Enterprise recommendation method and device based on feature fusion

Abstract

The invention provides an enterprise recommendation method and device based on feature fusion, and relates to the technical field of big data. The feature fusion-based enterprise recommendation method comprises the steps of obtaining a policy training set, an enterprise training set, a news training set and a public opinion training set; generating a multivariate heterogeneous graph according to the policy training set, the enterprise training set, the news training set and the public opinion training set; obtaining a heterogeneous graph neural network; training the heterogeneous graph neural network according to a multivariate heterogeneous graph to obtain a trained heterogeneous graph neural network; obtaining policy information to be predicted; and inputting the policy information to be predicted into the trained heterogeneous graph neural network so as to obtain predicted enterprise information. According to the enterprise recommendation method based on feature fusion, a multigraph structure of enterprise-policy-third-party institution published news-public opinion information is introduced, more comprehensive information integration is achieved, and complex nonlinear relations and potential modes in data can be captured.

Description

Technical Field

[0001] The present invention belongs to the technical field of large models, and specifically relates to a method for recommending an enterprise based on feature fusion and an apparatus for recommending an enterprise based on feature fusion. Background Art

[0002] The main goal of policy recommendation is to proactively provide valuable and targeted information based on the situation of the enterprise. The performance of policy recommendation in existing technologies is usually limited by cold start and data sparsity. In order to solve these problems, the use of auxiliary information is a very promising direction. In recent years, many deep learning-based methods have attempted to integrate auxiliary information for recommendation, which has achieved performance improvements compared to traditional methods, but few models can process graph structure information. Therefore, how to provide personalized and accurate service recommendations to enterprises based on the graph structure information of enterprise data resources and make good use of enterprise information and service resource information has become a problem to be solved. The enterprise recommendation methods in existing technologies usually use a bipartite graph approach, which generally only considers the relationship between enterprises and policies, while ignoring the influence of other factors. Summary of the invention

[0003] In view of the deficiencies of the prior art, the present invention provides an enterprise recommendation method based on feature fusion to solve the above technical problems.

[0004] To achieve the above object, the present invention provides the following technical solution: a method for recommending enterprises based on feature fusion, the method for recommending enterprises based on feature fusion comprising:

[0005] Obtain policy training sets, enterprise training sets, news training sets, and public opinion training sets;

[0006] Generate a multivariate heterogeneous graph based on the policy training set, the enterprise training set, the news training set, and the public opinion training set;

[0007] Obtain heterogeneous graph neural network;

[0008] Training the heterogeneous graph neural network according to the multivariate heterogeneous graph, thereby obtaining a trained heterogeneous graph neural network;

[0009] Obtain policy information to be predicted;

[0010] The policy information to be predicted is input into the trained heterogeneous graph neural network to obtain the predicted enterprise information.

[0011] Optionally, the obtaining of a policy training set, an enterprise training set, a news training set, and a public opinion training set includes:

[0012] Obtain policy text datasets, enterprise text datasets, news text datasets, and public opinion text datasets;

[0013] Cleaning the policy text dataset, the enterprise text dataset, the news text dataset, and the public opinion text dataset;

[0014] Through word embedding feature extraction technology, features are extracted from the cleaned policy text dataset, enterprise text dataset, news text dataset, and public opinion text dataset respectively, so as to obtain policy text vector training set, policy text vector training set, enterprise text vector training set, and news text vector training set.

[0015] Optionally, generating a multivariate heterogeneous graph according to the policy training set, the enterprise training set, the news training set, and the public opinion training set includes:

[0016] Acquire a policy node according to the policy training set;

[0017] Acquire enterprise nodes according to the enterprise training set;

[0018] Acquire news nodes according to the news training set;

[0019] Acquire public opinion nodes according to the public opinion training set;

[0020] Construct edge information between enterprise nodes and policy nodes;

[0021] Construct side information between enterprise nodes and news nodes;

[0022] Construct edge information between policy nodes and news nodes;

[0023] Construct side information between opinion nodes, policy nodes, news nodes, and enterprise nodes;

[0024] A multivariate heterogeneous graph is generated based on the obtained policy nodes, enterprise nodes, news nodes, public opinion nodes, edge information between enterprise nodes and policy nodes, edge information between enterprise nodes and news nodes, edge information between policy nodes and news nodes, and edge information between public opinion nodes and policy nodes, news nodes, and enterprise nodes.

[0025] Optionally, the step of constructing the side information between the enterprise node and the news node includes:

[0026] Get the release time of each piece of news in the news training set respectively;

[0027] Perform the following operations for each enterprise:

[0028] Calculate the similarity between the enterprise vector in the enterprise text vector training set and each news vector in each news text vector training set as the basic weight of the enterprise news;

[0029] The basic weight is adjusted through the time decay function of corporate news to obtain the final weight of corporate news.

