Data matching service method for different stages of enterprise based on AI artificial intelligence

Abstract

The invention discloses an enterprise different-stage data matching service method based on AI artificial intelligence. The method comprises the following steps: S1, collecting operation data of an enterprise; s2, preprocessing the operation data; s3, extracting a time characteristic index sequence, and constructing an enterprise development stage evolution path; s4, constructing an enterprise development stage identification model, and generating an enterprise development stage label and a matching trend vector; s5, taking the enterprise development stage label and the matching trend vector as query conditions, and retrieving candidate service resource items; and S6, calculating a basic matching score by using the service resource matching model, performing directional adjustment, and outputting a recommended service resource list. According to the method, AI modeling and a comparative learning mechanism are fused, enterprise stage identification and trend-driven recommendation are realized, and the method has the advantages of high intelligence, accurate matching and high adaptability.

Description

Technical Field

[0001] The present invention relates to the field of enterprise service recommendation technology, and in particular to an AI-based data matching service method for enterprises at different stages. Background Art

[0002] With the development of the digital economy and intelligent industries, more and more enterprise service platforms and SaaS-type enterprise operation systems have begun to explore the use of artificial intelligence technology to model enterprise development dynamics and automatically recommend corresponding policies, services, products and resources based on the stage of the enterprise. However, the systems currently on the market for enterprise development stage identification and intelligent resource recommendation still generally have problems such as single modeling methods, one-sided data utilization, static recommendation logic, and strong reliance on lifecycle labels. These problems make it difficult to meet the dual needs of enterprises for accurate stage identification and intelligent service matching in actual operations.

[0003] In traditional enterprise service management systems, enterprise life cycle stages are mostly classified based on manual labels or predefined rules. For example, an enterprise is considered to be in the start-up phase after registration, in the growth phase after obtaining the first round of financing, and in the mature phase after the number of employees or revenue exceeds a certain threshold. This approach relies heavily on single threshold judgments of individual fields in structured data (such as registration years, revenue scale, etc.), ignoring the complex evolutionary path of enterprise development and lacking in-depth analysis of unstructured data, such as corporate news, policy interaction records, industry comments, and user feedback. This stage classification method not only has problems with poor timeliness and highly subjective labels, but also relies heavily on manually defined life cycle boundaries, making it difficult to adapt to the diversity and uncertainty of enterprise development rhythms.

[0004] At the same time, most existing recommendation systems are based on static matching rules or shallow collaborative filtering, and often only consider the company's current labels or industry information for resource matching, and lack the ability to predict the company's future development trends. For example, if the current company is labeled as "growth stage", the recommendation system will call the "growth stage applicable services" template for recommendation, without considering whether the company is in the transformation or rapid leap stage, nor whether it is about to enter a recession or strategic contraction stage. The lack of trend perception makes it impossible for the recommendation system to proactively guide the company to access resources that are in line with the future evolution path, thereby reducing the timeliness and accuracy of the recommendation service. In addition, many systems still use structured data as the main source of recommendation input, and have not effectively modeled and integrated unstructured data such as text announcements, interactive comments, external industry documents, etc., resulting in information source fragmentation and a single modeling dimension.

[0005] At the AI ​​model level, some studies have attempted to use deep learning methods to model corporate behavior, such as LSTM modeling based on time series data and corporate association analysis based on graph networks. However, most of these models are aimed at goals such as risk assessment and financing forecasting, and lack a deep coupling modeling framework for enterprise service adaptation. At the same time, the few models that introduce semantic vectors or Transformer structures are mostly used on the consumer side (such as product recommendations, user portraits, etc.), and have not yet been systematically applied to the complex task scenarios of enterprise development stage identification and trend-guided recommendations.

[0006] Furthermore, existing technologies typically use static labels or fixed stage divisions to model different enterprise stages, lacking the ability to express dynamic evolution. For example, even when time series modeling is employed, the output stage states are still treated as standard classification results, without being integrated with the enterprise service recommendation mechanism. Furthermore, no consideration is given to constructing stage semantic representations or trend vectors to enhance model learning capabilities through the continuous sequence of changes in enterprise stages. This "label-driven" lifecycle analysis approach struggles to support multi-stage behavioral change modeling and trend prediction.

[0007] For modeling and matching service resources, existing systems mostly rely on content retrieval or rule-based matching, failing to align service vectors with the semantic space of enterprise behavior. This lack of a unified representation space prevents end-to-end semantic alignment between enterprise characteristics and service resources, forcing recommendations to be made based on manual rules or empirical templates. Furthermore, existing models fail to incorporate historical matching behavior as feedback signals into model optimization, lacking reinforcement learning or comparative learning mechanisms, making it difficult to achieve dynamic adaptation between resources and enterprises.

[0008] Therefore, how to design an end-to-end enterprise service matching method that integrates structured and unstructured enterprise data, combines stage evolution modeling with service trend perception, and optimizes service matching accuracy based on joint training has become an important technical challenge facing current enterprise intelligent service platforms. Summary of the Invention

[0009] One purpose of the present invention is to propose a data matching service method for enterprises at different stages based on AI artificial intelligence. The present invention integrates AI deep learning, stage evolution modeling and joint comparative learning mechanism to construct an enterprise development stage identification and service resource matching system that integrates structured and unstructured data. It can achieve accurate identification of the current stage of the enterprise and intelligent prediction of future trends, and generate forward-looking and personalized service recommendation plans. It has the advantages of strong intelligence, accurate matching, high trend adaptability and a wide range of application scenarios.

