CN113408564A - Graph processing method, network training method, device, equipment and storage medium - Google Patents

Abstract

The embodiments of the present application disclose a graph processing method, a network training method, an apparatus, a device, and a storage medium, which are applicable to fields such as artificial intelligence. The method includes: acquiring a graph to be processed; acquiring node features of each node in the graph to be processed; for each node, according to the similarity between the node feature of the node and the graph category feature corresponding to each candidate graph category, from each candidate graph category The category of the node is determined in the graph category; based on the category of each node, corresponding processing is performed on at least one node in the graph to be processed. Using the embodiment of the present application, the category of each node in the graph to be processed can be determined based on each candidate graph category, the prediction accuracy of the node category can be improved, and the nodes in the to-be-processed graph can be effectively processed, with high applicability.

Description

Graph processing method, network training method, apparatus, device and storage medium

technical field

The present application relates to the field of artificial intelligence, and in particular, to a graph processing method, a network training method, an apparatus, a device, and a storage medium.

Background technique

In the current field of artificial intelligence for the classification task of nodes in the graph, the node classification model is obtained by labeling the categories of each node in the sample graph, and the network model is trained according to the node characteristics of the node, and then the node classification model is used to predict each node classification model. The class of the node. Among them, the node features and marked categories of each node in the sample graph are used as the input information of the network model, and the category of the node to be tested is output through the trained node classification model.

However, in the actual scene, since the category of each node in the graph will be affected by the graph category to a certain extent, the existing node classification model training methods and the existing node category determination methods often ignore the graph category as the node category. This will seriously affect the accuracy of the node classification model and the node classification method.

To sum up, how to further improve the accuracy of node classification has become an urgent problem to be solved.

SUMMARY OF THE INVENTION

The embodiments of the present application provide a graph processing method, a network training method, an apparatus, a device, and a storage medium. By adopting the embodiments of the present application, the accuracy of determining the category of the nodes in the graph to be processed can be improved, and the nodes in the graph to be processed can be effectively processed. Handling, high applicability.

In a first aspect, an embodiment of the present application provides a graph processing method, which includes:

Get the pending graph;

Obtain the node characteristics of each node in the above to-be-processed graph;

For each node, according to the similarity between the node feature of the node and the graph category feature corresponding to each candidate graph category, the category of the node is determined from each of the candidate graph categories, wherein each of the candidate graph categories includes the above-mentioned candidate graph categories. Graph class for processing graphs;

Based on the category of each of the above-mentioned nodes, corresponding processing is performed on at least one node in the above-mentioned to-be-processed graph.

In a second aspect, an embodiment of the present application provides a method for training a node feature extraction network, the method comprising:

obtaining an initial graph classification network, where the initial graph classification network includes a node feature extraction module, a graph feature extraction module and a graph classification module that are cascaded in sequence;

Obtain training data, each sample image in the training data is marked with a sample label, and the sample label represents the real image category of the sample image;

Inputting each of the above-mentioned sample graphs into the above-mentioned node feature extraction module to obtain node features of each node of each of the above-mentioned sample graphs;

Input the node features of each of the above-mentioned nodes into the above-mentioned graph feature extraction module to obtain the graph features of each of the above-mentioned sample graphs;

Input the graph features of each of the above-mentioned sample graphs into the above-mentioned graph classification module to obtain the predicted graph category of the above-mentioned sample graphs;

Determine the total training loss value according to the predicted graph category of each of the above-mentioned sample graphs and the sample label of each of the above-mentioned sample graphs;

The initial graph classification network is iteratively trained according to the above-mentioned total training loss value and the above-mentioned training data, until the above-mentioned total training loss value meets the training end condition, the node feature extraction module in the initial graph classification network at the end of training is determined as a node Feature extraction network.

In a third aspect, an embodiment of the present application provides a graph processing apparatus, the apparatus comprising:

The graph acquisition module is used to obtain the graph to be processed;

A feature acquisition module, configured to acquire the node features of each node in the above-mentioned to-be-processed graph;

A classification module, for each node, according to the similarity between the node feature of the node and the graph class feature corresponding to each candidate graph class, determine the class of the node from each of the above-mentioned candidate graph classes, wherein each of the above-mentioned candidate graph classes The graph category includes the graph category of the above-mentioned to-be-processed graph;

The graph processing module is configured to perform corresponding processing on at least one node in the graph to be processed based on the category of each of the above-mentioned nodes.

In a fourth aspect, an embodiment of the present application provides a node feature extraction network training device, which is used for:

a network acquisition module, configured to acquire an initial graph classification network, where the initial graph classification network includes a node feature extraction module, a graph feature extraction module and a graph classification module that are cascaded in sequence;

a data acquisition module, used for acquiring training data, each sample image in the training data is marked with a sample label, and the sample label represents the real image category of the sample image;

an input module, configured to input each of the above-mentioned sample graphs into the above-mentioned node feature extraction module to obtain node features of each node of each of the above-mentioned sample graphs;

The above-mentioned input module is used for inputting the node features of each of the above-mentioned nodes to the above-mentioned graph feature extraction module to obtain the graph features of each of the above-mentioned sample graphs;

The above-mentioned input module is used to input the graph features of each of the above-mentioned sample graphs into the above-mentioned graph classification module to obtain the predicted graph categories of the above-mentioned sample graphs;

The loss determination module is used to determine the total training loss value according to the prediction map category of each of the above-mentioned sample maps and the sample labels of each of the above-mentioned sample maps;

The network determination module is used to iteratively train the above-mentioned initial graph classification network according to the above-mentioned total training loss value and the above-mentioned training data, until the above-mentioned total training loss value meets the training end condition, classify the nodes in the network with the initial graph at the end of training The feature extraction module is determined as a node feature extraction network.

In a fifth aspect, an embodiment of the present application provides an electronic device, including a processor and a memory, where the processor and the memory are connected to each other;

The above-mentioned memory is used to store computer programs;

The above-mentioned processor is configured to execute the method provided by any one of the embodiments of the above-mentioned first aspect and/or the second aspect when the above-mentioned computer program is invoked.

In a sixth aspect, an embodiment of the present application provides a computer-readable storage medium, where a computer program is stored in the computer-readable storage medium, and the computer program is executed by a processor to implement any one of the first aspect and/or the second aspect. The method provided by the embodiment.

In a seventh aspect, embodiments of the present application provide a computer program product or computer program, where the computer program product or computer program includes computer instructions, and the computer instructions are stored in a computer-readable storage medium. The processor of the electronic device reads the computer instructions from the computer-readable storage medium, and the processor executes the computer instructions, so that the computer device executes the method provided by any embodiment of the first aspect and/or the second aspect.

In this embodiment of the present application, by determining the similarity between the node feature of each node in the graph to be processed and the graph category feature corresponding to the candidate graph category, the category of each node is determined based on the similarity, not only in the determination of the node category In the process, the node characteristics of each node itself are fully considered, and the influence of the graph category on the category of each node can be combined, and then the category of each node can be determined through the candidate graph category, which can further improve the determination of the category of each node in the graph to be processed. accuracy and high applicability.

Description of drawings

In order to illustrate the technical solutions in the embodiments of the present application more clearly, the following briefly introduces the drawings required in the embodiments. Obviously, the drawings in the following description are only some embodiments of the present application. For those of ordinary skill in the art, other drawings can also be obtained from these drawings without any creative effort.

1 is a schematic diagram of the principle of a graph processing method provided by an embodiment of the present application;

2 is a schematic flowchart of a graph processing method provided by an embodiment of the present application;

3 is a schematic diagram of the principle of a method for determining a node feature provided by an embodiment of the present application;

4 is a schematic flowchart of a method for determining a node feature provided by an embodiment of the present application;

5 is a schematic diagram of a scene for processing a to-be-processed graph provided by an embodiment of the present application;

6 is a schematic diagram of another scenario for processing a to-be-processed graph provided by an embodiment of the present application;

7 is a schematic flowchart of a training method for a node feature extraction network provided by an embodiment of the present application;

8 is a schematic structural diagram of a graph classification network provided by an embodiment of the present application;

9 is a schematic diagram of a method for determining a characteristic distance provided by an embodiment of the present application;

10 is a schematic diagram of a training method for a node feature extraction network provided by an embodiment of the present application;

11 is a schematic structural diagram of a graph processing apparatus provided by an embodiment of the present application;

12 is a schematic structural diagram of a node feature extraction network training device provided by an embodiment of the present application;

FIG. 13 is a schematic structural diagram of an electronic device provided by an embodiment of the present application.

manual

The technical solutions in the embodiments of the present application will be clearly and completely described below with reference to the drawings in the embodiments of the present application. Obviously, the described embodiments are only a part of the embodiments of the present application, but not all of the embodiments. Based on the embodiments in the present application, all other embodiments obtained by those of ordinary skill in the art without creative efforts shall fall within the protection scope of the present application.

