Enterprise valuation method and system based on heterogeneous graph neural network

Abstract

The present invention discloses a method and system for enterprise valuation based on heterogeneous graph neural network, which belongs to the field of graph neural network. The method includes the following steps: building an industry heterogeneous graph; sampling the neighbors of the enterprise to be valued; focusing on the sampling node attention aggregation learning; weighted fusion of the aggregation features of heterogeneous nodes; predicting enterprise valuation. The present invention adopts graph neural network as the basic model, introduces the connection information of company nodes and personnel nodes in the aggregation learning of heterogeneous graph neural network and fuses them through the attention mechanism; due to the introduction of graph neural network algorithm, the efficiency is greatly improved compared with traditional expert valuation, and the enterprise valuation can be quickly updated when the company and personnel information changes and the market fluctuates; focusing on company valuation from a high-level perspective from the perspective of industry field information, relying on massive data to effectively solve the cognitive bias and subjective limitations caused by experts' direct modeling of enterprises.

Description

Technical Field

[0001] The present invention belongs to the field of graph neural networks, and in particular relates to an enterprise valuation method and system based on heterogeneous graph neural networks. Background Art

[0002] Enterprise valuation technology is aimed at listed or unlisted companies. It relies on the company's valuation model to evaluate the intrinsic value of the company based on the company's assets and profitability. By analyzing the valuation model and exploring relevant influencing factors, the understanding of the industry and the company can be transformed into specific investment suggestions. In addition, the model can predict the impact of corporate strategies, capital transactions, etc. on corporate value. Reasonable evaluation of corporate value is the basis for investment, financing and various transactions.

[0003] Existing enterprise valuation methods rely on domain experts to comprehensively analyze industry information, company core resources and other data to estimate the company's market value. Traditional valuation methods are mainly divided into two types: cash discount method and relative valuation method. The former uses the expected cash flow of the company in a specific future period to restore the current present value. Starting from the basic situation of the company, the evaluation experts analyze various factors that affect the company's future profitability to make reasonable predictions on cash flow, and give a discount rate based on the company's future risk judgment, and make enterprise valuations through cash flow and discount rate; the latter finds comparable assets or companies, uses price indicators such as price-earnings ratio, price-to-book ratio, price-to-sales ratio, and price-to-cash ratio to compare with comparable assets or companies, and uses the value of comparable assets or companies to estimate the value of the company. After research, it was found that the main problems with existing valuation methods are:

[0004] 1) It requires domain experts to fully understand the industry and company situation and build models on this basis, which involves many subjective factors.

[0005] 2) When the market experiences large fluctuations, the relevant reference indicators will change significantly, which will mislead the company's value assessment.

[0006] 3) Due to limitations in labor costs and efficiency, it is difficult to respond promptly to short-term market fluctuations.

[0007] Deep learning is an algorithm in machine learning that is based on learning to represent data. Related models provide solutions for various industries as SAAS, and have become an important factor in promoting the upgrading and development of traditional industries. In recent years, the application of deep learning in graph structures has been growing. Researchers have borrowed the ideas of convolutional networks and recurrent networks to design models to process graph structure data, and a series of graph neural network algorithms for non-Euclidean data have emerged.

[0008] The valuation modeling of an enterprise not only needs to consider the characteristics of the company itself, such as basic information and operating conditions, but also requires complex industry field data formed by the interrelationships between companies. In addition to the tangible assets of the enterprise, talent resources and intellectual property rights with technology as capital are also becoming the main determinants of the company's future profitability. For this reason, companies and personnel are regarded as nodes and an industry heterogeneous graph structure is constructed. Summary of the invention

[0009] In order to solve the problem of subjectivity and inefficiency of expert methods for enterprise valuation, the present invention proposes an enterprise valuation method and system based on heterogeneous graph neural network, which effectively solves the influence of human subjective factors, can quickly capture market fluctuations and abnormal indicators and give timely valuation feedback.