[0010] According to an embodiment of the present invention, a data matching service method for enterprises at different stages based on AI artificial intelligence includes the following steps:

[0011] S1. Collect the business data of the enterprise;

[0012] S2. Preprocess the business data to generate structured feature vectors and semantic feature vectors;

[0013] S3. Extract the time characteristic indicator sequence from the pre-processed structured data and construct the evolution path of the enterprise development stage;

[0014] S4. Input the structural feature vector, semantic feature vector, and enterprise development stage evolution path into the enterprise development stage recognition model to generate a fusion representation vector, an enterprise development stage label, and a matching trend vector;

[0015] S5. Using the enterprise development stage label and the matching trend vector as query conditions, searching for candidate service resource items in the service resource database and generating a service resource feature vector;

[0016] S6. Input the fusion representation vector and the service resource feature vector into the service resource matching model. The service resource matching model uses a matching offset mechanism and combines the fusion representation vector to calculate the basic matching score. The matching score is directionally adjusted based on the basic matching trend vector, and a list of recommended service resources is output.

[0017] Optionally, the business data includes structured data and unstructured data.

[0018] Optionally, the structured data includes financial data, human resources data and sales data, and the unstructured data includes corporate web page text, policy document text, industry news text and user evaluation text.

[0019] Optionally, the preprocessing of the structured data includes missing value filling, standardization, normalization and structured feature coding, the preprocessing of the unstructured data includes text cleaning, word segmentation and Transformer coding, and the structured feature coding uses a multi-layer perceptron.

[0020] Optionally, the enterprise development stage evolution path includes a stage path representation sequence, a jump indicator sequence and a path direction vector.

[0021] Optionally, the S3 specifically includes:

[0022] S31. Selecting time series indicators related to enterprise development dynamics from the pre-processed structured data to form a feature set;

[0023] S32, constructing the feature set into a time series matrix;

[0024] S33. Perform a stationary test on each indicator sequence in the time series matrix. If the ADF test statistic of the indicator sequence is greater than the stationary significance threshold, perform differential processing:

[0025]

[0026] in, represents the difference value of the jth indicator at time step t, represents the original value of the jth indicator at time step t, represents the original value of the jth indicator at time step t-1, and N represents the length of the original time series;

[0027] S34. Apply the sliding window mechanism to the stabilized time series matrix, set the window length and step size, and generate a set of stage window sequences:

[0028]

[0029] in, Represents the set of stage window sequences, W i represents the i-th sliding window subsequence, and q represents the number of stage windows;

[0030] S35. Input each sliding window subsequence into the GRU network, extract the stage representation vector, and obtain the stage path representation sequence:

[0031] ε={e1,e2,...,e q};

[0032] Among them, ε represents the complete stage path representation sequence of the enterprise, e q The phase representation vector representing the qth phase window;

[0033] S36. Construct a stage jump indicator sequence based on the stage path representation sequence:

[0034] φ i =||e i -e i-1 ||2;

[0035] Among them, φ i represents the Euclidean distance between the vector of the i-th stage and the vector of the previous stage, measuring the degree of jump in the development status of the enterprise, ||·||2 represents the L2 norm, e i represents the phase representation vector of the i-th phase window, e i-1 The phase representation vector representing the i-1th phase window;

[0036] S37. Calculate the path direction vector:

[0037]

[0038] in, It represents the path direction vector of the overall stage of the enterprise, reflecting the development trend of the enterprise, q represents the number of stage windows, e i+1 The phase representation vector representing the i+1th phase window.

[0039] Optionally, the S4 specifically includes:

[0040] S41, obtaining a structural feature vector, a semantic feature vector, a stage path sequence, a jump indicator sequence, and a path direction vector;

[0041] S42, inputting the structured feature vector into a structured feature extraction layer to obtain a structured deep representation vector, wherein the structured feature extraction layer adopts a two-layer Transformer network;

[0042] Inputting the semantic feature vector into the semantic feature extraction layer to obtain a semantic deep representation vector, wherein the semantic feature extraction layer is based on the pre-trained BERT model;

[0043] S43. Construct a stage enhancement sequence based on the stage path sequence, jump index sequence and path direction vector:

[0044]

[0045] Among them, e′ i represents the stage enhancement vector in the stage enhancement sequence, e i represents the phase representation vector of the i-th phase window, φ represents the phase jump indicator sequence, represents the path direction vector;

[0046] S44. Enhance the input vector for each stage, construct positive sample vectors and negative sample vectors, and select the positive matching service vectors and historical negative matching service vectors matched by the enterprise in the current stage from the service matching log. The historical negative matching service vectors represent services that the enterprise encountered but did not adopt in the stage, or services that are obviously not suitable for the stage.

[0047] S45. Establish a joint contrastive learning loss function and train a stage encoding network. The stage encoding network adopts a bidirectional GRU network:

[0048]

[0049] in, represents the joint contrastive learning loss function, L e represents the service matching guided loss function, sim(·) represents the cosine similarity, τ represents the temperature coefficient, λ represents the matching loss weighting coefficient, K represents the number of negative samples, s i Indicates a positive matching service vector, Represents the historical negative matching service vector, e′ i represents the stage enhancement vector in the stage enhancement sequence, represents the positive sample vector, Represents a negative sample vector, exp represents a natural exponential function;

[0050] S46, inputting the stage enhancement sequence into the trained stage encoding network, and outputting the stage evolution representation vector;

[0051] S47. Input the structured deep representation vector, semantic deep representation vector, and stage evolution representation vector into the fusion module to calculate the attention weights of the structural and semantic features:

[0052]

[0053] Among them, α s and α t Represent the attention weights of structural and semantic features respectively, W represents the training fusion matrix, V s represents the structured deep representation vector, V t Represents the semantic deep representation vector, V p represents the phase evolution representation vector, and T represents the transpose operation;

[0054] S48. Based on the attention weight, the structural and semantic features are fused to obtain the fused representation vector:

[0055] V f =α s ·V s +α t ·V t ;

[0056] Among them, V f represents the fused representation vector;

[0057] S49. Use the fused representation vector to make predictions and output the enterprise development stage label and matching trend vector:

[0058] y=argmax(Softmax(W c V f +b c ));

[0059]

[0060] Among them, y represents the enterprise development stage label, Represents the matching trend vector, W c and b c Represents the classification parameter, W t and b tIt represents the parameters of trend prediction, argmax represents the variable value when the function value is maximized, and Softmax represents layer normalization.