The graph processing method provided by the embodiments of the present application relates to the field of machine learning (ML) in artificial intelligence (Artificial Intelligence, AI) and the field of computer vision (Computer Vision, CV) and other fields. Among them, machine learning (ML) is the study of how computers simulate or realize human learning behaviors to acquire new knowledge or skills, and to reorganize existing knowledge structures to continuously improve their performance. Computer vision technology can use cameras and computers instead of human eyes to identify, track and measure targets, and further process them.

Artificial intelligence is a 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. For example, in the embodiments of the present application, the nodes in the graph can be identified through computer vision technology, and based on the neural network, the machine has the performance of classifying the nodes through machine learning, so that the graph can be processed according to the category of the nodes.

The graph processing method provided by the embodiment of the present application also relates to the fields of cloud computing (cloud computing), artificial intelligence cloud service, etc. in cloud technology (Cloud technology). In the embodiment of the present application, the computing tasks involved in the graph processing method are distributed on a resource pool formed by a large number of computers through cloud computing to improve the efficiency of graph processing. In addition, the graph processing method can also be used as an artificial intelligence service, and the corresponding artificial intelligence cloud service of graph processing can be provided through the artificial intelligence platform.

Referring to FIG. 1 , FIG. 1 is a schematic schematic diagram of a graph processing method provided by an embodiment of the present application. As shown in FIG. 1 , node 101 , node 102 , node 103 and node 104 are the nodes in FIG. 10 to be processed, wherein the connection between the nodes can represent the connection relationship between the nodes, and the specific relationship type can be based on the actual application The scene is determined, and there is no restriction here. When the to-be-processed Figure 10 is a user relationship diagram, each node in the to-be-processed Figure 10 can represent each user, and the lines between the nodes represent the user relationship (such as a friend relationship) between the users, such as the nodes in the to-be-processed Figure 10 101 has a friend relationship with node 102, node 103, and node 104. When the to-be-processed figure 10 is a knowledge graph, each node in the to-be-processed figure 10 can represent each knowledge point, and the continuity between the nodes represents the association relationship between the various knowledge points, as represented by the node 103 in the to-be-processed figure 10 There is an association relationship between the knowledge points and the knowledge points represented by the nodes 101 and 102 .

Further, the node features of each node in Fig. 10 to be processed can be obtained through the graph neural network, that is, the feature of each node in the to-be-processed Fig. 10 is represented by a vector. For each node (taking the node 104 as an example), the node type of the node 104 is determined from each candidate graph category according to the similarity between the node feature of the node 104 and the graph category feature corresponding to each candidate graph category. Wherein, the above candidate graph categories are preset graph categories, so the category of each node in Fig. 10 to be processed can be determined from the category level of the graph, and then based on the list of each node in Fig. 10 to be processed, any one of them can be determined. Or multiple nodes will be processed accordingly.

Referring to FIG. 2, FIG. 2 is a schematic flowchart of a graph processing method provided by an embodiment of the present application. As shown in FIG. 2 , the graph processing method provided by the embodiment of the present application may include the following steps:

Step S21, acquiring the graph to be processed.

In some feasible implementations, the graph to be processed may represent objects and their relationships. The objects in the graph to be processed are called nodes, and the relationships between the objects are described by edges. When acquiring the graph to be processed, the objects within a certain range and the relationship between the objects can be determined, and then each object is used as a node, and the relationship between the objects is mapped to the edge between the objects, so as to be based on the relationship between the nodes and the objects. The edges between the nodes constitute the graph to be processed.

The to-be-processed graph can be used to represent social networks, information dissemination networks, knowledge graphs, chemical molecular structures, etc., which can be determined based on actual application scenarios, and is not limited here. For example, within the scope of a social application, for a user, the user, other users who have a social relationship with the user (the user's application contacts), and users who have a social relationship with the user's application contacts can be As each node in the graph to be processed, the user maps the social relationship between the user represented by any node and other users to the edges corresponding to the node, and then based on the nodes representing each user and the edges between the nodes Constitutes the graph to be processed.

For another example, for the organic 2,4,6-trinitrotoluene, the chemical molecular structure of 2,4,6-trinitrotoluene can also be represented by a diagram, that is, 2,4,6-trinitrotoluene The functional group (benzene ring) and the substituent (nitro group) are used as nodes in the Each node and each edge constitute the graph to be processed.

It should be particularly noted that the object selection when acquiring the to-be-processed graph and the selection range of the object may be determined based on the requirements of the actual application scenario, which is not limited here. In order to reflect the friend relationship of a user in a social application, the user and the user's application contacts can be regarded as nodes, the friend relationship between the user and its application contacts, and the mutual relationship among the user's application contacts. The friend relationship can be used as the edge between the corresponding nodes. For example, to reflect the propagation path of a message, each user through which the message is propagated within a certain time range can be used as each node, and the edge between each node is The propagation path of the message.

Step S22, acquiring node features of each node in the graph to be processed.

In some feasible implementations, the implementation of determining the node characteristics of each node in the to-be-processed graph is further described with reference to FIG. 3 , which is a schematic diagram of the principle of the method for determining the node characteristics provided by the embodiment of the present application. As shown in FIG. 3 , the node feature extraction network includes an initial feature extraction sub-network, and multiple node feature extraction sub-networks cascaded with the initial feature extraction sub-network. The initial feature of each node in the graph to be identified can be obtained through the initial feature extraction sub-network, and the initial feature of each node is further processed through each node feature extraction sub-network to obtain the node feature of each node. Wherein, each node feature extraction sub-network corresponds to a candidate graph category, and the specific method for determining the node feature of each node can be found in FIG. The method for determining node characteristics shown in FIG. 4 may include the following steps:

Step S221: Extract the initial features of each node in the graph to be processed through the initial feature extraction sub-network.

In some feasible implementations, the initial features of each node in the graph to be processed can be obtained through the initial feature extraction sub-network. Specifically, when the initial feature extraction sub-network is used to extract the initial features of each node in the graph to be processed, the convolution operation can be directly defined on each node in the graph to be processed by using the spatial domain convolution method through the initial feature extraction sub-network. In order to extract the spatial features on the graph to be processed, and then obtain the preliminarily processed features of each node in the graph to be processed (for the convenience of description, hereinafter referred to as the first feature), and the first feature of each node is A feature is determined to be the initial feature of the node.

Optionally, through the initial feature extraction sub-network, based on the method of graphic signal processing, a signal processing transformation, such as Fourier transform or Laplace transform, is performed on the to-be-processed graph, and then the convolution of the to-be-processed graph is performed to extract The first feature of each node in the graph, and the first feature of each node is determined as the initial feature of the node. It should be noted that the above-mentioned specific implementation of extracting the node features of each node in the graph to be processed through the initial feature extraction sub-network can be determined based on the actual application scenario requirements and the network mechanism of the initial feature extraction sub-network, which is not limited here. .

In some feasible implementations, since the importance of the connection relationship between the nodes in the graph to be processed is different, the relative importance between a node and different adjacent nodes is also different. Therefore, in order to consider the above problems, the The graph attention network processes the first feature of each node in the graph to be processed, and uses the processed feature as the initial feature of each node.

待处理图中各节点的初始特征为

n为待处理图中节点的个数，此时对于待处理图中的节点i来说，节点i的自相关系数为

其中，α^k是第k个注意力机制的权重系数，权重系数α^k由一个单层的前向神经网络计算得到，即：In a specific implementation, after obtaining the first feature of each node in the graph to be processed through the initial feature extraction sub-network, the attention autocorrelation coefficient of the node can be determined according to the first feature of the node. Assume that the first feature of each node in the graph to be processed is

The initial features of each node in the graph to be processed are

n is the number of nodes in the graph to be processed. For node i in the graph to be processed, the autocorrelation coefficient of node i is

Among them, α ^k is the weight coefficient of the k-th attention mechanism, and the weight coefficient α ^k is calculated by a single-layer forward neural network, namely:

为该层前向神经网络对应的权重向量，LeakyReLU为激活函数，W₁为待处理图中各节点所共享的权值矩阵，用于表示待处理图中各节点的第一特征

和初始特征

之间的转换关系，N(i)为节点i的相邻节点的集合。同理，可根据待处理图中的节点i的第一特征

及节点i的各相邻节点的第一特征

得到节点i与各相邻节点的互相关系数为

其中：in,

is the weight vector corresponding to the forward neural network of this layer, LeakyReLU is the activation function, and W ₁ is the weight matrix shared by each node in the graph to be processed, which is used to represent the first feature of each node in the graph to be processed

and initial features

The transformation relationship between N(i) is the set of adjacent nodes of node i. Similarly, according to the first feature of node i in the graph to be processed

and the first feature of each adjacent node of node i

The cross-correlation coefficient between node i and each adjacent node is obtained as

in:

即节点i在其自身以及相邻节点影响下的初始特征。其中，K表示注意力机制的数量，||表示拼接过程，σ()表示非线性激活函数。Further, for node i, the initial feature of node i can be obtained through the initial feature extraction network

That is, the initial features of node i under the influence of itself and its neighbors. where K represents the number of attention mechanisms, || represents the splicing process, and σ() represents the nonlinear activation function.