[0030] Optionally, the constructing of the side information between the policy node and the news node includes:

[0031] Get the release time of each piece of news in the news training set respectively;

[0032] Perform the following operations for each policy:

[0033] Calculate the similarity between the policy vector in the policy text vector training set and each news vector in each news text vector training set as the policy news basic weight;

[0034] The basic weight is adjusted through the time decay function of policy news to obtain the final weight of policy news.

[0035] Optionally, the step of training the heterogeneous graph neural network according to the multivariate heterogeneous graph to obtain a trained heterogeneous graph neural network for training includes:

[0036] Constructing a heterogeneous graph neural network, wherein the heterogeneous graph neural network includes a graph convolution layer, a heterogeneous interaction layer, a contrastive learning layer, and an output layer;

[0037] The heterogeneous graph neural network is jointly trained through a joint training method.

[0038] Optionally, the joint training of the heterogeneous graph neural network by joint training includes:

[0039] Establishing joint training tasks, including predicting the connection relationship between enterprises and policies and their weights, predicting the connection relationship between enterprises and news and their weights, predicting the connection relationship between policies and news and their weights, predicting the connection relationship between public opinion and enterprises, policies, news and their weights, public opinion sentiment classification tasks and news influence prediction tasks;

[0040] Get the total loss function;

[0041] The heterogeneous graph neural network is trained, the total loss function is minimized by a gradient descent algorithm, and the parameters of the heterogeneous graph neural network are updated by a back propagation algorithm.

[0042] Optionally, the total loss function formula is as follows:

[0043] ;in,

[0044] Total Loss is the total loss function; Loss 1is the connection relationship between enterprises and policies and their weight loss function; Loss 2 is the connection relationship between enterprises and news and its weight loss function; Loss 3 is the connection relationship between policy and news and its weight loss function; Loss 4 is the connection relationship between public opinion and enterprises, policies, and news and its weight loss function; Loss 5 is the loss function for the public opinion sentiment classification task; Loss 6 A is the loss function for the news influence prediction task; α , B β , C γ , D δ 、E ϵ and F ζ is a weight coefficient used to balance the impact of different tasks on the total loss.

[0045] Optionally, the connection relationship between the enterprise and the policy and its weight loss function are obtained by the following formula:

[0046] ;in,

[0047] Loss 1 is the connection relationship between enterprises and policies and its weight loss function; N is the total number of samples; is the sample index; ω i For the Dynamic weight of each sample, used to adjust the impact of the sample on the total loss; For the The true value of samples; For the The predicted value of samples;

[0048] The connection relationship between the enterprise and the news and its weight loss function are obtained by the following formula:

[0049] ;in,

[0050] is the balance factor, which is set according to the ratio of positive and negative samples; γ is the adjustment factor; Loss 2 is the connection relationship between enterprises and news and its weight loss function; N is the total number of samples; is the sample index; For the The true value of samples; For the The predicted value of samples;

[0051] The connection relationship between the policy and the news and its weight loss function are obtained by the following formula:

[0052] ;in,

[0053] Loss 3 is the connection relationship between policy and news and its weight loss function; δ is the preset threshold; N is the total number of samples; is the sample index; For the The true value of samples; For the The predicted value of samples;

[0054] The connection relationship between public opinion and enterprises, policies, and news and their weight loss function are obtained by the following formula:

[0055] ;in,

[0056] Loss 4 is the connection relationship between public opinion and enterprises, policies, and news, as well as their weight loss function; is the smoothing factor; N is the total number of samples; is the sample index; For the The true value of samples; For the The predicted value of samples; C is the total number of categories; c is the category index; For the The true label of samples in the cth category; For the The predicted probability of a sample in the cth category;

[0057] The loss function of the public opinion sentiment classification task includes:

[0058] ;in,

[0059] Loss 5 is the loss function of public opinion sentiment classification task; λ is the balance factor; N is the total number of samples; is the sample index; For the The true label of samples in the cth category; For the The predicted probability of a sample in the cth category; For the The embedding vector of each sample, For the The center vector of the category to which the sample belongs;

[0060] The news influence prediction task loss function includes:

[0061] ;in,

[0062] Loss 6 is the loss function for the news influence prediction task; τ is the quantile; N is the total number of samples; i is the sample index; For the The true value of samples; For the The predicted value of samples; is the quantile loss function, where u is .

