A service network-aware learning method based on multi-level graph variational inference

Abstract

The present application provides a service network perception learning method based on multi-level graph variational inference, which relates to the field of big data processing technology. The present application calls a target hierarchical graph variational inference network model that matches the target service network, builds an adapted actual hierarchical latent space graph structure based on the actual graph structure data of the target service network, and performs hierarchical interaction learning on the actual hierarchical latent space graph structure to obtain the actual Gaussian distribution parameters of the latent space hierarchical interaction variables at the actual hierarchical latent space graph structure, and then performs reparameterization processing based on the actual Gaussian distribution parameters to obtain the actual latent space hierarchical interaction variable data that can characterize the actual interaction relationship status of each network node in the target service network with other network nodes, so as to achieve accurate perception of the complex node interaction relationship within the target service network, so as to improve the downstream task execution quality of the target service network.

Description

Technical Field

[0001] The present application relates to the technical field of big data processing, and in particular to a service network perception learning method based on multi-level graph variational inference. Background Art

[0002] With the development of modern service industries, service industries have become increasingly complex and diversified. They are no longer simply one-way dependent relationships with primary industries (e.g., agriculture, fisheries, etc.) and secondary industries (e.g., mining, manufacturing, etc.), but rather complex and coupled interaction relationships. Therefore, the service networks of various service industries (e.g., freight transportation, tourism, etc.) have gradually begun to show a trend of complex connections of all elements and have non-Euclidean topological structures. Among them, a generalized complex service network can include node elements such as service sites, resources, supply stations, users, manufacturers, and delivery centers, and through various connection forms between numerous node elements, a service network topological structure containing attribute information is jointly constructed.

[0003] In today's digital and complex development context, various service industries have a large amount of multi-source, heterogeneous, redundant, and sparse general service data. The emergence of a large amount of data makes it difficult for traditional model-driven methods to meet the data processing needs of today's service networks in the aspects of perception, decision-making, planning, and control. Among them, information perception tasks are the basic premise for service networks to perform various downstream tasks (for example, intelligent decision-making tasks, intelligent control, etc.), and the quality of information perception tasks determines the execution quality of various downstream tasks.

[0004] At the same time, because the service industry is highly related to the primary and secondary industries and information is coupled, the corresponding service network actually has inherent system characteristics such as the diversity of service entity nodes, the variability of the system's internal dynamics, and the multi-dimensional correlation between various types of big data, which leads to very complex interaction characteristics between various network nodes in the service network. Influencing factors such as technological progress, human decision-making, and changes in political and social policies will further aggravate the complex evolution of interaction characteristics. Different influencing factors can act alone or be coupled with each other, resulting in complex properties of the entire service network system. Therefore, for today's service networks, it is difficult to quantify the interaction status of network nodes caused by each influencing factor, and it is difficult to effectively perceive the coupling and interaction relationship between various network nodes in the service network, which seriously affects the execution quality of downstream tasks of the corresponding service network. Summary of the invention

[0005] In view of this, the purpose of this application is to provide a service network perception learning method based on multi-level graph variational inference, which can decouple and characterize the coupled node interaction relationships within the target service network, so as to achieve accurate perception of the complex node interaction relationships within the target service network, thereby facilitating the improvement of the downstream task execution quality of the target service network.

[0006] In order to achieve the above purpose, the technical solution adopted in the embodiment of the present application is as follows:

[0007] In a first aspect, the present application provides a service network-aware learning method based on multi-level graph variational inference, the method comprising:

[0008] Obtain the actual graph structure data of the target service network;

[0009] Calling a target hierarchical graph variational inference network model that matches the target service network, constructing an actual hierarchical latent space graph structure of the target service network based on the actual graph structure data, and performing hierarchical interaction learning on the actual hierarchical latent space graph structure to obtain actual Gaussian distribution parameters of latent space hierarchical interaction variables at the actual hierarchical latent space graph structure, wherein the latent space hierarchical interaction variables are used to represent the interaction relationship between each network node in the corresponding service network;

[0010] The target hierarchical graph variational inference network model is called to perform reparameterization processing based on the actual Gaussian distribution parameters to obtain actual latent space hierarchical interaction variable data of the target service network that matches the actual graph structure data.

[0011] In an optional embodiment, the method further comprises:

[0012] Obtain graph structure sample data of any service network type and initialize the hierarchical graph variational inference network model to be trained;

[0013] Calling the to-be-trained hierarchical graph variational inference network model to construct a corresponding sample hierarchical latent space graph structure based on the graph structure sample data, and performing hierarchical interaction learning on the sample hierarchical latent space graph structure to obtain estimated Gaussian distribution parameters of latent space hierarchical interaction variables at the sample hierarchical latent space graph structure;

[0014] Calling the to-be-trained hierarchical graph variational inference network model to perform reparameterization processing based on the actual Gaussian distribution parameters to obtain latent space hierarchical interaction variable estimation data, and calculating the actual loss function value of the to-be-trained hierarchical graph variational inference network model according to the latent space hierarchical interaction variable estimation data and the estimated Gaussian distribution parameters;

[0015] According to the estimated data of the latent space hierarchical interaction variables and the actual loss function value, with the purpose of minimizing the target loss function, the hierarchical graph variational inference network model to be trained is back-propagated and trained to obtain a target hierarchical graph variational inference network model that matches the service network type, wherein the target loss function is the loss function of the hierarchical graph variational inference network model to be trained.

[0016] In an optional embodiment, the objective loss function is expressed by the following equation:

[0017] ;

[0018] in, It is used to represent the hierarchical graph variational inference network model to be trained. The objective loss function under the parameter state, used to represent the variational lower limit of the to-be-trained hierarchical graph variational inference network model for the latent space hierarchical interaction variables, Used to indicate The to-be-trained hierarchical graph variational inference network model in a parameter state, It is used to represent the estimated data of the latent space hierarchical interaction variables, It is used to represent the mathematical expectation function of the estimated data of the latent space hierarchical interaction variables, For representing the graph structure sample data, A sample adjacency matrix used to represent the sample data of the graph structure. is used to represent the forward inference result of the sample adjacency matrix by the latent space hierarchical interaction variable estimation data, It is used to represent the relative entropy solution function, Used to represent Gaussian noise with mean 0 and standard deviation 1. is used to represent the variable mean of the latent space hierarchical interaction variable at the estimated Gaussian distribution parameter, Used to represent the variable standard deviation of the latent space hierarchical interaction variable at the estimated Gaussian distribution parameters.