Optionally, for any node in the graph to be processed, in order to further aggregate the node features of its adjacent nodes, after obtaining the initial features of the node, based on the above-mentioned attention mechanism and the initial features of the node. The initial feature of the node is updated one or more times to obtain the final initial feature of the node. And in each update, the initial feature before the update can be used as the first feature in the above calculation process, and then the updated initial feature is obtained based on the above process. The specific update times can be determined based on the actual application scenario, which is not limited here. For example, taking the above-mentioned one-time calculation process of the initial feature as the calculation process of a graph convolution layer, the initial feature update times of each node in the graph to be processed can be determined based on the number of layers of the graph convolution layer in the initial feature extraction sub-network, wherein, The number and weight of the attention mechanism between each convolutional layer, the corresponding weight vector of the forward neural network, and the weight matrix representing the transformation relationship between the initial features before updating and the initial features after updating of each node, It can be the same or different, and it can also be determined based on the number of graph convolution layers and the requirements of the actual application scenario, which is not limited here.

Step S222: For each node, based on the initial feature of the node, extract a sub-network through the node feature corresponding to each candidate graph category, and obtain the feature of the node corresponding to each candidate graph category.

In some feasible implementations, since only the initial features under the influence of the connection relationship between each node itself and its adjacent nodes can be obtained based on the attention mechanism, if the node category of each node is directly determined based on the initial features of each node , it will lead to the deviation of the node category of the node and the inaccurate determination of the node category. Therefore, for each node in the to-be-processed graph, the node feature extraction sub-network corresponding to each candidate graph category in the network can be extracted according to the initial feature of the node, and the feature of the node corresponding to each candidate graph category can be obtained. In other words, when the initial feature of the node itself includes the connection relationship between the nodes, the initial feature of each node is further represented by each candidate graph category, and the feature of each node corresponding to each candidate graph category is obtained, To further improve the accuracy of node category determination.

Specifically, in the node feature extraction network, the node feature extraction sub-network corresponding to each candidate graph category is connected to the initial feature extraction sub-network, and each candidate graph category is a preset of a different graph in the process of training the node feature extraction network. category. Based on this, for each node, through each node feature extraction sub-network, and according to the graph category feature corresponding to each candidate graph category, the feature of the node corresponding to each candidate graph category is obtained. That is, through the feature extraction sub-network of each node, the initial features of the node are further represented under the influence of each candidate graph category.

The graph category feature corresponding to a candidate graph category is the network parameter of the node feature extraction sub-network corresponding to the candidate graph category. The specific network parameters can be determined based on the actual network of the feature extraction sub-network, which is not limited here.

For example, each node feature extraction sub-network is a network based on a Gaussian mixture model, that is, a node feature extraction sub-network corresponds to a Gaussian function in the Gaussian mixture model, and a graph category feature corresponding to a candidate graph category is the candidate graph category corresponding to Gaussian distribution parameters for Gaussian functions in Gaussian mixture Gaussian models. Wherein, the above-mentioned Gaussian distribution parameters include but are not limited to the mean vector and covariance matrix corresponding to the Gaussian function, etc., which can be specifically determined based on the requirements of the actual application scenario, and are not limited here. Specifically, for node i in the graph to be processed, through the feature extraction sub-network of each node, the feature of node i corresponding to each candidate graph category can be obtained

为节点i的初始节点特征，W₂为混合高斯模型对应的权重，ω_c()为第c个候选图类别对应的高斯函数，

表示第c个候选图类别对应的高斯函数的协方差矩阵，

表示第c个候选图类别对应的高斯函数的均值向量，即表征该候选图类别的中心。in,

is the initial node feature of node i, W ₂ is the weight corresponding to the mixture Gaussian model, ω _c () is the Gaussian function corresponding to the c-th candidate graph category,

represents the covariance matrix of the Gaussian function corresponding to the c-th candidate graph category,

Represents the mean vector of the Gaussian function corresponding to the c-th candidate image category, that is, characterizes the center of the candidate image category.

Step S223, fuse the features of the node corresponding to each candidate graph category to obtain the node feature of the node.

其中C表示各候选图类别的数量，还可基于其他预设融合网络、模型等将该节点对应于各候选图类别的特征进行融合得到该节点的节点特征。In some feasible implementations, for each node, the features of the node corresponding to each candidate graph category may be fused to obtain the node features of the node under the influence of different candidate graph categories. For example, it is obtained based on obtaining the class weight of each candidate graph category, based on the weight of each candidate graph category, and the feature of the node corresponding to each candidate graph category by weighted summation. Or, the mean feature of the feature corresponding to each candidate graph category of the node is taken as the final node feature of the node, that is, the node feature of the node

Among them, C represents the number of each candidate graph category, and the node feature of the node can be obtained by fusing the features of the node corresponding to each candidate graph category based on other preset fusion networks, models, etc.

It should be particularly noted that the specific implementation of the above-mentioned fusion of the features of each node in the graph to be processed corresponding to each candidate graph category is only an example, which can be determined based on actual application scenario requirements, and is not limited.

Step S23: For each node, according to the similarity between the node feature of the node and the graph class feature corresponding to each candidate graph class, determine the class of the node from each candidate graph class.

In some feasible implementations, after the node features of each node in the graph to be processed are obtained, the class of each node can be determined from each candidate graph class. For each node in the graph to be processed, the similarity between the node feature of the node and the graph category feature corresponding to each candidate graph category can be determined, and then the candidate graph category corresponding to the highest similarity can be determined as the node’s feature. category. That is to say, the higher the similarity between the node feature of the node and the graph category feature corresponding to the candidate graph category, the closer the category of the node is to the candidate graph category. Among them, each candidate graph category includes the graph category of the graph to be processed, so when determining the node category of each node, the influence of the graph category of the graph to be processed on the category of each node is fully considered, and the determination of the category of the node is improved. Accuracy.

可由公式

确定，即将最大余弦相似度对应的图类别特征对应的候选图类别，确定为节点的类别。Among them, the similarity between the node features of the above-mentioned nodes and the graph category features of each candidate graph category, including but not limited to Euclidean distance, cosine similarity, Manhattan distance, Chebyshev distance, and Jaccard coefficient, etc. It is determined based on the actual application scenario requirements, and there is no restriction here. For example, when the above similarity is cosine similarity, the category of each node

by the formula

Determine, that is, determine the candidate graph category corresponding to the graph category feature corresponding to the maximum cosine similarity as the category of the node.

In some feasible implementations, the graph category in the embodiments of the present application is used to represent the characteristics of the graph in a certain field and a certain dimension. Among them, the same graph may have different graph categories in different fields and different dimensions. Different graphs may have the same graph category or different graph categories, which can be determined based on the graph composition and actual application scenarios, which are not limited here. .

For example, if the graph to be processed represents a payment network, and each node in the graph to be processed is a payment user in the payment network. In the field of transfer risk control, the graph category of the graph to be processed may be a money laundering network, that is, there is money laundering in the payment network represented by the graph to be processed, or the graph category of the graph to be processed may be a normal payment network, that is, the user's payment behavior is normal Behavior.

Optionally, for the payment network, the graph category of the graph to be processed may also be a mobile payment network, which means that the payment user in the payment network adopts mobile payment, or the graph category of the graph to be processed may also be bank card transfer, which means that the paying user Pay by bank card transfer.

For another example, if the graph to be processed represents a social network, and each node in the graph to be processed is a social user in the social network, in the field of information dissemination, the graph category of the graph to be processed may be a rumor relay network, that is, the graph represented by the graph to be processed. There is a phenomenon of rumor spreading in the social network, or the graph category of the graph to be processed may be a normal social network (a social network without rumors).

Optionally, for a social network, the graph category of the graph to be processed may also be classified based on the age of the user, such as a social network for middle-aged and elderly people, a social network for teenagers, and the like.

For another example, if the graph to be processed is a task topology graph or a program control graph, and each node in the graph to be processed is each task node in the task topology graph (program control graph). For the running perspective, the graph category of the graph to be processed may be that the task (program) failed to execute, that is, there is a task (program) node that failed to execute in the task topology graph, or the graph category of the graph to be processed may be that the task (program) was successfully executed. That is, the complete task (program) corresponding to the task topology map can be executed successfully.

