Task processing method, device, equipment and storage medium
By introducing the hyper-edge dependence vertex weights of vertices in the hypergraph neural network model, the probability transfer matrix of the hypergraph is determined, which solves the problem of low task accuracy in the existing model and achieves more efficient task processing.
Patent Information
- Application Number
- CN202210220485.X
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- Filing Date
- 2022-03-08
- Publication Date
- 2024-11-26
- Estimated Expiration
- 2042-03-08
AI Technical Summary
The existing hypergraph neural network models have low accuracy in classification, clustering, or regression tasks, mainly because their probability transfer matrix fails to make full use of the fine-grained information of hypergraph topology.
When building a hypergraph neural network model, the hyper-edge-dependent vertex weight (EDVW) of the vertex is introduced, and the probability transfer matrix of the hypergraph is determined through the vertex set, hyper-edge set, hyper-edge weight set and EDVW set, and then a weighted adjacency relationship is constructed Matrix and Laplace matrices capture richer hypergraph topology information.
The accuracy of the hypergraph neural network model in target task processing is improved, the ability to represent the hypergraph topology structure is enhanced, and the performance of task processing is improved.
Smart Images

Figure CN114625926B_ABST
Abstract
Description
Technical Field
[0001] The embodiments of the present application relate to the field of artificial intelligence technology, and in particular to a task processing method, apparatus, device and storage medium. Background Art
[0002] Graphs are topologies used to describe discrete data in the real world. Their elements include vertices, edges, and edge weights. Many real-world problems require the use of graphs or hypergraph structures to organize data, such as citation networks, social networks, and protein molecular structure networks. With the development of graph neural networks, the representation and information aggregation capabilities of graph structured data have also developed rapidly. In an ordinary graph, an edge can connect two vertices and represent a certain relationship between the two vertices. The degree of each edge in an ordinary graph is 2. A hypergraph is a topological structure with stronger representation capabilities. Each hyperedge in a hypergraph can contain multiple vertices.
[0003] Graph Neural Network (GNN) is a neural network that directly acts on graph structures and mainly processes graph structured data. In related technologies, hypergraph learning is introduced into the field of deep learning, and a hypergraph neural network (Hypergraph Neural Network, HGNN) model is derived using the hypergraph Laplacian matrix (a matrix representation of a hypergraph). The convolutional layer in the hypergraph neural network model can use the "vertex-hyperedge-vertex" information flow in the hypergraph to achieve information aggregation of vertices in the hypergraph to connect downstream classification, clustering or regression tasks, etc. The information aggregation function of the hypergraph convolutional network in the hypergraph neural network model can be understood as a simple hypergraph random walk model in mathematical concepts. According to the hypergraph random walk model, the probability transfer matrix of the hypergraph neural network model can be obtained as: Where H is the vertex-hyperedge adjacency matrix, W is the edge weight diagonal matrix, and D e is the degree diagonal matrix, D v is a diagonal matrix of point degrees, where the edge degree is the number of vertices adjacent to a hyperedge, and the point degree is the number of hyperedges adjacent to a vertex.
[0004] It can be seen that the probability transfer matrix of the above hypergraph neural network model utilizes the number of adjacent hyperedges of vertices and the number of adjacent vertices of hyperedges. The performance of the hypergraph neural network model is low, which leads to low accuracy in classification, clustering or regression tasks using the hypergraph neural network model. Summary of the invention
[0005] The present application provides a task processing method, apparatus, device and storage medium to improve the accuracy of classification, clustering or regression tasks.
[0006] In a first aspect, the present application provides a task processing method, comprising:
[0007] Obtaining original features of a target vertex, wherein the target vertex is any vertex in a pre-constructed hypergraph, wherein the hypergraph includes a vertex set, a hyperedge set, a hyperedge weight set, and a hyperedge dependent vertex weight EDVW set of the vertex;
[0008] Inputting the original features of the target vertex into a hypergraph neural network model, and outputting the embedded features of the target vertex, wherein the hypergraph neural network model is determined according to a probability transfer matrix of the hypergraph, and the probability transfer matrix of the hypergraph is determined according to the vertex set, the hyperedge set, the hyperedge weight set, and the EDVW set of the vertex;
[0009] According to the embedded features of the target vertex, a target task is processed, and the target task includes any one of vertex classification, vertex clustering and regression tasks.
[0010] In a second aspect, the present application provides a task processing device, comprising:
[0011] An acquisition module, used to acquire the original features of a target vertex, wherein the target vertex is any vertex in a pre-constructed hypergraph, wherein the hypergraph includes a vertex set, a hyperedge set, a hyperedge weight set, and a hyperedge dependent vertex weight EDVW set of the vertex;
[0012] A processing module, configured to input the original features of the target vertex into a hypergraph neural network model, and output the embedded features of the target vertex, wherein the hypergraph neural network model is determined according to a probability transfer matrix of the hypergraph, and the probability transfer matrix of the hypergraph is determined according to the vertex set, the hyperedge set, the hyperedge weight set, and the EDVW set of the vertex;
[0013] The processing module is further used to: process a target task according to the embedded features of the target vertex, and the target task includes any one of vertex classification, vertex clustering and regression tasks.
[0014] In a third aspect, a task processing device is provided, comprising: a processor and a memory, the memory being used to store a computer program, the processor being used to call and run the computer program stored in the memory to execute the method of the first aspect.
[0015] According to a fourth aspect, a computer-readable storage medium is provided for storing a computer program, wherein the computer program enables a computer to execute the method according to the first aspect.
[0016] According to a fifth aspect, a computer program product is provided, comprising a computer program, wherein the computer program implements the method according to the first aspect when executed by a processor.
[0017] In summary, in the present application, the original features of the target vertex are first obtained, and the target vertex is any vertex in the pre-constructed hypergraph. The original features of the target vertex are input into the hypergraph neural network model, and the embedded features of the target vertex are output. According to the embedded features of the target vertex, the target task is processed. The hypergraph neural network model is determined according to the probability transfer matrix of the hypergraph, and the probability transfer matrix of the hypergraph is determined according to the vertex set, the hyperedge set, the hyperedge weight set and the EDVW set of the vertex. Since the determination factor of the probability transfer matrix of the hypergraph adds the EDVW of the vertex, that is, the fine-grained information that the hypergraph topological structure can represent is considered, the hypergraph random walk under the probability transfer matrix of the hypergraph can capture richer hypergraph topological information. Therefore, the performance of the hypergraph neural network model of the hypergraph obtained according to the probability transfer matrix of the hypergraph is better, and then the use of the hypergraph neural network model for target task processing can improve the accuracy of target task processing. BRIEF DESCRIPTION OF THE DRAWINGS
[0018] In order to more clearly illustrate the technical solutions in the embodiments of the present invention, the following briefly introduces the drawings required for use in the description of the embodiments. Obviously, the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without creative work.
[0019] Figure 1 A schematic diagram of a hypergraph constructed based on a citation network;
[0020] Figure 2 A flowchart of a task processing method provided in an embodiment of the present application;
[0021] Figure 3 A flowchart of a task processing method provided in an embodiment of the present application;
[0022] Figure 4 A flowchart of a process for acquiring a hypergraph neural network model in a task processing method provided in an embodiment of the present application;
[0023] Figure 5 A flowchart of a task processing method provided in an embodiment of the present application;
[0024] Figure 6 A flowchart of a task processing method provided in an embodiment of the present application;
[0025] Figure 7 A schematic diagram of the structure of a task processing device provided in an embodiment of the present application;
[0026] Figure 8 It is a schematic block diagram of a task processing device 800 provided in an embodiment of the present application. DETAILED DESCRIPTION
[0027] The following will be combined with the drawings in the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative work are within the scope of protection of the present invention.
[0028] It should be noted that the terms "first", "second", etc. in the specification and claims of the present invention and the above-mentioned drawings are used to distinguish similar objects, and are not necessarily used to describe a specific order or sequence. It should be understood that the data used in this way can be interchangeable where appropriate, so that the embodiments of the present invention described herein can be implemented in an order other than those illustrated or described herein. In addition, the terms "including" and "having" and any variations thereof are intended to cover non-exclusive inclusions, for example, a process, method, system, product, or server that includes a series of steps or units is not necessarily limited to those steps or units clearly listed, but may include other steps or units that are not clearly listed or inherent to these processes, methods, products, or devices.
[0029] Before introducing the technical solution of this application, the following is an introduction to the relevant knowledge of this application:
[0030] Artificial Intelligence (AI): It is the theory, method, technology and application system that uses digital computers or machines controlled by digital computers to simulate, extend and expand human intelligence, perceive the environment, acquire knowledge and use knowledge to obtain the best results. In other words, artificial intelligence is a comprehensive technology of computer science, which attempts to understand the essence of intelligence and produce a new intelligent machine that can respond in a similar way to human intelligence. Artificial intelligence is to study the design principles and implementation methods of various intelligent machines, so that machines have the functions of perception, reasoning and decision-making.
[0031] Machine Learning (ML): It is a multi-disciplinary interdisciplinary subject, involving probability theory, statistics, approximation theory, convex analysis, algorithm complexity theory and other disciplines. It specializes in studying how computers simulate or implement human learning behavior to acquire new knowledge or skills and reorganize existing knowledge structures to continuously improve their performance. Machine learning is the core of artificial intelligence and the fundamental way to make computers intelligent. Its applications are spread across all areas of artificial intelligence. Machine learning and deep learning usually include artificial neural networks, belief networks, reinforcement learning, transfer learning, inductive learning, and self-learning.