[0019] In an optional embodiment, the objective loss function The self-supervised learning loss function during model training is expressed as follows:

[0020] ;

[0021] in, It is used to represent the self-supervised learning loss function, It is used to represent the total number of network nodes in the service network corresponding to the graph structure sample data. Used to indicate the corresponding service network The network node and The node connection relationship of the network nodes in the sample adjacency matrix, Used to indicate the corresponding service network The network node and The link prediction probability of the topological links of network nodes, Used to indicate the corresponding service network The latent space hierarchical interaction characteristics of the network nodes in the latent space hierarchical interaction variable estimation data, Used to indicate the corresponding service network The latent space hierarchical interaction characteristics of the network nodes in the latent space hierarchical interaction variable estimation data, Used to represent the node link relationship prediction function.

[0022] In an optional embodiment, for any hierarchical latent space graph structure, the hierarchical latent space graph structure includes a multi-layer latent space graph structure, the hierarchical latent space graph structure is subjected to hierarchical interaction learning to obtain Gaussian distribution parameters of latent space hierarchical interaction variables at the hierarchical latent space graph structure, including:

[0023] For each parameter type in the variable mean and variable standard deviation of the latent space hierarchical interaction variable when it obeys Gaussian distribution, calculate the intra-layer interaction attention weight coefficient matching the parameter type for each network node and its neighboring nodes in the corresponding service network in each layer of the latent space graph structure;

[0024] For each layer of latent space graph structure, according to the intra-layer interaction attention weight coefficient of each network node and its neighboring nodes in the latent space graph structure that matches the parameter type, and the feature vector of the corresponding neighboring node in the latent space graph structure, the cross-layer attention learning weight of each network node matching the parameter type in the latent space graph structure is calculated;

[0025] Performing exponential normalization on the cross-layer attention learning weights of all network nodes that match the parameter type in the latent space graph structure, and obtaining the hierarchical interaction attention weight matrix of the corresponding service network that matches the parameter type in the latent space graph structure;

[0026] According to the hierarchical interaction attention weight matrix of the corresponding service network that matches the parameter type at each layer of the latent space graph structure, a weighted sum operation is performed on the feature matrix of each layer of the latent space graph structure to obtain the parameter statistics of the latent space hierarchical interaction variable that matches the parameter type at the hierarchical latent space graph structure.

[0027] In an optional implementation, for each parameter type in the variable mean and the variable standard deviation, the corresponding service network The network node and its neighbor nodes are in The intra-layer interaction attention weight coefficient matching the parameter type in the layer hidden space graph structure is calculated using the following formula:

[0028] ;

[0029] in, Used to indicate the corresponding service network The network node and The network node is The intra-layer interaction attention weight coefficient matching the parameter type in the layer latent space graph structure, Used to indicate the number of The node set consists of all neighbor nodes of a network node. Used to represent exponential functions, The first model weight matrix that matches the parameter type is used to represent the corresponding hierarchical graph variational inference network model, A second model weight matrix that matches the parameter type corresponding to the hierarchical graph variational inference network model is used to represent the second model weight matrix that matches the parameter type. Used to indicate the corresponding service network The network node is The feature vector at the layer hidden space graph structure, Used to indicate the corresponding service network The network node is Feature vector at the layer hidden space graph structure.

[0030] In an optional implementation, for each parameter type in the variable mean and the variable standard deviation, the corresponding service network The network node is The cross-layer attention learning weights matching the parameter type in the layer hidden space graph structure are calculated using the following formula:

[0031] ;

[0032] in, Used to indicate the corresponding service network The network node is The cross-layer attention learning weights in the layer hidden space graph structure that match the parameter type, Used to indicate the corresponding service network The network node and The network node is The intra-layer interaction attention weight coefficient matching the parameter type in the layer latent space graph structure, Used to indicate the Attention learning vector of the layer-hidden space graph structure, Used to represent the hyperbolic tangent activation function, Used to indicate the number of The node set consists of all neighbor nodes of a network node. A third model weight matrix that matches the parameter type and represents the corresponding hierarchical graph variational inference network model. Used to indicate the corresponding service network The network node is Feature vector at the layer hidden space graph structure.

[0033] In an optional embodiment, for each parameter type in the variable mean and the variable standard deviation, the corresponding service network is in the first The hierarchical interaction attention weight matrix matching the parameter type at the layer hidden space graph structure is calculated using the following equation:

[0034] ;

[0035] in, Used to indicate that the corresponding service network is in The hierarchical interaction attention weight matrix matching the parameter type at the layer hidden space graph structure, Used to indicate the total number of network nodes of the corresponding service network. Used to indicate the corresponding service network The network node is The hierarchical interaction attention weight values ​​in the hierarchical interaction attention weight matrix of the layer latent space graph structure that matches the parameter type, Used to indicate the corresponding service network The network node is The cross-layer attention learning weights in the layer hidden space graph structure that match the parameter type, Used to represent exponential functions.

[0036] In an optional embodiment, the reparameterization operation of the Gaussian distribution parameters of the latent space hierarchical interaction variables is expressed by the following formula:

[0037] ;

[0038] in, The reparameterization result used to represent the latent space hierarchical interaction variables, It is used to represent the variable mean of the latent space hierarchical interaction variable when it obeys Gaussian distribution, It is used to represent the standard deviation of the latent space hierarchical interaction variable when it obeys Gaussian distribution, Used to represent random Gaussian noise.

[0039] In an optional embodiment, any hierarchical graph variational inference network model consists of a feature learning encoder, a hierarchical interaction learning module and a generative decoder;

[0040] The feature learning encoder is used to realize the hierarchical latent space graph structure creation function;

[0041] The hierarchical interaction learning module is used to perform hierarchical interaction learning on the hierarchical latent space graph structure constructed by the feature learning encoder, wherein the hierarchical interaction learning module includes an intra-layer attention learning unit corresponding to the intra-layer interaction attention learning operation, and a cross-layer attention learning unit corresponding to the cross-layer interaction attention learning operation;

[0042] The generative decoder is used to reparameterize the Gaussian distribution parameters of the latent space hierarchical interaction variables obtained by the hierarchical interaction learning module.