[0063] The present application also provides a device for recommending enterprises based on feature fusion, and the device for recommending enterprises based on feature fusion includes:

[0064] A training set acquisition module, which is used to acquire a policy training set, an enterprise training set, a news training set, and a public opinion training set;

[0065] A multivariate heterogeneous graph generation module, the multivariate heterogeneous graph generation module is used to generate a multivariate heterogeneous graph according to the policy training set, the enterprise training set, the news training set and the public opinion training set;

[0066] A heterogeneous graph neural network acquisition module, wherein the heterogeneous graph neural network acquisition module is used to acquire a heterogeneous graph neural network;

[0067] A training module, wherein the training module is used to train the heterogeneous graph neural network according to the multivariate heterogeneous graph, so as to obtain a trained heterogeneous graph neural network;

[0068] A module for acquiring information about policies to be predicted, wherein the module is used to acquire information about policies to be predicted;

[0069] A predicted enterprise information acquisition module is used to input the policy information to be predicted into a trained heterogeneous graph neural network, thereby acquiring predicted enterprise information.

[0070] The enterprise recommendation method based on feature fusion in this application introduces a multi-part graph structure of enterprise-policy-news and public opinion information released by third-party agencies to achieve more comprehensive information integration and capture complex nonlinear relationships and potential patterns in the data. It can solve the problem of inaccurate recommendation results in the existing technology when the amount of policy data is small and there is a lack of comprehensive consideration of external information of enterprises (such as news and public opinion information released by third-party agencies). BRIEF DESCRIPTION OF THE DRAWINGS

[0071] Figure 1A flowchart of a method for recommending enterprises based on feature fusion according to an embodiment of the present application. DETAILED DESCRIPTION

[0072] The following will be combined with the drawings in the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0073] like Figure 1 The enterprise recommendation methods based on feature fusion shown include:

[0074] Step 1: Obtain policy training sets, enterprise training sets, news training sets, and public opinion training sets;

[0075] Step 2: Generate a multivariate heterogeneous graph based on the policy training set, the enterprise training set, the news training set, and the public opinion training set;

[0076] Step 3: Obtain heterogeneous graph neural network;

[0077] Step 4: training the heterogeneous graph neural network according to the multivariate heterogeneous graph, thereby obtaining a trained heterogeneous graph neural network;

[0078] Step 5: Obtain the policy information to be predicted;

[0079] Step 6: Input the policy information to be predicted into the trained heterogeneous graph neural network to obtain the predicted enterprise information.

[0080] The enterprise recommendation method based on feature fusion in this application introduces a multi-part graph structure of enterprise-policy-news and public opinion information released by third-party agencies to achieve more comprehensive information integration and capture complex nonlinear relationships and potential patterns in the data. It can solve the problem of inaccurate recommendation results in the existing technology when the amount of policy data is small and there is a lack of comprehensive consideration of external information of enterprises (such as news and public opinion information released by third-party agencies).

[0081] In this embodiment, the acquisition of the policy training set, the enterprise training set, the news training set, and the public opinion training set includes:

[0082] Obtain policy text datasets, enterprise text datasets, news text datasets, and public opinion text datasets;

[0083] Cleaning the policy text dataset, the enterprise text dataset, the news text dataset, and the public opinion text dataset;

[0084] Through word embedding feature extraction technology, features are extracted from the cleaned policy text dataset, enterprise text dataset, news text dataset, and public opinion text dataset respectively, so as to obtain policy text vector training set, policy text vector training set, enterprise text vector training set, and news text vector training set.

[0085] In this embodiment, the policy text dataset, enterprise text dataset, news text dataset, and public opinion text dataset are all text information, among which the policy text may include applicable industries, laws and regulations, issuing agencies, etc., and the text information in the enterprise text dataset may include the company name, number of employees, total profit, etc. of each company.

[0086] In this embodiment, the news information in the news text data set may include various information of various enterprises, which may be obtained through news websites and the like.

[0087] The public opinion information in the public opinion text dataset can be obtained through various news websites, social media and other channels.

[0088] In this embodiment, cleaning the policy text dataset, the enterprise text dataset, the news text dataset, and the public opinion text dataset refers to cleaning the collected text data, such as removing irrelevant characters, HTML tags, stop words, etc.