[0061] Optionally, the enterprise development stage identification model includes:

[0062] Feature extraction layer, extracting deep expressions of structured feature vectors and semantic feature vectors;

[0063] The stage evolution modeling layer generates the stage evolution representation vector based on the evolution path of the enterprise development stage;

[0064] The stage feature fusion layer uses the stage evolution representation vector as the attention query vector to guide the dynamic weighted fusion of the deep expression of the structural feature vector and the semantic feature vector, and outputs the fused representation vector;

[0065] The stage identification and trend prediction layer jointly models the fusion representation vector and outputs the enterprise development stage label and the matching trend vector. The matching trend vector represents the future evolution direction of the enterprise stage.

[0066] Optionally, the S5 specifically includes:

[0067] S51. Obtain an enterprise development stage label, a matching trend vector, and a fusion representation vector. The enterprise development stage label indicates the current life cycle stage of the enterprise, the matching trend vector indicates the direction of the enterprise's evolution from the current stage to the future stage, and the fusion representation vector indicates the comprehensive semantic representation of the enterprise at the current stage.

[0068] S52. Retrieve the service resource feature vectors of the corresponding stage from the service resource database according to the enterprise development stage label to construct a candidate set. The service resource database records the service resource feature vectors corresponding to each development stage.

[0069] S53. For each service resource feature vector in the candidate set, calculate a directional consistency measure with the matching trend vector, wherein the consistency measure uses the L2 norm;

[0070] S54. Set a consistency threshold, retain only service resource feature vectors whose direction consistency metric is greater than the consistency threshold, and construct a set of trend-related candidate service resources;

[0071] S55. Calculate semantic similarity between the fusion representation vector and each service resource feature vector in the candidate set using cosine similarity to obtain a matching score.

[0072] Optionally, the S6 specifically includes:

[0073] S61, obtaining a fusion representation vector, a matching trend vector, and a service resource feature vector;

[0074] S62. Build a service resource matching model, receive the fusion representation vector and the service resource feature vector as input, calculate the basic matching score, and introduce the matching trend vector. Adjust the basic matching score based on the direction information of the matching trend vector to guide the matching towards the direction of enterprise evolution, and generate a final matching score after trend correction:

[0075]

[0076] Among them, V f represents the fusion representation vector, s j Represents the jth service resource feature vector, score j represents the final matching score under the combined effect of fusion semantics and trend offset, γ represents the trend offset control coefficient, Represents the matching trend vector;

[0077] S63. Arrange all service resource feature vectors in descending order according to the final matching scores, and construct a sorted list of recommended service resources;

[0078] S64. Output the sorted recommended service resource list as a target recommendation list for the enterprise at the current stage and development trend, and provide it to the service platform or user end for display, calling or access to the service processing flow.

[0079] The beneficial effects of the present invention are:

[0080] First of all, the present invention provides a data matching service method for enterprises at different stages based on AI artificial intelligence. It addresses the problems in the existing technology such as the reliance on manual labels for enterprise life cycle identification, the separation of service recommendation models, the separation of structured and unstructured information processing, and the lack of trend prediction capabilities. It constructs a unified, end-to-end enterprise development stage perception and resource recommendation framework with significant technical improvements and practical application value. The method integrates the enterprise's structured and unstructured data and adopts a deep feature extraction strategy based on Transformer and pre-trained models, which significantly enhances the ability to understand the enterprise's multi-source business data and solves the problem of low utilization of unstructured text information in traditional systems. At the same time, it introduces sliding windows and GRU networks to model time series indicators, and combines differential smoothing processing to effectively extract the evolutionary characteristics of the enterprise at different time stages, so that the stage representation no longer relies on static labels or manual rules, but is dynamically generated through data-driven methods.

[0081] Secondly, the present invention significantly improves the expressive ability of stage evolution modeling by constructing a stage-enhanced representation direction vector, which can more accurately reflect the enterprise growth sequence, introduce jump indicators and path and state transition characteristics, and further, the present invention adopts a joint contrastive learning loss function, which not only optimizes the enterprise stage recognition accuracy in the training stage, but also introduces service matching feedback signals to achieve end-to-end alignment of the enterprise stage vector and the service semantic space, effectively solving the problem of separation of enterprise representation and service resource representation in the existing technology, and improving the context relevance and dynamic adaptability of service resource recommendation.

[0082] In addition, during the recommendation stage, the present invention adjusts the trend direction of the service resource matching score through the guidance mechanism of the matching trend vector, forming a trend-aware recommendation strategy that is different from the static semantic similarity calculation. This mechanism enables the recommendation system to not only focus on the current needs of the enterprise, but also has the ability to predict the future demand trends of the enterprise and actively guide service recommendations, thereby improving the timeliness, foresight and scenario adaptability of service recommendations. Compared with the traditional method of relying solely on current tags for recommendation, the present invention can dynamically respond to changes in the development trajectory of the enterprise and realize the technological leap of the service resource matching mechanism from "current matching" to "trend-guided matching".

[0083] Finally, the present invention also introduces a stage attention mechanism in the fusion layer, using the stage evolution vector as the attention query item to guide the fusion of structured and semantic features. This fusion strategy not only improves the information coupling quality between different modal data, but also realizes the adaptive modeling of stage differences in the feature fusion process, effectively avoiding semantic redundancy or information imbalance problems, thereby embedding the information prior of the company's development situation in the representation learning stage. BRIEF DESCRIPTION OF THE DRAWINGS

[0084] The accompanying drawings are used to provide a further understanding of the present invention and constitute a part of the specification. Together with the embodiments of the present invention, they are used to explain the present invention and do not constitute a limitation of the present invention. In the accompanying drawings:

[0085] Figure 1 This is a flow chart of a data matching service method for enterprises at different stages based on AI artificial intelligence proposed by the present invention;

[0086] Figure 2 This is a structural diagram of an enterprise development stage identification model for an AI-based enterprise data matching service method at different stages proposed by the present invention;

[0087] Figure 3 This is a schematic diagram of the process of generating stage path vectors, jump indicator sequences and path direction vectors by the stage evolution modeling module of the AI ​​artificial intelligence-based enterprise data matching service method proposed by the present invention. DETAILED DESCRIPTION

[0088] The present invention will now be described in further detail with reference to the accompanying drawings, which are simplified schematic diagrams that illustrate the basic structure of the present invention in a schematic manner.