[0043] In this case, the beneficial effects of the embodiments of the present application may include the following:

[0044] The present application calls a target hierarchical graph variational inference network model that matches the target service network, constructs an adapted actual hierarchical latent space graph structure based on the actual graph structure data of the target service network, and performs hierarchical interaction learning on the actual hierarchical latent space graph structure to obtain the actual Gaussian distribution parameters of the latent space hierarchical interaction variables at the actual hierarchical latent space graph structure, and then performs reparameterization processing based on the actual Gaussian distribution parameters to obtain the actual latent space hierarchical interaction variable data of the target service network that matches the actual graph structure data, so as to effectively characterize the actual interaction relationship status of each network node with other network nodes in the target service network through the actual latent space hierarchical interaction variable data, thereby completing the effective decoupling and characterization of the node interaction relationship coupled inside the target service network, realizing the accurate perception of the complex node interaction relationship inside the target service network, and facilitating the improvement of the downstream task execution quality of the target service network.

[0045] In order to make the above-mentioned objects, features and advantages of the present application more obvious and easy to understand, preferred embodiments are specifically cited below and described in detail with reference to the attached drawings. BRIEF DESCRIPTION OF THE DRAWINGS

[0046] In order to more clearly illustrate the technical solutions of the embodiments of the present application, the drawings required for use in the embodiments will be briefly introduced below. It should be understood that the following drawings only show certain embodiments of the present application and therefore should not be regarded as limiting the scope. For ordinary technicians in this field, other related drawings can be obtained based on these drawings without paying creative work.

[0047] Figure 1 A schematic diagram of the composition of a computer device provided in an embodiment of the present application;

[0048] Figure 2 One of the flow diagrams of the service network perception learning method based on multi-level graph variational inference provided in the embodiment of the present application;

[0049] Figure 3 A schematic diagram of a parameter solution process for providing Gaussian distribution parameters of latent space hierarchical interaction variables at any hierarchical latent space graph structure in an embodiment of the present application;

[0050] Figure 4 The second flowchart of the service network perception learning method based on multi-level graph variational inference provided in the embodiment of the present application. DETAILED DESCRIPTION

[0051] In order to make the purpose, technical solution and advantages of the embodiments of the present application clearer, the technical solution in the embodiments of the present application will be clearly and completely described below in conjunction with the drawings in the embodiments of the present application. Obviously, the described embodiments are part of the embodiments of the present application, rather than all the embodiments. The components of the embodiments of the present application described and shown in the drawings here can be arranged and designed in various different configurations.

[0052] Therefore, the following detailed description of the embodiments of the present application provided in the accompanying drawings is not intended to limit the scope of the present application for which protection is sought, but merely represents selected embodiments of the present application. Based on the embodiments in the present application, all other embodiments obtained by ordinary technicians in the field without creative work are within the scope of protection of the present application.

[0053] It should be noted that similar reference numerals and letters denote similar items in the following drawings, and therefore, once an item is defined in one drawing, further definition and explanation thereof is not required in subsequent drawings.

[0054] In the description of the present application, it should be understood that the terms "center", "up", "down", "left", "right", "vertical", "horizontal", "inside", "outside", etc. indicate orientations or positional relationships based on the orientations or positional relationships shown in the accompanying drawings, or are the orientations or positional relationships in which the product of the application is conventionally placed when in use, or are the orientations or positional relationships conventionally understood by those skilled in the art. They are only for the convenience of describing the present application and simplifying the description, and do not indicate or imply that the device or element referred to must have a specific orientation, be constructed and operated in a specific orientation, and therefore should not be understood as a limitation on the present application.

[0055] In the description of this application, it should also be noted that, unless otherwise clearly specified and limited, the terms "set", "install", "connect", and "connect" should be understood in a broad sense, for example, it can be a fixed connection, a detachable connection, or an integral connection; it can be a mechanical connection or an electrical connection; it can be a direct connection, or it can be indirectly connected through an intermediate medium, or it can be the internal communication of two elements. For ordinary technicians in this field, the specific meanings of the above terms in this application can be understood according to specific circumstances.

[0056] In addition, in the description of the present application, it can also be understood that the relational terms such as the term "first" and "second" are only used to distinguish one entity or operation from another entity or operation, and do not necessarily require or imply that there is any such actual relationship or order between these entities or operations. Moreover, the term "include", "comprise" or any other variant thereof is intended to cover non-exclusive inclusion, so that the process, method, article or equipment including a series of elements includes not only those elements, but also includes other elements that are not clearly listed, or also includes elements inherent to such process, method, article or equipment. In the absence of more restrictions, the elements defined by the sentence "comprise one..." do not exclude the existence of other identical elements in the process, method, article or equipment including the elements. For those of ordinary skill in the art, the specific meanings of the above terms in the present application can be understood by specific circumstances.

[0057] In conjunction with the accompanying drawings, some embodiments of the present application are described in detail below. In the absence of conflict, the following embodiments and features in the embodiments can be combined with each other.

[0058] Please refer to Figure 1 , Figure 11 is a schematic diagram of the composition of the computer device 10 provided in the embodiment of the present application. In the embodiment of the present application, the computer device 10 can obtain the graph structure data of the service network of any service industry type, and decouple and characterize the node interaction relationship coupled within the service network based on the obtained graph structure data, so as to achieve accurate perception of the complex node interaction relationship within the service network, so as to improve the quality of downstream task execution of the service network. Among them, the graph structure data includes the feature matrix and adjacency matrix of the corresponding service network, and the feature matrix can be used The adjacency matrix can be expressed as The characteristic matrix is ​​represented by the vector length of all network nodes in the corresponding service network. The adjacency matrix is ​​used to represent the connection relationship between all network nodes and other network nodes in the corresponding service network, where Used to represent the total number of network nodes of the corresponding service network; the arbitrary service industry type may be, but not limited to, freight transportation industry, tourism industry, etc.; the downstream tasks of the corresponding service network may include, but not limited to, service network node classification, service network node prediction, service community detection, service network link prediction and other tasks; the computer device 10 may be, but not limited to, a laptop computer, a personal computer, a server and other computing devices.

[0059] In the embodiment of the present application, the computer device 10 may include a memory 11, a processor 12 and a communication unit 13. The memory 11, the processor 12 and the communication unit 13 are electrically connected to each other directly or indirectly to achieve data transmission or interaction.

[0060] In this embodiment, the memory 11 is used to store computer programs, and the processor 12 can execute the computer programs accordingly after receiving the execution instruction. In addition, the memory 11 is also used to store target hierarchical graph variational inference network models corresponding to multiple service network types, wherein each service network type corresponds to a service industry type, and the target hierarchical graph variational inference network model is a neural network model trained by combining latent space (multi-) hierarchical mapping technology and variational inference technology. The target hierarchical graph variational inference network model can be used to mine and characterize the interaction relationship between each network node in the adapted service network.

