Enterprise recommendation method based on multi-dimensional cosine similarity algorithm

Abstract

The invention discloses an enterprise recommendation method based on a multi-dimensional cosine similarity algorithm, and relates to the technical field of data processing. Comprising the steps of collecting enterprise information data, and performing normalization and tagging processing to form enterprise basic data and skill data; vectorizing the enterprise basic data and the skill data to form an enterprise comprehensive strength matrix A, a skill feature matrix B and an inter-enterprise skill influence model matrix C, and constructing data sets D1, D2 and D3; aiming at different dimensions, using CNN, Transform-XL and GNN models to respectively learn enterprise skill features, comprehensive strength features and skill influence features, optimizing model parameters and improving prediction performance; an SSA algorithm is adopted to optimize hyper-parameters of the XGBoost, and optimal parameters are obtained through MAE and MSE performance evaluation; and generating a recommendation list for the enterprises with relatively high matching degrees according to the sequence, and outputting a final recommendation enterprise list. According to the invention, the features and scores of the target enterprise can be obtained in a short time.

Description

Technical Field

[0001] The present invention relates to the technical field of data processing, and in particular to an enterprise recommendation method based on a multi-dimensional cosine similarity algorithm. Background Art

[0002] As society enters the era of big data with rapid information expansion, the government's investment promotion work is also facing unprecedented challenges. The traditional investment promotion method manually screens information that matches the investment promotion needs from a large number of enterprises, which has problems such as low efficiency, information asymmetry, and imperfect resources. It is also difficult to tap the value of resources suitable for the development of regional industries, resulting in investment promotion errors. In addition, by directly calculating the similarity between the enterprise characteristics and the demand feature vector space model, the recommendation results only consider the text similarity, without considering other influencing factors. If the enterprise has poor credit and weak comprehensive capabilities, it will lead to the enterprise's inability to successfully complete the task, which will also reduce the accuracy of the recommendation system and the user's trust. Summary of the invention

[0003] In view of the problems existing in the above background technology, the present invention proposes an enterprise recommendation method based on a multi-dimensional cosine similarity algorithm.

[0004] Therefore, the problem to be solved by the present invention is how to obtain the characteristics and scores of the target enterprise in a shorter time.

[0005] In order to solve the above technical problems, the present invention provides the following technical solutions: In the first aspect, an embodiment of the present invention provides an enterprise recommendation method based on a multi-dimensional cosine similarity algorithm, which includes collecting enterprise information data, normalizing and labeling it to form enterprise basic data and skill data; vectorizing the enterprise basic data and skill data to form an enterprise comprehensive strength matrix A, a skill feature matrix B, and an inter-enterprise skill influence model matrix C, and constructing data sets D1, D2, and D3; for different dimensions, using CNN, Transformer-XL, and GNN models to respectively learn enterprise skill characteristics, comprehensive strength characteristics, and skill influence characteristics, optimize model parameters, and improve prediction performance; using the SSA algorithm to optimize the hyperparameters of XGBoost, and obtaining the optimal parameters through MAE and MSE performance evaluation; generating a recommendation list by sorting the enterprises with higher matching degrees, and outputting the final recommended enterprise list.

[0006] As a preferred solution of the enterprise recommendation method based on the multi-dimensional cosine similarity algorithm described in the present invention, the formation of the enterprise comprehensive strength matrix A includes: standardizing and vectorizing the basic attribute data of the enterprise; using normalization processing to scale the data to the same dimension for different numerical features of the enterprise; using One-Hot encoding to generate discrete feature vectors for different categorical features of the enterprise; forming the skill feature matrix B includes: using TF-IDF to vectorize the text skill data to measure the weight of the enterprise skill features; combining the processed skill feature vectors into the skill feature matrix B; forming the inter-enterprise skill influence model matrix C includes: using cosine similarity to measure the similarity of the skill vectors between enterprises; filling the similarity values ​​between all enterprise pairs into the inter-enterprise skill influence model matrix C.

[0007] As a preferred solution of the enterprise recommendation method based on the multi-dimensional cosine similarity algorithm of the present invention, the method of vectorizing text-based skill data using TF-IDF includes:

[0008] in, is the frequency of word t in document d, N is the total number of enterprises, is the number of enterprise documents containing word t.