[0089] refer to Figure 1-3 , a data matching service method for enterprises at different stages based on AI artificial intelligence, including the following steps:

[0090] S1. Collect the business data of the enterprise;

[0091] S2. Preprocess the business data to generate structured feature vectors and semantic feature vectors;

[0092] S3. Extract the time characteristic indicator sequence from the pre-processed structured data and construct the evolution path of the enterprise development stage;

[0093] S4. Input the structural feature vector, semantic feature vector, and enterprise development stage evolution path into the enterprise development stage recognition model to generate a fusion representation vector, an enterprise development stage label, and a matching trend vector;

[0094] S5. Using the enterprise development stage label and the matching trend vector as query conditions, searching for candidate service resource items in the service resource database and generating a service resource feature vector;

[0095] S6. Input the fusion representation vector and the service resource feature vector into the service resource matching model. The service resource matching model uses a matching offset mechanism and combines the fusion representation vector to calculate the basic matching score. The matching score is directionally adjusted based on the basic matching trend vector, and a list of recommended service resources is output.

[0096] The present invention proposes an AI-based data matching service method for enterprises at different stages. Through six core steps, it effectively integrates enterprise multi-source data with development stage identification and service recommendation tasks, and establishes a full-chain intelligent recommendation process of data-driven, trend-aware, semantic fusion and dynamic matching. This method opens up a complete closed loop from data collection, feature extraction, stage identification to service resource output, effectively solving the problems of unclear division of enterprise development stages, static and rough service matching, and delayed response of recommendation systems. It has the advantages of high degree of automation, strong adaptability to enterprise heterogeneity, and high recommendation accuracy, and improves the intelligence level of platform services.

[0097] In this embodiment, the business data includes structured data and unstructured data.

[0098] This invention improves the integrity of data source utilization and the depth of modeling by clearly segmenting enterprise operating data into structured and unstructured data. Structured data reflects a company's basic operating behavior, while unstructured data contains external dynamics and semantic information about the company. The combination of the two helps depict the company's development status from multiple perspectives. Compared with existing methods that rely solely on a small number of structural indicators for rough judgment, this invention significantly broadens the data dimension and enhances the input semantic density, providing a richer modeling foundation for subsequent models. This improves the contextual awareness of enterprise stage identification and service matching, and enhances the generalization and practical adaptability of the recommendation system.

[0099] In this embodiment, the structured data includes financial data, human resources data and sales data, and the unstructured data includes corporate web page text, policy document text, industry news text and user evaluation text.

[0100] Based on the detailed definition of structured and unstructured data, the present invention further provides clear data sources, such as financial data, human resources information, web page text and industry news, which effectively supports the model's detailed characterization of the enterprise status. After introducing multi-source heterogeneous information, the system can simultaneously capture the internal operation signals of the enterprise and the influence of external public opinion, so that the enterprise stage identification model has stronger upstream and downstream perception capabilities. Compared with traditional methods that only rely on single features such as finance or registration time, the present invention significantly enhances the model's ability to express the multi-dimensional status of the enterprise, and improves the credibility and dynamic tracking capabilities of the identification results.

[0101] In this embodiment, the preprocessing of the structured data includes missing value filling, standardization, normalization and structured feature coding, and the preprocessing of the unstructured data includes text cleaning, word segmentation and Transformer coding, and the structured feature coding uses a multi-layer perceptron.

[0102] The present invention systematically defines the preprocessing flow of structured and unstructured data, introduces key operations such as missing value filling, standardization, text segmentation, and BERT encoding, and realizes deep modeling through a structured feature coding network, so that the input features have a unified vector expression form. Through the linkage processing of standardized cleaning and embedded modeling, it effectively solves the problems of inconsistent dimensions between structural fields and the inability of text information to directly participate in modeling. In particular, the use of the Transformer network to extract unstructured semantic features significantly improves the expression ability of the model when processing complex text data outside the enterprise, providing high-quality semantic input for the recognition and recommendation models in the subsequent stages.

[0103] In this embodiment, the enterprise development stage evolution path includes a stage path representation sequence, a jump indicator sequence and a path direction vector.

[0104] In this embodiment, S3 specifically includes:

[0105] S31. Selecting time series indicators related to enterprise development dynamics from the pre-processed structured data to form a feature set;

[0106] S32, constructing the feature set into a time series matrix;

[0107] S33. Perform a stationary test on each indicator sequence in the time series matrix. If the ADF test statistic of the indicator sequence is greater than the stationary significance threshold, perform differential processing:

[0108]

[0109] in, represents the difference value of the jth indicator at time step t, represents the original value of the jth indicator at time step t, represents the original value of the jth indicator at time step t-1, and N represents the length of the original time series;

[0110] S34. Apply the sliding window mechanism to the stabilized time series matrix, set the window length and step size, and generate a set of stage window sequences:

[0111]

[0112] in, Represents the set of stage window sequences, W i represents the i-th sliding window subsequence, and q represents the number of stage windows;

[0113] S35. Input each sliding window subsequence into the GRU network, extract the stage representation vector, and obtain the stage path representation sequence:

[0114] ε={e1,e2,...,e q};

[0115] Among them, ε represents the complete stage path representation sequence of the enterprise, e q The phase representation vector representing the qth phase window;

[0116] S36. Construct a stage jump indicator sequence based on the stage path representation sequence:

[0117] φ i =||e i -e i-1 ||2;

[0118] Among them, φ irepresents the Euclidean distance between the vector of the i-th stage and the vector of the previous stage, measuring the degree of jump in the development status of the enterprise, ||·||2 represents the L2 norm, e i represents the phase representation vector of the i-th phase window, e i-1 The phase representation vector representing the i-1th phase window;

[0119] S37. Calculate the path direction vector:

[0120]

[0121] in, It represents the path direction vector of the overall stage of the enterprise, reflecting the development trend of the enterprise, q represents the number of stage windows, e i+1 The phase representation vector representing the i+1th phase window.

