of a secure data sharing mechanism prevents the effective aggregation and utilization of cross-institutional data, which not only limits data scale but also weakens model training effectiveness. When faced with the cold start problem, the failure rate of recommendation systems can reach as high as 58%, severely impacting the system's generalization and recommendation quality. Finally, the relevant technology lacks the ability to handle long-tail queries. In complex scenarios like multi-entity combination queries (such as "person name + company name"), the existing system has an error rate as high as 42%. This results in poor performance in core metrics like DCG, significantly impairing the user experience and making it difficult to meet diverse user needs.

[0007] In general, related technologies rely too much on fixed rules and simple indicators (such as registered capital or regional distance) for sorting, which directly leads to poor sorting results and generally low relevance indicators (such as DCG). In addition, the high cost of system maintenance also makes it difficult to keep up with the pace of rapid business iteration. What is particularly critical is that the generalization ability of related technologies is insufficient, making it difficult to effectively respond to long-tail queries. At the same time, the mining and utilization of high-value features such as corporate graph associations and text matching are also stretched, which seriously restricts the improvement of recommendation quality. These defects together constitute technical problems that need to be urgently solved in this field. Summary of the Invention

[0008] In order to solve the problems that related technologies rely on fixed rules and simple indicators, have high maintenance costs and weak generalization capabilities, this application provides a set of enterprise recommendation information system and enterprise recommendation method.

[0009] In one aspect, the present application provides an enterprise recommendation information system, comprising: Data acquisition module, used to obtain user query terms, enterprise-related data and competitor enterprise ranking information; A graph repository, connected to the data acquisition module, is used to generate enterprise graph data based on the enterprise-related data; a feature generation module, connected to the data acquisition module and the graph repository, is used to generate a feature vector of the target enterprise and the corresponding upstream and downstream enterprises based on the enterprise graph data, the user query term, and the enterprise-related data, wherein the feature vector includes enterprise graph association feature information, text matching feature information, and business attribute feature information; A ranking model module is connected to the data acquisition module, the feature generation module, and the intervention engine. The ranking model module has a ranking model for training the ranking model and using the trained ranking model to score and rank the upstream and downstream enterprises based on the feature vector to generate ranking data. The ranking model includes a wide model and a deep model. The ranking model module generates linear prediction data using the wide model and generates deep prediction data using the deep model, and fuses the linear prediction data with the deep prediction data to generate the ranking data. The deep model includes a time-series weighted graph attention network. The time-series weighted graph attention network introduces a time decay factor when calculating the attention weights of neighbor nodes of enterprise nodes in the enterprise graph association feature information, so that the attention weights representing enterprise relationships decay as the time difference increases. An intervention engine is connected to the ranking model module and is used to determine whether the ranking data contains regionally sensitive information. When it is determined that the regionally sensitive information exists, a geographic distance factor is superimposed on the ranking data to generate enterprise recommendation information. The intervention engine also determines whether the ranking data is long-tail data. When it is determined that the ranking data is long-tail data, an intervention operation matching from a long-tail case library is performed on the ranking data to form enterprise recommendation information.

[0010] Using this technical solution, the data acquisition module acquires competitor ranking information, the feature generation module constructs multi-dimensional feature vectors, and the ranking model module uses a wide and deep model for ranking. Furthermore, a temporal weighted graph attention network (TW-GAT) and an intervention engine are introduced for post-processing. This solution comprehensively addresses the problems of traditional recommender systems, such as their reliance on fixed rules, single feature dimensions, inability to capture temporal dynamics, and weak generalization. By combining the wide and deep model with TW-GAT, it effectively integrates multimodal features and captures the dynamic timeliness of business relationships, addressing issues such as underutilization of features and delayed dynamic updates. The intervention engine accurately handles geographically sensitive and long-tail query scenarios, significantly improving bad case resolution and user experience. Ultimately, the system achieves significant improvements in ranking performance (DCG increased by 14.8%), maintenance costs (reduced by 87.5%), and generalization, thanks to the use of advanced technologies such as vector embedding, combined weighting, and graph neural networks.

[0011] Optionally, the deep model also includes a multi-layer perceptron, wherein the deep model performs graph dynamic representation learning on the enterprise nodes in the enterprise graph associated feature information through the temporal weighted graph attention network to generate an enterprise dynamic representation vector, and fuses the enterprise dynamic representation vector with the text matching feature information and the business attribute feature information to form a comprehensive feature vector, and also maps the comprehensive feature vector to the deep prediction data through the multi-layer perceptron.

[0012] Using the above technical solution, in the deep model, a temporal weighted graph attention network (TW-GAT) is used to generate a dynamic representation vector of the enterprise containing the dynamic structural information of the graph, which is then fused with the text and business attribute feature information and input into the multi-layer perceptron (MLP) for high-order interactive learning. This solution solves the problem of the mutual separation of multimodal features such as graphs, texts, and businesses, and the serious loss of related information in related technologies. In this way, the system can not only understand the features of a single dimension, but also learn complex, cross-modal deep patterns such as "companies in a specific technical field, which have recently cooperated with well-known institutions and are highly relevant to user queries are more valuable to recommend", thereby greatly improving the accuracy and generalization ability of recommendations.

[0013] Optionally, the deep model performs graph dynamic representation learning on the enterprise nodes in the enterprise graph associated feature information through the time-series weighted graph attention network to generate an enterprise dynamic representation vector, wherein a time decay factor is introduced when calculating the attention weight of the neighboring nodes of the enterprise node, which is expressed as the following formula: Among them, W ij represents the dynamic attention weight of node j relative to node i, exp(.) represents the exponential function, h i and h j Represents the embedding vectors of node i and node j respectively, W represents the learnable weight matrix, a T represents the learnable attention vector, [Wh i ||Wh j ] represents the concatenation of the transformed embedding vectors of node i and node j, LeakyReLU represents the activation function, Represents all neighbor nodes k∈N of node i i Sum the attention scores, e -λΔt represents the time attenuation factor, Δt represents the time difference, and λ represents the attenuation coefficient.

[0014] Using the above technical solution, the time decay factor e is innovatively introduced when calculating the graph attention weight. -λΔt This solution directly addresses the technical pain point of traditional graph models' inability to measure the "timeliness" of relationships, which leads to delayed updates of dynamic features. This design enables the model to assign higher weights to recent relationships (such as recent investments and new collaborations), while weakening the influence of outdated relationships. This allows the generated enterprise representation vector to more accurately reflect the company's current activity and market position, achieving a dynamic update delay of less than 1 hour for node representations, thereby improving the timeliness and accuracy of recommendations in a rapidly changing business environment.

[0015] Optionally, the deep model performs graph dynamic representation learning on the enterprise nodes in the enterprise graph associated feature information through the temporal weighted graph attention network to generate an enterprise dynamic representation vector, wherein multi-hop relationship aggregation is performed on the multi-hop neighbor nodes of the enterprise node, which is expressed as the following formula: in, represents the embedding vector learned by node i in the l+1th layer of TW-GAT, σ represents the nonlinear activation function, Represents the one-hop neighbor node j∈N of the aggregation node i i Information, represents the embedding vector of neighbor node j at layer l, W (l) represents the learnable weight matrix of layer l, represents the attention weight of node j to node i, β·READOUT(.) represents the aggregation of information from the two-hop neighbor nodes of node i, represents the embedding vector of neighbor node k in layer l, represents the set of two-hop neighbor nodes of node i, READOUT(.) represents the readout function, and β represents the learnable coefficient.

[0016] Using the above technical solution, through multi-layer TW-GAT stacking and multi-hop relationship aggregation formula, the node's two-hop (or multi-hop) neighbor information is integrated into the representation learning of the current node. This solution aims to solve the problem of insufficient utilization of technology graph features and improve resource conversion efficiency by optimizing the switching and dynamic management of frequency domain resources. By aggregating multi-hop neighbor information, the model can capture deeper and more complex indirect enterprise connections (such as "supplier's supplier"), not just direct connections. This enables the generated enterprise embedding vector to more comprehensively and robustly reflect the enterprise's true position and influence in the industrial chain and cooperation network, thereby improving the depth and breadth of recommendations.

[0017] Optionally, the intervention engine also annotates the enterprise recommendation information to generate annotated proofreading data; the ranking model module performs first-stage training and second-stage training on the ranking model, and updates the ranking model through a federated learning framework. During the first-stage training, the ranking model is trained based on the ranking information of the competing enterprises, and during the second-stage training, the ranking model is adjusted based on the annotated proofreading data fed back from the intervention engine.

[0018] Using this technical solution, a two-stage training model was implemented, combining "first-stage training based on competitive information" with "second-stage training based on annotated and proofread data," and continuously updating the model using a federated learning framework. This solution addressed issues such as poor cold start performance, slow iteration speed, and data silos. The first-stage training enabled the model to quickly reach industry benchmarks; the second-stage fine-tuning and closed-loop feedback from the intervention engine continuously optimized the model, addressing the long-tail problem and increasing the bad case resolution rate to 90%. The application of federated learning aggregated data from multiple partner institutions without sharing the original data, effectively addressing the data silo issue, improving the model's AUC by 11.3%, and significantly enhancing the model's generalization capabilities.

[0019] Optionally, the enterprise feature system further includes: An encryption communication module is connected to the sorting model module and is used to encrypt communication between the sorting model module and the server under the federated learning architecture, wherein the encryption communication module uses the Paillier homomorphic encryption algorithm to encrypt the model update information of the sorting model and decrypt the encrypted data from the server.