[0009] As a preferred solution of the enterprise recommendation method based on the multi-dimensional cosine similarity algorithm described in the present invention, the use of CNN, Transformer-XL and GNN models to respectively learn the enterprise skill characteristics, comprehensive strength characteristics and skill influence characteristics includes the following steps: inputting the enterprise's skill feature matrix B into the convolutional neural network model to learn the enterprise's characteristics in the skill dimension; using the Transformer-XL model to model the enterprise's comprehensive strength feature matrix A to capture the temporal correlation and long-term dependency of the enterprise's comprehensive strength in multiple aspects; for the skill influence matrix C between enterprises, using a graph neural network to model the similarity relationship of skills between enterprises, and learn the skill influence between different enterprises.

[0010] As a preferred solution of the enterprise recommendation method based on the multi-dimensional cosine similarity algorithm described in the present invention, the step of inputting the enterprise's skill feature matrix B into the convolutional neural network model includes the following steps: each row in the skill feature matrix B represents a skill feature vector of an enterprise; n-gram features of different lengths in the enterprise's skills are extracted through a one-dimensional convolutional layer; the maximum pooling layer is used to reduce the dimension of the feature map and extract the most representative features in the skill features. ; Input the convolution and pooling feature vector into the fully connected layer to calculate the skill prediction value of each enterprise :

[0011] in, and are the weights and biases of the fully connected layer; the mean square error is selected as the loss function to measure the difference between the predicted value and the true skill feature value.

[0012] As a preferred solution of the enterprise recommendation method based on the multi-dimensional cosine similarity algorithm of the present invention, wherein: the use of the Transformer-XL model to model the enterprise comprehensive strength feature matrix A includes the following steps: the feature vector of each enterprise in the comprehensive strength feature matrix A is used as an input sequence; a multi-head self-attention layer is used to calculate the attention weight for the input feature vector to obtain the global dependency relationship of the enterprise comprehensive strength features:

[0013] Among them, Q, K and V are query, key and value respectively, and the matrix is the scaling factor; the relative position encoding mechanism of Transformer-XL is used to retain the sequential information of enterprise features.

[0014] As a preferred solution of the enterprise recommendation method based on the multi-dimensional cosine similarity algorithm described in the present invention, the steps of implementing the training of the XGBoost algorithm by the SSA algorithm are: normalizing the data; initializing the parameters of the sparrow search algorithm, the parameters of the XGBoost model and the thresholds of the boosted parameters; the fitness function Fi is the root mean square error between the fitting value and the measured value of the training set, and Fi is used to calculate the fitness value fi of the sparrow population; the sparrow fitness values ​​fi are sorted and the optimal value fg is selected; the positions of the discoverers, followers and guards in the sparrow population are updated; if the optimal fitness value fg obtained in the current iteration is less than the accuracy set by the model, the iteration is terminated; or if the number of iterations is exhausted, the iterative search is terminated; the global optimal position xbj at the end of the iteration is used to optimize the XGBoost model boosting parameters; the XGBoost algorithm is used to train the data and predict, and the prediction results are output.

[0015] As a preferred solution of the enterprise recommendation method based on the multi-dimensional cosine similarity algorithm described in the present invention, the generation process of the recommendation list is as follows: using the enterprise skill feature model, the inter-enterprise skill influence model and the enterprise comprehensive capability model as well as the data of three dimensions, the feature vectors of each dimension of the corresponding target enterprise are obtained; the feature vectors of each dimension calculated for each enterprise are looped with the feature vectors of each dimension of the target enterprise, and the cosine similarity is compared according to the weights obtained by ensemble learning, and the enterprises with relatively large similarity are selected for recommendation.

[0016] In a second aspect, an embodiment of the present invention provides a computer device comprising a memory and a processor, wherein the memory stores a computer program, wherein: when the computer program instructions are executed by the processor, the steps of the enterprise recommendation method based on a multi-dimensional cosine similarity algorithm as described in the first aspect of the present invention are implemented.

[0017] In a third aspect, an embodiment of the present invention provides a computer-readable storage medium having a computer program stored thereon, wherein: when the computer program instructions are executed by a processor, the steps of the enterprise recommendation method based on a multi-dimensional cosine similarity algorithm as described in the first aspect of the present invention are implemented.