[0032] Deep Learning (DL): It is a branch of machine learning and an algorithm that attempts to use multiple processing layers containing complex structures or composed of multiple nonlinear transformations to perform high-level abstraction of data. Deep learning is to learn the inherent laws and representation levels of training sample data. The information obtained in the learning process is of great help in the interpretation of data such as text, images and sounds. The ultimate goal of deep learning is to enable machines to have analytical learning capabilities like humans and to be able to recognize data such as text, images and sounds. Deep learning is a complex machine learning algorithm that has achieved results in speech and image recognition that far exceed previous related technologies.
[0033] Neural Network (NN): A deep learning model in the field of machine learning and cognitive science that imitates the structure and function of biological neural networks.
[0034] Graph Neural Network (GNN): A neural network that acts directly on the graph structure, mainly for processing data with non-Euclidean spatial structure (graph structure). It has the characteristics of ignoring the input order of vertices; during the calculation process, the representation of the vertex is affected by its surrounding neighboring vertices, while the connection of the graph itself remains unchanged; the representation of the graph structure enables graph-based reasoning. Usually, a graph neural network consists of two modules: a propagation module and an output module. The propagation module is used to transfer information between vertices in the graph and update the state, and the output module is used to define the objective function based on the vector representation of the vertices and edges of the graph according to different tasks. Graph neural networks include: Graph Convolutional Networks (GCNs), Gated Graph Neural Networks (GGNNs), and Graph Attention Networks (GAT) based on the attention mechanism.
[0035] Hypergraph is a topological structure with stronger representation capability. Each hyperedge in a hypergraph can contain multiple vertices. An ordinary graph can be regarded as a special hypergraph with a hyperedge degree limited to 2. Therefore, a hypergraph has a more generalized representation capability than a graph.
[0036] Edge-Dependent Vertex Weight (EDVW): In a hypergraph, the weight of the same vertex depends on different hyperedges, that is, the vertex has different weights in different hyperedges. EDVW can be called high-order, fine-grained cross-information between vertices in a hypergraph.
[0037] Hypergraphs have many application scenarios, such as citation networks, social networks, protein molecular structure networks, and 3D-Visual Object Classification, all of which can be represented by constructing hypergraphs. The following uses citation networks and 3D-Visual Object Classification to illustrate the construction of hypergraphs and the importance of EDVW in hypergraphs.
[0038] First, let's introduce the citation network. In academic literature, mutual citations between papers or co-authors of papers constitute a certain relationship between papers. In the citation network, vertices are defined as papers, and edges are composed of citation relationships or co-authors between papers. However, it is difficult to directly express some higher-order implicit interaction relationships simply by the citation relationship between two papers. Therefore, constructing hyperedges through some prior knowledge can obtain stronger representation capabilities. Figure 1 is a schematic diagram of a hypergraph constructed based on a citation network, such as Figure 1 As shown in , a paper is a vertex, and the hyperedge is composed of the citation relationship or co-authors between papers. For example, all papers signed by the same author A in the paper database constitute a hyperedge a, that is, Figure 1 The hyperedge 1 shown in , includes vertices 1, 2, 3 and 4 and the connection relationship between the four vertices. For another example, all papers in the paper database signed by the same author B constitute a hyperedge b, that is, Figure 1 Hyperedge 2 shown in the figure includes vertices 3, 4, 5, and 6, as well as the connection relationship between the four vertices. All papers that cite the same paper can also constitute a hyperedge. It is understandable that the same vertex (paper) has different importance for its different authors (hyperedges). For example, a paper is of relatively important significance in determining the research field of its first author (hyperedge), but it may be relatively less important to other authors (hyperedges). This information is EDVW information, that is, in a hypergraph, the weight of a vertex is reflected differently in different hyperedges. For example, the weight of vertex 3 in hyperedge 1 (which can be called the weight of vertex 3 relying on hyperedge 1) is different from the weight of vertex 3 in hyperedge 2 (which can be called the weight of vertex 3 relying on hyperedge 2).
[0039] Next, we introduce 3D-visual object classification. In 3D-visual object classification, when constructing a hypergraph, each 3D object is a vertex. These 3D objects are photographed from multiple perspectives, and different image processing techniques are used to extract the features of the 3D objects (i.e., vertex features). In order to extract high-order similarity information between 3D objects, a hypergraph needs to be constructed to represent them. The general method for constructing a hyperedge is: define a distance in the vertex feature space, take any object j (corresponding to vertex j) as the center point, use the K nearest neighbor algorithm, select the K nearest vertices, use the feature distance between vertex j and its K neighbor vertices as the weight of vertex j, and construct a hyperedge with a hyperedge degree (i.e., the number of vertices in the hyperedge) of K. It can be seen that EDVW information is naturally introduced into the constructed hypergraph by such a hypergraph construction method. Because for the same vertex j, the feature distances between vertex j and its K different neighbor vertices are generally different on a fine-grained scale, that is, in different hyperedges constructed with other different objects as the center, the weight of vertex j is different, that is, the EDVW of vertex j is different.
[0040] In the hypergraph neural network model proposed in the related art, from the perspective of hypergraph random walk, the probability transfer matrix of the hypergraph neural network model is Where H is the vertex-hyperedge adjacency matrix, W is the edge weight diagonal matrix, and D e is the degree diagonal matrix, D v is the point degree diagonal matrix, the edge degree is the number of vertices adjacent to the hyperedge, and the point degree is the number of adjacent hyperedges of the vertex. It can be seen that the probability transfer matrix of the hypergraph neural network model uses the number of adjacent hyperedges of the vertex and the number of adjacent vertices of the hyperedge. The performance of the hypergraph neural network model is low, which leads to low accuracy in classification, clustering or regression tasks using the hypergraph neural network model.
[0041] To solve this problem, the present application provides a task processing method, device, equipment and storage medium. In the process of constructing a hypergraph neural network model, the EDVW of the vertex is introduced when constructing the hypergraph, and the probability transfer matrix of the hypergraph is determined according to the vertex set, hyperedge set, hyperedge weight set and EDVW set of the vertex in the hypergraph, and then the weighted adjacency matrix of the vertices in the hypergraph is determined according to the probability transfer matrix of the hypergraph, and the Laplacian matrix of the hypergraph is determined according to the weighted adjacency matrix of the vertices in the hypergraph, and finally, the hypergraph neural network model of the hypergraph is constructed according to the Laplacian matrix of the hypergraph. Since the acquisition of the probability transfer matrix of the hypergraph adds the EDVW of the vertex, that is, the fine-grained information that the hypergraph topological structure can represent is taken into account, the hypergraph random walk under the probability transfer matrix of the hypergraph can capture richer hypergraph topological information. Therefore, the performance of the hypergraph neural network model of the hypergraph obtained according to the probability transfer matrix of the hypergraph is better, and then when the hypergraph neural network model is used to process the target task, the accuracy of the target task processing can be improved.
[0042] The following is a brief introduction to the application scenarios to which the technical solutions of the embodiments of the present application can be applied. It should be noted that the application scenarios introduced below are only used to illustrate the embodiments of the present application and are not limited. In specific implementation, the technical solutions provided by the embodiments of the present application can be flexibly applied according to actual needs.
[0043] The task processing method provided in the embodiment of the present application can be applied to scenarios such as citation networks, social networks, protein molecular structure networks, and 3D-Visual Object Classification. All of the above scenarios can construct hypergraphs to represent them. After the hypergraph is constructed, a hypergraph neural network model of the hypergraph can be constructed using the task processing method provided in the embodiment of the present application. The target task can be further processed through the hypergraph neural network model, and the target task includes any one of vertex classification, vertex clustering, and regression tasks. Taking the combination of a hypergraph neural network model for downstream classification as an example, for example, after a hypergraph and a hypergraph neural network model of a hypergraph are constructed for a citation network, Figure 2 A flowchart of a task processing method provided in an embodiment of the present application is shown as follows: Figure 2 As shown, if you want to obtain the classification of a vertex in the hypergraph (for example, you want to obtain the field to which a paper belongs), first obtain the original features of the target vertex, which is any vertex in the constructed hypergraph, input the original features of the target vertex into the hypergraph neural network model, output the embedded features of the target vertex, input the embedded features of the target vertex into the downstream classifier, output the classification label of the target vertex, and finally determine the classification of the target vertex based on the classification label of the target vertex, thereby finally obtaining the classification of the target vertex.
[0044] The technical solution of this application will be described in detail below:
[0045] Figure 3 The flowchart of a task processing method provided in an embodiment of the present application can be executed by a task processing device, and the task processing device can be implemented by software and / or hardware. Optionally, the task processing device can be a server. Figure 3 As shown, the method may include the following steps:
[0046] S101. Obtain original features of a target vertex, where the target vertex is any vertex in a pre-constructed hypergraph, and the hypergraph includes a vertex set, a hyperedge set, a hyperedge weight set, and an EDVW set of the vertex.
[0047] Specifically, in an actual application scenario, a hypergraph is first constructed based on the data of the application scenario. The constructed hypergraph includes a vertex set, a hyperedge set, a hyperedge weight set, and an EDVW set. It can be understood that the vertex set is the set of vertices, the hyperedge set is the set of hyperedges, the hyperedge weight set is the set of hyperedge weights, and the EDVW set is the set of EDVWs of vertices. The specific hypergraph construction method is not limited in this embodiment.