Optionally, for task topology graphs, from the perspective of task execution, the graph categories of the graphs to be processed may be front-end tasks and back-end tasks, and from the perspective of task types, the graph categories of the graphs to be processed may also be recognition tasks, classification tasks, and translation tasks. tasks, etc.

The candidate graph category is a graph category including the graph to be processed, and a category set of other graph categories possible for the graph to be processed. The specific number and division of graph categories can be determined based on actual application scenario requirements, which are not limited here.

Step S24, based on the category of each node, perform corresponding processing on at least one node in the to-be-processed graph.

In some feasible implementations, after the category of each node in the graph to be processed is determined, one or more nodes in the graph to be processed may be processed accordingly. Specifically, the nodes in the graph to be processed may be clustered and segmented according to the types of the nodes, so as to realize the segmentation of the graph to be processed. If the graph to be processed is a payment network, the payment network represented by the graph to be processed can be divided into a normal payment network and a money laundering network based on the category of each node. Among them, for the payment users corresponding to the money laundering network, measures such as risk warning and transaction restriction can be taken in a timely manner. For payment users corresponding to normal payment networks, risk reminders can be given to them to improve their awareness of risk prevention.

Optionally, each node may be screened based on the category of each node in the to-be-processed graph. Specifically, the screening may be performed based on the number of nodes in each category, a preset screening category, etc., which is not limited here. For example, the to-be-processed graph is a task topology graph, and different task testing methods are adopted based on different node categories (task types).

The above-mentioned method for processing the nodes in the to-be-processed graph based on the category of each node in the to-be-processed graph is merely an example, and is not limited in the embodiments of the present application.

In some feasible implementations, after the category of each node in the graph to be processed is determined, based on the graph category of the graph to be processed, corresponding processing can also be performed on nodes whose category is different from that of the graph to be processed in each node , the specific processing method can be determined based on the actual application scenario requirements and the category of the node and the graph category of the graph to be processed, which is not limited here.

Among them, the category of the node is the same as the category of the graph to be processed. It can be understood that the two correspond to the same feature, the same behavior, etc. The category of the graph is the rumor propagation network, and the category of the node is the rumor relay. Nodes correspond to rumor propagation, that is It can be stated that the class of the node is the same as the class of the graph. In the same way, nodes whose types of nodes in the graph to be processed are different from those of the graph to be processed can be determined based on the same method.

For example, if the graph category of the graph to be processed indicates that the corresponding program control graph is running normally, if there is a node whose category is failed to run in the category of each node in the graph to be processed determined based on the above implementation manner, the node can be processed Measures such as program testing, program changes, or program replacements.

For example, if the graph category of the graph to be processed indicates that its corresponding social network is a normal social network (a social network without rumors), if the category of each node in the to-be-processed graph determined based on the above implementation manner, there is a rumor propagating node. That is, the rumor propagation node is a node with a different category of the node in the graph to be processed and the graph category of the graph to be processed. At this time, the node can be refuted, or other non-rumor nodes can be reminded accordingly.

As shown in FIG. 5 , FIG. 5 is a schematic diagram of a scenario of processing a graph to be processed according to an embodiment of the present application. In FIG. 5 , a graph to be processed can be obtained based on each user in the information dissemination network and the information dissemination relationship among the users. Among them, each node in the graph to be processed corresponds to each user in the dissemination network, and the edge between each node represents the corresponding information dissemination relationship. When each candidate graph category includes rumors (rumor dissemination network) and real information (rumor-free social network), the category of each node in the graph to be processed can be determined based on the graph processing method provided by the embodiment of the present application, that is, Fig. 5 The categories of node A, node B, node C, node E and node F are rumors (rumor propagation nodes), that is, the information spread by the users corresponding to node A, node B, node C, node E and node F are rumors . The categories of node D, node G, and node H are real information (normal user nodes), that is, the information spread by users corresponding to node D, node G, and node H is real information.

If in the process of determining the graph to be processed, the graph category of the graph to be processed is marked as a rumor-free social network, and the categories of node A, node B, node C, node E, and node F are different from the graph category of the graph to be processed, then User A, user B, user C, user E, and user F can be processed to refute rumors. Further, based on the information dissemination relationship between the nodes, it can be determined that the node A in the graph to be processed is the information dissemination source, and then measures such as restricting information dissemination and information dissemination alarms can be taken for user A.

Optionally, based on the graph category of the graph to be processed and the node category of each node in the graph to be processed, the target node in the graph to be processed is determined, and the target node is processed. The target node can be a node that changes the structural relationship, operation logic, composition, etc. of the to-be-processed graph represented by the graph category, or can be a node that is not related to the graph category of the to-be-processed graph, which may be based on practical applications. Scenario requirements are determined, and there are no restrictions here.

For example, the graph to be processed represents the chemical molecular structure, each node represents the functional group of the chemical molecular structure, the graph category of the to-be-processed graph represents the chemical name of the chemical molecular structure, and the category of each node represents the name of each functional group. Then, based on the graph category of the graph to be processed and the category of each node, the wrong functional group in the molecular structure under the chemical name is determined, and then the wrong functional group is changed based on the chemical molecular structure formed by each functional group. The chemical names are the same.

As shown in FIG. 6 , FIG. 6 is a schematic diagram of another scenario of processing a to-be-processed graph provided by an embodiment of the present application. Figure 6 shows a chemical molecular structure whose name is labeled as 2,4-dinitroethylbenzene. By using the functional groups in the molecular structure based on the connection relationship of each functional group, the chemical molecular structure can be represented as a to-be-processed map. express. When each candidate graph category includes each functional group and each substituent, the graph processing method provided by the embodiment of the present application can determine the node category of each node in the graph to be processed, that is, the category of node A in FIG. 6 is a benzene ring, and the category of node B The class is methyl, the class of node C is nitro and the class of node E is nitro. Since the chemical molecular structure is initially marked as 2,4-dinitroethylbenzene, that is, the graph class of the graph to be processed is 2,4-dinitroethylbenzene, in the process of determining the class of each node, the class of node B is It is determined to be methyl, that is, the functional group corresponding to node B is methyl, so it can be determined that node B is the target node. Further, the functional group corresponding to node B can be changed from methyl to ethyl, so that the final molecular structure is the molecular structure of 2,4-dinitroethylbenzene.

In this embodiment of the present application, by determining the similarity between the node feature of each node in the graph to be processed and the graph category feature corresponding to the candidate graph category, the category of each node is determined based on the similarity, not only in the determination of the node category In the process, the node characteristics of each node itself are fully considered, and the influence of the graph category on the category of each node can be combined, and then the category of each node can be determined through the candidate graph category, which can further improve the determination of the category of each node in the graph to be processed. accuracy. At the same time, since the application of the graph category for the node category is effectively combined in the process of determining the category of the nodes in the graph to be processed, the applicable scenarios and applicability of the graph processing method provided by the embodiments of the present application can be further improved. high.

In some feasible implementation manners, the node feature extraction network used for obtaining the node features of each node in the graph to be processed in the embodiments of the present application is obtained by pre-training according to the training data, and the specific training process can be seen in FIG. 7 . FIG. 7 is a schematic flowchart of a training method for a node feature extraction network provided by an embodiment of the present application. The training method for a node feature extraction network shown in FIG. 7 may include the following steps:

Step S71 , acquiring an initial graph classification network and training data.

In some feasible implementations, when training the node feature extraction network, the initial graph classification network and training data for training can be obtained first, wherein each sample graph in the training data is marked with a sample label, and each sample graph is marked with a sample label. The corresponding sample label characterizes the real image category of the sample image. The initial graph classification network includes a node feature extraction module, a graph feature extraction module and a graph classification module that are cascaded in sequence. The specific functions of the above modules in the training process can be found in step S72, which is not described here.

The training data in the embodiments of the present application may be obtained from existing data in databases and cloud storage, or may be obtained based on big data technology for training the node feature extraction network, which may be determined based on actual application scenario requirements. , which is not limited here.

Step S72, input each sample graph into the node feature extraction module, obtain the node feature of each node of each sample graph, input the node feature of each node into the graph feature extraction module, obtain the graph feature of each sample graph, The graph features are input to the graph classification module to obtain the predicted graph category of the sample graph.