[0018] The beneficial effects of the present invention are as follows: the present invention can obtain the characteristics and scores of the target enterprise in a relatively short time. Among them, the integrated model can reduce the deviation and variance of a single model by combining the prediction results of multiple basic models, thereby improving the generalization ability of the overall model, and can capture the advantages of different models, reduce the erroneous predictions of individual models, thereby improving the overall prediction accuracy, and improving the prediction accuracy, generalization ability and robustness of the overall model recommended by the present invention; and according to the needs of regional industrial development, using the model of the present invention, the investment promotion party can carry out more targeted investment promotion activities; using the model of the present invention, the trust of the target enterprise can be high, the accuracy of investment promotion can be greatly improved, time and cost can be saved, the success rate and efficiency of investment promotion projects can be improved, and the development of the regional economy and industrial upgrading can be promoted. BRIEF DESCRIPTION OF THE DRAWINGS

[0019] In order to more clearly illustrate the technical solutions of the embodiments of the present invention, the following briefly introduces the drawings required for use in the description of the embodiments. Obviously, the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without creative work.

[0020] Figure 1 Flowchart of a business recommendation method based on a multi-dimensional cosine similarity algorithm. DETAILED DESCRIPTION

[0021] In order to make the above-mentioned objects, features and advantages of the present invention more obvious and easy to understand, the specific implementation methods of the present invention are described in detail below in conjunction with the accompanying drawings.

[0022] In the following description, many specific details are set forth to facilitate a full understanding of the present invention, but the present invention may also be implemented in other ways different from those described herein, and those skilled in the art may make similar generalizations without violating the connotation of the present invention. Therefore, the present invention is not limited to the specific embodiments disclosed below.

[0023] Secondly, the term "one embodiment" or "embodiment" as used herein refers to a specific feature, structure, or characteristic that may be included in at least one implementation of the present invention. The term "in one embodiment" that appears in different places in this specification does not necessarily refer to the same embodiment, nor does it refer to a separate or selective embodiment that is mutually exclusive with other embodiments.

[0024] Example 1 Reference Figure 1 , which is the first embodiment of the present invention, and provides a method for recommending enterprises based on a multi-dimensional cosine similarity algorithm, such as Figure 1 As shown, including, S1: Collect information such as the company’s basic attributes, skills, performance history, and operating conditions, and normalize and label these data to form basic company data and skill data.

[0025] Specifically, we collect public information of each company, including the company's years of operation, industry prospects, and shareholder background; the company's performance history, including financing performance, judicial performance, and commercial performance; the company's management level, including stable operations, litigation disputes, and investment and financing; the company's evaluation, including regulatory evaluation, qualification certification, and honorary titles; the company's operating conditions, including tax payment, bidding, and financial performance; as well as the company's registered capital, staff size, number of insured persons, company profile, business scope, corporate assets, corporate net profit and other data as the company's comprehensive strength data.

[0026] Collect data such as the title and abstract of the enterprise's software copyright, the title and abstract of the patent, the title and abstract of the bidding document, the title and abstract of the enterprise or outstanding achievements as the enterprise's technical data.

[0027] The Qicha score and science and technology innovation score of each enterprise are crawled from Qichacha, and the enterprise rating of Tianyancha is used as label data.

[0028] The data is split into enterprise, three-dimensional related text data and corresponding label data, and the processed data is divided into training set, training set 2, validation set and test set according to 6:2:1:1.

[0029] S2: Quantify the basic data and skill data of enterprises to form the enterprise comprehensive strength matrix A, skill feature matrix B and inter-enterprise skill influence model matrix C, and construct data sets D1, D2 and D3.

[0030] S2.1: Standardize and vectorize the basic attribute data of the enterprise to form the enterprise comprehensive strength matrix A: For the numerical characteristics of the enterprise, such as the operating years, registered capital, and staff size, normalization is used to scale the data to the same dimension; For categorical features such as enterprise evaluation and qualification certification, One-Hot encoding is used to generate discrete feature vectors. Finally, the comprehensive strength matrix A can be expressed as:

[0031] in, Represents the jth comprehensive strength characteristic of the ith enterprise, m is the number of enterprises, and n is the number of characteristics.