[0089] In this embodiment, generating a multivariate heterogeneous graph according to the policy training set, the enterprise training set, the news training set, and the public opinion training set includes:

[0090] Acquire a policy node according to the policy training set. It can be understood that the policy training set includes multiple policies, and one policy corresponds to one policy node;

[0091] Acquire an enterprise node according to the enterprise training set. It can be understood that the enterprise training set includes multiple enterprises, and one enterprise corresponds to one enterprise node;

[0092] Acquire a news node according to the news training set. It can be understood that the news training set includes a plurality of news, and one news corresponds to one news node;

[0093] Obtaining a public opinion node according to the public opinion training set. It can be understood that the public opinion training set includes multiple public opinions, and one public opinion corresponds to one public opinion node;

[0094] Construct edge information between enterprise nodes and policy nodes. In this embodiment, the enterprise names or enterprise attributes (such as industry, scale, etc.) mentioned in the text of the policy training set are matched with the enterprise nodes to construct edges between enterprises and policies. A similarity function (such as cosine similarity) can be used to calculate the similarity between enterprise and policy text vectors as the weight of the edge.

[0095] Constructing the side information between the enterprise node and the news node. In this embodiment, constructing the side information between the enterprise node and the news node includes:

[0096] Get the release time of each piece of news in the news training set respectively;

[0097] Perform the following operations for each enterprise:

[0098] Calculate the similarity between the enterprise vector in the enterprise text vector training set and each news vector in each news text vector training set as the basic weight of the enterprise news;

[0099] The basic weight is adjusted through the time decay function of enterprise news to obtain the final weight of enterprise news. In order to more accurately reflect the correlation between enterprises and news, and take into account the time sensitivity of news (that is, newer news may have a greater impact on enterprises), we introduce a dynamic weight mechanism and use an exponential decay function to introduce a time factor into the edge weight.

[0100] Specifically, a similarity function (such as cosine similarity) is used to calculate the similarity between the enterprise and news text vectors as the enterprise news basic weight. The enterprise news basic weight is obtained by the following formula:

[0101] ;in,

[0102] is the vector representation of the enterprise node; is the vector representation of the news node; Basic weight for corporate news.

[0103] In this embodiment, an enterprise news exponential decay function is introduced to adjust the edge weights, taking into account the impact of news release time on the current state of the enterprise.

[0104] In this embodiment, the final weight is obtained by the following formula:

[0105] ;in,

[0106] is the final weight, As the basic weight, is the attenuation function; the attenuation function formula is as follows:

[0107] ; where t is the current time, t 0 is the news release time, and λ is the decay rate (a parameter that controls the decay speed).

[0108] Constructing side information between policy nodes and news nodes; in this embodiment, constructing side information between policy nodes and news nodes includes:

[0109] Get the release time of each piece of news in the news training set respectively;

[0110] Perform the following operations for each policy:

[0111] Calculate the similarity between the policy vector in the policy text vector training set and each news vector in each news text vector training set as the policy news basic weight;

[0112] The basic weight is adjusted through the time decay function of policy news to obtain the final weight of policy news.

[0113] In this embodiment, the edge weight between policy and news depends not only on the degree of interpretation and evaluation of the policy by the news, but also on the relationship between the time of news release and the release or validity period of the policy. Therefore, the time decay function of policy news is as follows:

[0114] ; where t is the current time, t p It is the policy release time. is the policy effective date or end time, is the decay rate. This function ensures that the influence of news gradually decays during the policy validity period, but no longer has any influence after the expiration.

[0115] In this embodiment, the basic weight of policy news is obtained by the following formula:

[0116] ; where f is a mapping function that converts the news’ attention to policy (or relevance score) into a weight.

[0117] In this embodiment, the news's attention to the policy can be obtained by the following method:

[0118] Obtain the weight of the number of times the policy is mentioned in the news and the weight of the evaluation tone, and then weight them together to obtain the basic weight of the policy news.

[0119] It is understandable that the method for obtaining the final weight of policy news is the same as the method for obtaining the final weight of corporate news, and will not be repeated here.

[0120] Construct the side information of public opinion nodes and policy nodes, news nodes and enterprise nodes; in this embodiment, this can be achieved by analyzing the sentiment tendency of public opinion (positive, negative, neutral), the scope of dissemination, the credibility of the source and other factors to obtain the public opinion basis weight. Specifically, the public opinion basis weight is obtained by the following formula:

[0121] ;in,

[0122] g is a mapping function that converts the influence factors of public opinion into weights, o represents the public opinion node, and p / e / n represent the policy, enterprise, and news nodes respectively.

[0123] In this embodiment, considering the timeliness of public opinion, a public opinion time decay function is used to adjust the public opinion basic weight. The adjustment method is the same as the adjustment methods of other basic weights, which will not be repeated here.

[0124] A multivariate heterogeneous graph is generated based on the obtained policy nodes, enterprise nodes, news nodes, public opinion nodes, edge information between enterprise nodes and policy nodes, edge information between enterprise nodes and news nodes, edge information between policy nodes and news nodes, and edge information between public opinion nodes and policy nodes, news nodes, and enterprise nodes.