In some feasible implementations, in the graph processing method provided by the embodiments of the present application, the node feature extraction network can extract the node features of each node in the graph to be processed, and then based on the node features of each node and the corresponding The graph class feature determines the class of a node. Therefore, in the training process, the node features of each node in each sample graph obtained by the node feature extraction module can be processed based on the graph feature extraction module and the graph classification module to obtain the predicted graph category of each sample graph, so as to obtain the predicted graph category of each sample graph according to the The graph category measures whether the node features of each node extracted by the node feature extraction module are valid and accurate. Therefore, when the predicted graph category of each sample graph is consistent with the real graph category represented by the sample label, it can be shown that the node features of each node extracted by the node feature extraction module are accurate, so the graph classification network at the end of training can be classified into the network. The node feature extraction module is used as the node feature extraction network obtained by training.

For details, refer to FIG. 8 , which is a schematic structural diagram of a graph classification network provided by an embodiment of the present application. Figure 8 shows the first training process of the initial graph classification network based on the training data. As shown in Figure 8, after the training data is obtained, each sample graph in the training data can be input into the initial graph classification network The node feature extraction module of , obtains the node features of each node in each sample graph. The above-mentioned node feature extraction module also includes an initial feature extraction sub-network and a node feature extraction sub-network corresponding to each candidate graph category connected to the initial feature extraction sub-network. Specifically, each node in each sample graph can be extracted through the initial feature extraction sub-network. , and further obtain the node features of each node through the node feature extraction sub-network. Wherein, for the specific implementation of extracting the node features of each node in each sample graph through the node feature extraction module in the initial graph classification network, please refer to step S22 in FIG. The implementation of the node feature will not be repeated here.

Further, the node features of each node in each sample graph obtained by the node feature extraction module are input into the graph feature extraction module to obtain graph features of each sample graph. Specifically, for each sample graph, the node features of each node can be compressed through the attention pooling network in the graph feature extraction module to reduce the amount of data processing and aggregate the node features of each node. The attention score corresponding to each node is obtained, and the attention score of each node is normalized by the softmax function to obtain the attention distribution that can clearly represent the importance of each node. For each sample graph, according to the node features of each node in the sample graph and the corresponding attention distribution of each node in the sample graph, the node features of each node in the sample graph can be converted into a fixed-length graph feature representation, That is, the graph features of the sample graph are obtained. Through the attention pooling network in the graph feature extraction module, the nodes whose category is the same as the real category of the sample graph can get more attention, and the nodes whose category is different from the real category of the sample graph can get less attention , so that the obtained graph features of the sample graph can effectively distinguish the categories of each node.

Wherein, when the attention score corresponding to each node is obtained by the perceptron, the number of layers of the perceptron and the corresponding activation function can be determined based on the actual application scenario, which is not limited here. For example, for a sample graph, the node features of each node in the sample graph can be represented by a matrix H. When a two-layer multilayer perceptron is used to obtain the attention score corresponding to each node, the sample graph obtained based on the above implementation manner The graph feature E=Attn(H)=softmax(Θ ₂ tanh(Θ ₁ H ^T ))H. Among them, tanh() is the activation function corresponding to the perceptron, Θ ₁ and Θ ₂ are the weights corresponding to each layer in the two-layer perceptron, respectively, Θ ₂ tanh(Θ ₁ H ^T ) is the corresponding weight of each node in the sample graph Attention score, the attention distribution corresponding to each node obtained by the softmax function is softmax(Θ ₂ tanh(Θ ₁ H ^T )).

Based on the above implementation manner, the graph features of each sample graph can be obtained through the graph feature extraction model, and the graph features of each sample graph with different number of nodes can be made to have the same length through the attention pooling network in the graph feature extraction model.

Finally, the graph features of each sample graph are input into the graph classification module to obtain the predicted graph category of each sample graph. Specifically, the graph classification module can independently predict the predicted graph category of each sample graph based on the graph features of each sample graph. The specific prediction method can be determined by a classification algorithm, a related neural network, etc., which is not limited here.

Optionally, when each sample graph has an association relationship, in order to improve the prediction efficiency of the prediction graph category of each sample graph, and further improve the prediction accuracy of the prediction graph category of each sample graph in combination with the association relationship between the sample graphs The graph classification module can be used to construct hierarchical graph features based on the relationship between the sample graphs. Among them, a feature in the feature of the hierarchical graph can represent the graph feature of a sample graph. In other words, if the feature of the hierarchical graph is analogized to the graph feature of a new graph, the graph feature of each sample graph can be the graph feature of the new graph. The node characteristics of each node in .

Among them, the relationship between the sample graphs can be realized as: each sample graph corresponds to different parts of the same social network, and each sample graph corresponds to different parts of the same chemical molecular structure, or different social networks based on the same user, etc. , which can be determined based on the actual application scenario requirements, and is not limited here.

Further, the hierarchical graph features can be updated through the message passing network, graph convolutional network, and graph attention network in the graph classification module, so as to strengthen the relationship between the graph features of each sample graph in the hierarchical graph feature so far, and get updated. Later hierarchical graph features. For example, the specific update process of updating hierarchical graph features through a graph convolutional network is as follows:

F=hierGCN(E')=Concat(E',σ(A ^hier E'Θ ₃ ));

Among them, E' is the hierarchical graph feature constructed based on the graph feature E of each sample graph, _Θ3 is the corresponding weight of the graph convolutional network, A ^hier is the relationship matrix constructed based on the relationship between the sample graphs, σ() is the activation function corresponding to the graph convolutional network, and Concat represents the concatenation operation.

The updated hierarchical graph feature F is processed by the fully connected layer and the softmax function, and the probability that each sample graph is a candidate graph class can be obtained, and then the candidate graph class with the highest probability is used as the predicted graph class of the corresponding sample graph. .

Optionally, when there is no association between the sample graphs, or the association between the sample graphs is unknown, the features of the hierarchical graph can also be updated through the perceptron in the graph classification module to reduce the amount of data processing. At the same time, the updated hierarchical graph feature is obtained, and then the predicted graph category of each sample graph is obtained according to the updated hierarchical graph feature.

Step S73: Determine the total training loss value according to the predicted graph category of each sample graph and the sample label of each sample graph.

In some feasible implementations, the total training loss value in the training process can be used to measure whether the graph class prediction ability of the graph classification network is stable and accurate. It is determined that the node feature extraction module in the graph classification network has a stable and accurate ability to obtain the node features of each node.

Specifically, the first training loss value in the training process of the graph classification network can be determined according to the predicted graph category of each sample graph and the sample label of each sample graph, and the first training loss value is determined as the total training loss value. Among them, the first training loss value represents the difference between the predicted image category of each sample image and its real image category, and the smaller the first training loss value is, the better the prediction accuracy and stability of the image classification network for the image category of the sample image. high. Wherein, the first training loss value may be calculated by using a cross-entropy function loss function or other classification loss functions, which may be specifically determined based on the requirements of an actual application scenario, which is not limited here.

In some feasible implementations, since the embodiment of the present application mainly obtains a node feature extraction network that can accurately and stably obtain the node features of each node by training a stable graph classification network, the prediction graph of each sample graph is measured as a whole. On the basis of the difference between the category and the real graph category, the similarity between the graph features obtained by each sample graph based on the node features extracted by the node feature extraction network and the graph category features of each candidate graph category can be further evaluated to further evaluate the graph. Stability and accuracy of classification networks and node extraction networks.

其中，n表示各样本图的图特征的索引，c表示和候选图特征的索引。为进一步凸显各样本图的图特征

与各候选图类别对应的图类别特征

之间的第一特征距离d_nc，可先确定各图特征对应的目标候选图类别，其中，对于一个样本图的图特征来说，其对应的目标候选图类别为与该样本图的真实图类别一致的候选类别。进一步可基于每一图特征与相对应的目标候选图类别的图类别特征之间的第二特征距离来增强相对应的第一特征距离，具体过程如下：For the graph features of each sample graph, the first feature distance between the graph feature of each sample graph and the graph class feature corresponding to each candidate graph class can be determined

Among them, n represents the index of the graph feature of each sample graph, and c represents the index of the candidate graph feature. In order to further highlight the graph features of each sample graph

Graph class features corresponding to each candidate graph class

The first feature distance d _nc between them can first determine the target candidate image category corresponding to each image feature. For the image feature of a sample image, the corresponding target candidate image category is the real image of the sample image. Class-consistent candidate classes. Further, the corresponding first feature distance can be enhanced based on the second feature distance between each image feature and the image category feature of the corresponding target candidate image category, and the specific process is as follows:

表示索引为n的样本图的图特征，与其相对应的目标候选图类别对应的图类别特征之间的第二特征距离，(d_nc)_enhance表示每一图特征与各候选图类别对应的图类别特征之间的特征距离。Among them, μ is the discount hyperparameter in the above distance enhancement process, which can be determined based on actual application requirements, and is not limited here. y ⁽ⁿ⁾ represents the true graph category of the graph feature of the sample graph with index n,

Indicates the second feature distance between the graph features of the sample graph with index n and the graph category features corresponding to the corresponding target candidate graph categories, (d _nc ) _enhance indicates the graph corresponding to each graph feature and each candidate graph category Feature distance between category features.