[0032] Furthermore, the skill data of the enterprise is preprocessed and vectorized to generate the skill feature matrix B, which includes the following steps: Use TF-IDF (term frequency-inverse document frequency) to vectorize text-based skill data to measure the weight of enterprise skill features:

[0033] in, is the frequency of word t in document d, N is the total number of enterprises, is the number of enterprise documents containing word t.

[0034] Combine the processed skill feature vectors into a skill feature matrix B:

[0035] in, represents the weight of the jth skill feature of the ith enterprise, and p is the total number of skill features.

[0036] Calculate the inter-enterprise skill influence model matrix C, which represents the similarity or influence of inter-enterprise skill characteristics. The specific steps are as follows: The cosine similarity is used to measure the similarity of the skill vectors between enterprises, and the similarity between enterprise e and enterprise f in the skill matrix B is calculated:

[0037] Fill the similarity values ​​between all enterprise pairs into the inter-enterprise skill influence model matrix C:

[0038] in, Represents the skill similarity between enterprise e and enterprise f.

[0039] S3: For different dimensions, CNN, Transformer-XL and GNN models are used to learn enterprise skill characteristics, comprehensive strength characteristics and skill influence characteristics respectively, optimize model parameters and improve prediction performance.

[0040] S3.1: Input the enterprise's skill feature matrix B into the convolutional neural network (CNN) model to learn the enterprise's features in the skill dimension, where each row in the skill feature matrix B represents a skill feature vector of an enterprise. The specific operations are as follows: Extract n-gram features of different lengths from enterprise skills through a one-dimensional convolutional layer; Use the maximum pooling layer to reduce the dimension of the feature map and extract the most representative features of the skill features. ; The convolution and pooling feature vectors are input into the fully connected layer to calculate the skill prediction value of each enterprise. :

[0041] in, and are the weights and biases of the fully connected layers.

[0042] The mean square error (MSE) is selected as the loss function to measure the difference between the predicted value and the true skill feature value.

[0043] S3.2: Use the Transformer-XL model to model the enterprise comprehensive strength feature matrix A to capture the temporal correlation and long-term dependency of the enterprise's comprehensive strength in multiple aspects. The specific operations are as follows: The characteristic vector of each enterprise in the comprehensive strength characteristic matrix A is used as the input sequence; Using a multi-head self-attention layer, the attention weight is calculated for the input feature vector to obtain the global dependency relationship of the comprehensive strength characteristics of the enterprise:

[0044] Among them, Q, K and V are query, key and value respectively, and the matrix is the scaling factor.

[0045] By utilizing the relative position encoding mechanism of Transformer-XL, the sequential information of enterprise characteristics is retained and the ability to model the sequence is strengthened.

[0046] Finally, the mean square error (MSE) is also used as the loss function to calculate the comprehensive strength prediction error, and the Transformer-XL model parameters are optimized through back propagation.

[0047] S3.3: For the inter-enterprise skill impact matrix C, a graph neural network (GNN) is used to model the similarity relationship between inter-enterprise skills and learn the skill impact between different enterprises. The details are as follows: Construct an undirected graph with enterprise nodes and edges between enterprises, where each enterprise is a node and the edge weights between nodes are the skill similarities between enterprises (elements in matrix C).

[0048] Through the message passing mechanism of GNN, enterprise skill information is aggregated from neighboring nodes and the feature vector of each enterprise node is updated :

[0049] in, is the node feature of the l+1th layer, and are weights and biases, is the edge weight.

[0050] After multiple layers of message passing, the enterprise node feature vector is mapped to the skill impact prediction value through a fully connected layer; MSE is also used as the loss function to minimize the prediction error of the impact of inter-firm skills, and the GNN model is optimized through back propagation.

[0051] S4: Use the SSA algorithm to optimize the hyperparameters of XGBoost, obtain the optimal parameters through MAE and MSE performance evaluation, and improve the accuracy of the recommendation model.

[0052] The sparrow search algorithm (SSA) is introduced to optimize the hyperparameters of the XGBoost model to obtain the best hyperparameter combination of the recommendation model. The optimization process improves the overall accuracy of the model by minimizing the mean square error (MSE) and mean absolute error (MAE).