[0048] S102. Input the original features of the target vertex into the hypergraph neural network model, and output the embedded features of the target vertex. The hypergraph neural network model is determined according to the probability transfer matrix of the hypergraph. The probability transfer matrix of the hypergraph is determined according to the vertex set, hyperedge set, hyperedge weight set and vertex EDVW set.
[0049] S103. Processing a target task according to the embedded features of the target vertex, where the target task includes any one of vertex classification, vertex clustering and regression tasks.
[0050] Optionally, the target task may also be other tasks processed using a hypergraph neural network model, which is not limited in this embodiment.
[0051] Specifically, if the target task is vertex classification, the target task is processed according to the embedded features of the target vertex in S103, which may be:
[0052] S1031. Input the embedded features of the target vertex into the downstream classifier, and output the classification label of the target vertex.
[0053] S1032. Obtain the category to which the target vertex belongs according to the category label of the target vertex.
[0054] Optionally, the hypergraph neural network model in this embodiment is determined according to the probability transfer matrix of the hypergraph in the following manner:
[0055] According to the probability transfer matrix of the hypergraph, a weighted adjacency matrix of the vertices in the hypergraph is determined, and according to the weighted adjacency matrix of the vertices in the hypergraph, a Laplacian matrix of the hypergraph is determined;
[0056] According to the Laplace matrix of the hypergraph, a hypergraph neural network model of the hypergraph is constructed.
[0057] As an implementable approach, the probability transfer matrix of the hypergraph can be determined according to the vertex set, the hyperedge set, the hyperedge weight set, and the EDVW set of the vertices in the following manner:
[0058] S1. Calculate the transition probability from vertex u to vertex v based on the prior weight of hyperedge e, the EDVW of vertex u in the process of entering hyperedge e from vertex u, the EDVW of vertex u in the process of flowing out of hyperedge e from vertex u, the EDVW of vertex v in the process of entering hyperedge e from vertex v, the EDVW of vertex v in the process of flowing out of hyperedge e from vertex v, the out-degree of hyperedge e, the in-degree of hyperedge e, the degree of vertex u and the objective function. Vertex u and vertex v are any two vertices in the vertex set, and hyperedge e is the hyperedge between vertex u and vertex v.
[0059] Among them, the out-degree of hyperedge e is the sum of the EDVW of all vertices adjacent to hyperedge e that flow out of hyperedge e, the in-degree of hyperedge e is the sum of the EDVW of all vertices adjacent to hyperedge e that enter hyperedge e, and the degree of vertex u is determined according to the prior weight of hyperedge e, the EDVW of vertex u in the process of entering hyperedge e from vertex u, the EDVW of vertex u in the process of flowing out of hyperedge e from vertex u, the out-degree of hyperedge e, the in-degree of hyperedge e and the objective function. The objective function is used to represent the relationship between the out-degree of hyperedge e and the degree of vertex u.
[0060] In this embodiment, the objective function is used to express the relationship between the out-degree of hyperedge e and the degree of vertex u. Specifically, it can be used to reflect the different forms of influence of the out-degree of hyperedge e on the degree of vertex u. For example, the objective function is ρ(·). If the objective function is ρ(x)=x, it reflects that the out-degree of hyperedge e has a positive correlation with the degree of vertex u. If ρ(x)=1 / x, it reflects that the out-degree of hyperedge e has a negative correlation with the degree of vertex u. This influence is different on hypergraphs constructed with different data and can be regarded as a "hyperparameter".
[0061] Among them, the degree of vertex u is determined according to the prior weight of hyperedge e, the EDVW of vertex u in the process of entering hyperedge e from vertex u, the EDVW of vertex u in the process of flowing out of hyperedge e from vertex u, the out-degree of hyperedge e, the in-degree of hyperedge e and the objective function. Optionally, the degree of vertex u can be: the sum of the first results of all hyperedges e between vertex u and vertex v, the first result is the product of the prior weight of hyperedge e and the second result, the second result is the sum of the third result and the fourth result, the third result is the out-degree of hyperedge e, the objective function value with the out-degree of hyperedge e as the variable, and the product of the EDVW of vertex u in the process of entering hyperedge e from vertex u, and the fourth result is the in-degree of hyperedge e, the objective function value with the in-degree of hyperedge e as the variable, and the product of the EDVW of vertex u in the process of flowing out of hyperedge e from vertex u.
[0062] Specifically, in an implementable manner, the degree d(u) of vertex u can be expressed by the following formula (1):
[0063]
[0064] Among them, w(e) is the prior weight of hyperedge e, δ out (e) is the out-degree of hyperedge e, ρ(δ out (e)) is the objective function value with the out-degree of hyperedge e as the variable, Q in (u, e) is the EDVW of vertex u in the process of entering hyperedge e from vertex u, δ in (e) is the in-degree of hyperedge e, ρ(δ in (e)) is the objective function value with the in-degree of hyperedge e as the variable, Q out (u, e)) is the EDVW of vertex u in the process of hyperedge e flowing out from vertex u, and ε is the hyperedge set.
[0065] Among them, the out-degree δ of hyperedge e out (e) is the sum of the EDVWs of all vertices adjacent to hyperedge e that flow out of hyperedge e, which can be expressed by the following formula (2):
[0066]
[0067] Among them, Q out (v, e) is the EDVW of vertex v in the process of hyperedge e flowing out from vertex v, and V is the vertex set.
[0068] Among them, the in-degree δ of the hyperedge e in (e) is the sum of the EDVWs of all vertices adjacent to hyperedge e that enter hyperedge e, which can be expressed by the following formula (3):
[0069]
[0070] Among them, Q in(v, e) is the EDVW of vertex u in the process of entering hyperedge e from vertex v.
[0071] Optionally, the transition probability from vertex u to vertex v can be:
[0072] The target result is divided by twice the degree of vertex u. The target result is the sum of the fifth results of all hyperedges e between vertex u and vertex v. The fifth result is the product of the prior weight of hyperedge e and the sixth result. The sixth result is the sum of the seventh result and the eighth result.
[0073] Among them, the seventh result is the product of the EDVW of vertex u in the process of entering hyperedge e from vertex u, the EDVW of vertex u in the process of flowing out of hyperedge e from vertex u, and the objective function value with the out-degree of hyperedge e as the variable; the eighth result is the product of the EDVW of vertex v in the process of entering hyperedge e from vertex v, the EDVW of vertex v in the process of flowing out of hyperedge e from vertex v, and the objective function value with the in-degree of hyperedge e as the variable.
[0074] Specifically, in an implementable manner, according to the above formulas (1), (2) and (3), the transition probability P(u, v) from vertex u to vertex v can be expressed by the following formula (4):
[0075]
[0076] Among them, w(e) is the prior weight of hyperedge e, and ε is the hyperedge set.
[0077] S2. According to the transition probability from vertex u to vertex v, the probability transfer matrix of the hypergraph is obtained. The transition probability from vertex u to vertex v is an element in the probability transfer matrix of the hypergraph.
[0078] Specifically, after obtaining the transition probability of any two vertices in the vertex set through S1, the probability transition matrix of the hypergraph can be obtained. According to the above formula (4), the probability transition matrix P of the hypergraph can be expressed by the following formula (5):
[0079]
[0080] Where W is the diagonal matrix of hyperedge weights, as well as are all diagonal matrices, and the elements on the diagonal correspond to the out-degree of the hyperedge, the in-degree of the hyperedge, and the degree of the vertex defined above, respectively. ρ(·) is the objective function. out For Q out (u, e)) corresponds to the matrix, Q in For Q in The matrix corresponding to (u, e).
[0081] Specifically, according to the probability transfer matrix of the hypergraph, the weighted adjacency matrix of the vertices in the hypergraph can be determined. According to the above formula (5), the weighted adjacency matrix A of the vertices in the hypergraph can be determined, and A can be expressed by the following formula (6):
[0082]
[0083] Optionally, the Laplacian matrix of the hypergraph is determined according to the weighted adjacency matrix of the vertices in the hypergraph, which can be specifically:
[0084] The weighted adjacency matrix of the vertices in the hypergraph is symmetrically regularized to obtain the first target matrix.
[0085] Subtracting the first target matrix from the identity matrix yields the Laplacian matrix of the hypergraph.
[0086] Specifically, according to formula (6), the weighted adjacency matrix A of the vertices in the hypergraph is symmetric regularized, and the first target matrix can be expressed by the following formula (7):
[0087]
[0088] The definitions of the relevant parameters in formula (7) can refer to the definitions of the same parameters in the previous formula and will not be repeated here.
[0089] Subtracting the first target matrix from the identity matrix, the Laplacian matrix L of the hypergraph can be expressed as follows:
[0090]
[0091] In the embodiment of the present application, the eigenvalue range of L is [0, 2], which provides a theoretical guarantee for the stability of the hypergraph neural network model of the hypergraph.
[0092] In an embodiment of the present application, after obtaining the Laplace matrix of the hypergraph, a hypergraph neural network model of the hypergraph can be constructed based on the Laplace matrix of the hypergraph.
[0093] As an implementable method, a hypergraph neural network model of a hypergraph is constructed according to the Laplacian matrix of the hypergraph, which may specifically include:
[0094] S1. Determine the Fourier transform on the hypergraph and the inverse Fourier transform on the hypergraph according to the characteristic matrix of the Laplacian matrix of the hypergraph and the frequency domain signals of the vertices in the hypergraph.