为一个样本图的图特征，图特征

与图类别特征

之间的第一特征距离为d₁，与图类别特征

之间的第一特征距离为d₂，与图类别特征

之间的第一特征距离为d₃。当图类别特征

对应的候选图类别与图特征

对应的样本图的真实类别相同时，图类别特征

对应的候选图类别为图特征

对应的目标候选图类别，图类别特征

与目标候选图类别对应的图类别特征之间的第二特征距离同样为d₁，此时将图类别特征

与目标候选图类别对应的图类别特征之间的第二特征距离表示为δd_y。因此，对于图特征

来说，其与图类别特征

至今的特征距离为d₁+δd_y，与图类别特征

之间的特征距离为d₂+δd_y，与图类别特征

之间的特征距离为d₃+δd_y。For example, referring to FIG. 9 , FIG. 9 is a schematic diagram of a method for determining a characteristic distance provided by an embodiment of the present application. In Figure 9,

is the graph feature of a sample graph, graph feature

with graph category features

The distance between the first feature is d ₁ , and the graph category feature

The distance between the first feature is d ₂ , and the graph category feature

The first characteristic distance between is d ₃ . When graph class features

Corresponding candidate graph categories and graph features

When the real categories of the corresponding sample images are the same, the graph category features

The corresponding candidate graph category is the graph feature

Corresponding target candidate graph category, graph category feature

The second feature distance between the graph category features corresponding to the target candidate graph category is also d ₁ .

The second feature distance between graph class features corresponding to the target candidate graph class is denoted as δd _y . Therefore, for graph features

For example, it is related to the graph category feature

The feature distance so far is d ₁ +δd _y , which is related to the graph category feature

The feature distance between is d ₂ +δd _y , and the graph category feature

The characteristic distance between them is d ₃ +δd _y .

Further, the distance matrix D _enhance is normalized by the softmax function to obtain the similarity matrix S=softmax(-D _enhance ). Among them, the distance matrix D _enhance is determined by (d _nc ) _enhance , each row element in the similarity matrix represents the similarity between the graph feature of a sample graph and the graph category feature corresponding to each candidate graph category, and each column element represents each sample graph. The similarity between the graph feature of and the graph class feature corresponding to a candidate graph class. That is to say, any element in the similarity matrix can represent the similarity between a graph feature of a sample graph and a graph class feature corresponding to a candidate graph class.

Further, the second training loss value can be calculated according to the above-mentioned similarity matrix and the real picture category of each sample picture, and then the total training loss value can be obtained according to the first training loss value and the second training loss value. Wherein, according to the real image category of each sample image, a relationship matrix for representing the real image category of each sample image and each candidate image category can be constructed. Among them, each element in the above relationship matrix represents the relationship between the real image category of a sample image and a candidate image category (if the two are the same, the value of the element is 1; if the two are different, the value of the element is 1. The value is 0), and each element can represent the real similarity between the graph features of a sample graph and the graph category features of each candidate graph category. Therefore, the above-mentioned second training loss value represents the similarity between each sample graph and the graph category feature corresponding to each candidate category during the training process, and the difference between the actual similarity between each graph feature and each graph category feature .

Wherein, the second training loss value may be calculated by using a cross-entropy function loss function or other classification loss function, which may be specifically determined based on actual application scenario requirements, and is not limited herein.

Optionally, after obtaining the first training loss value and the second training loss value, the first weight corresponding to the first training loss value and the second weight corresponding to the second training loss value may be obtained. value, the second training loss value, the first weight and the second weight, the total training loss value L=αL _cls +(1−α)L _con is determined. where L is the total training loss value, L _cls is the first training loss value, L _con is the second training loss value, α is the first weight, (1-α) is the second weight, and α∈[0,1 ].

Step S74: Perform iterative training on the initial graph classification network according to the total training loss value and the training data, until the total training loss value meets the training end condition, determine the node feature extraction module in the graph classification network at the end of the training as the node feature extraction module. network.

In some feasible implementations, according to the total training loss value and each sample graph in the training data, the initial graph classification network is trained based on the above implementation manner, and each candidate graph class is updated through the backpropagation network in each training process The corresponding graph category features and other network and model parameters make the node feature representation obtained by the node feature extraction module more reasonable. Referring to FIG. 10, FIG. 10 is a schematic diagram of a training method for a node feature extraction network provided by an embodiment of the present application. As shown in Figure 10, for the first training, the input of the initial graph classification network is each sample graph, and the node feature of each node in each sample graph can be extracted through the node feature extraction module, in which the graph attention network can be used as the node feature The initial feature extraction sub-network in the module is used to extract the initial features of each node in the sample graph. The node feature extraction sub-network can obtain the node features of each node based on the initial features of each node. Through the attention pooling network in the graph feature module, the graph features corresponding to each sample graph can be obtained based on the node features of each node in each sample graph, so as to further obtain the hierarchical graph features through the graph classification module, and pass the graph volume The hierarchical graph features of the product network are updated to obtain the predicted graph categories of each sample graph according to the updated hierarchical graph features.

When the total training loss value meets the training end condition, the node feature extraction model in the graph classification network at the end of the training can be used as the node feature extraction network, and then the node feature extraction network obtained by training can determine the category of each node in the graph to be processed. . Wherein, the above-mentioned training end condition may be that the above-mentioned total training loss value reaches a convergent state, or that the above-mentioned total training loss value is lower than a preset threshold, etc., which may be determined based on actual application scenario requirements, and is not limited here.

In the embodiment of the present application, since the node feature extraction network determines the category of each node according to the graph category of the candidate graph category, by training the graph classification network, the node feature extraction module in the graph classification network at the end of the training is used as the function The node feature extraction network used to determine the category of the node. On the one hand, the accuracy of the graph category prediction can be used to measure the accuracy of the node feature extraction module in the process of training the graph classification network. On the other hand, the total training loss in the training process can be used. The value of the graph category corresponding to the candidate graph category is continuously adjusted, so that the final node feature extraction network can determine the category of each node according to the accurate and appropriate graph category features, which has high applicability.

Referring to FIG. 11 , FIG. 11 is a schematic structural diagram of a graph processing apparatus provided by an embodiment of the present application. The graph processing apparatus provided by the embodiment of the present application includes:

The graph obtaining module 111 is used to obtain the graph to be processed;

A feature acquisition module 112, configured to acquire node features of each node in the above-mentioned to-be-processed graph;

The classification module 113 is configured to, for each node, determine the class of the node from the above candidate graph classes according to the similarity between the node feature of the node and the graph class feature corresponding to each candidate graph class, wherein each of the above The candidate graph category includes the graph category of the above-mentioned to-be-processed graph;

The graph processing module 114 is configured to perform corresponding processing on at least one node in the graph to be processed based on the category of each of the above-mentioned nodes.

In some feasible implementation manners, the above-mentioned feature acquisition module 112 is used for:

The node features of each node in the graph to be processed are obtained through a node feature extraction network, wherein the node feature extraction network includes an initial feature extraction sub-network, and node feature extraction corresponding to each candidate graph category connected to the initial feature extraction sub-network subnet;

For each node, through the above-mentioned initial feature extraction sub-network, extract the initial feature of the node in the above-mentioned to-be-processed graph;

Based on the initial feature of the node, a sub-network is extracted through the node feature corresponding to each candidate graph category, and the feature of the node corresponding to each of the above candidate graph categories is obtained;

The features of the node corresponding to each of the above candidate graph categories are fused to obtain the node feature of the node.

In some feasible implementation manners, the above-mentioned feature acquisition module 112 is used for:

Obtain the first feature of the node, and determine the attention autocorrelation coefficient of the node according to the first feature of the node;

According to the first feature of the node and the first feature of each adjacent node of the node, determine the attention cross-correlation coefficient between the node and each of the above-mentioned adjacent nodes;

The initial feature of the node is determined according to the above-mentioned attention autocorrelation coefficient and the above-mentioned attention cross-correlation coefficient.

In some feasible implementations, a graph class feature corresponding to a candidate graph class is used to extract network parameters of the sub-network for the node feature corresponding to the candidate graph class.

In some feasible implementations, the node feature extraction sub-network is a network based on a Gaussian mixture model, and the graph class feature corresponding to a candidate graph category is a Gaussian distribution parameter of the Gaussian mixture model corresponding to the candidate graph category.

In some feasible implementation manners, the above-mentioned classification module 113 is used for:

For each node, determine the highest similarity from the similarity between the node feature of the node and the graph category feature corresponding to each candidate graph category;

The category of the candidate graph corresponding to the above highest similarity is determined as the category of the node.