[0053] The goal of hyperparameter optimization includes determining the optimal number of iterations, tree depth, and learning rate so that the model can better adapt to the enterprise feature data set. The specific steps are: The steps for implementing the SSA algorithm to train the XGBoost algorithm are as follows: S4.1: Data preprocessing, normalize the data; S4.2: Parameter setting, initializing the parameters of the sparrow search algorithm, the parameters of the XGBoost model and the threshold of the boosting parameters.

[0054] S4.3: The fitness function Fi (objective function) is the root mean square error between the fitted value and the measured value of the training set. Fi is used to calculate the fitness value fi of the hemp population.

[0055] S4.4: Sort the sparrow fitness values ​​fi and select the optimal value fg S4.5: Update the locations of spotters, followers, and sentinels in a sparrow population.

[0056] S4.6: Search iterative termination condition. If the optimal fitness value fg obtained in the current iteration is less than the accuracy set by the model, the iteration is terminated. This is because the expected optimization effect has been achieved, and further iterations can no longer bring more optimization benefits, but will waste computing resources; or if the number of iterations is exhausted, the iterative search is terminated.

[0057] S4.7: The global optimal position xbj at the end of the algorithm iteration is used to optimize the XGBoost model parameters, that is, to obtain the optimal number of iterations, tree depth and learning rate and assign them to the XGBoost model.

[0058] S4.8: Use the XGBoost algorithm to train the data and make predictions, and output the prediction results.

[0059] By comparing the prediction results output by the model with the real labeled data set, the model performance is evaluated based on MAE and MSE respectively to verify the applicability and accuracy of the recommendation model in the enterprise recommendation scenario.

[0060] S5: Generate a recommendation list by sorting the enterprises with higher matching degrees, and output the final list of recommended enterprises.

[0061] Using the enterprise skill feature model, inter-enterprise skill impact model, enterprise comprehensive capability model and data from three dimensions, we obtain the feature vectors of each dimension of the corresponding target enterprise; loop through the feature vectors of each dimension calculated for each enterprise and the feature vectors of each dimension of the target enterprise, perform cosine similarity comparison according to the weights learned from the ensemble, and select enterprises with greater similarity for recommendation. The specific operations are as follows: The enterprise comprehensive strength matrix A, skill feature matrix B and inter-enterprise skill influence model matrix C generated in step S2 are used to extract the characteristic vectors of each dimension of the target enterprise, that is, the characteristic vector of the target enterprise is ; For each enterprise i, calculate its corresponding comprehensive strength vector according to the feature vector obtained in step S2 :

[0062] in, They represent the performance of enterprise i in comprehensive strength, skill characteristics and skill influence characteristics respectively.

[0063] The weight parameter W obtained by ensemble learning is used to adjust the influence of each dimension and define the weighted feature vector:

[0064] Calculate the cosine similarity between the weighted feature vector of enterprise i and the feature vector of the target enterprise The matching scores obtained will be used as the sorting basis to screen the recommended companies.

[0065] The calculated matching scores are arranged in descending order to ensure that enterprises with higher matching scores are displayed first in the recommendation list, and threshold filtering conditions are set to filter out low matching enterprises, thereby obtaining a list of high matching enterprises that meet the needs.

[0066] Based on the sorting and screening results, a company recommendation list is generated, which contains the basic information, matching score and corresponding feature description (skills, strength, influence, etc.) of each recommended company.

[0067] Output the final recommended enterprise list, which can be divided into layers such as high match and second highest match for further use and display.

[0068] This embodiment also provides a computer device, which is suitable for the enterprise recommendation method based on the multi-dimensional cosine similarity algorithm, including a memory and a processor; the memory is used to store computer-executable instructions, and the processor is used to execute the computer-executable instructions to implement the enterprise recommendation method based on the multi-dimensional cosine similarity algorithm proposed in the above embodiment.