[0010] The technical solution adopted by the present invention is as follows:

[0011] In a first aspect, a method for enterprise valuation based on a heterogeneous graph neural network is provided, comprising the following steps:

[0012] Step 1: Build an industry heterogeneous graph;

[0013] Step 2: Sampling of neighbors of the enterprise to be valued;

[0014] Step 3: Use attention aggregation learning on nodes;

[0015] Step 4: weighted fusion of aggregated features of heterogeneous nodes;

[0016] Step 5: Predict corporate valuation.

[0017] Furthermore, the industry heterogeneous graph is composed of two types of nodes: companies and personnel, which connect companies and personnel based on investment and employment relationships.

[0018] Furthermore, the random walk algorithm is used to perform normalized sampling of neighbor nodes of different types and numbers, as follows:

[0019] (1) Starting from the enterprise to be valued, a random neighbor of the current node is selected with probability r each time, and the initial enterprise to be valued is returned with probability 1-r, generating a node sequence of length N;

[0020] (2) The generated series of associated neighbor node sequences are divided into two groups according to type: company and personnel. For each type, the n nodes with the highest occurrence frequency are selected from the sequence to generate heterogeneous neighbor node sets of two types of companies to be valued.

[0021] Furthermore, attention aggregation learning is used for the sampling nodes, including the following steps:

[0022] (1) Starting from the enterprise to be valued, a set of company nodes and a set of personnel nodes are generated by random walks and input into the attention aggregation network respectively;

[0023] (2) A hierarchical multi-head attention weighted network that integrates relevant information is used to aggregate and learn each type of node. There are a total of L layers in the network, and each layer updates the current node through the attention mechanism. The update form of the node feature is:

[0024]

[0025] Among them, H l is the number of attention heads in layer l, N t (i) is the neighbor set of node i of category t, U is the union, N t (i)∪{i} is a set N t (i) Union operation with set {i}; representation of node k at layer l It is calculated by weighted summation of node attention in layer l-1. is the weight of nodes k and j calculated by attention head h at layer l, and the formula is as follows:

[0026]

[0027] in is the weight matrix of the k-th layer attention head h, Ψ(p k，j ) is the connection type between nodes k and j, at k，j is the connection attribute, and is the node feature of k, j, II represents feature concatenation, and ⊙ represents vector dot product;

[0028] (3) Calculate the mean of the node representations of the Lth layer to obtain the aggregate representation of nodes of the same type.

[0029] Furthermore, the aggregated features of heterogeneous nodes are weighted and fused as follows:

[0030] The company and personnel aggregate features obtained in the previous step are weighted and fused with the characteristics of the enterprise node to be valued, and the final company representation that integrates the heterogeneous graph structure and the enterprise's own information is obtained. The corresponding weighted fusion module in the framework is in the form of:

[0031]

[0032] in Indicates the heterogeneous node representation of the fusion enterprise's own information, f2(v i ) is the final company representation; β j is the adaptive weight, calculated as:

[0033]

[0034] LeakyReLU is a nonlinear activation function, uT is the parameter weight of the fully connected network, The calculation formula is as follows:

[0035]

[0036] Where Φ(x i ) represents the embedding representation of the original company, It is calculated by the aggregated learned representation of different types of nodes and the original company embedding representation.

[0037] In the second aspect, a corporate valuation system based on a heterogeneous graph neural network is provided, comprising:

[0038] The first module is used to build industry heterogeneous graphs;

[0039] The second module is used to sample the neighbors of the enterprise to be valued;

[0040] The third module is used to use attention aggregation learning on the sampling nodes;

[0041] The fourth module is used to weight and fuse the aggregate features of heterogeneous nodes;

[0042] The fifth module is used to predict corporate valuation.

[0043] According to a third aspect, an electronic device is provided, comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor implements the above-mentioned enterprise valuation method based on heterogeneous graph neural network when executing the program.

[0044] In a fourth aspect, a computer-readable storage medium is provided, on which a computer program is stored, and when the program is executed by a processor, the above-mentioned enterprise valuation method based on heterogeneous graph neural network is implemented.

[0045] The present invention adopts graph neural network as the basic model, introduces the connection information of company nodes and personnel nodes in the aggregate learning of heterogeneous graph neural network and fuses them through the attention mechanism. Compared with the prior art, the present invention has the following features:

[0046] First, the present invention realizes the modeling of enterprise valuation tasks, uses graph neural network technology to mine the complex feature representation of the company from the heterogeneous graph of companies and personnel in the industry field, and completes the value estimation of the enterprise.