[0122] The present invention generates the evolution path of the enterprise development stage by modeling the time series of structured data, combining sliding windows, stationarity detection and GRU encoder, and further calculates the jump index and development direction vector to comprehensively reflect the evolution trend of the enterprise life cycle. This method not only retains the historical behavior trajectory, but also can dynamically measure state transitions and direction deviations. Compared with the static labeling method, it is more timely and predictive. Through the triple modeling of stage path representation, stage jump degree and path direction, the system realizes continuous modeling from temporal behavior to stage expression, effectively improving the robustness of the stage recognition model and its ability to explain the development model.

[0123] In this embodiment, the S4 specifically includes:

[0124] S41, obtaining a structural feature vector, a semantic feature vector, a stage path sequence, a jump indicator sequence, and a path direction vector;

[0125] S42, inputting the structured feature vector into a structured feature extraction layer to obtain a structured deep representation vector, wherein the structured feature extraction layer adopts a two-layer Transformer network;

[0126] Inputting the semantic feature vector into the semantic feature extraction layer to obtain a semantic deep representation vector, wherein the semantic feature extraction layer is based on the pre-trained BERT model;

[0127] S43. Construct a stage enhancement sequence based on the stage path sequence, jump index sequence and path direction vector:

[0128]

[0129] Among them, e′ i represents the stage enhancement vector in the stage enhancement sequence, e irepresents the phase representation vector of the i-th phase window, φ represents the phase jump indicator sequence, represents the path direction vector;

[0130] S44. Enhance the input vector for each stage, construct positive sample vectors and negative sample vectors, and select the positive matching service vectors and historical negative matching service vectors matched by the enterprise in the current stage from the service matching log. The historical negative matching service vectors represent services that the enterprise encountered but did not adopt in the stage, or services that are obviously not suitable for the stage.

[0131] S45. Establish a joint contrastive learning loss function and train a stage encoding network. The stage encoding network adopts a bidirectional GRU network:

[0132]

[0133] in, represents the joint contrastive learning loss function, L e represents the service matching guided loss function, sim(·) represents the cosine similarity, τ represents the temperature coefficient, λ represents the matching loss weighting coefficient, K represents the number of negative samples, s i Indicates a positive matching service vector, Represents the historical negative matching service vector, e′ i represents the stage enhancement vector in the stage enhancement sequence, represents the positive sample vector, Represents a negative sample vector, exp represents a natural exponential function;

[0134] S46, inputting the stage enhancement sequence into the trained stage encoding network, and outputting the stage evolution representation vector;

[0135] S47. Input the structured deep representation vector, semantic deep representation vector, and stage evolution representation vector into the fusion module to calculate the attention weights of the structural and semantic features:

[0136]

[0137] Among them, α s and α t Represent the attention weights of structural and semantic features respectively, W represents the training fusion matrix, V s represents the structured deep representation vector, V t Represents the semantic deep representation vector, V p represents the phase evolution representation vector, and T represents the transpose operation;

[0138] S48. Based on the attention weight, the structural and semantic features are fused to obtain the fused representation vector:

[0139] Vf =α s ·V s +α t ·V t ;

[0140] Among them, V f represents the fused representation vector;

[0141] S49. Use the fused representation vector to make predictions and output the enterprise development stage label and matching trend vector:

[0142] y=arg max(Softmax(W c V f +b c ));

[0143]

[0144] Among them, y represents the enterprise development stage label, Represents the matching trend vector, W c and b c Represents the classification parameter, W t and b t It represents the parameters of trend prediction, argmax represents the variable value when the function value is maximized, and Softmax represents layer normalization.

[0145] The present invention performs multi-layer screening in the service resource database based on stage labels and trend vectors. It first performs preliminary matching based on stage labels, then calculates directional consistency through trend vectors, and finally guides the semantic matching process of service resources, effectively avoiding the common misunderstanding of recommendation systems that "discount the present but not the future." By introducing directional thresholds and trend vector calculation mechanisms, the recommendation process has trend perception capabilities, which not only meets the current needs of enterprises, but also has forward-looking guidance capabilities. This method enhances the dynamic nature and contextual adaptability of service screening, and greatly improves the matching degree between recommendation results and enterprise development goals.

[0146] In this embodiment, the enterprise development stage identification model includes:

[0147] Feature extraction layer, extracting deep expressions of structured feature vectors and semantic feature vectors;

[0148] The stage evolution modeling layer generates the stage evolution representation vector based on the evolution path of the enterprise development stage;

[0149] The stage feature fusion layer uses the stage evolution representation vector as the attention query vector to guide the dynamic weighted fusion of the deep expression of the structural feature vector and the semantic feature vector, and outputs the fused representation vector;

[0150] The stage identification and trend prediction layer jointly models the fusion representation vector and outputs the enterprise development stage label and the matching trend vector. The matching trend vector represents the future evolution direction of the enterprise stage.

[0151] The present invention modularizes the enterprise development stage identification model structure, which includes four major modules: feature extraction, stage modeling, feature fusion and trend prediction, forming a model system with clear division of labor and logical closed loop. This structure makes full use of the stage path information to guide the feature fusion process, enhances the expression consistency between the enterprise behavior semantics and the life cycle status, and guides the weighted fusion of structured and semantic features by using the stage vector as the attention query item, thus achieving context-aware enhanced deep representation, improving the joint modeling capability of stage identification and trend prediction, and providing a strong semantic basis for subsequent recommendations.