[0069] The computer device may be a terminal, and the computer device includes a processor, a memory, a communication interface, a display screen and an input device connected through a system bus. Among them, the processor of the computer device is used to provide computing and control capabilities. The memory of the computer device includes a non-volatile storage medium and an internal memory. The non-volatile storage medium stores an operating system and a computer program. The internal memory provides an environment for the operation of the operating system and the computer program in the non-volatile storage medium. The communication interface of the computer device is used to communicate with an external terminal in a wired or wireless manner, and the wireless manner can be achieved through WIFI, an operator network, NFC (near field communication) or other technologies. The display screen of the computer device may be a liquid crystal display screen or an electronic ink display screen, and the input device of the computer device may be a touch layer covered on the display screen, or a key, trackball or touchpad provided on the housing of the computer device, or an external keyboard, touchpad or mouse, etc.

[0070] This embodiment also provides a storage medium on which a computer program is stored. When the program is executed by a processor, the enterprise recommendation method based on a multi-dimensional cosine similarity algorithm as proposed in the above embodiment is implemented.

[0071] In summary, the present invention can obtain the characteristics and scores of target enterprises in a relatively short period of time. Among them, the integrated model can reduce the deviation and variance of a single model by combining the prediction results of multiple basic models, thereby improving the generalization ability of the overall model, and can capture the advantages of different models, reduce the erroneous predictions of individual models, thereby improving the overall prediction accuracy, and improving the prediction accuracy, generalization ability and robustness of the overall model recommended by the present invention. And according to the needs of regional industrial development, using the model of the present invention, the investment promotion party can carry out more targeted investment promotion activities; using the model of the present invention, the trust of the target enterprise is high, the accuracy of investment promotion is greatly improved, and time and cost are saved; using the model of the present invention, the success rate and efficiency of investment promotion projects are improved, and the development of the regional economy and industrial upgrading are promoted.

[0072] Example 2 Referring to Table 1, which is a second embodiment of the present invention, this embodiment provides an enterprise recommendation method based on a multi-dimensional cosine similarity algorithm. In order to verify the beneficial effects of the present invention, a scientific demonstration is carried out through simulation experiments.

[0073] In this embodiment, the goal is to verify a business recommendation method based on a multi-dimensional cosine similarity algorithm, by constructing a multi-dimensional data model of the comprehensive strength, skill characteristics and skill influence of the enterprise, and providing a recommendation list based on matching degree for different enterprises.

[0074] The experiment mainly includes data collection and preprocessing, model training and verification, parameter optimization and output of recommendation results. The experimental process is divided into the following steps: First, we obtain the basic attributes, skill characteristics and label data of enterprises from multiple public data platforms, including information such as operating years, registered capital, industry prospects, shareholder background, bidding and litigation records. Data sources include enterprise credit platforms, market monitoring platforms, etc., and these information are uniformly labeled and normalized. Numerical data (such as registered capital) are normalized and scaled to [0,1], and One-Hot encoding is used for categorical data (such as industry categories) to generate discrete feature vectors to form the enterprise comprehensive strength data set A. The TF-IDF method is used to vectorize skill-related data to form a skill feature data set B. The crawled third-party enterprise rating data (such as science and technology innovation points) is used as label data, and the enterprise's performance in different dimensions is marked and classified.

[0075] The basic data and skill data of enterprises are quantified to generate comprehensive strength matrix A, skill feature matrix B and skill impact model matrix C. Matrix A includes numerical characteristics such as operating years, registered capital, and litigation dispute records. Matrix B contains the skill feature weights of enterprises, such as data on technical patents and software copyrights. The cosine similarity algorithm is used to construct the skill impact model C between enterprises. Each element of the matrix represents the degree of similarity of enterprises in skill characteristics.

[0076] The CNN, Transformer-XL and GNN models are used to model and optimize the skill characteristics, comprehensive strength characteristics and inter-enterprise skill influence characteristics of enterprises, and extract enterprise feature relationships and time series information. The sparrow search algorithm (SSA) is used to optimize the hyperparameters of the XGBoost model. The goal is to minimize the mean square error (MSE) by adjusting parameters such as the number of iterations, tree depth, and learning rate, thereby improving the accuracy of the recommendation model. Finally, the optimal feature vectors of each dimension are obtained through multi-level model ensemble learning.