[0047] Secondly, the present invention introduces the consideration of the relationship between nodes in the node aggregation learning process in the heterogeneous graph neural network, and learns the complex associations between company and personnel nodes by integrating the relationship features when calculating the attention between the source node and the target node in the update node features. This heterogeneous graph neural network solution focuses on company valuation from a high-level industry perspective, relying on massive data and deep learning models, and effectively solves the cognitive bias and subjective limitations caused by experts' direct modeling of enterprises.

[0048] Finally, thanks to the high efficiency of the graph neural network algorithm, the present invention can quickly update the enterprise valuation when company and personnel information changes or the market fluctuates. BRIEF DESCRIPTION OF THE DRAWINGS

[0049] Figure 1 It is the overall flow chart of the present invention.

[0050] Figure 2 The framework diagram of heterogeneous graph neural network.

[0051] Figure 3 This is a schematic diagram of the company's personnel heterogeneity.

[0052] Figure 4 Sub-flowchart for the attention aggregation learning steps. DETAILED DESCRIPTION

[0053] The specific implementation of the present invention is described below with reference to the accompanying drawings.

[0054] like Figure 1 As shown, the present invention discloses a method for enterprise valuation based on heterogeneous graph neural network. Figure 2 This is a diagram of the heterogeneous network framework designed by the present invention, from left to right are the neighbor sampling module, the aggregation learning module and the feature fusion module. The specific description is as follows:

[0055] Step 1: Build an industry heterogeneous graph:

[0056] like Figure 3 The figure shows a heterogeneous graph structure composed of two types of nodes: company and personnel. The company nodes are represented by data such as basic information of the enterprise and operating information. The personnel nodes represent the company's talent resources, intellectual property rights and other intangible resources. The connection between company nodes represents the investment situation. Through these associations, industry information can be constructed. The connection between personnel nodes depicts the talent and technology resource network of the enterprise. The present invention connects companies and personnel according to relationships such as investment and appointment. The companies, personnel and relationships in the heterogeneous graph use numerical vectors mined from the above data as the initial feature representation of nodes and edges.

[0057] Step 2: Conduct neighbor sampling for the enterprise to be valued:

[0058] In order to mine the company's related information from the industry heterogeneous graph, it is necessary to obtain a series of neighbor related nodes for the enterprise to be valued through random sampling. The present invention uses a random walk algorithm to perform standardized sampling of neighbor nodes of different types and quantities, corresponding to the neighbor sampling module in the framework, and the specific method is as follows:

[0059] ① Starting from the enterprise to be valued, a random neighbor of the current node is selected with a probability of r each time, and the initial enterprise to be valued is returned with a probability of 1-r to generate a node sequence of length N;

[0060] ② The generated series of associated neighbor node sequences are divided into two groups according to type: company and personnel. For each type, the n nodes with the highest frequency of occurrence are selected from the sequence to generate two types of heterogeneous neighbor node sets for the enterprise to be valued.

[0061] Step 3: Use attention aggregation learning for nodes:

[0062] Since different companies or people’s neighbors contribute differently to the value assessment of the valued enterprise, such as some mature companies have more powerful core resources, and industry leaders represent more advanced technical resources, these nodes undoubtedly have a greater impact on the valued enterprise. For this reason, the present invention designs a hierarchical multi-head attention weighted network that integrates related information for aggregate learning, corresponding to the aggregate learning module in the framework.