[0125] In this embodiment, the acquired node and edge information are integrated into a multivariate heterogeneous graph, the node set of the multivariate heterogeneous graph includes an enterprise node set, a policy node set, a news node set and a public opinion node set; the edge set includes all the edges constructed above.

[0126] You can use graph databases or graph processing frameworks (such as Neo4j, NetworkX, etc.) to store multivariate heterogeneous graphs.

[0127] In this embodiment, the step of training the heterogeneous graph neural network according to the multivariate heterogeneous graph to obtain the trained heterogeneous graph neural network for training includes:

[0128] Constructing a heterogeneous graph neural network, wherein the heterogeneous graph neural network includes a graph convolution layer, a heterogeneous interaction layer, a contrastive learning layer, and an output layer;

[0129] The heterogeneous graph neural network is jointly trained through a joint training method.

[0130] In this embodiment, the input of the graph convolution layer is a multivariate heterogeneous graph, and the information of neighbor nodes is aggregated through graph convolution operations (such as GCN, GraphSAGE, etc.) to update the representation of the nodes. Considering the heterogeneity of the four-part graph, different convolution kernels or aggregation functions can be designed for different types of nodes and edges.

[0131] In this embodiment, the input of the heterogeneous interaction layer is the output of the graph convolution layer, which is used to capture the heterogeneous interaction relationship between enterprises, policies, news and public opinion information. This can be achieved through attention mechanisms, bilinear transformations, or heterogeneous message passing mechanisms in graph neural networks. Specifically, a specific interaction function can be designed for each pair of node types (such as enterprise and policy, enterprise and news, etc.) to learn the complex relationship between them.

[0132] In this embodiment, the input of the contrastive learning layer is the output of the heterogeneous interaction layer, and the contrastive learning layer is used to optimize the representation of the node through contrastive learning. This can be achieved by generating different views of the node (such as through data enhancement technology), maximizing the consistency of positive pairs (representations of the same node in different views), and minimizing the consistency of negative pairs (representations of different nodes in different views). The contrastive loss function can be used to guide the training of the model.

[0133] In this embodiment, the input of the output layer is the output of the contrastive learning layer. The output layer is used to calculate the matching degree between each policy node and the enterprise node according to the input policy information to be predicted, and output a list of recommended results.

[0134] In this embodiment, the present application trains the heterogeneous graph neural network of the present application by means of joint training. Specifically, the joint training of the heterogeneous graph neural network by means of joint training includes:

[0135] Establishing joint training tasks, including predicting the connection relationship between enterprises and policies and their weights, predicting the connection relationship between enterprises and news and their weights, predicting the connection relationship between policies and news and their weights, predicting the connection relationship between public opinion and enterprises, policies, news and their weights, public opinion sentiment classification tasks and news influence prediction tasks;

[0136] Get the total loss function;

[0137] The heterogeneous graph neural network is trained, the total loss function is minimized by a gradient descent algorithm, and the parameters of the heterogeneous graph neural network are updated by a back propagation algorithm.

[0138] In this embodiment, a multi-task joint training method is used to train a heterogeneous graph neural network to capture the complex relationships between different types of nodes and edges in the graph, and further improve the generalization ability and matching accuracy of the model.

[0139] In this embodiment, each task has its own loss function, so the total loss function formula is as follows:

[0140] ;in,

[0141] Total Loss is the total loss function, Loss 1 is the connection relationship between enterprises and policies and their weight loss function, Loss 2 is the connection relationship between enterprises and news and its weight loss function, Loss 3 is the connection relationship between policy and news and its weight loss function, Loss 4 is the connection relationship between public opinion and enterprises, policies, and news, as well as its weight loss function, Loss 5 is the loss function of the public opinion sentiment classification task, Loss 6 A is the loss function for the news influence prediction task; α , B β , C γ , D δ 、E ϵ and F ζ is a weight coefficient used to balance the impact of different tasks on the total loss.

[0142] In this embodiment, the connection relationship between the enterprise and the policy and its weight loss function are obtained by the following formula:

[0143] ;in,

[0144] Loss 1 is the connection relationship between enterprises and policies and its weight loss function; N is the total number of samples; is the sample index, indicating the samples; i For the Dynamic weight of each sample, used to adjust the impact of the sample on the total loss; For the The true value of samples; For the The predicted value of a sample.

[0145] In this embodiment, the weight ω i It can be adjusted according to the prediction difficulty of the sample. For example, you can set ,in is a small positive number to prevent the denominator from being zero. This method gives greater weight to samples with larger prediction errors.