[0095] Specifically, in this embodiment, the characteristic matrix U of the Laplacian matrix L of the hypergraph is defined as (u1, u2, ..., u n ), where u iis the eigenvector of L. According to the eigenmatrix of the Laplacian matrix of the hypergraph and the frequency domain signal of the vertices in the hypergraph, the Fourier transform on the hypergraph and the inverse Fourier transform on the hypergraph can be expressed by the following formula (9):
[0096]
[0097] Where x∈R n is a frequency domain signal at a hypergraph vertex, is the transformed spectral domain signal.
[0098] S12. Use the second-order Chebyshev polynomials of the eigenvalues of the second target matrix to estimate the filter from any input vertex to any output vertex on the hypergraph, and obtain the estimated result of the filter from any input vertex to any output vertex on the hypergraph. The second target matrix is the matrix obtained by subtracting the unit matrix from the Laplace matrix of the hypergraph.
[0099] Specifically, the second target matrix Since the eigenvalues of L range from [0, 2] and the eigenvalues of the second target matrix range from [-1, 1], it is suitable for Chebyshev polynomial estimation.
[0100] S13. Determine the convolution operation of the hypergraph according to the Fourier transform on the hypergraph, the inverse Fourier transform on the hypergraph, and the estimated results of the filter from any input vertex to any output vertex on the hypergraph.
[0101] Optionally, the convolution operation of the hypergraph can be in the form of a filter from any input vertex to any output vertex on the hypergraph, a first matrix and the product of the frequency domain signal of the input vertex, where the first matrix is the sum of the unit matrix and the first target matrix.
[0102] Wherein, the first target matrix is shown in the above formula (7), and the first matrix can be:
[0103] Optionally, the form of the convolution operation of the hypergraph can be expressed as follows:
[0104]
[0105] Where θ0 = -θ1 = θ g Therefore, a θ g represents a filter between a pair of input vertices and an output vertex, θ g is a learnable parameter.
[0106] S14. Determine the signal convolution operator of the hypergraph according to the convolution operation of the hypergraph and the convolution kernel formed by all filters on the hypergraph.
[0107] Optionally, according to the convolution operation of the hypergraph and the convolution kernel formed by all filters on the hypergraph, a signal convolution operator of the hypergraph is determined, which can be specifically:
[0108] According to the first matrix and convolution kernel in the convolution operation of the hypergraph, the relationship between the input vertex and the output vertex on the hypergraph is determined as follows: the output vertex is equal to the product of the first matrix, the input vertex and the convolution kernel. The relationship between the input vertex and the output vertex on the hypergraph is determined as the signal convolution operator of the hypergraph.
[0109] For example, if the input vertex on the hypergraph There is C in Channel, output vertex X out There is C out channels, you can use Represents the convolution kernel composed of all filters (the convolution kernel is a learnable parameter matrix). The relationship between the input vertices and the output vertices on the hypergraph (i.e., the signal convolution operator of the hypergraph) can be expressed by the following formula (11):
[0110]
[0111] S15. Determine a hypergraph neural network model of the hypergraph based on the signal convolution operator and activation function of the hypergraph.
[0112] Specifically, in the above formula 11, since the eigenvalue range of L is [0, 2], then The eigenvalue range of is also [0, 2], which will cause numerical instability and gradient explosion problems in the hypergraph neural network. To solve this problem, in the embodiment of the present application, the hypergraph neural network model of the hypergraph is determined according to the signal convolution operator and activation function of the hypergraph, which can be specifically:
[0113] S151. Calculate a row-normalized diagonal matrix according to the first matrix in the signal convolution operator of the hypergraph.
[0114] Among them, the formula of the first matrix is expressed as:
[0115] The row-normalized diagonal matrix calculated from the first matrix can be expressed as follows:
[0116]
[0117] The first matrix can also be called a self-loop adjacency matrix. is the row sum of the first matrix.
[0118] S152. Renormalize the first matrix using a row-normalized diagonal matrix to obtain a second matrix.
[0119] Specifically, the row-normalized diagonal matrix The first matrix is re-normalized, and the obtained second matrix can be expressed as follows:
[0120]
[0121] in, is a normalized diagonal matrix. The definitions of other parameters are the same as those in the previous formula and will not be repeated here.
[0122] S153. Determine the relationship between the input of the lth layer and the output of the lth layer in the hypergraph neural network model of the hypergraph according to the second matrix, the convolution kernel parameters and the activation function.
[0123] S154. According to the relationship between the input of the lth layer and the output of the lth layer, a hypergraph neural network model of the hypergraph is obtained.
[0124] Optionally, the hypergraph neural network model of the hypergraph obtained by S154 can be expressed by the following formula (14):
[0125]
[0126] Among them, X (l) is the input of the network layer l, X (l+1) is the output of the lth layer of the network, Θ is the learnable convolution kernel parameter in the network, and ReLu is the activation function.
[0127] The task processing method provided in this embodiment first obtains the original features of the target vertex, which is any vertex in the pre-constructed hypergraph, inputs the original features of the target vertex into the hypergraph neural network model, outputs the embedded features of the target vertex, and processes the target task according to the embedded features of the target vertex. The hypergraph neural network model is determined according to the probability transfer matrix of the hypergraph, and the probability transfer matrix of the hypergraph is determined according to the vertex set, the hyperedge set, the hyperedge weight set and the EDVW set of the vertex. Since the determination factor of the probability transfer matrix of the hypergraph adds the EDVW of the vertex, that is, the fine-grained information that the hypergraph topological structure can represent is taken into account, the hypergraph random walk under the probability transfer matrix of the hypergraph can capture richer hypergraph topological information. Therefore, the performance of the hypergraph neural network model of the hypergraph obtained according to the probability transfer matrix of the hypergraph is better, and then the use of the hypergraph neural network model for target task processing can improve the accuracy of target task processing.
[0128] The hypergraph neural network model obtained by the above method can not only give play to and reflect the advantages of the hypergraph topological structure by modeling the real "graph" problem as a hypergraph, and better capture the high-order and fine-grained cross-information between the hypergraph vertices, but also has good theoretical analytic properties because it is rooted in EDVW-Hypergraph Random Walk (Edge-Dependent Vertex Weight Hypergraph RandomWalk). The hypergraph Laplace matrix established in this application can cover most of the existing hypergraph Laplace matrices, which means that the hypergraph neural network model proposed in this application has stronger scalability and higher performance.
[0129] The task processing process is described in detail below with reference to a specific embodiment.
[0130] Figure 4 A flowchart of the process of obtaining a hypergraph neural network model in a task processing method provided in an embodiment of the present application, such as Figure 4 As shown, the method may include the following steps:
[0131] S201. Construct a hypergraph, where the hypergraph includes a vertex set, a hyperedge set, a hyperedge weight set, and an EDVW set of vertices.
[0132] S202: Determine a probability transfer matrix of the hypergraph according to the vertex set, hyperedge set, hyperedge weight set and vertex EDVW set included in the hypergraph.
[0133] Specifically, the transition probability P(u,v) from vertex u to vertex v defined by the “Uniform Hypergraph Random Walk” in this embodiment can be expressed by the following formula (4):
[0134]
[0135] Among them, w(e) is the prior weight of hyperedge e. The definitions of other parameters can be found in the description of the above formula and will not be repeated here. ε is the hyperedge set, and d(u) is the degree of vertex u.
[0136] The degree d(u) of vertex u can be expressed by the following formula (1):
[0137]
[0138] The out-degree δ of hyperedge e out (e) can be expressed by the following formula (2):
[0139]
[0140] The in-degree δ of hyperedge e in (e) can be expressed by the following formula (3):
[0141]
[0142] After obtaining the transition probability of any two vertices in the vertex set according to formula (4), the probability transition matrix of the hypergraph can be obtained. According to the above formula (4), the probability transition matrix P of the hypergraph can be expressed by the following formula (5):
[0143]
[0144] in, is the diagonal matrix of vertex degrees, W is the diagonal matrix of hyperedge weights, is the diagonal matrix of the out-degrees of the hyperedges, is the diagonal matrix of the in-degree of the hyperedges, and ρ(·) is the objective function.
[0145] S203. Determine a weighted adjacency matrix of vertices in the hypergraph according to the probability transfer matrix of the hypergraph, and determine a Laplacian matrix of the hypergraph according to the weighted adjacency matrix of vertices in the hypergraph.
[0146] Specifically, according to the probability transfer matrix of the hypergraph, the weighted adjacency matrix of the vertices in the hypergraph can be determined. According to formula (5), the weighted adjacency matrix A of the vertices in the hypergraph can be determined, and A can be expressed by the following formula (6):
[0147]
[0148] According to formula (6), the weighted adjacency matrix of the vertices in the hypergraph is symmetric regularized, and the first target matrix can be expressed by the following formula (7):
[0149]
[0150] The definitions of the relevant parameters in formula (7) can refer to the definitions of the same parameters in the previous formula and will not be repeated here.
[0151] Subtracting the first target matrix from the identity matrix, the Laplacian matrix L of the hypergraph can be expressed as follows:
[0152]
[0153] In the embodiment of the present application, the eigenvalue range of L is [0, 2], which provides a theoretical guarantee for the stability of the hypergraph neural network model of the hypergraph.
[0154] S204. Determine the Fourier transform on the hypergraph and the inverse Fourier transform on the hypergraph according to the characteristic matrix of the Laplacian matrix of the hypergraph and the frequency domain signals of the vertices in the hypergraph.