[0061] In this embodiment, the communication unit 13 is used to establish a communication connection between the computer device 10 and other electronic devices through a network, and to send and receive data through the network, wherein the network includes a wired communication network and a wireless communication network. For example, the computer device 10 can obtain the graph structure data of a service network that needs to perceive the internal complex node interaction relationship through the communication unit 13, so as to accurately perceive the node interaction relationship based on the obtained graph structure data.

[0062] In this embodiment, the computer device 10 may pre-store a specific computer program related to the service network perception learning function in the memory 11, and decouple and characterize the node interaction relationship coupled within any service network by driving the processor 12 to execute the specific computer program stored in the memory 11, so as to achieve accurate perception of the complex node interaction relationship within the service network, thereby facilitating the improvement of the downstream task execution quality of the corresponding service network.

[0063] Understandably, Figure 1 The block diagram shown is only a schematic diagram of a composition of the computer device 10. The computer device 10 may also include Figure 1 More or fewer components as shown, or with Figure 1 Different configurations are shown. Figure 1 Each component shown in the figure can be implemented by hardware, software or a combination thereof.

[0064] In this application, in order to ensure that the computer device 10 decouples and characterizes the node interaction relationship coupled within any service network, so as to achieve accurate perception of the complex node interaction relationship within the service network, and to facilitate the improvement of the downstream task execution quality of the corresponding service network, the embodiment of this application provides a service network perception learning method based on multi-level graph variational inference to achieve the above purpose. The service network perception learning method based on multi-level graph variational inference provided by this application is described in detail below.

[0065] Please refer to Figure 2 , Figure 2 This is one of the flow diagrams of the service network perception learning method based on multi-level graph variational inference provided in the embodiment of the present application. In the embodiment of the present application, Figure 2 The service network awareness learning method shown may include steps S210 to S230.

[0066] Step S210, obtaining actual graph structure data of the target service network.

[0067] The target service network is a service network that needs to perceive the interaction relationship between complex internal nodes; the actual graph structure data includes the actual feature matrix and the actual adjacency matrix of the target service network, wherein the actual feature matrix is ​​composed of all network nodes in the target service network with a preset vector length (for example, the vector length in the above text). ), and the actual adjacency matrix is ​​used to represent the actual connection relationship between all network nodes in the target service network and other network nodes.

[0068] Step S220, call the target hierarchical graph variational inference network model that matches the target service network, construct the actual hierarchical latent space graph structure of the target service network based on the actual graph structure data, and perform hierarchical interaction learning on the actual hierarchical latent space graph structure to obtain the actual Gaussian distribution parameters of the latent space hierarchical interaction variables at the actual hierarchical latent space graph structure.

[0069] In this embodiment, the latent space hierarchical interaction variables can be used To express, For the corresponding service network The latent space hierarchical interaction characteristics of network nodes, Used to indicate the corresponding service network The specific interaction relationship between a network node and other network nodes, where It is used to represent the total number of network nodes at the corresponding service network, so the latent space hierarchical interaction variable It can be used to represent the interaction relationship between various network nodes in the corresponding service network.

[0070] When the computer device 10 obtains the actual graph structure data of the target service network, it will identify the actual service network type of the target service network, and then according to the service network type identification result, in the target hierarchical graph variational inference network models corresponding to the multiple service network types stored in the memory 11, find out the target hierarchical graph variational inference network model that actually matches the target service network, and then call the found target hierarchical graph variational inference network model to accurately perceive the complex node interaction relationship within the target service network.

[0071] Among them, because it is difficult to quantify the network node interaction caused by various influencing factors that may be involved in a complex service network, the target hierarchical graph variational inference network model provided in this application will be based on latent space (multi-) hierarchical mapping technology. From the perspective of latent space multi-view, the corresponding service network will be reconstructed into a discretized multi-layer feature latent space graph structure in latent space, and a hierarchical latent space graph structure matching the service network will be obtained, so that the mutually coupled node interaction relationship caused by different influencing factors in the corresponding service network can be described through the hierarchical latent space graph structure. At this time, the hierarchical latent space graph structure can be used To express, is the first Hidden space graph structure, It is used to represent the total number of layers of the corresponding hierarchical latent space graph structure; each layer of the latent space graph structure is a feature latent space graph structure, and each layer of the latent space graph structure is a graph feature learning layer of the corresponding hierarchical latent space graph structure that represents a certain relationship description angle, so as to ensure that the corresponding hierarchical latent space graph structure can provide a multi-dimensional description of the internal coupled node interaction relationship in the service network.

[0072] Therefore, the target hierarchical graph variational inference network model provided by the present application can use the hierarchical interaction process to represent the information coupling relationship between each network node in the corresponding service network based on the constructed hierarchical latent space graph structure, and the hierarchical interaction process describes the actual correlation between different influencing factors in the multi-layer latent space graph structure. Therefore, the target hierarchical graph variational inference network model provided by the present application can obtain the parameter information related to the latent space hierarchical interaction variables of the corresponding service network by learning the hierarchical interaction process.

[0073] In this process, since the target hierarchical graph variational inference network model provided in this application is implemented using variational inference technology, the process of solving the latent space hierarchical interaction variables can be regarded as "the process of obtaining the feature vector in the low-dimensional feature space using approximate true posterior inference technology". At this time, the parameter information related to the latent space hierarchical interaction variables is the Gaussian distribution parameters of the latent space hierarchical interaction variables when they obey the Gaussian distribution, and the parameter types involved in the Gaussian distribution parameters include the variable mean and variable standard deviation of the latent space hierarchical interaction variables when they obey the Gaussian distribution.

[0074] Therefore, the actual Gaussian distribution parameters determined by the computer device 10 for the target service network may include actual parameter statistics of the variable means of the latent space hierarchical interaction variables and actual parameter statistics of the variable standard deviations of the latent space hierarchical interaction variables.

[0075] Optionally, see Figure 3, Figure 3 It is a schematic diagram of the parameter solution process provided by the embodiment of the present application for the Gaussian distribution parameters of the latent space hierarchical interaction variables at any hierarchical latent space graph structure. In the embodiment of the present application, after the computer device 10 constructs a matching hierarchical latent space graph structure for a certain service network through any hierarchical graph variational inference network model (including a trained target hierarchical graph variational inference network model and an untrained hierarchical graph variational inference network model to be trained), the specific method flow of "calling the hierarchical graph variational inference network model to perform hierarchical interaction learning on the hierarchical latent space graph structure to obtain the Gaussian distribution parameters of the latent space hierarchical interaction variables at the hierarchical latent space graph structure" may include steps S310 to S340.