[0020] Using the above technical solution, an encrypted communication module is added to the federated learning framework, and the Paillier homomorphic encryption algorithm is used to encrypt and protect model update information (such as gradients). This solution directly addresses the core issues of lack of privacy protection and data security concerns in cross-institutional data collaboration. Using homomorphic encryption technology, the server can complete the aggregation of model gradients of all parties without exposing the original data. This encryption method allows data to be calculated in an encrypted state, thereby ensuring the data privacy and security of the participants. In many fields such as finance, healthcare, and the Internet of Things, homomorphic encryption technology has been proven to effectively protect data privacy while promoting data sharing and analysis, which is of great significance for solving the problem of data silos.

[0021] On the other hand, the present application also provides an enterprise recommendation method, comprising the following steps: S1. Obtain user query terms, enterprise-related data, and competitor ranking information; S2. generating enterprise graph data based on the enterprise-related data; S3. Generate a feature vector for the target enterprise and its corresponding upstream and downstream enterprises based on the enterprise graph data, the user query term, and the enterprise-related data. The feature vector includes enterprise graph association feature information, text matching feature information, and business attribute feature information. S4, training ranking model; S5. Using the ranking model to score and rank the upstream and downstream enterprises based on the feature vectors to generate ranking data, the ranking model includes a wide model and a deep model, wherein the wide model generates linear prediction data, the deep model generates deep prediction data, and the linear prediction data and the deep prediction data are fused to generate the ranking data; the deep model includes a time-series weighted graph attention network, which introduces a time decay factor when calculating the attention weights of neighbor nodes of enterprise nodes in the enterprise graph association feature information, so that the attention weights representing enterprise relationships decay as the time difference increases; S6. Determine whether the sorting data contains any regionally sensitive information. S7. When it is determined that the ranking data contains regionally sensitive information, superimposing a geographic distance factor on the ranking data to form enterprise recommendation information; S8. Determine whether the sorted data is long-tail data; S9. When it is determined that the sorted data is long-tail data, an intervention operation matching from a long-tail case library is performed on the sorted data to form enterprise recommendation information.

[0022] Using the above technical solution, by obtaining competitor information, constructing a vector that integrates multimodal features, using Wide & Deep and Temporal Weighted Graph Attention Networks (TW-GAT) for training and ranking, and combining it with a dynamic intervention engine to correct the results. This method systematically solves a series of technical difficulties faced by traditional recommendation methods, such as relying on static rules, poor ranking effects, high maintenance costs, and weak generalization capabilities. Through multi-dimensional feature engineering and deep learning models, this method significantly improves the ranking effect of the recommendation system, for example, the DCG index is improved to 3.33. In addition, the introduction of a dynamic intervention mechanism greatly enhances the system's adaptability to complex scenarios, for example, the bad case resolution rate is increased to 90%. These improvements are particularly important in the field of e-commerce, because the recommendation system not only improves shopping efficiency and user experience by analyzing user behavior data, but also effectively increases sales and conversion rates.

[0023] Optionally, the deep model in step S6 also includes a multi-layer perceptron, wherein generating deep prediction data through the deep model includes: performing graph dynamic representation learning on the enterprise nodes in the enterprise graph associated feature information by the temporal weighted graph attention network to generate an enterprise dynamic representation vector; fusing the enterprise dynamic representation vector with the text matching feature information and the business attribute feature information to form a comprehensive feature vector; and mapping the comprehensive feature vector to the deep prediction data by the multi-layer perceptron.

[0024] Using this technical solution, the recommendation method specifically defines a path for generating deep prediction data: first, using TW-GAT to learn a dynamic graph representation of the enterprise, then integrating this representation with text and business features, and finally mapping it to a final score using a multi-layer perceptron (MLP). This method addresses the technical issue of effectively integrating different types of features, which can lead to information loss. By integrating structured graphs, unstructured text, and business data, and using deep learning and nonlinear transformations, it captures high-order feature interactions, providing a more comprehensive and insightful basis for recommendations and significantly improving recommendation accuracy.

[0025] Optionally, the time-series weighted graph attention network performs graph dynamic representation learning on the enterprise nodes in the enterprise graph associated feature information to generate an enterprise dynamic representation vector, wherein a time decay factor is introduced when calculating the attention weight of the neighboring nodes of the enterprise node, which is expressed as the following formula: Among them, W ijrepresents the dynamic attention weight of node j relative to node i, exp(.) represents the exponential function, h i and h j Represents the embedding vectors of node i and node j respectively, W represents the learnable weight matrix, a T represents the learnable attention vector, [Wh i ||Wh j ] represents the concatenation of the transformed embedding vectors of node i and node j, LeakyReLU represents the activation function, Represents all neighbor nodes k∈N of node i i Sum the attention scores, e -λΔt represents the time attenuation factor, Δt represents the time difference, and λ represents the attenuation coefficient; Furthermore, multi-hop relationship aggregation is performed on the multi-hop neighbor nodes of the enterprise node, which is expressed as the following formula: in, represents the embedding vector learned by node i in the l+1th layer of TW-GAT, σ represents the nonlinear activation function, Represents the one-hop neighbor node j∈N of the aggregation node i i Information, represents the embedding vector of neighbor node j at layer l, W (l) represents the learnable weight matrix of layer l, represents the attention weight of node j to node i, β·READOUT(.) represents the aggregation of information from the two-hop neighbor nodes of node i, represents the embedding vector of neighbor node k in layer l, represents the set of two-hop neighbor nodes of node i, READOUT(.) represents the readout function, and β represents the learnable coefficient.

[0026] Using the above technical solution, the attention weight calculation formula with a time decay factor and the multi-hop neighbor aggregation formula are specifically applied in the process of graph dynamic representation learning. This method solves the problems of graph models being insensitive to the timeliness of relationships and insufficiently utilizing indirect relationships in representation learning. The application of the time decay factor ensures that recommendations are more timely, while multi-hop aggregation allows the model to see deeper and farther, and can make judgments based on more comprehensive industry chain information. By combining advanced technical means such as deep learning and natural language processing, companies can generate more accurate and comprehensive representation vectors, thereby significantly improving the quality of graph features and further improving the accuracy and personalization of recommendation systems.

[0027] Optionally, step S4 includes performing first-stage training and second-stage training on the ranking model, and updating the ranking model through a federated learning framework, wherein the ranking model is trained based on the ranking information of the competing enterprises during the first-stage training, and the ranking model is adjusted based on the annotation proofreading data generated by the annotation of the enterprise recommendation information during the second-stage training; wherein communication with the server is encrypted under the federated learning framework, the model update information of the ranking model is encrypted using the Paillier homomorphic encryption algorithm, and the encrypted data from the server is decrypted.

[0028] Using the above technical solution, two-stage training, federated learning updates, and Paillier homomorphic encryption communication are integrated into the model training and update process. This method not only solves the problems of slow iteration and low performance ceiling of a single model, but also addresses the data privacy and security barriers in multi-institutional collaboration. Two-stage training enables rapid iteration and continuous optimization; through federated learning and homomorphic encryption, a wider range of data sources are securely aggregated, breaking down data silos and greatly improving the model's generalization ability and performance in cold start scenarios (model AUC increased by 11.3%), achieving a balance between performance and security.

[0029] In summary, this application includes at least one of the following beneficial technical effects: 1. The data acquisition module acquires competitor ranking information, the feature generation module constructs multi-dimensional feature vectors, the ranking model module uses a wide and deep model for ranking, and introduces a temporal weighted graph attention network (TW-GAT) and an intervention engine for post-processing. This solution comprehensively overcomes the shortcomings of traditional recommendation systems, such as reliance on static rules, limited feature dimensions, insufficient capture of temporal dynamics, and poor generalization. The combination of the wide and deep model and TW-GAT achieves deep fusion of multimodal features and accurately captures the time-sensitive dynamics of corporate relationships, thereby solving the problems of insufficient feature utilization and delayed dynamic updates. With the help of the intervention engine, precise optimization is carried out for regionally sensitive and long-tail query scenarios, significantly improving the efficiency of resolving bad cases and optimizing the user experience. Ultimately, the system has significantly improved ranking results (DCG increased by 14.8%), maintenance costs (reduced by 87.5%), and generalization capabilities.

[0030] 2. In the deep model, a temporal weighted graph attention network (TW-GAT) is used to capture the graph structure information of corporate dynamics. By calculating the attention weights between nodes, a vector representing the dynamic structural information of the graph is dynamically generated. These vectors are then fused with text and business attribute feature information, and high-level interactive learning is performed through a multi-layer perceptron (MLP). This solves the problem in related technologies where multimodal features such as graphs, text, and business are separated from each other and associated information is severely lost. The system can not only analyze single features, but also grasp complex cross-modal deep patterns such as "companies in a specific technology field that have recently cooperated with well-known institutions and are highly matched to user queries are more valuable for recommendation," significantly improving the accuracy and generalization of recommendations.

[0031] 3. Through multi-layer TW-GAT stacking and multi-hop relationship aggregation formula, the node's two-hop (or multi-hop) neighbor information is integrated into the representation learning of the current node. This solves the problem of insufficient utilization of related technical graph features (on average, only 32% of potential association paths are utilized). By aggregating multi-hop neighbor information, the model is able to capture deeper and more complex indirect enterprise connections (such as "supplier's supplier"), rather than just direct connections. This allows the generated enterprise embedding vector to more comprehensively and robustly reflect the enterprise's true position and influence in the industrial chain and cooperation network, thereby improving the depth and breadth of recommendations.