[0146] In this embodiment, the weight ω i This can be determined by:

[0147] Train the model using the initial weight calculation method and evaluate the performance on the validation set.

[0148] Try different value, adjust the weight calculation method, and retrain the model.

[0149] Compare the performance of the models on the validation set under different weight adjustment methods and select the weight adjustment method with the best performance.

[0150] The connection relationship between the enterprise and the news and its weight loss function are obtained by the following formula:

[0151] ;in,

[0152] is the balance factor, which is set according to the ratio of positive and negative samples; γ is the adjustment factor; Loss 2 is the connection relationship between enterprises and news and its weight loss function; N is the total number of samples; is the sample index, indicating the samples; For the The true value of samples; For the The predicted value of a sample.

[0153] In this embodiment, for example, if there are few positive samples, we can set For positive samples, =1 for negative samples.

[0154] In this embodiment, the value of γ is between 0 and 5 and can be set as needed.

[0155] The connection relationship between the policy and the news and its weight loss function are obtained by the following formula:

[0156] ;in,

[0157] Loss 3 is the connection relationship between policy and news and its weight loss function; δ is the preset threshold; N is the total number of samples; is the sample index, indicating the samples; For the The true value of samples; For the The predicted value of samples;

[0158] In this embodiment, the value range of δ can be set manually, and after setting, the most appropriate δ is selected by evaluating the performance on the validation set.

[0159] The connection relationship between public opinion and enterprises, policies, and news and their weight loss function are obtained by the following formula:

[0160] ;in,

[0161] Loss 4 is the connection relationship between public opinion and enterprises, policies, and news, as well as their weight loss function; is the smoothing factor; N is the total number of samples; is the sample index, indicating the samples; For the The true value of samples; For the The predicted value of samples; C is the total number of categories; c is the category index, indicating the cth category; For the The true label of samples in the cth category, For the The predicted probability of a sample in the cth category;

[0162] In this embodiment, The value range is between 0 and 0.2.

[0163] The loss function of the public opinion sentiment classification task includes:

[0164] ;in,

[0165] Loss 5 is the loss function of public opinion sentiment classification task; λ is the balance factor; N is the total number of samples; is the sample index; For the The true label of samples in the cth category; For the The predicted probability of a sample in the cth category; For the The embedding vector of each sample, For the The center vector of the category to which the sample belongs;

[0166] In this embodiment, λ is set according to the importance of the cross entropy loss and the center loss.

[0167] The news influence prediction task loss function includes:

[0168] ;in,

[0169] Loss 6 is the loss function for the news influence prediction task; τ is the quantile; N is the total number of samples; is the sample index; For the The true value of samples; For the The predicted value of samples; is the quantile loss function, where u is .

[0170] In this embodiment, τ can be set as needed. For example, if the median prediction performance is of concern, τ=0.5 is set.

[0171] The above loss functions have the following advantages:

[0172] (1) The connection between enterprises and policies and their weight loss function add dynamic weights, which can be adjusted according to the prediction difficulty of the sample. For example, for samples with large prediction errors, larger weights can be given to encourage the model to pay more attention to these difficult-to-predict samples.

[0173] (2) The connection relationship between enterprises and news and their weight loss function are added with a balancing factor and an adjustment factor. The balancing factor is used to balance the ratio of positive and negative samples; the adjustment factor is used to control the loss contribution of easy-to-classify samples.

[0174] (3) The connection between policy and news and its weighted loss function control the switching point between quadratic loss and linear loss through a threshold to enhance the robustness to outliers.

[0175] (4) The connection between public opinion and enterprises, policies, and news and their weight loss function converts hard labels (one-hot encoding) into soft labels through a smoothing factor, thereby preventing the model from overfitting.

[0176] (5) The loss function of the public opinion sentiment classification task balances the ratio of cross entropy loss and center loss by setting a balance factor.

[0177] (6) The loss function of the news influence prediction task uses quantile loss to capture errors at different quantiles, thereby more comprehensively evaluating the prediction performance of the model.

[0178] Through the coordination of the above-mentioned loss models, the generalization ability and prediction accuracy of the model can be improved. These improvement strategies are optimized in terms of sample difficulty, class imbalance, outlier robustness, overfitting prevention, intra-class compactness, and error capture at different quantiles.