[0063] Given a set of neighboring companies and a set of neighboring personnel for a company to be valued, such as Figure 4 As shown in Figure 2, the process of attention aggregation learning is as follows:

[0064] ① Randomly walk the company to be valued and generate a set of company nodes and a set of personnel nodes, which are input into the attention aggregation network respectively;

[0065] ② Considering that the connections between nodes contain different information, a hierarchical multi-head attention weighted network that integrates the associated information is then used to aggregate and learn each type of node. There are a total of L layers in the network, and each layer updates the current node through the attention mechanism. The update form of the node feature is:

[0066]

[0067] Among them, H l is the number of attention heads in the first layer. Using different attention heads is aimed at learning multi-angle feature attributes, improving model performance and optimizing stability. t (i) is the set of neighbors of node i of category t, and ∪ is the union. The representation of node k in the lth layer is It is calculated by weighted summation of node attention in layer l-1. is the weight of nodes k and j calculated by attention head h at layer l, and the formula is as follows:

[0068]

[0069] in is the weight matrix of the k-th layer attention head h, Ψ(p k，j ) is the connection representation between nodes k and j, such as investment, holding, appointment, cooperation, etc. k，j The attributes of the connection, such as the proportion of investment, position held, and other specific information, and is the node feature of k and j, II represents feature concatenation, ⊙ represents vector dot product, and the attention of the node needs to be normalized during calculation.

[0070] ③After multi-head attention aggregation learning, nodes of the same type update their own representations based on information transmission, and finally the aggregation features of nodes of the same type are obtained through the following formula:

[0071]

[0072] in is the feature representation of the t-th type of node, which is calculated by the mean of the node features in the network output layer.

[0073] Step 4: Weighted fusion of aggregated features of heterogeneous nodes:

[0074] The present invention weightedly fuses the company and personnel aggregation features obtained in description 3 with the characteristics of the enterprise node to be valued, and obtains the final company representation that integrates the heterogeneous graph structure and the enterprise's own information. The weighted fusion module in the corresponding framework is in the form of:

[0075]

[0076] in Indicates the heterogeneous node representation of the fusion enterprise's own information, f2(v i ) is the final company representation; β j is the adaptive weight, calculated as:

[0077]

[0078] LeakyReLU is a nonlinear activation function, uT is the parameter weight of the fully connected network, The calculation formula is as follows:

[0079]

[0080] Where Φ(x i ) represents the embedding representation of the original company, It is calculated by the aggregated learned representation of different types of nodes and the original company embedding representation.

[0081] Step 5: Predict company valuation:

[0082] Use the final company feature representation to predict corporate valuation through a fully connected network.

[0083] The present invention also provides an enterprise valuation system based on heterogeneous graph neural network, which is used to implement the above-mentioned enterprise valuation method based on heterogeneous graph neural network, and the system includes:

[0084] The first module is used to build industry heterogeneous graphs;

[0085] The second module is used to sample the neighbors of the enterprise to be valued;

[0086] The third module is used to use attention aggregation learning on the sampling nodes;

[0087] The fourth module is used to weight and fuse the aggregate features of heterogeneous nodes;

[0088] The fifth module is used to predict corporate valuation.

[0089] The specific implementation method of the above modules is the same as the specific implementation steps of the aforementioned enterprise valuation method based on heterogeneous graph neural network, which will not be repeated here.

[0090] Furthermore, the present invention also provides an electronic device, including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor implements the above-mentioned enterprise valuation method based on heterogeneous graph neural network when executing the program.

[0091] A computer-readable storage medium stores a computer program, which, when executed by a processor, implements the above-mentioned enterprise valuation method based on heterogeneous graph neural network.

[0092] In the embodiments provided in the present application, it should be understood that the disclosed methods, devices, and equipment can be implemented in other ways. For example, the system embodiments described above are only schematic, for example, the division of the modules is only a logical function division, and there may be other division methods in actual implementation, such as multiple modules or components can be combined or integrated into another system, or some features can be ignored or not executed.

[0093] In addition, each functional unit in each embodiment of the present invention may be integrated into one processing unit, or each unit may exist physically separately, or two or more units may be integrated into one unit. The above-mentioned integrated unit may be implemented in the form of hardware or in the form of software functional units.

[0094] If the integrated unit is implemented in the form of a software functional unit and sold or used as an independent product, it can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of the present invention is essentially or the part that contributes to the prior art or all or part of the technical solution can be embodied in the form of a software product. The computer software product is stored in a storage medium, including several instructions for a computer device (which can be a personal computer, a server, or a network device, etc.) to perform all or part of the steps of the method described in each embodiment of the present invention. The aforementioned storage medium includes: U disk, mobile hard disk, read-only memory (ROM, Read-Only Memory), random access memory (RAM, Random Access Memory), disk or optical disk and other media that can store program codes.