[0076] Step S310, for each parameter type in the variable mean and variable standard deviation of the latent space hierarchical interaction variables when they obey the Gaussian distribution, calculate the intra-layer interaction attention weight coefficient that matches the parameter type for each network node and its neighbor nodes in the corresponding service network in each layer of the latent space graph structure.

[0077] In this embodiment, the hierarchical interaction learning operation of a certain hierarchical latent space graph structure can be divided into an intra-layer interaction attention learning operation and a cross-layer interaction attention learning operation. For the intra-layer interaction attention learning operation, it is necessary to obtain the intra-layer interaction attention weight learning kernel corresponding to the variable mean and the variable standard deviation of the latent space graph structure through the intra-layer attention learning mechanism for each layer of the latent space graph structure included in the hierarchical latent space graph structure, so as to effectively describe the intra-layer interaction relationship of all network nodes in the corresponding service network matching the parameter type at the latent space graph structure through the intra-layer interaction attention weight learning kernel of the two parameter types.

[0078] Among them, for each parameter type in the variable mean and variable standard deviation, any service network in the The intra-layer interaction attention weight learning kernel matching the parameter type at the layer hidden space graph structure can be adopted To express, Used to indicate the corresponding service network The network node and The network node is The layer-interaction attention weight coefficient in the layer hidden space graph structure that matches the parameter type. A network node can be A network node itself, or The neighbor node of the network node can also be A non-neighbor node of a network node.

[0079] Among them, when The network node is The network node itself, Can be directly set to 1; The network node is When a network node is a non-neighbor node, Can be directly set to 0; The network node is When the neighbor node of a network node is a node, for each parameter type in the variable mean and variable standard deviation, the corresponding node in the service network The network node and its neighbor nodes are in The intra-layer interaction attention weight coefficient matching the parameter type in the layer hidden space graph structure is calculated using the following formula:

[0080] ;

[0081] in, Used to indicate the corresponding service network The network node and The network node is The intra-layer interaction attention weight coefficient matching the parameter type in the layer latent space graph structure, Used to indicate the number of The node set consists of all neighbor nodes of a network node. Used to represent exponential functions, The first model weight matrix that matches the parameter type is used to represent the corresponding hierarchical graph variational inference network model, A second model weight matrix that matches the parameter type corresponding to the hierarchical graph variational inference network model is used to represent the second model weight matrix that matches the parameter type. Used to indicate the corresponding service network The network node is The feature vector at the layer hidden space graph structure, Used to indicate the corresponding service network The network node is Feature vector at the layer hidden space graph structure.

[0082] Step S320, for each layer of the latent space graph structure, according to the intra-layer interaction attention weight coefficient of each network node and its neighbor nodes in the latent space graph structure that matches the parameter type, and the feature vector of the corresponding neighbor node in the latent space graph structure, calculate the cross-layer attention learning weight of each network node in the latent space graph structure that matches the parameter type.

[0083] In this embodiment, for the cross-layer interaction attention learning operation, it is necessary to obtain the cross-layer attention learning weights corresponding to the variable mean and the variable standard deviation of each network node in the corresponding service network at the latent space graph structure for each layer of the latent space graph structure included in the hierarchical latent space graph structure through a cross-layer attention learning mechanism, so as to effectively describe the cross-layer interaction relationship between each network node in the corresponding service network that matches the parameter type at the hierarchical latent space graph structure through the cross-layer attention learning weights of each network node of each of the two parameter types.

[0084] Among them, for each parameter type in the variable mean and variable standard deviation, the corresponding service network The network node is The cross-layer attention learning weights matching the parameter type in the layer hidden space graph structure are calculated using the following formula:

[0085] ;

[0086] in, Used to indicate the corresponding service network The network node is The cross-layer attention learning weights in the layer hidden space graph structure that match the parameter type, Used to indicate the corresponding service network The network node and The network node is The intra-layer interaction attention weight coefficient matching the parameter type in the layer latent space graph structure, Used to indicate the Attention learning vector of the layer-hidden space graph structure, Used to represent the hyperbolic tangent activation function, Used to indicate the number of The node set consists of all neighbor nodes of a network node. A third model weight matrix that matches the parameter type and represents the corresponding hierarchical graph variational inference network model. Used to indicate the corresponding service network The network node is Feature vector at the layer hidden space graph structure.

[0087] Step S330, exponentially normalize the cross-layer attention learning weights of all network nodes that match the parameter type in the latent space graph structure to obtain a hierarchical interaction attention weight matrix of the corresponding service network that matches the parameter type at the latent space graph structure.

[0088] In this embodiment, for each layer of latent space graph structure included in any hierarchical latent space graph structure, a Softmax function can be used to perform exponential normalization processing on the cross-layer attention learning weights of all network nodes in the service network in the latent space graph structure to obtain the hierarchical interaction attention weight matrix of the service network at the latent space graph structure.

[0089] In this process, for each parameter type in the variable mean and variable standard deviation, the corresponding service network is The hierarchical interaction attention weight matrix matching the parameter type at the layer hidden space graph structure can be calculated using the following equation:

[0090] ;

[0091] in, Used to indicate that the corresponding service network is in The hierarchical interaction attention weight matrix matching the parameter type at the layer hidden space graph structure, Used to indicate the total number of network nodes of the corresponding service network. Used to indicate the corresponding service network The network node is The hierarchical interaction attention weight values ​​in the hierarchical interaction attention weight matrix of the layer latent space graph structure that matches the parameter type, Used to indicate the corresponding service network The network node is The cross-layer attention learning weights in the layer hidden space graph structure that match the parameter type, Used to represent exponential functions.

[0092] Step S340, according to the hierarchical interaction attention weight matrix of the corresponding service network that matches the parameter type at each layer of the latent space graph structure, a weighted sum operation is performed on the feature matrix of each layer of the latent space graph structure to obtain the parameter statistics of the latent space hierarchical interaction variables that match the parameter type at the hierarchical latent space graph structure.

[0093] Among them, the characteristic matrix of each layer of the latent space graph structure is composed of the characteristic vectors of all network nodes in the corresponding service network at the latent space graph structure; the parameter statistics of the latent space hierarchical interaction variables that match the variable mean are composed of the latent space hierarchical interaction feature means of all network nodes in the corresponding service network; the parameter statistics of the latent space hierarchical interaction variables that match the variable standard deviation are composed of the latent space hierarchical interaction feature standard deviations of all network nodes in the corresponding service network.