[0032] 4. A two-stage training model, combining "first-stage training based on competitive information" with "second-stage training based on annotated and proofread data," and utilizing a federated learning framework for continuous model updates, addressed issues such as poor cold start performance, slow iteration, and data silos. The first-stage training enabled the model to quickly reach industry benchmarks; during the second-stage fine-tuning process, we continuously optimized the model through closed-loop feedback from the intervention engine. This process not only addressed the long-tail problem in the recommendation system but also significantly improved the ability to handle uncommon (bad) cases, increasing the resolution rate to 90%. The application of federated learning, through a distributed machine learning framework, aggregates data from multiple partner institutions without sharing the original data, effectively addressing data silos. For example, in risk control models, financial institutions achieve joint modeling through parameter exchange while meeting data privacy requirements. In the healthcare sector, federated learning enables medical institutions to integrate risk profiles across regions and industries while protecting patient privacy. For example, a smart diagnosis project involving a tertiary hospital and 12 primary healthcare institutions has been implemented. Furthermore, federated learning can improve the performance of machine learning models. For example, in the financial industry, financial institutions can analyze user behavior and provide more personalized financial services without sharing user data. These application cases demonstrate that federated learning not only improves model accuracy but also enhances the model's adaptability in specific environments. For example, the model's AUC has increased by 11.3%, significantly enhancing the model's generalization ability. BRIEF DESCRIPTION OF THE DRAWINGS

[0033] Figure 1 This is a module diagram of the enterprise recommendation information system according to an embodiment of the present application; Figure 2 1 is a schematic diagram of a data acquisition module according to an embodiment of the present application; Figure 3 It is a module diagram of the atlas repository of this application; Figure 4 1 is a schematic diagram of a feature generation module according to an embodiment of the present application; Figure 5 Schematic diagram of a sorting model module according to an embodiment of the present application; Figure 6 is a schematic diagram of a multi-hop aggregation mechanism according to an embodiment of the present application; Figure 7 is a schematic diagram of a module of an intervention engine according to an embodiment of the present application; Figure 8 1 is a schematic diagram of an encryption communication module according to an embodiment of the present application; Figure 9 Schematic diagram of the federated learning architecture of an embodiment of the present application; Figure 10It is a flowchart of the enterprise recommendation method of an embodiment of the present application.

[0034] Explanation of reference numerals: 10. Data acquisition module; 11. Multi-source data acquisition module; 12. User query term acquisition module; 13. Competitive data acquisition module; 14. Data preprocessing module; 20. Graph repository; 21. Graph construction and update module; 22. Graph database management module; 23. Graph indexing and query optimization module; 30. Feature generation module; 31. Graph feature extraction module; 32. Text feature extraction module; 33. Business attribute feature extraction module; 34. Feature fusion module; 40. Sorting module Model module; 41. Feature processing module; 42. Wide model; 43. Deep model; 44. Model training and optimization module; 45. Data fusion module; 46. Sorting optimization module; 47. Federated learning collaboration module; 50. Intervention engine; 51. Data reception and preliminary verification module; 52. Intervention operation matching module; 53. Data correction and adjustment module; 54. Long-tail case library maintenance and feedback module; 60. Encrypted communication module; 61. Encryption and decryption module; 62. Secure transmission module; 63. Key management module. DETAILED DESCRIPTION

[0035] The following is combined with Figure 1-10 This application is described in further detail.

[0036] The embodiment of the present application aims to disclose a new type of enterprise recommendation information system.

[0037] Figure 1 The following is a schematic diagram showing the module structure of the enterprise recommendation information system in the embodiment of the present application. Figure 1 The system consists of a data acquisition module (Data Acquisition Module) 10, a knowledge graph storage (Knowledge Graph Storage) 20, a feature generation module (Feature Engineering Module) 30, a ranking model module (Ranking Model Module) 40, an intervention engine (Intervention Engine) 50 and an encrypted communication module (Encrypted Communication Module) 60.

[0038] The data acquisition module 10 is responsible for collecting raw data from multiple data sources, including user query terms and competitor ranking information, and can obtain multi-source heterogeneous enterprise data from channels such as bidding websites, industrial and commercial systems, patent databases, and public opinion platforms.

[0039] The graph repository 20 is responsible for managing and storing enterprise graph data, and providing the required structured enterprise graph information for the feature generation module.

[0040] Based on the data provided by the data acquisition module 10 and the enterprise graph information in the graph repository 20, the feature generation module 30 generates a multidimensional feature vector for each pair of target enterprises and their upstream and downstream enterprises. These feature vectors are then fed into the ranking model module 40 for processing. The feature vectors include enterprise graph association features (Knowledge Graph Association Features), text matching features (Text Matching Features), and business attribute features (Business Attribute Features).

[0041] Specifically, enterprise graph association feature information includes Association Path Weight and Node Centrality assessment based on indicators such as financing rounds, industry influence index, and transaction frequency. Text matching feature information covers semantic similarity between query terms and target enterprises based on the BERT model, the degree of longest common subsequence matching, and the confidence level of named entity recognition. Business attribute features include 58 structured data indicators such as registered capital, number of insured persons, and bidding activity.

[0042] The ranking model module 40 uses the ranking model to train based on the competitor ranking information provided by the data acquisition module 10 and the annotated proofreading data fed back by the intervention engine 50. It then scores and ranks upstream and downstream companies in combination with the feature vectors to generate ranking information. The ranking model adopts a two-stage model training method: the first stage builds a preliminary model based on competitor data, and the second stage fine-tunes the data through manual proofreading to generate a high-precision final model. The ranking model module 40 adopts a federated learning framework and collaborates with the encrypted communication module 60 to aggregate multi-party gradient information through encrypted communication without sharing the original data, thereby realizing the update of the ranking model.

[0043] The intervention engine 50 is used to receive the ranking information of the ranking model module 40, and output the final enterprise recommendation information after adjustment according to the intervention strategy to solve the long-tail queries and bad cases in specific scenarios. The intervention engine 50 can also feed back the screened high-value samples to the ranking model module 40 for model fine-tuning. Among them, the intervention engine 50 will dynamically adjust the weight distribution of regions in combination with the geographic distance factor for regional sensitive queries. In addition, the intervention engine 50 also builds an abnormal scenario library by reverse crawling competitor data, and uses active learning technology to screen out boundary cases that require manual review, thereby forming a closed loop of continuous iteration of rules and models.

[0044] The encrypted communication module 60 is connected to the ranking model module and its main function is to ensure the data privacy of all participants in the federated learning environment. The encrypted communication module 60 uses Paillier homomorphic encryption technology to exchange and aggregate encrypted gradients instead of directly sharing raw data.

[0045] Figure 2 Schematic diagram of the data acquisition module of the embodiment of the present application. Figure 2 The data acquisition module 10 includes a multi-source data collection module 11, a user query collection module 12, a competitor data collection module 13 and a data preprocessing module 14.

[0046] The multi-source data acquisition module 11 is responsible for continuously and automatically collecting enterprise-related raw data from various internal and external data sources, including industrial and commercial system data, bidding network data, patent database data, public opinion platform data, and other public data sources.

[0047] The industrial and commercial system data covers key industrial and commercial registration information such as the company's registration details, equity structure, business scope, legal representative, registered capital, and company status (such as continued existence, cancellation, revocation, etc.). The bidding network data includes information on the company's bidding activities, such as winning bids, bidding projects, participation, historical performance, etc., which helps to evaluate the company's business activity and competitiveness. The patent database data shows the company's patent application status, authorization information and technology layout, which reflect the company's innovation capabilities and technical background. The public opinion platform data includes the company's public opinion information, media reports, social evaluations, etc., which helps to build a reputation and risk profile of the company. Other public data sources may include news media, industry newspapers, social media, etc., which are used to supplement and enrich corporate information.

[0048] The multi-source data acquisition module 11 can use crawlers or API interfaces to capture structured and unstructured data from the above data sources regularly or even in real time, and then transmit these data to the data preprocessing module 14 for cleaning and formatting.

[0049] The user query term acquisition module 12 is responsible for acquiring user search keywords in real time. It uses system event tracking or front-end logging technology to track user search behavior in the recommendation system in real time, such as searches for keywords such as "big data solutions" or "chip design companies." The collected user query terms are transmitted to the feature generation module 30 for query-related feature extraction.

[0050] The competitor data collection module 13 is designed to collect ranking information from competitor companies' recommendation information systems or search platforms. It accesses the competitor's platform regularly (e.g., once an hour) by simulating user requests and automated crawlers, capturing recommendation lists for specific queries, and passing this data to the data preprocessing module 14 for subsequent data verification and model training. For example, by simulating the specific query "AI robot company," the top 20 recommendation lists returned by the top competitor's system are captured for subsequent model training and performance comparison.

[0051] The data preprocessing module 14 is responsible for processing the raw data collected from various data sources, including steps such as cleaning, deduplication, standardization, conversion, and filling in missing values, in order to ensure the quality and consistency of the data and to preliminarily integrate and associate data from different sources. The data preprocessing module 14 can use data cleaning scripts, ETL (Extract, Transform, Load) tools or data pipelines to automatically process the raw data, and transmit the cleaned and preprocessed data to the feature generation module 30 and the graph repository 20, providing high-quality input for building and updating the enterprise graph. For example, the company name will be standardized to unify expressions such as "XX Technology Company" and "XX Technology Co., Ltd." At the same time, the missing registered capital data will be filled with the average value of similar companies in history.

[0052] In summary, the data acquisition module 10 serves as the "data lifeblood" of the entire system. Its various modules work together to ensure the system's continuous and efficient acquisition of diverse, high-quality enterprise data, providing a solid data foundation for subsequent feature engineering, model training, and dynamic intervention. In particular, the competitor data acquisition module 13 is one of the key components of the present invention's dynamic iteration and optimization capabilities.