[0155] Specifically, in this embodiment, the characteristic matrix U of the Laplacian matrix L of the hypergraph is defined as (u1, u2, ..., u n ), where u i is the eigenvector of L. According to the eigenmatrix of the Laplacian matrix of the hypergraph and the frequency domain signal of the vertices in the hypergraph, the Fourier transform on the hypergraph and the inverse Fourier transform on the hypergraph can be expressed by the following formula (9):
[0156]
[0157] Where x∈R n is a frequency domain signal on the hypergraph vertex, is the transformed spectral domain signal.
[0158] S205. Use the second-order Chebyshev polynomials of the eigenvalues of the second target matrix to estimate the filter from any input vertex to any output vertex on the hypergraph, and obtain the estimated result of the filter from any input vertex to any output vertex on the hypergraph. Determine the convolution operation of the hypergraph based on the Fourier transform on the hypergraph, the inverse Fourier transform on the hypergraph, and the estimated result of the filter from any input vertex to any output vertex on the hypergraph.
[0159] Among them, the second target matrix
[0160] The form of the convolution operation of the hypergraph is:
[0161]
[0162] Where θ0 = -θ1 = θ g Therefore, a θ g represents a filter between a pair of input vertices and an output vertex, θ g is a learnable parameter.
[0163] S206. Determine a signal convolution operator of the hypergraph according to the convolution operation of the hypergraph and the convolution kernel formed by all filters on the hypergraph.
[0164] Specifically, if the input vertex on the hypergraph There is C in Channel, output vertex X out There is C out channels, you can use Represents the convolution kernel composed of all filters (the convolution kernel is a learnable parameter matrix). The relationship between the input vertices and the output vertices on the hypergraph (i.e., the signal convolution operator of the hypergraph) can be expressed by the following formula (11):
[0165]
[0166] S207. Calculate a row-normalized diagonal matrix based on the first matrix in the signal convolution operator of the hypergraph, re-regularize the first matrix by the row-normalized diagonal matrix to obtain a second matrix, determine the relationship between the input of the lth layer and the output of the lth layer in the hypergraph neural network model of the hypergraph based on the second matrix, convolution kernel parameters and activation function, and obtain the hypergraph neural network model of the hypergraph based on the relationship between the input of the lth layer and the output of the lth layer.
[0167] Among them, the formula of the first matrix is expressed as:
[0168] The row-normalized diagonal matrix calculated from the first matrix can be expressed as follows:
[0169]
[0170] The first matrix can also be called a self-loop adjacency matrix. is the row sum of the first matrix.
[0171] Specifically, the row-normalized diagonal matrix The first matrix is re-normalized, and the obtained second matrix can be expressed as follows:
[0172]
[0173] in, is a normalized diagonal matrix. The definitions of other parameters are the same as those in the previous formula and will not be repeated here.
[0174] In this embodiment, the relationship between the input of the lth layer and the output of the lth layer in the hypergraph neural network model of the hypergraph is determined according to the second matrix, the convolution kernel parameters and the activation function, that is, the hypergraph neural network model can be expressed by the following formula (14):
[0175]
[0176] Among them, X (l) is the input of the lth layer of the network, Θ is the learnable convolution kernel parameter in the network, and ReLu is the activation function.
[0177] Figure 5 The flowchart of a task processing method provided in an embodiment of the present application can be executed by a task processing device, and the task processing device can be implemented by software and / or hardware. Optionally, the task processing device can be a server. Figure 5 As shown, the method may include the following steps:
[0178] S301. Obtain original features of a target vertex, where the target vertex is a vertex in a hypergraph. The hypergraph includes a vertex set, a hyperedge set, a hyperedge weight set, and an EDVW set of the vertex.
[0179] S302: Input the original features of the target vertex into the hypergraph neural network model, and output the embedded features of the target vertex.
[0180] S303: Input the embedded features of the target vertex into the downstream classifier, and output the classification label of the target vertex.
[0181] S304: Obtain the category to which the target vertex belongs according to the category label of the target vertex.
[0182] Among them, the hypergraph neural network model is based on Figure 4 The task processing method shown in the figure is obtained. The detailed process can be found in Figure 4 The embodiments shown will not be described in detail here.
[0183] The task processing method provided in this embodiment, by using a hypergraph neural network model, can improve the accuracy of vertex classification compared to the existing hypergraph convolutional neural network.
[0184] Combine the following Figure 6 The detailed process of the task processing method provided by the embodiment of the present application in an actual application scenario is described. In this embodiment, the application scenario is taken as a citation network as an example. In academic literature, mutual citations between papers or co-authors of papers constitute a certain relationship between papers.
[0185] Figure 6 A flowchart of a task processing method provided in an embodiment of the present application is shown in FIG. Figure 6 As shown, the method of this embodiment may include:
[0186] S301, construct a hypergraph, the hypergraph including a vertex set, a hyperedge set, a hyperedge weight set and an EDVW set of vertices.
[0187] Specifically, in a citation network, vertices are defined as papers, and edges are formed by citation relationships or co-authors between papers. In a paper database, all papers signed by the same author form a hyperedge, or all papers that cite the same paper form a hyperedge. For example, a hypergraph can be constructed based on the definitions of papers and hyperedges in a paper database, and the constructed hypergraph is the hypergraph corresponding to the paper database.
[0188] by Figure 1 Take the hypergraph shown as an example, Figure 1 is a schematic diagram of a hypergraph constructed based on a citation network, such as Figure 1The hypergraph shown in the figure includes two hyperedges, one paper is a vertex, and all papers signed by the same author A constitute a hyperedge a, that is, Figure 1 The hyperedge 1 shown in , hyperedge 1 includes vertices 1, 2, 3 and 4 and the connection relationship between the four vertices. All papers signed by the same author B constitute a hyperedge b, that is, Figure 1 The hyperedge 2 shown in , hyperedge 2 includes vertex 3, vertex 4, vertex 5 and vertex 6 and the connection relationship between the four vertices. Figure 1 In the hypergraph shown, the vertex set includes vertex 1, vertex 2, vertex 3, vertex 4, vertex 5 and vertex 6; the hyperedge set includes hyperedge 1 and hyperedge 2; the hyperedge weight set includes the weight of hyperedge 1 and the weight of hyperedge 2, the weight is the prior weight of the hyperedge, and the prior weight can be a pre-set weight; the vertex EDVW set includes: the EDVW of vertex 1 in the process of vertex 1 entering hyperedge 1, the EDVW of vertex 1 in the process of vertex 1 flowing out of hyperedge 1, the EDVW of vertex 2 in the process of vertex 2 entering hyperedge 1, the EDVW of vertex 2 in the process of vertex 2 flowing out of hyperedge 1, the EDVW of vertex 3 in the process of vertex 3 flowing out of hyperedge 1 The EDVW of vertex 3 in the process, the EDVW of vertex 4 in the process of vertex 4 entering hyperedge 1, the EDVW of vertex 4 in the process of vertex 4 flowing out of hyperedge 1, the EDVW of vertex 3 in the process of vertex 3 entering hyperedge 2, the EDVW of vertex 3 in the process of vertex 3 flowing out of hyperedge 2, the EDVW of vertex 4 in the process of vertex 4 entering hyperedge 2, the EDVW of vertex 4 in the process of vertex 4 flowing out of hyperedge 2, the EDVW of vertex 5 in the process of vertex 5 entering hyperedge 2, the EDVW of vertex 5 in the process of vertex 5 flowing out of hyperedge 2, the EDVW of vertex 6 in the process of vertex 6 entering hyperedge 2, and the EDVW of vertex 6 in the process of vertex 6 flowing out of hyperedge 2. Among them, the EDVW in the vertex EDVW set can be pre-set or calculated according to the hypergraph.
[0189] It is understandable that in practical applications, the hypergraph constructed based on the citation network includes more than two hyperedges.
[0190] S302: Determine a probability transfer matrix of the hypergraph according to the vertex set, hyperedge set, hyperedge weight set and vertex EDVW set included in the hypergraph.
[0191] S303: Determine a weighted adjacency matrix of vertices in the hypergraph according to the probability transfer matrix of the hypergraph, and determine a Laplacian matrix of the hypergraph according to the weighted adjacency matrix of vertices in the hypergraph.
[0192] S304: Determine the Fourier transform on the hypergraph and the inverse Fourier transform on the hypergraph according to the characteristic matrix of the Laplacian matrix of the hypergraph and the frequency domain signals of the vertices in the hypergraph.
[0193] S305. Use the second-order Chebyshev polynomials of the eigenvalues of the second target matrix to estimate the filter from any input vertex to any output vertex on the hypergraph, and obtain the estimated result of the filter from any input vertex to any output vertex on the hypergraph. Determine the convolution operation of the hypergraph based on the Fourier transform on the hypergraph, the inverse Fourier transform on the hypergraph, and the estimated result of the filter from any input vertex to any output vertex on the hypergraph.
[0194] Among them, the second target matrix
[0195] S306. Determine a signal convolution operator of the hypergraph according to the convolution operation of the hypergraph and the convolution kernel formed by all filters on the hypergraph.
[0196] S307. Calculate a row-normalized diagonal matrix based on the first matrix in the signal convolution operator of the hypergraph, re-regularize the first matrix by the row-normalized diagonal matrix to obtain a second matrix, determine the relationship between the input of the lth layer and the output of the lth layer in the hypergraph neural network model of the hypergraph based on the second matrix, convolution kernel parameters and activation function, and obtain the hypergraph neural network model of the hypergraph based on the relationship between the input of the lth layer and the output of the lth layer.