[0095] The above embodiments are only used to illustrate the technical solutions of the present invention, rather than to limit the same. Although the present invention has been described in detail with reference to the aforementioned embodiments, those skilled in the art should understand that the technical solutions described in the aforementioned embodiments may still be modified, or some of the technical features may be replaced by equivalents. However, these modifications or replacements do not deviate the essence of the corresponding technical solutions from the spirit and scope of the technical solutions of the embodiments of the present invention.

Claims

1. A method for enterprise valuation based on heterogeneous graph neural network, characterized in that: The steps include: Step 1: Build an industry heterogeneous graph, which consists of two types of nodes: company and personnel. Companies and personnel are connected to each other based on investment and position relationships. Step 2: Sampling the neighbors of the enterprise to be valued; using the random walk algorithm to perform standardized sampling of neighbor nodes of different types and numbers, as follows: (1) Starting from the enterprise to be valued, a random neighbor of the current node is selected with probability r each time, and the initial enterprise to be valued is returned with probability 1-r, generating a node sequence of length N; (2) The generated series of associated neighbor node sequences are divided into two groups according to type: company and personnel. For each type, the n nodes with the highest frequency of occurrence are selected from the sequence to generate heterogeneous neighbor node sets of two types of enterprises to be valued; Step 3: Use attention aggregation learning on the sampling nodes, including the following steps: (1) Starting from the enterprise to be valued, a set of company nodes and a set of personnel nodes are generated by random walks and input into the attention aggregation network respectively; (2) A hierarchical multi-head attention weighted network that integrates relevant information is used to aggregate and learn each type of node. There are a total of L layers in the network, and each layer updates the current node through the attention mechanism. The update form of the node feature is: Among them, H l is the number of attention heads in layer l, N t (i) is the neighbor set of node i of category t, ∪ is the union, N t (i)∪{i} is a set N t (i) Union operation with set {i}; representation of node k at layer l It is calculated by weighted summation of node attention in layer l-1. is the weight of nodes k and j calculated by attention head h at layer l, and the formula is as follows: in is the weight matrix of the k-th layer attention head h, Ψ(p k,j ) is the connection type between nodes k and j, at k,j is the connection attribute, and is the node feature of k, j, ∥ represents feature concatenation, ⊙ represents vector dot product; (3) Calculate the mean of the node representations at the Lth layer to obtain the aggregate representation of nodes of the same type; Step 4: Weighted fusion of the aggregated features of heterogeneous nodes, as follows: The company and personnel aggregate features obtained in the previous step are weighted and fused with the characteristics of the enterprise node to be valued, and the final company representation that integrates the heterogeneous graph structure and the enterprise's own information is obtained. The weighted fusion module in the corresponding framework is in the form of: in Indicates the heterogeneous node representation of the fusion enterprise's own information, f2(v i ) is the final company representation; β j is the adaptive weight, calculated as: LeakyReLU is a nonlinear activation function, u T is the parameter weight of the fully connected network, The calculation formula is as follows: Where Φ(x i ) represents the embedding representation of the original company, It is calculated by the aggregated learned representation of different types of nodes and the original company embedding representation; Step 5: Predict corporate valuation.

2. A business valuation system based on heterogeneous graph neural network, used to implement the business valuation method based on heterogeneous graph neural network according to claim 1, characterized in that: The system includes: The first module is used to build industry heterogeneous graphs; The second module is used to sample the neighbors of the enterprise to be valued; The third module is used to use attention aggregation learning on the sampling nodes; The fourth module is used to weight and fuse the aggregate features of heterogeneous nodes; The fifth module is used to predict corporate valuation.

3. An electronic device comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that: When the processor executes the program, the enterprise valuation method based on heterogeneous graph neural network as described in claim 1 is implemented.

4. A computer-readable storage medium having a computer program stored thereon, characterized in that: When the program is executed by a processor, the enterprise valuation method based on heterogeneous graph neural network as described in claim 1 is implemented.

Images