[0053] Figure 3 This is a schematic diagram of the module of the graph repository of this application. Figure 3 The graph repository 20 includes a knowledge graph construction and update module (Knowledge Graph Construction & Update Module) 21, a graph database management module (Graph Database Management Module) 22, and a knowledge graph indexing and query optimization module (Knowledge Graph Indexing & Query Optimization Module) 23.

[0054] The Graph Construction and Update Module 21 is responsible for receiving preprocessed enterprise-related data from the Data Acquisition Module 10 and constructing it into structured enterprise graph data based on predefined ontologies and architectures. Furthermore, the Graph Construction and Update Module 21 is responsible for continuously updating and maintaining enterprise graph data to ensure its timeliness and accuracy. For example, after a company completes a change in industrial and commercial equity, this module can quickly update the investment relationships in the graph within minutes, ensuring data traceability through version control.

[0055] The workflow of the graph construction and update module 21 is as follows: accurately identify core information such as enterprise entities, key personnel and locations from enterprise-related text data and structured data, and seamlessly connect this information to existing entities in the enterprise graph, or create new entity entries when necessary. Identify the relationships between entities, such as "Company A invests in Company B", "The legal representative of Company C is Mr. D", etc., and add these relationships to the enterprise graph data; carefully integrate information from multiple channels, effectively handle conflicting data and redundant information, and completely eliminate ambiguity, thereby ensuring the absolute consistency and integrity of the enterprise graph data. Record the change history of the enterprise graph data, support backtracking and auditing, and ensure data traceability and credibility; based on the real-time update requirements of the data source, flexibly adopt instant update methods (for example, with the help of message queue technology) or batch updates according to plan to ensure the timeliness and accuracy of the data. Finally, write the constructed and updated enterprise graph data into the graph database management module 22.

[0056] The graph database management module 22 is the core of the graph repository 20, and is responsible for the actual storage and management of the enterprise graph data. The graph database management module 22 uses a special graph database (such as Neo4j, ArangoDB, JanusGraph, etc.) to efficiently store nodes (enterprises, people, products, regions, industries, etc.) and the relationships between them (investment, appointment, supply, procurement, cooperation, address location, business scope, etc.). This structure can efficiently support complex relationship queries and path finding. The preferred solution is to use a graph database with a JanusGraph distributed architecture, which supports write operations of more than 5,000 nodes per second, and shortens the query response time to less than 80 milliseconds by combining a hybrid indexing technology that combines secondary indexes and full-text retrieval. The graph database management module 22 provides a data add, delete, modify, and query (CRUD) interface, supports efficient graph traversal and graph pattern matching operations, thereby greatly optimizing the storage and retrieval performance of large-scale graph data. The enterprise graph data will eventually be provided to the feature generation module 30 for feature calculation, and will also be provided to the ranking model module 40 for graph embedding and dynamic representation learning.

[0057] The graph indexing and query optimization module 23 is responsible for efficiently indexing the graph data and optimizing the query path to improve the speed and efficiency of graph queries, especially when facing complex multi-hop queries. The graph indexing and query optimization module 23 can create indexes for node attributes, relationship types, etc. to speed up data retrieval. It also parses external requests (such as graph query requests from the feature generation module 30 or the sorting model module 40) into query statements executable by the graph database of the graph database management module 22, and optimizes them to reduce query time. The results of frequent queries can also be cached to further improve the response speed. For example, for complex multi-hop queries (such as "C-level suppliers of Company B invested by Company A"), the query results can be quickly returned in a very short time (milliseconds).

[0058] In summary, the Graph Repository 20, as the core storage center for enterprise knowledge, leverages the collaborative operation of its various modules to achieve structured archiving, efficient management, continuous iteration, and rapid retrieval of enterprise knowledge. This provides a solid, flexible, and high-quality knowledge foundation for feature generation and model training. Its efficient operation is crucial for the intelligent recommendations of the entire recommendation system.

[0059] Figure 4 Schematic diagram of the feature generation module of the embodiment of the present application. Figure 4 The feature generation module 30 includes a knowledge graph feature extraction module 31, a text feature extraction module 32, a business attribute feature extraction module 33 and a feature fusion module 34.

[0060] The graph feature extraction module 31 is responsible for retrieving enterprise graph data from the graph repository 20, and then extracting correlation feature information that can quantify the degree of correlation between enterprises and the influence of the enterprise itself in the graph.

[0061] The association feature information of the enterprise graph includes association path weights and node centrality. The graph feature extraction module 31 uses an optimized PageRank algorithm to accurately calculate the association path weights between enterprises. This association path weight is not simply a simple path length; it also takes into account the quality and quantity of nodes in the path, reflecting the closeness of the relationship between enterprises (for example, if enterprise A invests in enterprise B, which invests in enterprise C, and enterprise C then collaborates with enterprise D, then a three-hop association path exists between enterprise A and enterprise D. Using the improved PageRank algorithm, the comprehensive weight of this path can be calculated, reflecting the closeness of the indirect relationship). Furthermore, the graph feature extraction module 31 constructs a dynamic node centrality based on 12 indicators, such as transaction frequency and financing rounds. This node centrality measures the importance and influence of an enterprise within the entire enterprise graph, as well as its ability to connect with other enterprises, dynamically reflecting the enterprise's activity and status in the market (for example, a technology enterprise that frequently participates in bidding and has just completed its Series C financing round will have a much higher node centrality than an enterprise with lower registered capital and long-term inactivity). Furthermore, the enterprise graph feature information also includes graph structural features such as path length, connectivity, and betweenness centrality.

[0062] The text feature extraction module 32 is used to extract text matching feature information from the user query and the text data such as the company name and business description in the company-related data. The text matching feature information includes semantic similarity, longest common subsequence match, and company name entity recognition confidence. The text feature extraction module 32 uses the BERT-based model to calculate the semantic similarity between the user query and the target company name, business description, and other text data, deeply capturing the deeper meaning behind the words. This means that even if the words are different but the concepts are the same, the system can accurately identify them (for example, when a user searches for "smart medical solutions," even if the company name does not directly contain "smart medical," as long as its business description is highly relevant, the BERT model will still give a high similarity score). Furthermore, the text feature extraction module 32 also calculates the longest common subsequence match between the user query and the company name. This is a string matching-based metric used to assess the similarity between the query and the company name at the character level, which is crucial for accurate matching (for example, the match between "XX Innovation Technology" and "XX Innovation Technology Development Co., Ltd." will be very high). The text feature extraction module 32 also calculates the confidence level of the enterprise name entity recognition, which helps to determine whether the enterprise entity is accurately recognized in the user query and the degree of accuracy of the recognition.

[0063] The business attribute feature extraction module 33 uses big data mining technology to extract business attribute feature information reflecting the business's operating status, scale, and activity from relevant enterprise data. Specifically, the business attribute feature extraction module 33 integrates 58 structured indicators, such as registered capital, number of insured persons, and bidding activity. These indicators are obtained by the data acquisition module 10 from databases such as industrial and commercial administration, bidding, and taxation.

[0064] Feature fusion module 34 fuses the feature information from graph feature extraction module 31, text feature extraction module 32, and business attribute feature extraction module 33 to construct a high-dimensional feature vector. Feature fusion module 34 can integrate feature information from different modalities using methods such as concatenation and weighted summation, preferably using a fusion method based on an attention mechanism.

[0065] In recommendation systems, feature engineering is a key step in building user interest in items. The goal of feature engineering is to extract valuable information from raw data and then combine item attributes with user historical behavior to predict user interest in items. For example, feature engineering can employ tools such as bag-of-words models, TFIDF, and clustering algorithms, using automated feature selection algorithms such as SVD or MF to predict users' potential preferences for items. Effective feature engineering can significantly improve the performance and user experience of recommendation systems, enhancing the accuracy, diversity, and interpretability of recommendations. Furthermore, recommendation algorithms based on feature representations perform in-depth analysis and extraction of user and item features, mapping them into a low-dimensional feature space to more accurately capture the underlying relationships between users and items. Such algorithms can fully tap into the valuable information in the data, improving the accuracy and reliability of recommendations.

[0066] Figure 5 Schematic diagram of the module of the sorting model module of the embodiment of the present application. Figure 5 The ranking model module 40 includes a feature processing module 41, a wide model 42, a deep model 43, a model training and optimization module 44, a data fusion module 45, a ranking optimization module 46 and a federated learning collaboration module 47.

[0067] The feature processing module 41 is responsible for receiving the feature vectors output by the feature generation module 30 and the competitor ranking information provided by the data acquisition module 10, further processing these feature vectors to meet the input standards of the ranking model, and performing feature engineering according to the distribution of the processed feature vectors and the competitor ranking information during the model training phase. Among them, the feature processing module 41 can standardize the feature vectors (such as Min-Max Scaling or Z-score Normalization) to eliminate the influence of different feature dimensions and numerical ranges, and can construct cross-features, such as combining industry feature information with regional feature information to reveal the nonlinear relationship between feature information. During the model training phase, the feature processing module 41 will adjust the feature processing strategy according to the statistical characteristics of the training samples (including positive and negative samples, especially negative samples constructed through competitor ranking information).

[0068] The ranking model module 40 uses a wide model 42 and a deep model 43 as ranking models.

[0069] The wide model 42 is primarily responsible for capturing the linear relationship between features and is able to memorize some sparse features. It has a particularly strong memory for sparse feature combinations that have appeared in historical data and verified in competitor ranking information. For example, it learns a strong linear association between the combined features of "Industry A + Region B" and high correlation. The wide model 42 receives feature vectors that have undergone feature processing and learns feature weights using a linear model (e.g., a logistic regression model).