[0077] By integrating the learned feature weights and using the cosine similarity algorithm, we calculate the similarity scores between each enterprise and the target enterprise in different dimensions, and the enterprise with the highest matching degree is recommended and output. The final recommendation list is displayed in descending order of matching degree for further use. The following table shows the test data of the enterprise matching recommendation model built based on this method: Table 1 Model test data

[0078] By analyzing the data in the above table, we can see that: First, from the perspective of comprehensive strength score, enterprise D has the highest score (0.88). Enterprises with strong comprehensive strength are given higher priority in the recommendation model. At the same time, its skill feature score (0.85) is also much higher than other enterprises, indicating that strong comprehensive capabilities have a higher weight priority in this model. Enterprise D has a long operating period (20 years), few litigation disputes (0 times), and its bidding and scientific and technological innovation scores are both at a high level. These factors show that this method not only matches in terms of skill features, but also fully considers the basic enterprise performance; Secondly, the skill characteristic score of enterprise B is 0.81. Although its overall strength is slightly lower (0.65), it has a higher score for scientific and technological innovation (88), so it has received a weighted bonus in the skill matching degree, indicating that the model also gives higher recommendation priority to enterprises with higher innovation capabilities. In contrast, the overall strength score of enterprise E is only 0.55, and its skill characteristic score is also at a low level (0.60). Its bidding and scientific and technological innovation scores are low (65 and 60), which led to its automatic elimination in the final recommendation. This process reflects the dual consideration of the comprehensive strength and skills of enterprises by this method; Finally, in terms of inter-enterprise skill similarity, the skill feature scores of Enterprise A and Enterprise F are relatively close (0.76 and 0.73, respectively), and their comprehensive strength scores are also similar (0.82 and 0.79), which means that this method can effectively screen out enterprises with similar skills and comprehensive strength as candidates, and the matching results have strong rationality and reference value. In traditional technologies, it is difficult to take into account both the skill characteristics and comprehensive strength characteristics of an enterprise at the same time. This method can ensure that the recommended list has higher objectivity and applicability through multi-dimensional analysis and matching, effectively making up for the shortcomings of existing methods in terms of refinement and diversity.

[0079] It should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention rather than to limit it. Although the present invention has been described in detail with reference to the preferred embodiments, those skilled in the art should understand that the technical solutions of the present invention may be modified or replaced by equivalents without departing from the spirit and scope of the technical solutions of the present invention, which should all be included in the scope of the claims of the present invention.

Claims

1. A method for recommending enterprises based on a multi-dimensional cosine similarity algorithm, characterized in that: include: Collect enterprise information data, normalize and label it, and form enterprise basic data and skill data; The basic data and skill data of enterprises are vectorized to form the enterprise comprehensive strength matrix A, skill feature matrix B and inter-enterprise skill influence model matrix C, and construct data sets D1, D2 and D3; For different dimensions, CNN, Transformer-XL and GNN models are used to learn enterprise skill characteristics, comprehensive strength characteristics and skill impact characteristics, optimize model parameters and improve prediction performance; The SSA algorithm is used to optimize the hyperparameters of XGBoost, and the optimal parameters are obtained through MAE and MSE performance evaluation; The enterprises with higher matching degrees are sorted to generate a recommendation list, and the final list of recommended enterprises is output.

2. The enterprise recommendation method based on the multi-dimensional cosine similarity algorithm according to claim 1, characterized in that: The formation of the enterprise comprehensive strength matrix A includes: Standardize and vectorize the basic attribute data of the enterprise; For different numerical characteristics of enterprises, normalization is used to scale the data to the same dimension; For different categorical features of enterprises, one-hot encoding is used to generate discrete feature vectors; The formation of the skill feature matrix B includes: Use TF-IDF to vectorize text-based skill data and measure the weight of enterprise skill features; Combining the processed skill feature vectors into a skill feature matrix B; The formation of the inter-enterprise skill impact model matrix C includes: Cosine similarity is used to measure the similarity of skill vectors between enterprises; Fill the similarity values ​​between all pairs of enterprises into the inter-enterprise skill influence model matrix C.

3. The enterprise recommendation method based on the multi-dimensional cosine similarity algorithm according to claim 2, characterized in that: The method of using TF-IDF to vectorize text-based skill data includes: in, is the frequency of word t in document d, N is the total number of enterprises, is the number of enterprise documents containing word t.