[0152] In this embodiment, the S5 specifically includes:

[0153] S51. Obtain an enterprise development stage label, a matching trend vector, and a fusion representation vector. The enterprise development stage label indicates the current life cycle stage of the enterprise, the matching trend vector indicates the direction of the enterprise's evolution from the current stage to the future stage, and the fusion representation vector indicates the comprehensive semantic representation of the enterprise at the current stage.

[0154] S52. Retrieve the service resource feature vectors of the corresponding stage from the service resource database according to the enterprise development stage label to construct a candidate set. The service resource database records the service resource feature vectors corresponding to each development stage.

[0155] S53. For each service resource feature vector in the candidate set, calculate a directional consistency measure with the matching trend vector, wherein the consistency measure uses the L2 norm;

[0156] S54. Set a consistency threshold, retain only service resource feature vectors whose direction consistency metric is greater than the consistency threshold, and construct a set of trend-related candidate service resources;

[0157] S55. Calculate semantic similarity between the fusion representation vector and each service resource feature vector in the candidate set using cosine similarity to obtain a matching score.

[0158] In the candidate service screening process, the present invention uses the dual constraints of "stage label + trend vector" and combines directional consistency calculation to effectively improve the accuracy and controllability of the initial screening link. In particular, in the trend judgment process, cosine directional similarity is used to measure whether the service is consistent with the future direction of the enterprise, which solves the adaptation gap problem caused by static service allocation. Compared with the traditional method of simple industry or stage screening, this method can more dynamically perceive the strategic tendencies and target evolution paths of the enterprise, improve the future adaptability of service recommendations, reduce resource waste and matching deviation, and enhance the platform service accuracy and enterprise acceptance.

[0159] In this embodiment, S6 specifically includes:

[0160] S61, obtaining a fusion representation vector, a matching trend vector, and a service resource feature vector;

[0161] S62. Build a service resource matching model, receive the fusion representation vector and the service resource feature vector as input, calculate the basic matching score, and introduce the matching trend vector. Adjust the basic matching score based on the direction information of the matching trend vector to guide the matching towards the direction of enterprise evolution, and generate a final matching score after trend correction:

[0162]

[0163] Among them, V f represents the fusion representation vector, s j Represents the jth service resource feature vector, score j represents the final matching score under the combined effect of fusion semantics and trend offset, γ represents the trend offset control coefficient, Represents the matching trend vector;

[0164] S63. Arrange all service resource feature vectors in descending order according to the final matching scores, and construct a sorted list of recommended service resources;

[0165] S64. Output the sorted recommended service resource list as a target recommendation list for the enterprise at the current stage and development trend, and provide it to the service platform or user end for display, calling or access to the service processing flow.

[0166] The service resource matching model of the present invention uses a trend offset mechanism to perform directional correction on the basic semantic matching score, thereby realizing the dynamic trend guidance of the recommendation mechanism. The system comprehensively considers the semantic proximity between the current fusion representation of the enterprise and the service resources, and introduces trend direction similarity for weighted adjustment, so that the recommendation is more in line with the enterprise's growth goals and future plans. This mechanism enables the system to maintain the effectiveness of recommendations when the enterprise's needs change rapidly or the stage jumps, avoiding recommendation "lag" or "mismatch". The trend correction strategy significantly improves the matching success rate of recommended services, the service adoption rate and the overall intelligence level of the system.

[0167] Example 1:

[0168] In order to verify the feasibility of the present invention in implementation, the present invention is applied to the enterprise service platform system of a certain artificial intelligence industrial park. The park is located in a city in the Yangtze River Delta. It began to promote the "Enterprise Full Life Cycle Service Intelligence" project in 2022, aiming to provide policy recommendations, technical services, financial docking, tax optimization and other types of resource support to more than 300 enterprises in the park by building an enterprise growth assistance platform. However, during the operation of the platform, it was found that the identification of the development stage of the enterprise is heavily dependent on manual declaration or enterprise-defined labels, and cannot dynamically reflect the true development status of the enterprise. At the same time, the service resource recommendation lacks foresight and matching accuracy, resulting in low conversion rate of resource push and low service reception rate.

[0169] In response to the above problems, the park platform introduced the "AI-based enterprise data matching service method at different stages" proposed in this invention in April 2023 and deployed it to the park cloud service platform, and began pilot application on 100 manufacturing, technology service and platform operation companies in the park. The official operation cycle is set to 6 months.

[0170] In the initial stage of platform operation, the system first collects structured data from the enterprise's operating data interface, such as the financial flow, human resource structure, customer sales records, etc. in the past 36 months; at the same time, it calls on unstructured text data such as web page content, policy application materials, public bidding announcements, intellectual property documents and social news reports maintained by the enterprise's operating entity to form a complete enterprise multi-dimensional information set. Through the structured preprocessing and text encoding method provided by the present invention, the system automatically generates standardized feature vectors and semantic feature vectors in the background, and inputs them into the enterprise development stage recognition model. Combining the 36-month operating time series of the enterprise to construct a sliding window sequence, extracting the stage representation through the bidirectional GRU network, and automatically deducing the enterprise stage evolution path and trend direction vector.

[0171] For example, a technology services company named A was established in early 2020. After reading its operating data, the platform automatically identified an accelerating upward trend in headcount, R&D spending, and contract signing amounts over the past 18 months. The system automatically shifted its development stage from "early growth" to "mid-to-late growth," and, based on its evolutionary trend vector, determined its intention to expand across regions within the next 3-6 months. Based on this stage label and trend vector, the system automatically selected policy resources, brand planning services, software copyright application guidance, and industrial investment and financing resources from the service resource database that matched this stage and aligned with its direction. The system then used the service matching model proposed in this invention to adjust the trend offset of the recommendation results and output a recommended list.