[0070] Deep Model 43 focuses on capturing complex nonlinear relationships between features and learning graph representations, demonstrating strong generalization capabilities. For example, it can capture complex patterns such as "high-growth companies in a specific technology field (business attributes), recent collaborations with well-known research institutions (graph features), and a high match with user query terms (text features)." This Deep Model 43 integrates a Temporal Weighted Graph Attention Network (TW-GAT) and a Multi-Layer Perceptron (MLP).

[0071] The deep model 43 first receives the feature vectors from the feature processing module 41, which have been normalized, discretized, and potentially cross-processed. Then, the model extracts enterprise graph association features, text matching features, and business attribute features.

[0072] The deep model 43 then learns the dynamic representation of the graph using a graph attention mechanism. First, each enterprise node in the enterprise graph's associated feature information is initialized as an embedding vector. Then, the attention mechanism is applied to each node using TW-GAT to calculate the contribution weights of each node's neighboring nodes to itself. This attention mechanism allows the model to dynamically focus on neighbors in the graph that are more important to the current node. To capture the dynamic nature of the graph, TW-GAT uses a time decay factor, similar to the time decay effect in the Hawkes process, to meet the needs of learning temporal graph representations. This means that newer relationships or events receive higher weights, while older relationships gradually lose their influence. For example, a recent investment relationship may better reflect the current activity of a company than an investment relationship from several years ago. Furthermore, TW-GAT not only considers directly connected neighbors but also aggregates information from multi-hop relationships (i.e., relationships connected by multiple intermediate nodes). This enables the model to capture deeper and more complex inter-enterprise relationships. After multi-layer TW-GAT processing, enterprise nodes generate dynamic representation vectors that encode structure, relationships, and temporal features.

[0073] Traditional graph neural network (GAT) nodes aggregate neighbor information and calculate static attention weights. In a dynamic enterprise graph, enterprise relationships (investment, cooperation, transactions) change over time, and their influence also changes. Therefore, in order to capture the dynamic nature of the graph, the deep model 43 introduces a time decay factor e in TW-GAT. -λΔt , its mathematical expression is as follows: Among them, W ij represents the dynamic attention weight of node j relative to node i, exp(.) represents the exponential function (used to convert the attention score into a positive value), h i and h j Represent the embedding vectors of node i and node j respectively, W represents the learnable weight matrix (used for linear transformation of node features), a T represents the learnable attention vector (used to calculate the attention score of the transformed embedding vector), [Wh i ||Wh j ] represents the concatenation of the transformed embedding vectors of node i and node j, LeakyReLU represents the activation function (used to introduce nonlinearity), Represents all neighbor nodes k∈N of node i i Sum the attention scores (used to normalize the attention weights), e -λΔtrepresents the introduced time decay factor, Δt represents the time difference (i.e., the time interval between the time point when the relationship occurs and the current time point (or a certain benchmark time point). The larger Δt is, the older the relationship is). λ represents the decay coefficient (the default value can be, for example, 0.15, which is used to control the speed of time decay. The larger the λ value is, the faster the time decay is and the more the weight of the old relationship decreases).

[0074] By combining traditional attention weights based on structural similarity with a temporal factor, the importance of neighbor node j to node i depends not only on their structural similarity and connection strength, but also on the timeliness of their relationship. By multiplying this factor by time decay, TW-GAT dynamically adjusts the contribution of neighbor nodes when aggregating information, ensuring that the node representations learned by the model reflect the latest dynamics and activity in the enterprise graph. For example, a recently established investment or partnership will have a higher influence weight than a similar relationship established many years ago.

[0075] Reference Figure 6 Multi-hop relationship aggregation means that when a node constructs its own representation, it not only considers its directly connected neighbors (i.e., one-hop neighbors), but also integrates information from two-hop, three-hop, and even more distant neighbors. This feature enables the model to capture deeper and more complex indirect connections between enterprises.

[0076] This process utilizes the multi-layer stacking structure of TW-GAT. In the first layer, each node aggregates information from its direct neighbors (1-hop neighbors) through the self-attention mechanism, calculates the attention weight, and incorporates time decay. In the second layer, each node now aggregates the representation of its 1-hop neighbors after the first layer update. The representation of these 1-hop neighbors already contains the information of their own 1-hop neighbors (that is, the 2-hop neighbors of the original node). Therefore, the second-layer aggregation actually contains the 2-hop information of the original node. Similarly, by stacking 1 layer of TW-GAT, it is possible to aggregate the neighbor information within a 1-hop range, which can be expressed as the following formula: in, represents the new embedding vector learned by node i in the l+1th layer of TW-GAT, σ represents the nonlinear activation function (such as ReLU), ∑ j∈Ni (.) represents the one-hop neighbor node j∈N of the aggregation node i i Information, represents the embedding vector of neighbor node j at layer l, W (l) represents the learnable weight matrix of layer l (used to linearly transform the neighbor embedding vectors), Represents the attention weight of node j to node i (which can be a dynamic attention weight W with a time decay factor ij ), β·READOUT(.) represents the aggregation of information from the two-hop neighbor nodes of node i, represents the embedding vector of neighbor node k in layer l, represents the set of two-hop neighbor nodes of node i (i.e., nodes that are connected to node i by two edges), READOUT(.) represents the readout function (used to aggregate the set of embedding vectors of two-hop neighbor nodes into a single vector. The readout function can be, for example, averaging, max pooling, or summing), and β represents the learnable coefficient (the default value can be, for example, 0.35, which is used to balance the influence of direct associations (one-hop relationships) and two-hop relationships. By adjusting β, the model can control the degree of dependence on more distant information).

[0077] Traditional graph neural networks usually only aggregate information from direct neighboring nodes, but this application introduces a multi-hop relationship aggregation mechanism that allows the model to consider not only the companies directly connected to the node when learning node representation, but also the information of companies indirectly connected through two edges (or more hops). This means that even if two companies do not have a direct cooperation or investment relationship, if they are indirectly connected through a common investor or customer, this indirect association will be captured in the representation of the node. Through the learnable coefficient β, the model can dynamically weigh the importance of these relationships with different hop counts (i.e., different distances), making the learned enterprise embedding vector more comprehensive and robust, and better able to reflect the influence, potential cooperation relationship or industrial chain position of the enterprise in a complex graph.

[0078] Finally, the deep model 43 performs multimodal feature fusion and high-order interaction learning through the multi-layer perceptron, fusing the enterprise dynamic representation vector with the text matching feature information and the business attribute feature information to ultimately form a comprehensive feature vector. The fused comprehensive feature vector is input into the multi-layer perceptron, which is a neural network composed of multiple fully connected layers and nonlinear activation functions (such as ReLU). The multi-layer perceptron uses multiple layers of nonlinear transformations to learn high-order interactions and complex patterns between fused features. For example, it can identify enterprises with specific business attributes, located in specific graph positions, and highly semantically related to the query term, and these enterprises have higher recommendation value. The last layer of the multi-layer perceptron is the output layer (with only one neuron and a Sigmoid activation function). The output layer maps the complexly transformed comprehensive feature vector to a single numerical value. This numerical value is the deep prediction data of the deep model 43, which represents the relevance or ranking score of the current candidate enterprise to the user's query or recommendation context. This score is a comprehensive evaluation obtained through deep learning and complex interactive analysis of multiple modal features (graphs, text, business attributes).

[0079] The model training and optimization module 44 is primarily responsible for managing the ranking model training process, encompassing two-stage training and online incremental learning, with a particular emphasis on iteratively optimizing the model using competitor data. The module receives pre-processed competitor ranking information from the data acquisition module 10, as well as annotated and proofread data from the intervention engine 50. It uses a pairwise ranking loss (e.g., LambdaRank, ListNet) or pointwise loss (e.g., cross-entropy) in conjunction with optimizers such as Adam and Adagrad to update the parameters of the wide and deep models.

[0080] During the first phase of training, the model training and optimization module 44 receives competitor ranking information from the data acquisition module and constructs a training set of "positive samples" (companies deemed relevant and highly ranked by the competitor platform) and "negative samples" (companies deemed irrelevant or ranked low by the competitor platform, or candidate companies that were not selected by the competitor platform) based on this competitor ranking information. The ranking models (broad model 42 and deep model 43) perform preliminary parameter training based on this data, aiming to quickly converge to a baseline model close to the competitor platform's performance.

[0081] During the second stage of fine-tuning, the model training and optimization module 44 uses the labeled and proofread data from the intervention engine 50 for fine-tuning. These data are high-value samples screened out through active learning and accurately labeled to improve the performance of the ranking model in specific scenarios and ensure that it surpasses competing platforms.

[0082] During the ongoing online incremental learning process, model training and optimization module 44 continuously incorporates new data provided by competitor data acquisition module 13 of data acquisition module 10. Through regular incremental training (e.g., weekly), the ranking model can quickly adapt to market dynamics, avoiding full training from scratch, significantly reducing maintenance costs and shortening update cycles.

[0083] The data fusion module 45 fuses the linear prediction data of the wide model 42 and the deep prediction data of the deep model 43 to obtain the final ranking score information. The data fusion module 45 can use weighted summation, stacking, or other ensemble learning methods to combine the memory capacity of the wide model 42 and the generalization ability of the deep model 43 to achieve a better overall prediction effect.

[0084] The ranking optimization module 46 arranges all candidate upstream and downstream enterprises in descending order according to the ranking score information to generate final ranking data, and transmits the ranking data to the intervention engine 50 for further correction and processing.