In some feasible implementation manners, the above-mentioned graph processing module 114 is further configured to:

Get the graph category of the above pending graph;

According to the category of each of the above-mentioned nodes, corresponding processing is performed on the nodes whose category of each of the above-mentioned nodes is different from that of the above-mentioned graph to be processed.

Referring to FIG. 12, FIG. 12 is a schematic structural diagram of a network training device for node feature extraction provided by an embodiment of the present application. The training device provided by the embodiment of the present application includes:

The network acquisition module 121 is configured to acquire an initial graph classification network, where the initial graph classification network includes a node feature extraction module, a graph feature extraction module and a graph classification module cascaded in sequence;

The data acquisition module 122 is used for acquiring training data, each sample image in the training data is marked with a sample label, and the sample label represents the real image category of the sample image;

The input module 123 is configured to input each of the above-mentioned sample graphs into the above-mentioned node feature extraction module to obtain the node features of each node of each of the above-mentioned sample graphs;

The above-mentioned input module 123 is used for inputting the node features of each of the above-mentioned nodes to the above-mentioned graph feature extraction module to obtain the graph features of each of the above-mentioned sample graphs;

The above-mentioned input module 123 is used to input the graph features of each of the above-mentioned sample graphs into the above-mentioned graph classification module to obtain the predicted graph categories of the above-mentioned sample graphs;

The loss determination module 124 is configured to determine the total training loss value according to the predicted graph category of each of the above-mentioned sample graphs and the sample label of each of the above-mentioned sample graphs;

The network determination module 125 is configured to iteratively train the above-mentioned initial graph classification network according to the above-mentioned total training loss value and the above-mentioned training data, until the above-mentioned total training loss value meets the training end condition, classify the initial graph classification network at the end of the training. The node feature extraction module is determined to be the above node feature extraction network.

In some feasible implementations, the above-mentioned loss determination module 124 is used to:

Determine the first training loss value according to the predicted graph category of each of the above-mentioned sample graphs and the sample label of each of the above-mentioned sample graphs;

Determine the feature distance between each image feature and the image category feature corresponding to each of the above candidate image categories;

Determine the second training loss value according to the feature distance corresponding to each of the above-mentioned graph features and each of the above-mentioned candidate graph categories;

The total training loss value is determined according to the above-mentioned first training loss value and the above-mentioned second training loss value.

In some feasible implementation manners, the above-mentioned input module 123 is used for:

The node features of each of the above nodes are input to the above graph feature extraction module, and the following operations are performed by the above graph feature extraction module:

For each sample graph, according to the node characteristics of each node in the sample graph, determine the attention distribution corresponding to each node in the sample graph;

According to the node features of each node in the sample graph and the above attention distribution, the graph features of the sample graph are determined.

In some feasible implementations, the above-mentioned loss determination module 124 is used to:

For each graph feature, determine the first feature distance between the graph feature and the graph category feature corresponding to each of the above candidate graph categories;

Determine the target candidate image category that is consistent with the real image category corresponding to the image feature from each of the above candidate image categories;

Determine the second feature distance between the graph feature and the graph category feature corresponding to the target graph category, and determine the difference between the graph feature and the graph category feature corresponding to each candidate graph category according to the first feature distance and the second feature distance. characteristic distance between.

In some feasible implementations, the above-mentioned loss determination module 124 is used to:

obtaining the first weight corresponding to the above-mentioned first training loss value, and the second weight corresponding to the above-mentioned second training loss value;

The total training loss value is determined according to the first training loss value, the second training loss value, the first weight, and the second weight.

In a specific implementation, the above-mentioned image processing apparatus can execute the implementation manners provided by the respective steps in the above-mentioned FIG. 2 , FIG. 4 and/or FIG. 7 through its built-in functional modules. For details, refer to the implementation manners provided by the above-mentioned respective steps. It is not repeated here.

In the embodiment of the present application, since the node feature extraction network determines the category of each node according to the graph category of the candidate graph category, by training the graph classification network, the node feature extraction module in the graph classification network at the end of the training is used as the function The node feature extraction network used to determine the category of the node. On the one hand, the accuracy of the graph category prediction can be used to measure the accuracy of the node feature extraction module in the process of training the graph classification network. On the other hand, the total training loss in the training process can be used. The value of the graph category corresponding to the candidate graph category is continuously adjusted, so that the final node feature extraction network can determine the category of each node according to the accurate and appropriate graph category features, which has high applicability.

Referring to FIG. 13 , FIG. 13 is a schematic structural diagram of an electronic device provided by an embodiment of the present application. As shown in FIG. 13 , the electronic device 1000 in this embodiment may include: a processor 1001 , a network interface 1004 and a memory 1005 . In addition, the above-mentioned electronic device 1000 may further include: a user interface 1003 , and at least one communication bus 1002 . Among them, the communication bus 1002 is used to realize the connection and communication between these components. The user interface 1003 may include a display screen (Display) and a keyboard (Keyboard), and the optional user interface 1003 may also include a standard wired interface and a wireless interface. Optionally, the network interface 1004 may include a standard wired interface and a wireless interface (eg, a WI-FI interface). The memory 1004 may be high-speed RAM memory or non-volatile memory, such as at least one disk memory. Optionally, the memory 1005 may also be at least one storage device located away from the aforementioned processor 1001 . As shown in FIG. 13 , the memory 1005 as a computer-readable storage medium may include an operating system, a network communication module, a user interface module, and a device control application program.

In the electronic device 1000 shown in FIG. 13 , the network interface 1004 can provide a network communication function; the user interface 1003 is mainly used to provide an input interface for the user; and the processor 1001 can be used to call the device control application stored in the memory 1005 program to achieve:

Get the pending graph;

Obtain the node characteristics of each node in the above to-be-processed graph;

For each node, according to the similarity between the node feature of the node and the graph category feature corresponding to each candidate graph category, the category of the node is determined from each of the candidate graph categories, wherein each of the candidate graph categories includes the above-mentioned candidate graph categories. Graph class for processing graphs;

Based on the category of each of the above-mentioned nodes, corresponding processing is performed on at least one node in the above-mentioned to-be-processed graph.

In some possible implementations, the above-mentioned processor 1001 is used for:

The node features of each node in the graph to be processed are obtained through a node feature extraction network, wherein the node feature extraction network includes an initial feature extraction sub-network, and node feature extraction corresponding to each candidate graph category connected to the initial feature extraction sub-network subnet;

For each node, through the above-mentioned initial feature extraction sub-network, extract the initial feature of the node in the above-mentioned to-be-processed graph;

Based on the initial feature of the node, a sub-network is extracted through the node feature corresponding to each candidate graph category, and the feature of the node corresponding to each of the above candidate graph categories is obtained;

The features of the node corresponding to each of the above candidate graph categories are fused to obtain the node feature of the node.

In some possible implementations, the above-mentioned processor 1001 is used for:

Obtain the first feature of the node, and determine the attention autocorrelation coefficient of the node according to the first feature of the node;

According to the first feature of the node and the first feature of each adjacent node of the node, determine the attention cross-correlation coefficient between the node and each of the above-mentioned adjacent nodes;

The initial feature of the node is determined according to the above-mentioned attention autocorrelation coefficient and the above-mentioned attention cross-correlation coefficient.

In some feasible implementations, a graph class feature corresponding to a candidate graph class is used to extract network parameters of the sub-network for the node feature corresponding to the candidate graph class.

In some feasible implementations, the node feature extraction sub-network is a network based on a Gaussian mixture model, and the graph class feature corresponding to a candidate graph category is a Gaussian distribution parameter of the Gaussian mixture model corresponding to the candidate graph category.

In some possible implementations, the above-mentioned processor 1001 is used for:

For each node, determine the highest similarity from the similarity between the node feature of the node and the graph category feature corresponding to each candidate graph category;

The category of the candidate graph corresponding to the above highest similarity is determined as the category of the node.

In some possible implementations, the above-mentioned processor 1001 is used for:

Get the graph category of the above pending graph;

According to the category of each of the above-mentioned nodes, corresponding processing is performed on the nodes whose category of each of the above-mentioned nodes is different from that of the above-mentioned graph to be processed.