[0094] Therefore, the present application can ensure that any hierarchical graph variational inference network model can perform hierarchical interaction learning for a certain hierarchical latent space graph structure by executing the above steps S310 to S340, and obtain the Gaussian distribution parameters of the latent space hierarchical interaction variables at the hierarchical latent space graph structure.

[0095] Step S230, calling the target hierarchical graph variational inference network model to perform reparameterization processing based on the actual Gaussian distribution parameters, to obtain the actual latent space hierarchical interaction variable data of the target service network that matches the actual graph structure data.

[0096] In this embodiment, since the sampling process of the latent space hierarchical interaction variable when it obeys Gaussian distribution is a classic random process, it is necessary to use the reparameterization technique to reparameterize the Gaussian distribution parameters of the latent space hierarchical interaction variable in order to redirect the sampling process of the latent space hierarchical interaction variable into a deterministic process, so as to ensure that the corresponding calculated latent space hierarchical interaction variable data has sufficiently high data reliability. Among them, the reparameterization operation based on the Gaussian distribution parameters of the latent space hierarchical interaction variable is expressed by the following formula:

[0097] ;

[0098] in, The reparameterization result used to represent the latent space hierarchical interaction variables, It is used to represent the variable mean of the latent space hierarchical interaction variable when it obeys Gaussian distribution, It is used to represent the standard deviation of the latent space hierarchical interaction variable when it obeys Gaussian distribution, Used to represent random Gaussian noise.

[0099] Therefore, the target hierarchical graph variational inference network model provided by the present application can use the above-mentioned reparameterization operation formula to ensure that the actual latent space hierarchical interaction variable data finally obtained can effectively characterize the node interaction relationship of the target service network, thereby realizing the accurate perception of the complex node interaction relationship within the target service network, and facilitating the improvement of the downstream task execution quality of the target service network. Among them, the actual latent space hierarchical interaction variable data includes the actual latent space hierarchical interaction characteristics of all network nodes in the target service network.

[0100] Therefore, the present application can decouple and characterize the node interaction relationship coupled within any service network by executing the above steps S210 to S230, thereby achieving accurate perception of the complex node interaction relationship within the service network, thereby facilitating improving the execution quality of downstream tasks of the corresponding service network.

[0101] Optionally, see Figure 4 , Figure 4 This is the second flow chart of the service network perception learning method based on multi-level graph variational inference provided in the embodiment of the present application. Figure 2 Compared with the service network-aware learning method shown in Figure 4 The service network awareness learning method shown may also include steps S410 to S440 to train an adaptive target hierarchical graph variational inference network model for different service network type distributions, and enable the corresponding target hierarchical graph variational inference network model to have a reliable perception function of complex node interaction relationships.

[0102] Step S410, obtaining graph structure sample data of any type of service network, and initializing the hierarchical graph variational inference network model to be trained.

[0103] In this embodiment, the graph structure sample data includes a sample feature matrix and a sample adjacency matrix of a sample service network of a corresponding service network type, wherein the sample feature matrix is ​​composed of all network nodes in the corresponding sample service network each having a preset vector length (for example, the vector length in the above text ), and the sample adjacency matrix is ​​used to represent the node connection relationship between all network nodes and other network nodes in the corresponding sample service network. The hierarchical graph variational inference network model to be trained is consistent with the model architecture of any target hierarchical graph variational inference network model provided in this application.

[0104] In one implementation of the present embodiment, any hierarchical graph variational inference network model provided by the present application is composed of a feature learning encoder, a hierarchical interaction learning module and a generative decoder, wherein the feature learning encoder is used to realize the hierarchical latent space graph structure creation function, the hierarchical interaction learning module is used to perform hierarchical interaction learning on the hierarchical latent space graph structure constructed by the feature learning encoder, and the generative decoder is used to reparameterize the Gaussian distribution parameters of the latent space hierarchical interaction variables obtained by the hierarchical interaction learning module, wherein the hierarchical interaction learning module may include an intra-layer attention learning unit corresponding to the intra-layer interaction attention learning operation, and a cross-layer attention learning unit corresponding to the cross-layer interaction attention learning operation.

[0105] Step S420, call the to-be-trained hierarchical graph variational inference network model to construct a corresponding sample hierarchical latent space graph structure based on the graph structure sample data, and perform hierarchical interaction learning on the sample hierarchical latent space graph structure to obtain the estimated Gaussian distribution parameters of the latent space hierarchical interaction variables at the sample hierarchical latent space graph structure.

[0106] In this embodiment, the specific execution process of the step "constructing the corresponding sample hierarchical latent space graph structure based on the graph structure sample data" executed by the to-be-trained hierarchical graph variational inference network model can refer to the detailed description of the hierarchical latent space graph structure creation operation involved in step S220 above, which will not be repeated here; the specific execution process of the step "performing hierarchical interaction learning on the sample hierarchical latent space graph structure to obtain the estimated Gaussian distribution parameters of the latent space hierarchical interaction variables at the sample hierarchical latent space graph structure" executed by the to-be-trained hierarchical graph variational inference network model can refer to the detailed description of steps S310 to S340 above, which will not be repeated here. Among them, the estimated Gaussian distribution parameters determined by the to-be-trained hierarchical graph variational inference network model can include the estimated parameter statistics of the variable mean of the latent space hierarchical interaction variables at the graph structure sample data, and the estimated parameter statistics of the variable standard deviation of the latent space hierarchical interaction variables at the graph structure sample data.

[0107] Step S430, call the hierarchical graph variational inference network model to be trained to perform reparameterization processing based on the actual Gaussian distribution parameters to obtain the estimated data of the latent space hierarchical interaction variables, and calculate the actual loss function value of the hierarchical graph variational inference network model to be trained based on the estimated data of the latent space hierarchical interaction variables and the estimated Gaussian distribution parameters.

[0108] In this embodiment, the hierarchical graph variational inference network model to be trained can use the above-mentioned reparameterization operation formula to calculate the latent space hierarchical interaction variable estimation data, wherein the latent space hierarchical interaction variable estimation data includes the estimated latent space hierarchical interaction characteristics of all network nodes in the corresponding service network.

[0109] In this embodiment, since the variational inference network model of the hierarchical graph to be trained is trained using variational inference technology, the sampling process of the latent space hierarchical interaction variables when they obey Gaussian distribution is a classic random process. In the training process of the variational inference network model of the hierarchical graph to be trained, the gradient back propagation algorithm cannot be directly used for model training. It is necessary to calculate the actual loss function value of the variational inference network model of the hierarchical graph to be trained, and then use the calculated actual loss function value to update the model gradient of the variational inference network model of the hierarchical graph to be trained before back propagation training. Among them, due to the use of variational inference technology, the target loss function of the variational inference network model of the hierarchical graph to be trained will be directly related to the variational lower limit of the variational inference network model of the hierarchical graph to be trained, and the model training goal "minimizing the target loss function" can be equated with "maximizing the variational lower limit", and the above-mentioned actual loss function value can be calculated based on the target loss function.