[0085] The federated learning collaboration module 47 is connected to the wide model 42 and the deep model 43. Within the federated learning framework, it is responsible for securely exchanging and aggregating model parameters with each partner platform, enabling global model training without leaking the original data. This mechanism enables the model to integrate a wider range of industry data, such as graphs and user behavior information from other platforms, thereby indirectly enhancing understanding of competitor behavior while avoiding direct access to sensitive data.

[0086] The federated learning collaboration module 47 encrypts the gradients (or parameter updates) of the local wide model 42 and deep model 43, such as using the Paillier homomorphic encryption algorithm, and then transmits them to the parameter server through the encrypted communication module 60. At the same time, it receives aggregated global model parameters from the parameter server to update the local wide model 42 and deep model 43. The federated learning collaboration module 47 supports asynchronous parameter servers, allowing for node delays of approximately 30%, achieving minute-level model updates (measured 48-second synchronization cycle), and can improve training efficiency by more than 45%.

[0087] In a sample deployment of a federated learning architecture, servers deployed on eight AWS c5.9xlarge high-performance instances are responsible for collecting, aggregating, and distributing encrypted model updates. Local servers at 12 partner institutions serve as participating nodes in the federated learning process, each holding its own unique copy of the enterprise data and models. All transmitted model update data (including gradients and parameter increments) is encrypted using the Paillier homomorphic encryption algorithm (with a 2048-bit key length), ensuring data confidentiality during network transmission and aggregation.

[0088] The entire federated learning training process strictly adheres to established security protocols: First, the server initializes a global ranking model and then securely distributes it to the 12 participating institutions. Each institution uses its own corporate data (for example, corporate graphs and user behavior data within its service area) to train the model locally and calculate local model updates (gradients). The original data never leaves the local area. After local training is completed, each institution performs Paillier homomorphic encryption on the local model update and then uploads the encrypted update to the server. The server directly performs homomorphic addition aggregation on these encrypted gradients without decrypting the encrypted gradients uploaded by each institution. The aggregated encrypted gradients are decrypted (by the private key of the parameter server) and used to update the global model. The updated global model is encrypted again and distributed to all participating institutions before entering the next training iteration cycle.

[0089] In summary, the ranking model module 40 leverages the advantages of federated learning, combining the memory capacity of wide models with the generalization capabilities of deep models. This allows for superior overall prediction results while protecting data privacy and reducing bandwidth consumption. By introducing the model training and optimization module 44, a baseline model is first trained using competitor ranking information. This model is then fine-tuned using labeled and proofread data, and combined with an online incremental learning mechanism, continuous optimization and dynamic adaptation of the model are achieved. Through federated learning technology, the recommendation system can process multi-source data and improve data utilization efficiency while protecting user privacy.

[0090] Figure 7 Schematic diagram of the module of the intervention engine of the embodiment of the present application. Figure 7The intervention engine 50 includes a data reception and preliminary verification module (Data Reception & Initial Validation Module) 51, an intervention operation matching module (Intervention Operation Matching Module) 52, a data correction and adjustment module (Data Correction & Adjustment Module) 53, and a long-tail case library maintenance and feedback module (Long-tail Case Management & Feedback Module) 54.

[0091] The data receiving and preliminary verification module 51 is responsible for receiving the ranking data (including the candidate enterprise list and ranking scores) from the ranking model module 40 and performing preliminary data format verification and integrity checks. Once the data receiving and preliminary verification module 51 confirms that the ranking data meets the processing specifications of the intervention engine 50, it is transmitted to the intervention operation matching module 52.

[0092] The intervention operation matching module 52 determines whether intervention is needed on the current ranking data and which intervention operation to apply based on predefined intervention benchmarks and real-time contextual information (such as user query intent, geographic location, etc.). The intervention operation matching module 52 is connected to the data acquisition module and analyzes the user query term to identify whether there is a specific intent such as regional sensitivity, industry specificity, or multi-entity combination. For example, when the user query term includes "Y city + cloud computing", "Y city" is identified as the regional intent. Furthermore, the intervention operation matching module 52 matches the corresponding intervention operation based on the identified intent and the internally maintained operation database. The intervention operation can be targeted at a specific industry, a specific region, a specific enterprise status (such as "restricted high-consumption enterprises are not recommended"), or a long-tail query mode. The data correction and adjustment module 53 performs corresponding adjustment operations on the ranking data based on the determination result of the intervention operation matching module 52.

[0093] Among them, when a regionally sensitive query is identified, a geographical distance factor will be added to the sorting data. For example, the distance between the candidate enterprise and the user or the specified query location is calculated, and the sorting weight is dynamically adjusted according to the distance to ensure that enterprises in the same city or nearby are given higher priority. When specific abnormal scenarios or low-quality data recorded in the long-tail case library are identified, direct adjustments are made. For example, if a multi-entity combination query (such as "person name + company name") was found to have poor recommendation effects in the past, the data correction and adjustment module 53 will flexibly make adjustments based on the preset intervention strategy, such as removing unnecessary data or introducing high-quality preset data. The long-tail case library covers 32 types of abnormal scenarios, and the bad case resolution rate is 90.2%. In addition, the data correction and adjustment module 53 can also perform correction operations defined by other business rules, such as forcibly excluding certain enterprises that do not meet the requirements, or improving the ranking of specific high-quality enterprises. The corrected and adjusted sorting data will be efficiently transmitted to the long-tail case library maintenance and feedback module 54 and the user interface.

[0094] The long-tail case library maintenance and feedback module 54 is integrated with a long-tail case library. The long-tail case library maintenance and feedback module 54 can maintain and update the long-tail case library according to the corrected and adjusted sorting data, and screen high-value samples through an active learning mechanism and push them to manual experts for annotation to form annotation proofreading data, and feed the annotation proofreading data back to the sorting model module 40 to form a rule-model iterative closed loop.

[0095] The long-tail case library maintenance and feedback module 54 is responsible for storing and managing various known long-tail query patterns, abnormal scenarios, and their corresponding optimal solutions or corrective measures in the long-tail case library. These cases may be constructed through "competitive data anti-crawling" or formed through manual annotation. In addition, the long-tail case library maintenance and feedback module 54 can analyze the model and the effect after intervention, and use active learning technology to screen out samples with a confidence level lower than 0.7. These samples are often boundary cases (hard samples) with low model prediction confidence but significant impact on user experience. These highly valuable samples are then handed over to manual experts for detailed annotation. The annotated and proofread data (such as the revised recommendation list or label) proofread by the manual expert is fed back to the ranking model module 40 as a supervisory signal for subsequent incremental model training or fine-tuning, thereby improving the model's generalization ability in long-tail and complex scenarios. Through active learning, 200+ high-value samples can be screened out every week, and the corresponding model update delay is within 4 hours.

[0096] The intervention process is further explained below through a typical intervention case of the intervention engine 50.

[0097] Case 1: The user queries "Construction Company A". The original ranking model may recommend "Bureau B (City X)" and "C Science and Engineering (City Y)" as highly relevant companies. However, if the user's current location or the query intention clearly points to City Y (for example, the user's IP address is in City Y), the intervention engine 50 will dynamically superimpose the geographical distance factor. After correction, even if the original score of "Bureau B (City X)" is slightly higher, due to the weight adjustment of geographical distance, "C Science and Engineering (City Y)" may be recommended first. According to actual measurements, by implementing a region-sensitive query optimization strategy, the CTR (click-through rate) has been significantly improved. Specifically, the CTR of such queries has increased by 37%.

[0098] Case 2: A user searches for "Logistics Company D." The original ranking might recommend "Logistics Company E (City Y)" and "Logistics Company F (City Z)" side by side. If the platform currently has specific business needs in City Y, or if the user's historical preference points to City Y, the intervention engine 50 will intervene. After this correction, the user conversion rate for "Logistics Company E (City Y)" increased by 28%.

[0099] Case 3: User queries "XX (name) e-commerce company". Traditional enterprise recommendation information systems may mistakenly recommend enterprises with "XX" in their names but unrelated businesses, such as "XX Building Materials Co., Ltd." due to the lack of understanding of the investment relationship between "XX" as a person name entity and "e-commerce company". The error rate of this multi-entity combination query is as high as 42% in traditional enterprise recommendation information systems. In this application, the intervention engine 50 can identify such complex query patterns of "person name + company name" and obtain correction rules from the long-tail case library. The long-tail case library has covered 32 types of abnormal scenarios. Through active learning screening and manual proofreading (for example, screening 200+ high-value boundary samples for manual labeling every week), the system can accurately identify "XX" as a well-known investor and accurately associate it with the well-known e-commerce companies it has invested in, thereby correcting the recommendation results. After correction, the bad case resolution rate of such long-tail queries can be greatly improved to 90.2%.

[0100] Figure 8 Schematic diagram of the encryption communication module of the embodiment of the present application. Figure 8 The encryption communication module 60 includes an encryption and decryption module (Encryption & Decryption Module) 61, a secure transmission module (Secure Transmission Module) 62 and a key management module (Key Management Module) 63.

[0101] The encryption and decryption module 61 is responsible for encrypting the model update information (such as gradients and model parameter increments) of the wide model 42 and deep model 43 in the sorting model module 40, ensuring that this information cannot be restored to the original data by eavesdroppers during transmission. In addition, the encryption and decryption module 61 is also responsible for decrypting the received encrypted information.

[0102] The encryption and decryption module 61 can use the Paillier homomorphic encryption algorithm for encryption and decryption. The Paillier homomorphic encryption algorithm is a semi-homomorphic encryption method that allows addition operations to be performed on encrypted data without decryption. When the ciphertext is added, the decrypted result is consistent with the result of performing the same operation on the plaintext. This is because the Paillier scheme meets the standard security definition of encryption schemes, namely semantic security. This process is particularly important in the framework of federated learning because it allows the central server to directly sum the encrypted gradients without decrypting the gradients of each client, thereby protecting data privacy.