[0197] In this embodiment, the relationship between the input of the lth layer and the output of the lth layer in the hypergraph neural network model of the hypergraph is determined according to the second matrix, the convolution kernel parameters and the activation function, that is, the hypergraph neural network model can be expressed by the following formula (14):
[0198]
[0199] Among them, X (l) is the input of the lth layer of the network, Θ is the learnable convolution kernel parameter in the network, and ReLu is the activation function.
[0200] The detailed process of S302-S307 can be found in Figure 4 The description of the illustrated embodiment will not be repeated here.
[0201] After obtaining the hypergraph neural network model of the hypergraph constructed by S301 through S302-S307, the hypergraph neural network model can be used to classify each vertex in the hypergraph to obtain the classification of the paper corresponding to each vertex. Specifically, for each vertex, the following S308-S311 can be used to obtain the classification of the vertex.
[0202] S308. Obtain original features of a target vertex, where the target vertex is a vertex in the hypergraph.
[0203] In this embodiment, a vertex is a paper, and the original feature of the vertex may be text information of the paper, for example, keywords that appear frequently in the paper, and so on.
[0204] S309: Input the original features of the target vertex into the hypergraph neural network model, and output the embedded features of the target vertex.
[0205] S310: Input the embedded features of the target vertex into a downstream classifier, and output the classification label of the target vertex.
[0206] In a citation network, a vertex is a paper, and the classification label of the vertex can include the field to which the paper belongs, such as social, scientific, and so on.
[0207] S311. Obtain the category to which the target vertex belongs according to the category label of the target vertex.
[0208] It should be noted that in the classification of 3D-visual objects, the specific processing process is similar to that of the citation network. The difference is that when constructing a hypergraph, each 3D object is a vertex, and the method for constructing a hyperedge can be: define a distance in the vertex feature space, take any 3D object j (corresponding to vertex j) as the center point, use the K nearest neighbor algorithm, select the K nearest vertices, use the feature distance between vertex j and its K neighbor vertices as the weight of vertex j, and construct a hyperedge with a hyperedge degree (that is, the number of vertices in the hyperedge) of K. After the hypergraph is constructed, the hypergraph neural network model of the hypergraph can be first obtained through the above method, and then the classification of each 3D object can be obtained.
[0209] The task processing method provided in this embodiment, in the citation network experiment and the 3D-visual object recognition experiment, by adopting the task processing method provided in this embodiment, the classification accuracy greatly surpasses the existing hypergraph convolutional neural network.
[0210] The following is an embodiment of the device of the present application, which can be used to execute the above method embodiment of the present application. For details not disclosed in the embodiment of the device of the present application, reference can be made to the above method embodiment of the present application.
[0211] Figure 7 A schematic diagram of a task processing device provided in an embodiment of the present application is shown in FIG. Figure 7 As shown, the device of this embodiment may include: an acquisition module 11 and a processing module 12.
[0212] The acquisition module 11 is used to acquire the original features of the target vertex, the target vertex is any vertex in the pre-constructed hypergraph, and the hypergraph includes a vertex set, a hyperedge set, a hyperedge weight set, and a hyperedge dependent vertex weight EDVW set of the vertex;
[0213] The processing module 12 is used to input the original features of the target vertex into the hypergraph neural network model and output the embedded features of the target vertex, the hypergraph neural network model is determined according to the probability transfer matrix of the hypergraph, and the probability transfer matrix of the hypergraph is determined according to the vertex set, the hyperedge set, the hyperedge weight set and the EDVW set of the vertex;
[0214] The processing module 12 is further used to process a target task according to the embedded features of the target vertex, where the target task includes any one of vertex classification, vertex clustering and regression tasks.
[0215] Optionally, the target task is vertex classification, and the processing module 12 is used to input the embedded features of the target vertex into a downstream classifier, output the classification label of the target vertex, and obtain the classification to which the target vertex belongs according to the classification label of the target vertex.
[0216] Optionally, the hypergraph neural network model is determined according to the probability transfer matrix of the hypergraph in the following manner:
[0217] According to the probability transfer matrix of the hypergraph, a weighted adjacency matrix of the vertices in the hypergraph is determined, and according to the weighted adjacency matrix of the vertices in the hypergraph, a Laplacian matrix of the hypergraph is determined;
[0218] According to the Laplace matrix of the hypergraph, a hypergraph neural network model of the hypergraph is constructed.
[0219] Optionally, the probability transfer matrix of the hypergraph is determined according to the vertex set, the hyperedge set, the hyperedge weight set and the EDVW set of the vertices in the following manner:
[0220] According to the prior weight of the hyperedge e, the EDVW of the vertex u in the process of entering the hyperedge e from the vertex u, the EDVW of the vertex u in the process of flowing out of the hyperedge e from the vertex u, the EDVW of the vertex v in the process of entering the hyperedge e from the vertex v, the EDVW of the vertex v in the process of flowing out of the hyperedge e from the vertex v, the out-degree of the hyperedge e, the in-degree of the hyperedge e, the degree of the vertex u and the objective function, calculate the transition probability from the vertex u to the vertex v, the vertex u and the vertex v are any two vertices in the vertex set, and the hyperedge e is the hyperedge between the vertex u and the vertex v;
[0221] The out-degree of the hyperedge e is the sum of the EDVWs of all vertices adjacent to the hyperedge e flowing out of the hyperedge e, the in-degree of the hyperedge e is the sum of the EDVWs of all vertices adjacent to the hyperedge e entering the hyperedge e, and the degree of the vertex u is determined according to the prior weight of the hyperedge e, the EDVW of the vertex u in the process of entering the hyperedge e from the vertex u, the EDVW of the vertex u in the process of flowing out of the hyperedge e from the vertex u, the out-degree of the hyperedge e, the in-degree of the hyperedge e, and the objective function, and the objective function is used to represent the relationship between the out-degree of the hyperedge e and the degree of the vertex u;
[0222] According to the transition probability from the vertex u to the vertex v, a probability transfer matrix of the hypergraph is obtained, and the transition probability from the vertex u to the vertex v is an element in the probability transfer matrix of the hypergraph.
[0223] Optionally, the degree of the vertex u is:
[0224] The sum of the first results of all hyperedges e between the vertex u and the vertex v, the first result being the product of the prior weight of the hyperedge e and the second result, the second result being the sum of the third result and the fourth result, the third result being the product of the out-degree of the hyperedge e, the objective function value with the out-degree of the hyperedge e as a variable, and the EDVW of the vertex u in the process of entering the hyperedge e from the vertex u, and the fourth result being the product of the in-degree of the hyperedge e, the objective function value with the in-degree of the hyperedge e as a variable, and the EDVW of the vertex u in the process of flowing out of the hyperedge e from the vertex u.
[0225] Optionally, the transition probability from the vertex u to the vertex v is:
[0226] The target result is divided by twice the degree of the vertex u, the target result is the sum of the fifth results of all hyperedges e between the vertex u and the vertex v, the fifth result is the product of the prior weight of the hyperedge e and the sixth result, and the sixth result is the sum of the seventh result and the eighth result;
[0227] Among them, the seventh result is the product of the EDVW of the vertex u in the process of entering the hyperedge e from the vertex u, the EDVW of the vertex u in the process of flowing out of the hyperedge e from the vertex u, and the objective function value with the out-degree of the hyperedge e as the variable; the eighth result is the product of the EDVW of the vertex v in the process of entering the hyperedge e from the vertex v, the EDVW of the vertex v in the process of flowing out of the hyperedge e from the vertex v, and the objective function value with the in-degree of the hyperedge e as the variable.
[0228] Optionally, a hypergraph neural network model of the hypergraph is constructed according to the Laplace matrix of the hypergraph, which may be specifically:
[0229] Determine a Fourier transform on the hypergraph and an inverse Fourier transform on the hypergraph according to a characteristic matrix of the Laplacian matrix of the hypergraph and frequency domain signals of vertices in the hypergraph;
[0230] Using the second-order Chebyshev polynomial of the eigenvalue of the second target matrix, estimating the filter from any input vertex to any output vertex on the hypergraph, to obtain an estimation result of the filter from any input vertex to any output vertex on the hypergraph, the second target matrix being a matrix obtained by subtracting the unit matrix from the Laplacian matrix of the hypergraph;
[0231] Determine a convolution operation of the hypergraph according to a Fourier transform on the hypergraph, an inverse Fourier transform on the hypergraph, and an estimation result of a filter from any input vertex to any output vertex on the hypergraph;
[0232] Determine a signal convolution operator of the hypergraph according to the convolution operation of the hypergraph and the convolution kernel formed by all the filters on the hypergraph;
[0233] According to the signal convolution operator and activation function of the hypergraph, a hypergraph neural network model of the hypergraph is determined.
[0234] Optionally, the convolution operation on the hypergraph can be of the form:
[0235] The filter from any input vertex to any output vertex on the hypergraph is the product of a first matrix and a frequency domain signal of the input vertex, wherein the first matrix is the sum of a unit matrix and the first target matrix.
[0236] Optionally, the signal convolution operator of the hypergraph is determined according to the convolution operation of the hypergraph and the convolution kernel formed by all the filters on the hypergraph, which may be specifically:
[0237] Determine, according to the first matrix and the convolution kernel in the convolution operation of the hypergraph, a relationship between an input vertex and an output vertex on the hypergraph: the output vertex is equal to a product of the first matrix, the input vertex and the convolution kernel;
[0238] The relationship between the input vertices and the output vertices on the hypergraph is determined as a signal convolution operator of the hypergraph.