[0179] The present application also provides a device for recommending enterprises based on feature fusion, which comprises a training set acquisition module, a multivariate heterogeneous graph generation module, a heterogeneous graph neural network acquisition module, a training module, a module for acquiring information about policies to be predicted, and a module for acquiring information about predicted enterprises, wherein:

[0180] The training set acquisition module is used to acquire policy training sets, enterprise training sets, news training sets and public opinion training sets;

[0181] The multivariate heterogeneous graph generation module is used to generate a multivariate heterogeneous graph according to the policy training set, the enterprise training set, the news training set and the public opinion training set;

[0182] The heterogeneous graph neural network acquisition module is used to acquire the heterogeneous graph neural network;

[0183] The training module is used to train the heterogeneous graph neural network according to the multivariate heterogeneous graph, so as to obtain a trained heterogeneous graph neural network;

[0184] The module for obtaining information on policies to be predicted is used to obtain information on policies to be predicted;

[0185] The predicted enterprise information acquisition module is used to input the policy information to be predicted into the trained heterogeneous graph neural network, so as to obtain the predicted enterprise information.

[0186] It should be noted that, in this article, relational terms such as first and second, etc. are only used to distinguish one entity or operation from another entity or operation, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Moreover, the terms "include", "comprise" or any other variants thereof are intended to cover non-exclusive inclusion, so that a process, method, article or device including a series of elements includes not only those elements, but also other elements not explicitly listed, or also includes elements inherent to such process, method, article or device.

[0187] Although embodiments of the present invention have been shown and described, it will be appreciated by those skilled in the art that various changes, modifications, substitutions and variations may be made to the embodiments without departing from the principles and spirit of the present invention, and that the scope of the present invention is defined by the appended claims and their equivalents.

Claims

1. A method for enterprise recommendation based on feature fusion, characterized in that: The enterprise recommendation method based on feature fusion includes: Obtain policy training sets, enterprise training sets, news training sets, and public opinion training sets; Generate a multivariate heterogeneous graph based on the policy training set, the enterprise training set, the news training set, and the public opinion training set; Obtain heterogeneous graph neural network; Training the heterogeneous graph neural network according to the multivariate heterogeneous graph, thereby obtaining a trained heterogeneous graph neural network; Obtain policy information to be predicted; The policy information to be predicted is input into the trained heterogeneous graph neural network to obtain the predicted enterprise information.

2. The enterprise recommendation method based on feature fusion according to claim 1, characterized in that: The acquisition of the policy training set, the enterprise training set, the news training set and the public opinion training set includes: Obtain policy text datasets, enterprise text datasets, news text datasets, and public opinion text datasets; Cleaning the policy text dataset, the enterprise text dataset, the news text dataset, and the public opinion text dataset; Through word embedding feature extraction technology, features are extracted from the cleaned policy text dataset, enterprise text dataset, news text dataset, and public opinion text dataset respectively, so as to obtain policy text vector training set, policy text vector training set, enterprise text vector training set, and news text vector training set.

3. The enterprise recommendation method based on feature fusion according to claim 2, characterized in that: Generating a multivariate heterogeneous graph according to the policy training set, the enterprise training set, the news training set, and the public opinion training set includes: Acquire a policy node according to the policy training set; Acquire enterprise nodes according to the enterprise training set; Acquire news nodes according to the news training set; Acquire public opinion nodes according to the public opinion training set; Construct edge information between enterprise nodes and policy nodes; Construct side information between enterprise nodes and news nodes; Construct edge information between policy nodes and news nodes; Construct side information between opinion nodes, policy nodes, news nodes, and enterprise nodes; A multivariate heterogeneous graph is generated based on the obtained policy nodes, enterprise nodes, news nodes, public opinion nodes, edge information between enterprise nodes and policy nodes, edge information between enterprise nodes and news nodes, edge information between policy nodes and news nodes, and edge information between public opinion nodes and policy nodes, news nodes, and enterprise nodes.

4. The enterprise recommendation method based on feature fusion according to claim 3, characterized in that: The construction of the side information between the enterprise node and the news node includes: Get the release time of each piece of news in the news training set respectively; Perform the following operations for each enterprise: Calculate the similarity between the enterprise vector in the enterprise text vector training set and each news vector in each news text vector training set as the basic weight of the enterprise news; The basic weight is adjusted through the time decay function of corporate news to obtain the final weight of corporate news.

5. The enterprise recommendation method based on feature fusion according to claim 4, characterized in that: The side information between the policy node and the news node is constructed as follows: Get the release time of each piece of news in the news training set respectively; Perform the following operations for each policy: Calculate the similarity between the policy vector in the policy text vector training set and each news vector in each news text vector training set as the policy news basic weight; The basic weight is adjusted through the time decay function of policy news to obtain the final weight of policy news.