[0103] During the encryption process, the encryption and decryption module 61 obtains the model update information to be encrypted from the federated learning collaboration module 47 of the ranking model module 40 and performs encryption processing. During the decryption process, the encryption and decryption module 61 receives the encrypted data from the secure transmission module 62, decrypts it, and obtains the aggregated plaintext gradient to update the ranking model.

[0104] The secure transmission module 62 is responsible for establishing a secure channel between the enterprise recommendation information system and the server, ensuring encrypted transmission of model updates to prevent tampering and theft. The secure transmission module 62 receives encrypted model update information from the encryption and decryption module 61 and uploads it to the server. It also receives encrypted data from the server and transmits it to the encryption and decryption module 61.

[0105] The secure transmission module 62 can use the Transport Layer Security (TLS / SSL) protocol to establish an encrypted communication tunnel, thereby ensuring the confidentiality (preventing eavesdropping), integrity (preventing tampering), and authentication (verifying the identities of both communicating parties) of data during network transmission. The enterprise recommendation information system and the server will perform identity authentication before communication to prevent unauthorized entities from participating in the federated learning process.

[0106] The key management module 63 connects the encryption and decryption module 61 and the secure transmission module 62, and is responsible for the key management of Paillier homomorphic encryption and the management of TLS / SSL digital certificates. It is the core of the encryption system security.

[0107] In general, under the federated learning architecture, Figure 8 and Figure 9The encrypted communication module 60 works collaboratively to achieve secure, private exchange and aggregation of model update information. This process involves the encryption and decryption module 61, which uses the Paillier homomorphic encryption algorithm to transform data, the secure transmission module 62 (for example, through the TLS / SSL protocol) to ensure the security of data during transmission, and the key management module 63, which is responsible for the secure lifecycle management of keys. This enables each collaborative platform to jointly contribute computing power without directly exposing sensitive raw data, improving the performance of shared models, thereby effectively solving the pain point of data privacy protection.

[0108] The implementation principle of an enterprise recommendation information system in an embodiment of the present application is as follows: the ranking information of competing enterprises is obtained through the data acquisition module 10, the feature generation module 30 constructs a multi-dimensional feature vector, the ranking model module 40 adopts a wide and deep (Wide & Deep) model for ranking, and introduces a temporal weighted graph attention network (TW-GAT) and the intervention engine 50 for post-processing. This solution comprehensively solves the problems of traditional recommendation systems relying on fixed rules, single feature dimensions, inability to capture temporal dynamics, and weak generalization capabilities. By combining the wide and deep model with TW-GAT, multimodal features are effectively integrated and the dynamic timeliness of enterprise relationships is captured, solving the problems of insufficient feature utilization and dynamic update delays; through the intervention engine, regionally sensitive and long-tail query scenarios are accurately processed, significantly improving the bad case resolution rate and user experience. Ultimately, the system achieved a 14.8% improvement in DCG in ranking results, reduced maintenance costs by 87.5%, and significantly improved generalization capabilities.

[0109] The embodiment of the present application also discloses a method for recommending an enterprise.

[0110] Figure 10 This is a flow chart of the enterprise recommendation method of the embodiment of the present application. Figure 10 , the method comprises the following steps: S1. Obtain user query terms, enterprise-related data, and competitor ranking information.

[0111] S2. Generate enterprise graph data based on the enterprise-related data.

[0112] S3. Generate a feature vector of the target enterprise and the corresponding upstream and downstream enterprises based on the enterprise graph data, the user query terms and the enterprise-related data. The feature vector includes enterprise graph association feature information, text matching feature information and business attribute feature information.

[0113] S4. Train the sorting model.

[0114] S5. Use the ranking model to score and rank the upstream and downstream enterprises based on the feature vector to generate ranking data, the ranking model includes a wide model and a deep model, wherein the wide model generates linear prediction data, the deep model generates deep prediction data, and the linear prediction data is fused with the deep prediction data to generate the ranking data; the deep model includes a time-series weighted graph attention network, and the time-series weighted graph attention network introduces a time decay factor when calculating the attention weight of the neighboring nodes of the enterprise node in the associated feature information of the enterprise graph, so that the attention weight representing the enterprise relationship decays as the time difference increases.

[0115] S6. Determine whether the sorting data contains regionally sensitive information.

[0116] S7. When it is determined that the ranking data contains region-sensitive information, a geographical distance factor is superimposed on the ranking data to form enterprise recommendation information.

[0117] S8. Determine whether the sorted data is long-tail data.

[0118] S9. When it is determined that the sorted data is long-tail data, an intervention operation matching from a long-tail case library is performed on the sorted data to form enterprise recommendation information.

[0119] The following describes each step of the method in detail.

[0120] Refer again Figure 2 In step S1, the user's search query, company-related data, and competitor ranking information are obtained. This information comes from industrial and commercial systems, bidding websites, patent databases, and public opinion platforms, and covers the company's registration information, business scope, equity structure, historical transaction records, innovation achievements, and media reputation. Simultaneously, the system captures the user's search query in real time and collects competitor ranking information from the competitor recommendation information system in real time or near real time for subsequent model training and effect evaluation.

[0121] Refer again Figure 3 In step S2, enterprise graph data is generated based on the enterprise-related data. The pre-processed enterprise-related data is constructed into structured enterprise graph data and continuously updated and maintained. This process includes entity recognition and linking, relationship extraction, and knowledge fusion, ensuring the timeliness and accuracy of the enterprise graph data.

[0122] Refer again Figure 4In step S3, a feature vector of the target enterprise and its corresponding upstream and downstream enterprises is generated based on the enterprise graph data, the user query term, and the enterprise-related data. The feature vector includes enterprise graph-related feature information, text matching feature information, and business attribute feature information. The enterprise graph data is processed to extract enterprise graph-related feature information, and the text data in the user query term and the enterprise-related data is processed to extract text matching feature information. By structuring the enterprise-related data, 58 key business attribute indicators can be extracted, such as registered capital, number of insured persons, and bidding activity. These indicators help to comprehensively measure the operating status of the enterprise.

[0123] Refer again Figure 5 In step S4, a ranking model is trained. In the first stage, the ranking model is trained based on the ranking information of the competing companies, and in the second stage, the ranking model is adjusted based on the annotated proofreading data. In step S5, the ranking model is used to score and rank the upstream and downstream companies based on the feature vectors to generate ranking data. The ranking model includes a wide model and a deep model. The wide model generates linear prediction data, and the deep model generates deep prediction data. The linear prediction data and the deep prediction data are fused to generate the ranking data.

[0124] Among them, the sorting model is updated using the federated learning framework. During this process, secure communication with the server is achieved through the federated learning architecture, and the Paillier homomorphic encryption algorithm is used to encrypt the update information of the sorting model and decrypt the encrypted data from the server.

[0125] This deep model combines a time-series weighted graph attention network and a multi-layer perceptron. The process of generating deep prediction data is as follows: First, the time-series weighted graph attention network learns the graph dynamic representation of enterprise nodes in the enterprise graph's associated feature information, generating enterprise dynamic representation vectors. These vectors are then combined with text matching feature information and business attribute feature information to form a comprehensive feature vector. Finally, the multi-layer perceptron maps the comprehensive feature vector into deep prediction data.

[0126] The time-series weighted graph attention network performs graph dynamic representation learning on the enterprise nodes in the enterprise graph associated feature information to generate an enterprise dynamic representation vector, wherein a time decay factor is introduced when calculating the attention weight of the neighboring nodes of the enterprise node, which is expressed as the following formula: Among them, W ij represents the dynamic attention weight of node j relative to node i, exp(.) represents the exponential function, h i and h jRepresents the embedding vectors of node i and node j respectively, W represents the learnable weight matrix, a T represents the learnable attention vector, [Wh i ||Wh j ] represents the concatenation of the transformed embedding vectors of node i and node j, LeakyReLU represents the activation function, Represents all neighbor nodes k∈N of node i i Sum the attention scores, e -λΔt represents the time attenuation factor, Δt represents the time difference, and λ represents the attenuation coefficient; Furthermore, multi-hop relationship aggregation is performed on the multi-hop neighbor nodes of the enterprise node, which is expressed as the following formula: in, represents the embedding vector learned by node i in the l+1th layer of TW-GAT, σ represents the nonlinear activation function, ∑ j∈Ni (.) represents the one-hop neighbor node j∈N of the aggregation node i i Information, represents the embedding vector of neighbor node j at layer l, W (l) represents the learnable weight matrix of layer l, represents the attention weight of node j to node i, β·READOUT(.) represents the aggregation of information from the two-hop neighbor nodes of node i, represents the embedding vector of neighbor node k in layer l, represents the set of two-hop neighbor nodes of node i, READOUT(.) represents the readout function, and β represents the learnable coefficient.

[0127] Refer again Figure 7 In step S6, it is determined whether the sorted data contains regionally sensitive information. In step S7, if it is determined that the sorted data contains regionally sensitive information, a geographic distance factor is superimposed on the sorted data to form enterprise recommendation information. In step S8, it is determined whether the sorted data is long-tail data. In step S9, if it is determined that the sorted data is long-tail data, an intervention operation matching from a long-tail case library is performed on the sorted data to form enterprise recommendation information.