[0239] Optionally, the hypergraph neural network model of the hypergraph is determined according to the signal convolution operator and activation function of the hypergraph, which may be:
[0240] Calculating a row-normalized diagonal matrix according to the first matrix in the signal convolution operator of the hypergraph;
[0241] Renormalize the first matrix by using the row-normalized diagonal matrix to obtain a second matrix;
[0242] Determine, according to the second matrix, the convolution kernel parameters and the activation function, a relationship between an input of the lth layer and an output of the lth layer in the hypergraph neural network model of the hypergraph;
[0243] According to the relationship between the input of the lth layer and the output of the lth layer, a hypergraph neural network model of the hypergraph is obtained.
[0244] The task processing device of the embodiment of the present application is described above from the perspective of the functional module in conjunction with the accompanying drawings. It should be understood that the functional module can be implemented in hardware form, can be implemented by instructions in software form, and can also be implemented by a combination of hardware and software modules. Specifically, the steps of the method embodiment in the embodiment of the present application can be completed by the hardware integrated logic circuit and / or software form instructions in the processor, and the steps of the method disclosed in the embodiment of the present application can be directly embodied as a hardware decoding processor to perform, or a combination of hardware and software modules in the decoding processor to perform. Optionally, the software module can be located in a mature storage medium in the field such as a random access memory, a flash memory, a read-only memory, a programmable read-only memory, an electrically erasable programmable memory, a register, etc. The storage medium is located in a memory, and the processor reads the information in the memory, and completes the steps in the above method embodiment in conjunction with its hardware.
[0245] Figure 8 It is a schematic block diagram of a task processing device 800 provided in an embodiment of the present application.
[0246] like Figure 8 As shown, the task processing device 800 may include:
[0247] The memory 810 and the processor 820, the memory 810 is used to store the computer program and transmit the program code to the processor 820. In other words, the processor 820 can call and run the computer program from the memory 810 to implement the method in the embodiment of the present application.
[0248] For example, the processor 820 may be configured to execute the above method embodiments according to instructions in the computer program.
[0249] In some embodiments of the present application, the processor 820 may include but is not limited to:
[0250] General-purpose processor, digital signal processor (DSP), application-specific integrated circuit (ASIC), field programmable gate array (FPGA) or other programmable logic device, discrete gate or transistor logic device, discrete hardware components, etc.
[0251] In some embodiments of the present application, the memory 810 includes but is not limited to:
[0252] Volatile memory and / or non-volatile memory. Among them, the non-volatile memory can be read-only memory (ROM), programmable ROM (PROM), erasable programmable read-only memory (EPROM), electrically erasable programmable read-only memory (EEPROM) or flash memory. The volatile memory can be random access memory (RAM), which is used as an external cache. By way of example and not limitation, many forms of RAM are available, such as static random access memory (SRAM), dynamic random access memory (DRAM), synchronous dynamic random access memory (SDRAM), double data rate synchronous dynamic random access memory (DDR SDRAM), enhanced synchronous dynamic random access memory (ESDRAM), synchronous link DRAM (SLDRAM), and direct RAM bus random access memory (DR RAM).
[0253] In some embodiments of the present application, the computer program may be divided into one or more modules, which are stored in the memory 810 and executed by the processor 820 to complete the method provided by the present application. The one or more modules may be a series of computer program instruction segments capable of completing specific functions, and the instruction segments are used to describe the execution process of the computer program in the device.
[0254] like Figure 8 As shown, the task processing device 800 may also include:
[0255] The transceiver 830 may be connected to the processor 820 or the memory 810 .
[0256] The processor 820 may control the transceiver 830 to communicate with other devices, specifically, to send information or data to other devices, or to receive information or data sent by other devices. The transceiver 830 may include a transmitter and a receiver. The transceiver 830 may further include an antenna, and the number of antennas may be one or more.
[0257] It should be understood that the various components in the device are connected via a bus system, wherein the bus system includes not only a data bus but also a power bus, a control bus and a status signal bus.
[0258] The present application also provides a computer storage medium on which a computer program is stored, and when the computer program is executed by a computer, the computer can perform the method of the above method embodiment. In other words, the present application embodiment also provides a computer program product containing instructions, and when the instructions are executed by a computer, the computer can perform the method of the above method embodiment.
[0259] When software is used for implementation, it can be implemented in whole or in part in the form of a computer program product. The computer program product includes one or more computer instructions. When the computer program instructions are loaded and executed on a computer, the process or function according to the embodiment of the present application is generated in whole or in part. The computer can be a general-purpose computer, a special-purpose computer, a computer network, or other programmable devices. The computer instructions can be stored in a computer-readable storage medium, or transmitted from one computer-readable storage medium to another computer-readable storage medium. For example, the computer instructions can be transmitted from a website site, computer, server or data center by wired (e.g., coaxial cable, optical fiber, digital subscriber line (digital subscriber line, DSL)) or wireless (e.g., infrared, wireless, microwave, etc.) mode to another website site, computer, server or data center. The computer-readable storage medium can be any available medium that a computer can access or a data storage device such as a server or data center that includes one or more available media integration. The available medium can be a magnetic medium (e.g., a floppy disk, a hard disk, a tape), an optical medium (e.g., a digital video disc (digital video disc, DVD)), or a semiconductor medium (e.g., a solid state drive (solid state disk, SSD)), etc.
[0260] Those of ordinary skill in the art will appreciate that the modules and algorithm steps of each example described in conjunction with the embodiments disclosed herein can be implemented in electronic hardware, or a combination of computer software and electronic hardware. Whether these functions are performed in hardware or software depends on the specific application and design constraints of the technical solution. Professional and technical personnel can use different methods to implement the described functions for each specific application, but such implementation should not be considered to be beyond the scope of this application.
[0261] In the several embodiments provided in the present application, it should be understood that the disclosed systems, devices and methods can be implemented in other ways. For example, the device embodiments described above are only schematic. For example, the division of the module is only a logical function division. 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. Another point is that the mutual coupling or direct coupling or communication connection shown or discussed can be through some interfaces, indirect coupling or communication connection of devices or modules, which can be electrical, mechanical or other forms.
[0262] The modules described as separate components may or may not be physically separated, and the components displayed as modules may or may not be physical modules, that is, they may be located in one place, or they may be distributed on multiple network units. Some or all of the modules may be selected according to actual needs to achieve the purpose of the scheme of this embodiment. For example, each functional module in each embodiment of the present application may be integrated into a processing module, or each module may exist physically separately, or two or more modules may be integrated into one module.
[0263] The above is only a specific implementation of the present application, but the protection scope of the present application is not limited thereto. Any technician familiar with the 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 task processing method, It is characterized in that include: Obtaining the original features of a target vertex, wherein the target vertex is any vertex in a pre-constructed hypergraph, wherein the hypergraph includes a vertex set, a hyperedge set, a hyperedge weight set, and a hyperedge dependent vertex weight EDVW set of the vertex, wherein the hypergraph is constructed according to a citation network, wherein a vertex in the hypergraph is a paper, and a hyperedge in the hypergraph is composed of all papers in a paper database that cite the same paper or all papers signed by the same author; or, wherein the hypergraph is constructed according to 3D-visual object classification, wherein a vertex in the hypergraph is a 3D object, and the construction process of the hyperedge in the hypergraph is as follows: defining a distance in the vertex feature space, taking any 3D object j as the center point, using a K nearest neighbor algorithm, selecting the K nearest vertices, and taking the feature distance between vertex j corresponding to the object j and K neighbor vertices as the weight of the vertex j, and constructing a hyperedge with a hyperedge degree of K; The original features of the target vertex are input into a hypergraph neural network model, and the embedded features of the target vertex are output. The hypergraph neural network model is determined according to the probability transfer matrix of the hypergraph, and the probability transfer matrix of the hypergraph is determined according to the vertex set, the hyperedge set, the hyperedge weight set and the EDVW set of the vertex in the following manner: According to the prior weight of hyperedge e, the EDVW of vertex u in the process of entering hyperedge e from vertex u, the EDVW of vertex u in the process of flowing out of hyperedge e from vertex u, the EDVW of vertex v in the process of entering hyperedge e from vertex v, the EDVW of vertex v in the process of flowing out of hyperedge e from vertex v, the out-degree of hyperedge e, the in-degree of hyperedge e, the degree of vertex u and the objective function, the transition probability from vertex u to vertex v is calculated, wherein the out-degree of hyperedge e is the degree of the adjacent hyperedge The in-degree of the hyperedge e is the sum of the EDVWs of all vertices of the hyperedge e flowing out of the hyperedge e, the in-degree of the hyperedge e is the sum of the EDVWs of all vertices adjacent to the hyperedge e entering the hyperedge e, the degree of the vertex u is determined according to the prior weight of the hyperedge e, the EDVW of the vertex u in the process of entering the hyperedge e from the vertex u, the EDVW of the vertex u in the process of flowing out of the hyperedge e from the vertex u, the out-degree of the hyperedge e, the in-degree of the hyperedge e and the objective function, and the objective function is used to represent the relationship between the out-degree of the hyperedge e and the degree of the vertex u; According to the transition probability from the vertex u to the vertex v, a probability transfer matrix of the hypergraph is obtained, wherein the transition probability from the vertex u to the vertex v is an element in the probability transfer matrix of the hypergraph, and the transition probability from the vertex u to the vertex v is: the target result divided by twice the degree of the vertex u, the target result is the sum of the fifth results of all hyperedges e between the vertex u and the vertex v, the fifth result is the product of the prior weight of the hyperedge e and the sixth result, and the sixth result is the sum of the seventh result and the eighth result; The seventh result is the product of the EDVW of the vertex u in the process of entering the hyperedge e from the vertex u, the EDVW of the vertex u in the process of flowing out of the hyperedge e from the vertex u, and the objective function value with the out-degree of the hyperedge e as the variable; the eighth result is the product of the EDVW of the vertex v in the process of entering the hyperedge e from the vertex v, the EDVW of the vertex v in the process of flowing out of the hyperedge e from the vertex v, and the objective function value with the in-degree of the hyperedge e as the variable; According to the embedded features of the target vertex, a target task is processed, and the target task includes any one of vertex classification, vertex clustering and regression tasks.