[0172] After the system was officially launched, the service behavior records of 100 pilot enterprises were tracked for 180 consecutive days. Table 1 shows the comparison data of key platform indicators before and after the implementation of the system.

[0173] Table 1 Comparison of comprehensive indicators of the platform before and after the implementation of the present invention

[0174] Indicator name Before implementation After implementation Improvement Service recommendation click rate 12.4％ 21.8％ +75.8％ Service application submission rate 5.7％ 13.3％ +133.3％ Recommendation click-through rate 11.2％ 19.4％ +73.2％ Negative feedback rate of enterprises 37.0％ 18.5％ -50.0％ Average recommended service matching score (out of 5) 3.4 4.3 +26.5％ Service recommendation update response time (hours) 48 8 -83.3％ Stage recognition model accuracy 62.1％ 91.4％ +47.2％ Service adoption rate (actual adoption among matching resources) 9.8％ 22.7％ +131.6％

[0175] According to the data analysis in Table 1, it can be clearly seen that after introducing the AI-based enterprise data matching service method at different stages proposed in this invention, the platform has achieved significant optimization in multiple core dimensions such as service recommendation efficiency, enterprise response behavior, model recognition accuracy, and user satisfaction, verifying the implementation effect and technical advantages of this method in real business scenarios.

[0176] In terms of the click-through rate of service recommendations, it increased significantly from 12.4% before implementation to 21.8% after implementation, an increase of 75.8%. This shows that through the semantic recommendation mechanism driven by the company's development stage and trends, the service content pushed by the system is more relevant and attractive, and can significantly stimulate the company's users' active click willingness. At the same time, the service application submission rate jumped from 5.7% to 13.3%, an increase of 133.3%. This not only reflects the high match between the recommended content and the actual needs of the company, but also shows that the company is more willing to take practical action to apply for service resources after receiving the recommendation.

[0177] The recommendation click-through conversion rate increased from 11.2% to 19.4%, an increase of 73.2%, indicating that the present invention has a significant ability to improve the key link "from click to conversion" in the recommendation chain. In particular, the trend shift mechanism is used to dynamically sort services, enabling enterprises to receive more appropriate service lists at different evolutionary stages, thereby enhancing the feasibility and practical value of recommendations.

[0178] From the perspective of user satisfaction, the negative feedback rate of enterprises dropped from 37.0% before implementation to 18.5%, a decrease of 50.0%, which greatly reduced the negative emotions of users caused by invalid or irrelevant recommendations. This result is consistent with the feedback from user interviews within the platform. Users generally said that "recommendations are more in line with development status", "service content is more accurate", and "recommendations are updated synchronously with changes in the enterprise". The average recommendation service matching score increased from 3.4 points to 4.3 points (out of 5 points), further reflecting that the overall recognition of enterprises for the quality of recommendations has increased significantly.

[0179] At the system level, the response speed of service recommendations has also been significantly improved, with the response time for recommendation updates reduced from 48 hours to 8 hours, an optimization rate of 83.3%. This improvement is attributed to the end-to-end modeling process and embedded stage identification mechanism in this invention, which breaks away from the limitations of manual updates and batch rule generation, and achieves rapid service response to enterprise data changes.

[0180] In terms of AI model capabilities, the accuracy of the enterprise stage identification model constructed by the present invention has increased from 62.1% to 91.4%, an increase of 47.2%. This shows that by introducing multi-dimensional modeling of stage path representation, jump indicators and trend vectors, the model is more detailed and robust in judging the stage of the enterprise, and can effectively avoid problems such as life cycle misjudgment and stage cross-ambiguity.

[0181] Finally, the adoption rate of services with the greatest business value increased from 9.8% to 22.7%, a 131.6% increase. This indicates a significant increase in the proportion of recommended services actually adopted by businesses and resulting in subsequent service actions, demonstrating that this invention has a real and quantifiable impact on the efficiency of matching supply and demand between businesses and resources.

[0182] The above description is only a preferred specific embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any technician familiar with the technical field, within the technical scope disclosed by the present invention, who makes equivalent replacements or changes based on the technical solution and inventive concept of the present invention, should be covered by the scope of protection of the present invention.

Claims

1. A data matching service method for enterprises at different stages based on AI artificial intelligence, characterized by: The steps include: S1. Collect the business data of the enterprise; S2. Preprocess the business data to generate structured feature vectors and semantic feature vectors; S3. Extract the time characteristic indicator sequence from the pre-processed structured data and construct the evolution path of the enterprise development stage; S4. Input the structural feature vector, semantic feature vector, and enterprise development stage evolution path into the enterprise development stage recognition model to generate a fusion representation vector, an enterprise development stage label, and a matching trend vector; S5. Using the enterprise development stage label and the matching trend vector as query conditions, searching for candidate service resource items in the service resource database and generating a service resource feature vector; S6. Input the fusion representation vector and the service resource feature vector into the service resource matching model. The service resource matching model uses a matching offset mechanism and combines the fusion representation vector to calculate the basic matching score. The matching score is directionally adjusted based on the basic matching trend vector, and a list of recommended service resources is output.

2. The data matching service method for enterprises at different stages based on AI artificial intelligence according to claim 1 is characterized in that: The business data includes structured data and unstructured data.

3. The data matching service method for enterprises at different stages based on AI artificial intelligence according to claim 2 is characterized in that: The structured data includes financial data, human resources data and sales data, and the unstructured data includes corporate web page text, policy document text, industry news text and user evaluation text.

4. The data matching service method for enterprises at different stages based on AI artificial intelligence according to claim 2 is characterized in that: The preprocessing of the structured data includes missing value filling, standardization, normalization and structured feature coding, and the preprocessing of the unstructured data includes text cleaning, word segmentation and Transformer coding. The structured feature coding uses a multi-layer perceptron.

5. The data matching service method for enterprises at different stages based on AI artificial intelligence according to claim 1 is characterized in that: The enterprise development stage evolution path includes a stage path representation sequence, a jump indicator sequence and a path direction vector.