[0128] The implementation principle of an enterprise recommendation method in an embodiment of the present application is: by obtaining competitive information, constructing a vector that integrates multimodal features, using wide depth (Wide & Deep) and temporal weighted graph attention network (TW-GAT) for training and sorting, and combining a dynamic intervention mechanism to correct the results. This method comprehensively overcomes a series of technical bottlenecks existing in traditional recommendation methods, such as reliance on static rules, low sorting efficiency, high maintenance costs, and insufficient generalization capabilities. By relying on the use of multi-dimensional feature engineering and deep learning models, this method significantly improves the sorting effect (the DCG value jumps to 3.33). At the same time, through the introduction of a dynamic intervention mechanism, it greatly enhances the adaptability to complex scenarios (the bad case resolution rate soars to 90%), thereby achieving a double leap in recommendation quality and maintenance efficiency.

[0129] The above are all preferred embodiments of the present application, and are not intended to limit the scope of protection of the present application. Therefore, any equivalent changes made based on the structure, shape, and principle of the present application should be included in the scope of protection of the present application.

Claims

1. An enterprise recommendation information system, characterized in that: include: A data acquisition module (10) is used to obtain user query terms, enterprise-related data and competitor enterprise ranking information; A graph repository (20), connected to the data acquisition module (10), for generating enterprise graph data based on the enterprise-related data; A feature generation module (30) is connected to the data acquisition module (10) and the graph storage library (20), and is used to generate a feature vector of the target enterprise and the corresponding upstream and downstream enterprises based on the enterprise graph data, the user query term and the enterprise-related data, wherein the feature vector includes enterprise graph association feature information, text matching feature information and business attribute feature information; A ranking model module (40) is connected to the data acquisition module (10) and the feature generation module (30). The ranking model module (40) has a ranking model for training the ranking model and using the trained ranking model to score and rank the upstream and downstream enterprises based on the feature vector to generate ranking data; the ranking model includes a wide model and a deep model, the ranking model module (40) generates linear prediction data through the wide model, generates deep prediction data through the deep model, and fuses the linear prediction data with the deep prediction data to generate the ranking data; the deep model includes a time-series weighted graph attention network, and the time-series weighted graph attention network introduces a time decay factor when calculating the attention weight of the neighboring nodes of the enterprise node in the enterprise graph association feature information, so that the attention weight representing the enterprise relationship decays as the time difference increases; An intervention engine (50) is connected to the ranking model module and is used to determine whether the ranking data contains regionally sensitive information. When it is determined that the regionally sensitive information exists, a geographic distance factor is superimposed on the ranking data to generate enterprise recommendation information. It also determines whether the ranking data is long-tail data. When it is determined that the ranking data is long-tail data, an intervention operation matching from a long-tail case library is performed on the ranking data to form enterprise recommendation information.

2. The enterprise recommendation information system according to claim 1, characterized in that: The deep model also includes a multi-layer perceptron, wherein the deep model performs graph dynamic representation learning on the enterprise nodes in the enterprise graph associated feature information through the temporal weighted graph attention network to generate an enterprise dynamic representation vector, and fuses the enterprise dynamic representation vector with the text matching feature information and the business attribute feature information to form a comprehensive feature vector, and also maps the comprehensive feature vector to the deep prediction data through the multi-layer perceptron.

3. The enterprise recommendation information system according to claim 2, characterized in that: The deep model performs graph dynamic representation learning on the enterprise nodes in the enterprise graph associated feature information through the time-series weighted graph attention network to generate an enterprise dynamic representation vector, wherein a time decay factor is introduced when calculating the attention weight of the neighboring nodes of the enterprise node, which is expressed as the following formula: Among them, W ij represents the dynamic attention weight of node j relative to node i, exp(.) represents the exponential function, h i and h j Represents the embedding vectors of node i and node j respectively, W represents the learnable weight matrix, a T represents the learnable attention vector, [Wh i ||Wh j ] represents the concatenation of the transformed embedding vectors of node i and node j, LeakyReLU represents the activation function, Represents all neighbor nodes k∈N of node i i Sum the attention scores, e -λΔt represents the time attenuation factor, Δt represents the time difference, and λ represents the attenuation coefficient.

4. The enterprise recommendation information system according to claim 2, characterized in that: The deep model performs graph dynamic representation learning on the enterprise nodes in the enterprise graph associated feature information through the time-series weighted graph attention network to generate an enterprise dynamic representation vector, wherein multi-hop relationship aggregation is performed on the multi-hop neighbor nodes of the enterprise node, which is expressed as the following formula: in, represents the embedding vector learned by node i in the l+1th layer of TW-GAT, σ represents the nonlinear activation function, Represents the one-hop neighbor node j∈N of the aggregation node i i Information, represents the embedding vector of neighbor node j at layer l, W (l) represents the learnable weight matrix of layer l, represents the attention weight of node j to node i, β·READOUT(.) represents the aggregation of information from the two-hop neighbor nodes of node i, represents the embedding vector of neighbor node k in layer l, represents the set of two-hop neighbor nodes of node i, READOUT(.) represents the readout function, and β represents the learnable coefficient.

5. The enterprise recommendation information system according to claim 1, characterized in that: The intervention engine (50) further annotates the enterprise recommendation information to generate annotated proofreading data; The ranking model module (40) performs first-stage training and second-stage training on the ranking model, and updates the ranking model through a federated learning framework. During the first-stage training, the ranking model is trained based on the ranking information of the competing enterprises, and during the second-stage training, the ranking model is adjusted based on the annotated proofreading data fed back from the intervention engine.

6. The enterprise recommendation information system according to claim 1, characterized in that: Also includes: An encryption communication module (60) is connected to the sorting model module (40) and is used to encrypt communication between the sorting model module (40) and the server under a federated learning architecture, wherein the encryption communication module (60) uses a Paillier homomorphic encryption algorithm to encrypt model update information of the sorting model and decrypt encrypted data from the server.

7. A method for recommending an enterprise, characterized in that: The following steps are involved: S1. Obtain user query terms, enterprise-related data, and competitor ranking information; S2. generating enterprise graph data based on the enterprise-related data; S3. Generate a feature vector for the target enterprise and its corresponding upstream and downstream enterprises based on the enterprise graph data, the user query term, and the enterprise-related data. The feature vector includes enterprise graph association feature information, text matching feature information, and business attribute feature information. S4, training ranking model; S5. Using the ranking model to score and rank the upstream and downstream enterprises based on the feature vectors to generate ranking data, the ranking model includes a wide model and a deep model, wherein the wide model generates linear prediction data, the deep model generates deep prediction data, and the linear prediction data and the deep prediction data are fused to generate the ranking data; the deep model includes a time-series weighted graph attention network, which introduces a time decay factor when calculating the attention weights of neighbor nodes of enterprise nodes in the enterprise graph association feature information, so that the attention weights representing enterprise relationships decay as the time difference increases; S6. Determine whether the sorting data contains any regionally sensitive information. S7. When it is determined that the ranking data contains regionally sensitive information, superimposing a geographic distance factor on the ranking data to form enterprise recommendation information; S8. Determine whether the sorted data is long-tail data; S9. When it is determined that the sorted data is long-tail data, an intervention operation matching from a long-tail case library is performed on the sorted data to form enterprise recommendation information.

8. The enterprise recommendation method according to claim 7, characterized in that: The deep model in step S6 also includes a multi-layer perceptron, wherein generating deep prediction data through the deep model includes: performing graph dynamic representation learning on the enterprise nodes in the enterprise graph associated feature information by the temporal weighted graph attention network to generate an enterprise dynamic representation vector; fusing the enterprise dynamic representation vector with the text matching feature information and the business attribute feature information to form a comprehensive feature vector; and mapping the comprehensive feature vector to the deep prediction data by the multi-layer perceptron.

9. The enterprise recommendation method according to claim 8, characterized in that: The time-series weighted graph attention network performs graph dynamic representation learning on the enterprise nodes in the enterprise graph associated feature information to generate an enterprise dynamic representation vector, wherein a time decay factor is introduced when calculating the attention weight of the neighboring nodes of the enterprise node, which is expressed as the following formula: Among them, W ij represents the dynamic attention weight of node j relative to node i, exp(.) represents the exponential function, h i and h j Represents the embedding vectors of node i and node j respectively, W represents the learnable weight matrix, a T represents the learnable attention vector, [Wh i ||Wh j ] represents the concatenation of the transformed embedding vectors of node i and node j, LeakyReLU represents the activation function, Represents all neighbor nodes k∈N of node i i Sum the attention scores, e -λΔt represents the time attenuation factor, Δt represents the time difference, and λ represents the attenuation coefficient; Furthermore, multi-hop relationship aggregation is performed on the multi-hop neighbor nodes of the enterprise node, which is expressed as the following formula: in, represents the embedding vector learned by node i in the l+1th layer of TW-GAT, σ represents the nonlinear activation function, Represents the one-hop neighbor node j∈N of the aggregation node i i Information, represents the embedding vector of neighbor node j at layer l, W (l) represents the learnable weight matrix of layer l, represents the attention weight of node j to node i, β·READOUT(.) represents the aggregation of information from the two-hop neighbor nodes of node i, represents the embedding vector of neighbor node k in layer l, represents the set of two-hop neighbor nodes of node i, READOUT(.) represents the readout function, and β represents the learnable coefficient.

10. The enterprise recommendation method according to claim 7, characterized in that: Step S4 includes performing first-stage training and second-stage training on the ranking model, and updating the ranking model through a federated learning framework. During the first-stage training, the ranking model is trained based on the ranking information of the competing enterprises, and during the second-stage training, the ranking model is adjusted based on the annotation proofreading data generated by the annotation of the enterprise recommendation information. In this case, communication with the server is encrypted under the federated learning framework, the model update information of the ranking model is encrypted using the Paillier homomorphic encryption algorithm, and the encrypted data from the server is decrypted.

Images