2. The method according to claim 1, It is characterized in that The target task is vertex classification, and the processing of the target task according to the embedded features of the target vertex includes: Inputting the embedded features of the target vertex into a downstream classifier, and outputting the classification label of the target vertex; According to the classification label of the target vertex, the classification to which the target vertex belongs is obtained.
3. The method according to claim 1, It is characterized in that The hypergraph neural network model is determined according to the probability transfer matrix of the hypergraph in the following manner: Determine a weighted adjacency matrix of vertices in the hypergraph according to a probability transfer matrix of the hypergraph, and determine a Laplacian matrix of the hypergraph according to the weighted adjacency matrix of vertices in the hypergraph; According to the Laplace matrix of the hypergraph, a hypergraph neural network model of the hypergraph is constructed.
4. The method according to claim 1, It is characterized in that The degree of the vertex u is: The sum of the first results of all hyperedges e between the vertex u and the vertex v, the first result being the product of the prior weight of the hyperedge e and the second result, the second result being the sum of the third result and the fourth result, the third result being the product of the out-degree of the hyperedge e, the objective function value with the out-degree of the hyperedge e as a variable, and the EDVW of the vertex u in the process of entering the hyperedge e from the vertex u, and the fourth result being the product of the in-degree of the hyperedge e, the objective function value with the in-degree of the hyperedge e as a variable, and the EDVW of the vertex u in the process of flowing out of the hyperedge e from the vertex u.
5. The method according to claim 3, It is characterized in that Determining the Laplacian matrix of the hypergraph according to the weighted adjacency matrix of vertices in the hypergraph includes: Performing symmetric regularization on the weighted adjacency relationship matrix of the vertices in the hypergraph to obtain a first target matrix; The Laplacian matrix of the hypergraph is obtained by subtracting the first target matrix from the identity matrix.
6. The method according to claim 3, It is characterized in that The step of constructing a hypergraph neural network model of the hypergraph according to the Laplace matrix of the hypergraph includes: Determine a Fourier transform on the hypergraph and an inverse Fourier transform on the hypergraph according to a characteristic matrix of the Laplacian matrix of the hypergraph and frequency domain signals of vertices in the hypergraph; Using the second-order Chebyshev polynomial of the eigenvalue of the second target matrix, estimating the filter from any input vertex to any output vertex on the hypergraph, to obtain an estimation result of the filter from any input vertex to any output vertex on the hypergraph, the second target matrix being a matrix obtained by subtracting the unit matrix from the Laplacian matrix of the hypergraph; Determine a convolution operation of the hypergraph according to a Fourier transform on the hypergraph, an inverse Fourier transform on the hypergraph, and an estimation result of a filter from any input vertex to any output vertex on the hypergraph; Determine a signal convolution operator of the hypergraph according to the convolution operation of the hypergraph and the convolution kernel formed by all the filters on the hypergraph; According to the signal convolution operator and activation function of the hypergraph, a hypergraph neural network model of the hypergraph is determined.
7. The method according to claim 6, It is characterized in that The form of the convolution operation of the hypergraph is: The filter from any input vertex to any output vertex on the hypergraph, the product of the first matrix and the frequency domain signal of the input vertex, wherein the first matrix is the sum of the unit matrix and the first target matrix.
8. The method according to claim 7, It is characterized in that Determining the signal convolution operator of the hypergraph according to the convolution operation of the hypergraph and the convolution kernel formed by all the filters on the hypergraph includes: Determine, according to the first matrix and the convolution kernel in the convolution operation of the hypergraph, a relationship between an input vertex and an output vertex on the hypergraph: the output vertex is equal to a product of the first matrix, the input vertex and the convolution kernel; The relationship between the input vertices and the output vertices on the hypergraph is determined as a signal convolution operator of the hypergraph.
9. The method according to claim 7, It is characterized in that Determining a hypergraph neural network model of the hypergraph according to a signal convolution operator and an activation function of the hypergraph includes: Calculating a row-normalized diagonal matrix according to the first matrix in the signal convolution operator of the hypergraph; Renormalize the first matrix by using the row-normalized diagonal matrix to obtain a second matrix; Determine the first hypergraph neural network model of the hypergraph according to the second matrix, the convolution kernel parameters and the activation function. The input of the layer is The relationship between the outputs of the layers; According to the said The input of the layer is The relationship between the outputs of the layers is used to obtain a hypergraph neural network model of the hypergraph.
10. A task processing device, It is characterized in that include: An acquisition module is used to acquire the original features of a target vertex, wherein the target vertex is any vertex in a pre-constructed hypergraph, wherein the hypergraph includes a vertex set, a hyperedge set, a hyperedge weight set, and a hyperedge dependent vertex weight EDVW set of the vertex, wherein the hypergraph is constructed according to a citation network, wherein a vertex in the hypergraph is a paper, and a hyperedge in the hypergraph is composed of all papers in a paper database that cite the same paper or all papers signed by the same author; or, wherein the hypergraph is constructed according to 3D-visual object classification, wherein a vertex in the hypergraph is a 3D object, and the construction process of the hyperedge in the hypergraph is as follows: a distance is defined in the vertex feature space, and any 3D object j is taken as the center point, and the K nearest neighbor algorithm is used to select the K nearest vertices, and the feature distance between the vertex j corresponding to the object j and the K neighbor vertices is used as the weight of the vertex j, and a hyperedge with a hyperedge degree of K is constructed; A processing module is used to input the original features of the target vertex into a hypergraph neural network model, and output the embedded features of the target vertex, wherein the hypergraph neural network model is determined according to the probability transfer matrix of the hypergraph, and the probability transfer matrix of the hypergraph is determined according to the vertex set, the hyperedge set, the hyperedge weight set, and the EDVW set of the vertex in the following manner: According to the prior weight of hyperedge e, the EDVW of vertex u in the process of entering hyperedge e from vertex u, the EDVW of vertex u in the process of flowing out of hyperedge e from vertex u, the EDVW of vertex v in the process of entering hyperedge e from vertex v, the EDVW of vertex v in the process of flowing out of hyperedge e from vertex v, the out-degree of hyperedge e, the in-degree of hyperedge e, the degree of vertex u and the objective function, the transition probability from vertex u to vertex v is calculated, wherein the out-degree of hyperedge e is the degree of the adjacent hyperedge The in-degree of the hyperedge e is the sum of the EDVWs of all vertices of the hyperedge e flowing out of the hyperedge e, the in-degree of the hyperedge e is the sum of the EDVWs of all vertices adjacent to the hyperedge e entering the hyperedge e, the degree of the vertex u is determined according to the prior weight of the hyperedge e, the EDVW of the vertex u in the process of entering the hyperedge e from the vertex u, the EDVW of the vertex u in the process of flowing out of the hyperedge e from the vertex u, the out-degree of the hyperedge e, the in-degree of the hyperedge e and the objective function, and the objective function is used to represent the relationship between the out-degree of the hyperedge e and the degree of the vertex u; According to the transition probability from the vertex u to the vertex v, a probability transfer matrix of the hypergraph is obtained, wherein the transition probability from the vertex u to the vertex v is an element in the probability transfer matrix of the hypergraph, and the transition probability from the vertex u to the vertex v is: the target result divided by twice the degree of the vertex u, the target result is the sum of the fifth results of all hyperedges e between the vertex u and the vertex v, the fifth result is the product of the prior weight of the hyperedge e and the sixth result, and the sixth result is the sum of the seventh result and the eighth result; The seventh result is the product of the EDVW of the vertex u in the process of entering the hyperedge e from the vertex u, the EDVW of the vertex u in the process of flowing out of the hyperedge e from the vertex u, and the objective function value with the out-degree of the hyperedge e as the variable; the eighth result is the product of the EDVW of the vertex v in the process of entering the hyperedge e from the vertex v, the EDVW of the vertex v in the process of flowing out of the hyperedge e from the vertex v, and the objective function value with the in-degree of the hyperedge e as the variable; The processing module is further used to: process a target task according to the embedded features of the target vertex, and the target task includes any one of vertex classification, vertex clustering and regression tasks.
11. A task processing device, It is characterized in that include: A processor and a memory, the memory being used to store a computer program, and the processor being used to call and run the computer program stored in the memory to execute the method according to any one of claims 1 to 9.
12. A computer-readable storage medium, It is characterized in that Used to store a computer program, wherein the computer program enables a computer to execute the method according to any one of claims 1 to 9.
13. A computer program product comprising a computer program, It is characterized in that When the computer program is executed by a processor, the steps of the method according to any one of claims 1 to 9 are implemented.