6. The data matching service method for enterprises at different stages based on AI artificial intelligence according to claim 1 is characterized in that: The S3 specifically includes: S31. Selecting time series indicators related to enterprise development dynamics from the pre-processed structured data to form a feature set; S32, constructing the feature set into a time series matrix; S33. Perform a stationary test on each indicator sequence in the time series matrix. If the ADF test statistic of the indicator sequence is greater than the stationary significance threshold, perform differential processing. S34, applying a sliding window mechanism to the stabilized time series matrix, setting the window length and step size, and generating a set of stage window sequences; S35. Input each sliding window subsequence into the GRU network, extract the stage representation vector, and obtain the stage path representation sequence; S36. Constructing a stage jump indicator sequence based on the stage path representation sequence; S37. Calculate the path direction vector: in, It represents the path direction vector of the overall stage of the enterprise, reflecting the development trend of the enterprise, q represents the number of stage windows, e i+1 The phase representation vector representing the i+1th phase window.

7. The data matching service method for enterprises at different stages based on AI artificial intelligence according to claim 1 is characterized in that: The S4 specifically includes: S41, obtaining a structural feature vector, a semantic feature vector, a stage path sequence, a jump indicator sequence, and a path direction vector; S42, inputting the structured feature vector into a structured feature extraction layer to obtain a structured deep representation vector, wherein the structured feature extraction layer adopts a two-layer Transformer network; Inputting the semantic feature vector into the semantic feature extraction layer to obtain a semantic deep representation vector, wherein the semantic feature extraction layer is based on the pre-trained BERT model; S43, constructing a stage enhancement sequence based on the stage path sequence, the jump index sequence and the path direction vector; S44. Enhance the input vector for each stage, construct a positive sample vector and a negative sample vector, and select the positive matching service vector and historical negative matching service vector matched by the enterprise in the current stage from the service matching log; S45. Establish a joint contrastive learning loss function and train a stage encoding network. The stage encoding network adopts a bidirectional GRU network: in, represents the joint contrastive learning loss function, L e represents the service matching guided loss function, sim(·) represents the cosine similarity, τ represents the temperature coefficient, λ represents the matching loss weighting coefficient, K represents the number of negative samples, s i Indicates a positive matching service vector, Represents the historical negative matching service vector, e′ i represents the stage enhancement vector in the stage enhancement sequence, represents the positive sample vector, Represents a negative sample vector, exp represents a natural exponential function; S46, inputting the stage enhancement sequence into the trained stage encoding network, and outputting the stage evolution representation vector; S47. Input the structured deep representation vector, semantic deep representation vector, and stage evolution representation vector into the fusion module to calculate the attention weights of the structural and semantic features: Among them, α s and α t Represent the attention weights of structural and semantic features respectively, W represents the training fusion matrix, V s represents the structured deep representation vector, V t Represents the semantic deep representation vector, V p represents the phase evolution representation vector, and T represents the transpose operation; S48, fusing structural and semantic features based on attention weights to obtain a fused representation vector; S49. Use the fused representation vector for prediction and output the enterprise development stage label and matching trend vector.

8. The data matching service method for enterprises at different stages based on AI artificial intelligence according to claim 1 is characterized in that: The enterprise development stage identification model includes: Feature extraction layer, extracting deep expressions of structured feature vectors and semantic feature vectors; The stage evolution modeling layer generates the stage evolution representation vector based on the evolution path of the enterprise development stage; The stage feature fusion layer uses the stage evolution representation vector as the attention query vector to guide the dynamic weighted fusion of the deep expression of the structural feature vector and the semantic feature vector, and outputs the fused representation vector; The stage identification and trend prediction layer jointly models the fusion representation vector and outputs the enterprise development stage label and the matching trend vector. The matching trend vector represents the future evolution direction of the enterprise stage.

9. The data matching service method for enterprises at different stages based on AI artificial intelligence according to claim 1 is characterized in that: The S5 specifically includes: S51. Obtain an enterprise development stage label, a matching trend vector, and a fusion representation vector. The enterprise development stage label indicates the current life cycle stage of the enterprise, the matching trend vector indicates the direction of the enterprise's evolution from the current stage to the future stage, and the fusion representation vector indicates the comprehensive semantic representation of the enterprise at the current stage. S52. Retrieve the service resource feature vectors of the corresponding stage from the service resource database according to the enterprise development stage label to construct a candidate set. The service resource database records the service resource feature vectors corresponding to each development stage. S53. For each service resource feature vector in the candidate set, calculate a directional consistency measure with the matching trend vector, wherein the consistency measure uses the L2 norm; S54. Set a consistency threshold, retain only service resource feature vectors whose direction consistency metric is greater than the consistency threshold, and construct a set of trend-related candidate service resources; S55. Calculate semantic similarity between the fusion representation vector and each service resource feature vector in the candidate set using cosine similarity to obtain a matching score.

10. The data matching service method for enterprises at different stages based on AI artificial intelligence according to claim 1 is characterized in that: The S6 specifically includes: S61, obtaining a fusion representation vector, a matching trend vector, and a service resource feature vector; S62. Build a service resource matching model, receive the fusion representation vector and the service resource feature vector as input, calculate the basic matching score, and introduce the matching trend vector. Adjust the basic matching score based on the direction information of the matching trend vector to guide the matching towards the direction of enterprise evolution, and generate a final matching score after trend correction: Among them, V f represents the fusion representation vector, s j Represents the jth service resource feature vector, score j represents the final matching score under the combined effect of fusion semantics and trend offset, γ represents the trend offset control coefficient, Represents the matching trend vector; S63. Arrange all service resource feature vectors in descending order according to the final matching scores, and construct a sorted list of recommended service resources; S64. Output the sorted recommended service resource list as a target recommendation list for the enterprise at the current stage and development trend, and provide it to the service platform or user end for display, calling or access to the service processing flow.

Images