[0110] In this process, the objective loss function is expressed by the following equation:

[0111] ;

[0112] in, It is used to represent the hierarchical graph variational inference network model to be trained. The objective loss function under the parameter state, used to represent the variational lower limit of the to-be-trained hierarchical graph variational inference network model for the latent space hierarchical interaction variables, Used to indicate The to-be-trained hierarchical graph variational inference network model in a parameter state, It is used to represent the estimated data of the latent space hierarchical interaction variables, It is used to represent the mathematical expectation function of the estimated data of the latent space hierarchical interaction variables, For representing the graph structure sample data, A sample adjacency matrix used to represent the sample data of the graph structure. is used to represent the forward inference result of the sample adjacency matrix by the latent space hierarchical interaction variable estimation data, It is used to represent the relative entropy solution function, It is used to represent Gaussian noise with mean 0 and standard deviation 1. is used to represent the variable mean of the latent space hierarchical interaction variable at the estimated Gaussian distribution parameter, Used to represent the variable standard deviation of the latent space hierarchical interaction variable at the estimated Gaussian distribution parameters.

[0113] It is worth noting that for the above expression of the variational lower limit For example, it can adopt a self-supervisory mechanism for learning during the model training process, and can use the network node link prediction results based on the latent space hierarchical interaction variable estimation data as the self-supervisory label of the known inherent attribute network topology structure of the corresponding service network to learn and optimize, so as to achieve the aforementioned model training goals. At this time, The self-supervised learning loss function during model training is expressed as follows:

[0114] ;

[0115] in, It is used to represent the self-supervised learning loss function, It is used to represent the total number of network nodes in the service network corresponding to the graph structure sample data. Used to indicate the corresponding service network The network node and The node connection relationship of the network nodes in the sample adjacency matrix, Used to indicate the corresponding service network The network node and The link prediction probability of the topological links of network nodes, Used to indicate the corresponding service network The latent space hierarchical interaction characteristics of the network nodes in the latent space hierarchical interaction variable estimation data, Used to indicate the corresponding service network The latent space hierarchical interaction characteristics of the network nodes in the latent space hierarchical interaction variable estimation data, It is used to represent the node link relationship prediction function. In one implementation of this embodiment, This can be implemented using a multi-layer perceptron.

[0116] Step S440, based on the estimated data of latent space hierarchical interaction variables and the actual loss function value, with the purpose of minimizing the target loss function, the hierarchical graph variational inference network model to be trained is back-propagated and trained to obtain a target hierarchical graph variational inference network model that matches the service network type.

[0117] In this embodiment, after obtaining the estimated data of the latent space hierarchical interaction variables and the actual loss function value, the model gradient of the hierarchical graph variational inference network model to be trained can be updated using the estimated data of the latent space hierarchical interaction variables and the actual loss function value, and then the process jumps to the above-mentioned step S420 to continue execution until the first model weight matrix, the second model weight matrix and the third model weight matrix of the finally trained hierarchical graph variational inference network model can meet the requirement of minimizing the target loss function (i.e., the target loss function converges). At this time, the finally trained hierarchical graph variational inference network model is an effective target hierarchical graph variational inference network model, and can have a reliable perception function of complex node interaction relationships.

[0118] Therefore, the present application can train an adaptive target hierarchical graph variational inference network model for different service network type distributions by executing the above steps S410 to S440, and enable the corresponding target hierarchical graph variational inference network model to have a reliable perception function of complex node interaction relationships.

[0119] The above are only various implementations of the present application, but the protection scope of the present application is not limited thereto. Any person skilled in the art who is familiar with the present technical field can easily think of changes or substitutions within the technical scope disclosed in the present application, which should be included in the protection scope of the present application. Therefore, the protection scope of the present application should be based on the protection scope of the claims.

Claims

1. A service network-aware learning method based on multi-level graph variational inference, characterized in that: The method comprises: Obtain the actual graph structure data of the target service network; Calling a target hierarchical graph variational inference network model that matches the target service network, constructing an actual hierarchical latent space graph structure of the target service network based on the actual graph structure data, and performing hierarchical interaction learning on the actual hierarchical latent space graph structure to obtain actual Gaussian distribution parameters of latent space hierarchical interaction variables at the actual hierarchical latent space graph structure, wherein the latent space hierarchical interaction variables are used to represent the interaction relationship between each network node in the corresponding service network; wherein any hierarchical graph variational inference network model is composed of a feature learning encoder, a hierarchical interaction learning module and a generative decoder, wherein the feature learning encoder is used to realize the hierarchical latent space graph structure creation function, the hierarchical interaction learning module is used to perform hierarchical interaction learning on the hierarchical latent space graph structure constructed by the feature learning encoder, and the generative decoder is used to reparameterize the Gaussian distribution parameters of the latent space hierarchical interaction variables obtained by the hierarchical interaction learning module, wherein the hierarchical interaction learning module includes an intra-layer attention learning unit corresponding to the intra-layer interaction attention learning operation, and a cross-layer attention learning unit corresponding to the cross-layer interaction attention learning operation; The target hierarchical graph variational inference network model is called to perform reparameterization processing based on the actual Gaussian distribution parameters to obtain the actual latent space hierarchical interaction variable data of the target service network that matches the actual graph structure data; wherein the reparameterization operation based on the Gaussian distribution parameters of the latent space hierarchical interaction variables adopts To express, The reparameterization result used to represent the latent space hierarchical interaction variables, It is used to represent the variable mean of the latent space hierarchical interaction variable when it obeys Gaussian distribution, It is used to represent the standard deviation of the latent space hierarchical interaction variable when it obeys Gaussian distribution, Used to represent random Gaussian noise; For any hierarchical latent space graph structure, the hierarchical latent space graph structure includes a multi-layer latent space graph structure, the hierarchical latent space graph structure is subjected to hierarchical interaction learning to obtain Gaussian distribution parameters of latent space hierarchical interaction variables at the hierarchical latent space graph structure, including: For each parameter type in the variable mean and variable standard deviation of the latent space hierarchical interaction variable when it obeys Gaussian distribution, calculate the intra-layer interaction attention weight coefficient matching the parameter type for each network node and its neighboring nodes in the corresponding service network in each layer of the latent space graph structure; For each layer of latent space graph structure, according to the intra-layer interaction attention weight coefficient of each network node and its neighboring nodes in the latent space graph structure that matches the parameter type, and the feature vector of the corresponding neighboring node in the latent space graph structure, the cross-layer attention learning weight of each network node matching the parameter type in the latent space graph structure is calculated; Performing exponential normalization on the cross-layer attention learning weights of all network nodes that match the parameter type in the latent space graph structure, and obtaining the hierarchical interaction attention weight matrix of the corresponding service network that matches the parameter type in the latent space graph structure; According to the hierarchical interaction attention weight matrix of the corresponding service network that matches the parameter type at each layer of the latent space graph structure, a weighted sum operation is performed on the feature matrix of each layer of the latent space graph structure to obtain the parameter statistics of the latent space hierarchical interaction variable that matches the parameter type at the hierarchical latent space graph structure.