6. The enterprise recommendation method based on feature fusion according to claim 5, characterized in that: The step of training the heterogeneous graph neural network according to the multivariate heterogeneous graph to obtain the trained heterogeneous graph neural network for training includes: Constructing a heterogeneous graph neural network, wherein the heterogeneous graph neural network includes a graph convolution layer, a heterogeneous interaction layer, a contrastive learning layer, and an output layer; The heterogeneous graph neural network is jointly trained through a joint training method.

7. The enterprise recommendation method based on feature fusion according to claim 6, characterized in that: The joint training of the heterogeneous graph neural network by joint training includes: Establishing joint training tasks, including predicting the connection relationship between enterprises and policies and their weights, predicting the connection relationship between enterprises and news and their weights, predicting the connection relationship between policies and news and their weights, predicting the connection relationship between public opinion and enterprises, policies, news and their weights, public opinion sentiment classification tasks and news influence prediction tasks; Get the total loss function; The heterogeneous graph neural network is trained, the total loss function is minimized by a gradient descent algorithm, and the parameters of the heterogeneous graph neural network are updated by a back propagation algorithm.

8. The enterprise recommendation method based on feature fusion according to claim 7, characterized in that: The total loss function formula is as follows: ;in, Total Loss is the total loss function; Loss1 is the connection between enterprises and policies and its weight loss function; Loss2 is the connection between enterprises and news and its weight loss function; Loss3 is the connection between policies and news and its weight loss function; Loss4 is the connection between public opinion and enterprises, policies, and news and its weight loss function; Loss5 is the loss function of public opinion sentiment classification task; Loss6 is the loss function of news influence prediction task; A α , B β , C γ , D δ 、E ϵ and F ζ is a weight coefficient used to balance the impact of different tasks on the total loss.

9. The enterprise recommendation method based on feature fusion according to claim 8, characterized in that: The connection relationship between the enterprise and the policy and its weight loss function are obtained by the following formula: ;in, Loss1 is the connection relationship between enterprises and policies and its weight loss function; N is the total number of samples; is the sample index; ω i For the Dynamic weight of each sample, used to adjust the impact of the sample on the total loss; For the The true value of samples; For the The predicted value of samples; The connection relationship between the enterprise and the news and its weight loss function are obtained by the following formula: ;in, is the balance factor, which is set according to the ratio of positive and negative samples; γ is the adjustment factor; Loss2 is the connection relationship between enterprises and news and its weight loss function; N is the total number of samples; is the sample index; For the The true value of samples; For the The predicted value of samples; The connection relationship between the policy and the news and its weight loss function are obtained by the following formula: ;in, Loss3 is the connection relationship between policy and news and its weight loss function; δ is the preset threshold; N is the total number of samples; is the sample index; For the The true value of samples; For the The predicted value of samples; The connection relationship between public opinion and enterprises, policies, and news and their weight loss function are obtained by the following formula: ;in, Loss4 is the connection relationship between public opinion and enterprises, policies, and news, and its weight loss function; is the smoothing factor; N is the total number of samples; is the sample index; For the The true value of samples; For the The predicted value of samples; C is the total number of categories; c is the category index; For the The true label of samples in the cth category; For the The predicted probability of a sample in the cth category; The loss function of the public opinion sentiment classification task includes: ;in, Loss5 is the loss function for the public opinion sentiment classification task; λ is the balance factor; N is the total number of samples; is the sample index; For the The true label of samples in the cth category; For the The predicted probability of a sample in the cth category; For the The embedding vector of each sample, For the The center vector of the category to which the sample belongs; The news influence prediction task loss function includes: ;in, Loss6 is the loss function for the news influence prediction task; τ is the quantile; N is the total number of samples; is the sample index; For the The true value of samples; For the The predicted value of samples; is the quantile loss function, where u is .

10. An enterprise recommendation device based on feature fusion, characterized in that: The enterprise recommendation device based on feature fusion includes: A training set acquisition module, which is used to acquire a policy training set, an enterprise training set, a news training set, and a public opinion training set; A multivariate heterogeneous graph generation module, the multivariate heterogeneous graph generation module is used to generate a multivariate heterogeneous graph according to the policy training set, the enterprise training set, the news training set and the public opinion training set; A heterogeneous graph neural network acquisition module, wherein the heterogeneous graph neural network acquisition module is used to acquire a heterogeneous graph neural network; A training module, wherein the training module is used to train the heterogeneous graph neural network according to the multivariate heterogeneous graph, so as to obtain a trained heterogeneous graph neural network; A module for acquiring information about policies to be predicted, wherein the module is used to acquire information about policies to be predicted; A predicted enterprise information acquisition module is used to input the policy information to be predicted into a trained heterogeneous graph neural network, thereby acquiring predicted enterprise information.

Images