4. The enterprise recommendation method based on the multi-dimensional cosine similarity algorithm according to claim 3, characterized in that: The use of CNN, Transformer-XL and GNN models to learn enterprise skill characteristics, comprehensive strength characteristics and skill influence characteristics respectively includes the following steps: Input the enterprise's skill feature matrix B into the convolutional neural network model to learn the enterprise's characteristics in the skill dimension; The Transformer-XL model is used to model the enterprise comprehensive strength feature matrix A, capturing the temporal correlation and long-term dependency of the enterprise's comprehensive strength in multiple aspects; For the inter-enterprise skill influence matrix C, a graph neural network is used to model the similarity relationship of inter-enterprise skills and learn the skill influence between different enterprises.

5. The enterprise recommendation method based on the multi-dimensional cosine similarity algorithm according to claim 4, characterized in that: The step of inputting the enterprise's skill feature matrix B into the convolutional neural network model comprises the following steps: Each row in the skill feature matrix B represents a skill feature vector of an enterprise; Extract n-gram features of different lengths from enterprise skills through a one-dimensional convolutional layer; Use the maximum pooling layer to reduce the dimension of the feature map and extract the most representative features of the skill features. ; The convolution and pooling feature vectors are input into the fully connected layer to calculate the skill prediction value of each enterprise. : in, and are the weights and biases of the fully connected layer; The mean square error is selected as the loss function to measure the difference between the predicted value and the true skill feature value.

6. The enterprise recommendation method based on the multi-dimensional cosine similarity algorithm according to claim 5, characterized in that: The use of the Transformer-XL model to model the enterprise comprehensive strength feature matrix A includes the following steps: The characteristic vector of each enterprise in the comprehensive strength characteristic matrix A is used as the input sequence; Using a multi-head self-attention layer, the attention weight is calculated for the input feature vector to obtain the global dependency relationship of the comprehensive strength characteristics of the enterprise: Among them, Q, K and V are query, key and value respectively, and the matrix is the scaling factor; The relative position encoding mechanism of Transformer-XL is used to retain the sequential information of enterprise features.

7. The enterprise recommendation method based on the multi-dimensional cosine similarity algorithm according to claim 6, characterized in that: The steps for implementing the SSA algorithm to train the XGBoost algorithm are as follows: Normalize the data; Initialize the parameters of the sparrow search algorithm, the parameters of the XGBoost model and the threshold of the boosting parameters; The fitness function Fi is the root mean square error between the fitted value and the measured value of the training set, and Fi is used to calculate the fitness value fi of the hemp population; Sort the sparrow fitness values ​​fi and select the optimal value fg; Updates the locations of spotters, followers, and sentinels in a sparrow population; If the optimal fitness value fg obtained in the current iteration is less than the accuracy set by the model, the iteration is terminated; or if the number of iterations is exhausted, the iterative search is terminated; The global optimal position xbj at the end of the iteration is used to optimize the parameters of the XGBoost model; Use the XGBoost algorithm to train data and make predictions, and output the prediction results.

8. The enterprise recommendation method based on the multi-dimensional cosine similarity algorithm according to claim 7, characterized in that: The generation process of the recommendation list is as follows: Using the enterprise skill feature model, the inter-enterprise skill impact model, the enterprise comprehensive capability model and the data of three dimensions, deriving the feature vectors of each dimension of the corresponding target enterprise; The feature vectors of each dimension calculated for each enterprise are looped with the feature vectors of each dimension of the target enterprise, and the cosine similarity is compared according to the weights learned by ensemble learning, and the enterprises with greater similarity are selected for recommendation.

9. A computer device comprising a memory and a processor, wherein the memory stores a computer program, wherein: When the processor executes the computer program, the steps of the enterprise recommendation method based on the multi-dimensional cosine similarity algorithm described in any one of claims 1 to 8 are implemented.

10. A computer-readable storage medium having a computer program stored thereon, characterized in that: When the computer program is executed by a processor, the steps of the enterprise recommendation method based on a multi-dimensional cosine similarity algorithm described in any one of claims 1 to 8 are implemented.

Images