2. The method according to claim 1, characterized in that The method further comprises: Obtain graph structure sample data of any service network type and initialize the hierarchical graph variational inference network model to be trained; Calling the to-be-trained hierarchical graph variational inference network model to construct a corresponding sample hierarchical latent space graph structure based on the graph structure sample data, and performing hierarchical interaction learning on the sample hierarchical latent space graph structure to obtain estimated Gaussian distribution parameters of latent space hierarchical interaction variables at the sample hierarchical latent space graph structure; Calling the to-be-trained hierarchical graph variational inference network model to perform reparameterization processing based on the actual Gaussian distribution parameters to obtain latent space hierarchical interaction variable estimation data, and calculating the actual loss function value of the to-be-trained hierarchical graph variational inference network model according to the latent space hierarchical interaction variable estimation data and the estimated Gaussian distribution parameters; According to the estimated data of the latent space hierarchical interaction variables and the actual loss function value, with the purpose of minimizing the target loss function, the hierarchical graph variational inference network model to be trained is back-propagated and trained to obtain a target hierarchical graph variational inference network model that matches the service network type, wherein the target loss function is the loss function of the hierarchical graph variational inference network model to be trained.

3. The method according to claim 2, characterized in that The objective loss function is expressed by the following equation: ； in, It is used to represent the hierarchical graph variational inference network model to be trained. The objective loss function under the parameter state, used to represent the variational lower limit of the to-be-trained hierarchical graph variational inference network model for the latent space hierarchical interaction variables, Used to indicate The to-be-trained hierarchical graph variational inference network model in a parameter state, It is used to represent the estimated data of the latent space hierarchical interaction variables, It is used to represent the mathematical expectation function of the estimated data of the latent space hierarchical interaction variables, For representing the graph structure sample data, A sample adjacency matrix used to represent the sample data of the graph structure. is used to represent the forward inference result of the sample adjacency matrix by the latent space hierarchical interaction variable estimation data, It is used to represent the relative entropy solution function, Used to represent Gaussian noise with mean 0 and standard deviation 1. is used to represent the variable mean of the latent space hierarchical interaction variable at the estimated Gaussian distribution parameter, Used to represent the variable standard deviation of the latent space hierarchical interaction variable at the estimated Gaussian distribution parameters.

4. The method according to claim 3, characterized in that The objective loss function The self-supervised learning loss function during model training is expressed as follows: ； in, It is used to represent the self-supervised learning loss function, It is used to represent the total number of network nodes in the service network corresponding to the graph structure sample data. Used to indicate the corresponding service network The network node and The node connection relationship of the network nodes in the sample adjacency matrix, Used to indicate the corresponding service network The network node and The link prediction probability of the topological links of network nodes, Used to indicate the corresponding service network The latent space hierarchical interaction characteristics of the network nodes in the latent space hierarchical interaction variable estimation data, Used to indicate the corresponding service network The latent space hierarchical interaction characteristics of the network nodes in the latent space hierarchical interaction variable estimation data, Used to represent the node link relationship prediction function.

5. The method according to any one of claims 1 to 4, characterized in that: For each parameter type in the variable mean and variable standard deviation, the corresponding service network The network node and its neighbor nodes are in The intra-layer interaction attention weight coefficient matching the parameter type in the layer hidden space graph structure is calculated using the following formula: ； in, Used to indicate the corresponding service network The network node and The network node is The intra-layer interaction attention weight coefficient matching the parameter type in the layer latent space graph structure, Used to indicate the number of The node set consists of all neighbor nodes of a network node. Used to represent exponential functions, The first model weight matrix that matches the parameter type is used to represent the corresponding hierarchical graph variational inference network model, A second model weight matrix that matches the parameter type corresponding to the hierarchical graph variational inference network model is used to represent the second model weight matrix that matches the parameter type. Used to indicate the corresponding service network The network node is The feature vector at the layer hidden space graph structure, Used to indicate the corresponding service network The network node is Feature vector at the layer hidden space graph structure.

6. The method according to any one of claims 1 to 4, characterized in that: For each parameter type in the variable mean and variable standard deviation, the corresponding service network The network node is The cross-layer attention learning weights matching the parameter type in the layer hidden space graph structure are calculated using the following formula: ； in, Used to indicate the corresponding service network The network node is The cross-layer attention learning weights in the layer hidden space graph structure that match the parameter type, Used to indicate the corresponding service network The network node and The network node is The intra-layer interaction attention weight coefficient matching the parameter type in the layer latent space graph structure, Used to indicate the Attention learning vector of the layer-hidden space graph structure, Used to represent the hyperbolic tangent activation function, Used to indicate the number of The node set consists of all neighbor nodes of a network node. A third model weight matrix that matches the parameter type and represents the corresponding hierarchical graph variational inference network model. Used to indicate the corresponding service network The network node is Feature vector at the layer hidden space graph structure.

7. The method according to any one of claims 1 to 4, characterized in that: For each parameter type in the variable mean and variable standard deviation, the corresponding service network is in The hierarchical interaction attention weight matrix matching the parameter type at the layer hidden space graph structure is calculated using the following equation: ； in, Used to indicate that the corresponding service network is in The hierarchical interaction attention weight matrix matching the parameter type at the layer hidden space graph structure, Used to indicate the total number of network nodes of the corresponding service network. Used to indicate the corresponding service network The network node is The hierarchical interaction attention weight values ​​in the hierarchical interaction attention weight matrix of the layer latent space graph structure that matches the parameter type, Used to indicate the corresponding service network The network node is The cross-layer attention learning weights in the layer hidden space graph structure that match the parameter type, Used to represent exponential functions.

Images