In some feasible implementation manners, when the electronic device 1000 is used for training the node feature extraction network, the above-mentioned processor 1001 is used for:

obtaining an initial graph classification network, where the initial graph classification network includes a node feature extraction module, a graph feature extraction module and a graph classification module that are cascaded in sequence;

Obtain training data, each sample image in the training data is marked with a sample label, and the sample label represents the real image category of the sample image;

Inputting each of the above-mentioned sample graphs into the above-mentioned node feature extraction module to obtain node features of each node of each of the above-mentioned sample graphs;

Input the node features of each of the above-mentioned nodes into the above-mentioned graph feature extraction module to obtain the graph features of each of the above-mentioned sample graphs;

Input the graph features of each of the above-mentioned sample graphs into the above-mentioned graph classification module to obtain the predicted graph category of the above-mentioned sample graphs;

Determine the total training loss value according to the predicted graph category of each of the above-mentioned sample graphs and the sample label of each of the above-mentioned sample graphs;

The initial graph classification network is iteratively trained according to the above-mentioned total training loss value and the above-mentioned training data, until the above-mentioned total training loss value meets the training end condition, the node feature extraction module in the initial graph classification network at the end of training is determined as a node Feature extraction network.

In some possible implementations, the above-mentioned processor 1001 is used for:

Determine the first training loss value according to the predicted graph category of each of the above-mentioned sample graphs and the sample label of each of the above-mentioned sample graphs;

Determine the feature distance between each image feature and the image category feature corresponding to each of the above candidate image categories;

Determine the second training loss value according to the feature distance corresponding to each of the above-mentioned graph features and each of the above-mentioned candidate graph categories;

The total training loss value is determined according to the above-mentioned first training loss value and the above-mentioned second training loss value.

In some possible implementations, the above-mentioned processor 1001 is used for:

The node features of each of the above nodes are input to the above graph feature extraction module, and the following operations are performed by the above graph feature extraction module:

For each sample graph, according to the node characteristics of each node in the sample graph, determine the attention distribution corresponding to each node in the sample graph;

According to the node features of each node in the sample graph and the above attention distribution, the graph features of the sample graph are determined.

In some possible implementations, the above-mentioned processor 1001 is used for:

For each graph feature, determine the first feature distance between the graph feature and the graph category feature corresponding to each of the above candidate graph categories;

Determine the target candidate image category that is consistent with the real image category corresponding to the image feature from each of the above candidate image categories;

Determine the second feature distance between the graph feature and the graph category feature corresponding to the target graph category, and determine the difference between the graph feature and the graph category feature corresponding to each candidate graph category according to the first feature distance and the second feature distance. characteristic distance between.

In some possible implementations, the above-mentioned processor 1001 is used for:

obtaining the first weight corresponding to the above-mentioned first training loss value, and the second weight corresponding to the above-mentioned second training loss value;

The total training loss value is determined according to the first training loss value, the second training loss value, the first weight, and the second weight.

It should be understood that, in some feasible implementation manners, the above-mentioned processor 1001 may be a central processing unit (central processing unit, CPU), and the processor may also be other general-purpose processors, digital signal processors (digital signal processor, DSP) , an application specific integrated circuit (ASIC), an off-the-shelf programmable gate array (field-programmable gate array, FPGA) or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, and the like. A general purpose processor may be a microprocessor or the processor may be any conventional processor or the like. The memory, which may include read-only memory and random access memory, provides instructions and data to the processor. A portion of the memory may also include non-volatile random access memory. For example, the memory may also store device type information.

In a specific implementation, the electronic device 1000 can execute the implementations provided by the steps in FIG. 2 , FIG. 4 and/or FIG. 7 through its built-in functional modules. For details, please refer to the implementations provided in the above steps. It is not repeated here.

In the embodiment of the present application, since the node feature extraction network determines the category of each node according to the graph category of the candidate graph category, by training the graph classification network, the node feature extraction module in the graph classification network at the end of the training is used as the function The node feature extraction network used to determine the category of the node. On the one hand, the accuracy of the graph category prediction can be used to measure the accuracy of the node feature extraction module in the process of training the graph classification network. On the other hand, the total training loss in the training process can be used. The value of the graph category corresponding to the candidate graph category is continuously adjusted, so that the final node feature extraction network can determine the category of each node according to the accurate and appropriate graph category features, which has high applicability.

Embodiments of the present application further provide a computer-readable storage medium, where a computer program is stored in the computer-readable storage medium, and is executed by a processor to implement the method provided by each step in FIG. 2 , FIG. 4 and/or FIG. 7 , specifically Reference may be made to the implementation manners provided by the foregoing steps, and details are not described herein again.

The above-mentioned computer-readable storage medium may be an internal storage unit of the task processing apparatus provided in any of the foregoing embodiments, such as a hard disk or a memory of an electronic device. The computer-readable storage medium may also be an external storage device of the electronic device, such as a plug-in hard disk, a smart media card (SMC), a secure digital (SD) card equipped on the electronic device, Flash card (flash card), etc. The above-mentioned computer-readable storage medium may also include a magnetic disk, an optical disk, a read-only memory (ROM) or a random access memory (RAM), and the like. Further, the computer-readable storage medium may also include both an internal storage unit of the electronic device and an external storage device. The computer-readable storage medium is used to store the computer program and other programs and data required by the electronic device. The computer-readable storage medium can also be used to temporarily store data that has been or will be output.

Embodiments of the present application provide a computer program product or computer program, where the computer program product or computer program includes computer instructions, and the computer instructions are stored in a computer-readable storage medium. The processor of the electronic device reads the computer instructions from the computer-readable storage medium, and the processor executes the computer instructions to cause the computer device to perform the method provided by each step in FIG. 2 , FIG. 4 and/or FIG. 7 .

The terms "first", "second" and the like in the claims and description of the present application and the drawings are used to distinguish different objects, rather than to describe a specific order. Furthermore, the terms "comprising" and "having" and any variations thereof are intended to cover non-exclusive inclusion. For example, a process, method, system, product or electronic device comprising a series of steps or units is not limited to the listed steps or units, but optionally also includes unlisted steps or units, or optionally also Include other steps or units inherent to these processes, methods, products or electronic devices. Reference herein to an "embodiment" means that a particular feature, structure, or characteristic described in connection with the embodiment can be included in at least one embodiment of the present application. The appearance of this phrase in various places in the specification is not necessarily all referring to the same embodiment, nor is it a separate or alternative embodiment that is mutually exclusive with other embodiments. It is explicitly and implicitly understood by those skilled in the art that the embodiments described herein may be combined with other embodiments. As used in this specification and the appended claims, the term "and/or" refers to and including any and all possible combinations of one or more of the associated listed items.

Those of ordinary skill in the art can realize that the units and algorithm steps of each example described in conjunction with the embodiments disclosed herein can be implemented in electronic hardware, computer software, or a combination of the two. Interchangeability, the above description has generally described the components and steps of each example in terms of function. Skilled artisans may implement the described functionality using different methods for each particular application, but such implementations should not be considered beyond the scope of this application.

The above disclosures are only the preferred embodiments of the present application, and cannot be used to limit the scope of the claims of the present application. Therefore, equivalent changes made according to the claims of the present application are still within the scope of the present application.

Claims (15)

1. A graph processing method, comprising:

acquiring a graph to be processed;

acquiring node characteristics of each node in the graph to be processed;

for each node, determining the class of the node from each candidate graph class according to the similarity between the node feature of the node and the graph class feature corresponding to each candidate graph class, wherein each candidate graph class comprises the graph class of the graph to be processed;

and performing corresponding processing on at least one node in the graph to be processed based on the category of each node.

2. The method according to claim 1, wherein the obtaining the node characteristics of each node in the graph to be processed comprises:

obtaining node characteristics of each node in the graph to be processed through a node characteristic extraction network, wherein the node characteristic extraction network comprises an initial characteristic extraction sub-network and node characteristic extraction sub-networks corresponding to candidate graph categories connected with the initial characteristic extraction sub-network;

the obtaining of the node characteristics of each node in the graph to be processed through the node characteristic extraction network includes:

for each node, extracting the initial feature of the node in the graph to be processed through the initial feature extraction sub-network;

extracting a sub-network through the node characteristics corresponding to each candidate graph category based on the initial characteristics of the node to obtain the characteristics of the node corresponding to each candidate graph category;

and fusing the characteristics of the node corresponding to each candidate graph category to obtain the node characteristics of the node.

3. The method of claim 2, wherein the extracting the initial feature of the node in the graph to be processed through the initial feature extraction sub-network comprises:

acquiring a first characteristic of the node, and determining an attention autocorrelation coefficient of the node according to the first characteristic of the node;

according to the first characteristic of the node and the first characteristics of each adjacent node of the node, determining the attention cross-correlation coefficient of the node and each adjacent node;

and determining the initial characteristics of the node according to the attention autocorrelation coefficient and the attention cross-correlation coefficient.

4. The method of claim 3, wherein a graph class feature corresponding to a candidate graph class extracts the network parameters of the sub-network for the node feature corresponding to the candidate graph class.

5. The method of claim 4, wherein the node feature extraction sub-network is a Gaussian mixture model-based network, and the feature of the candidate graph class is the Gaussian distribution parameter of the Gaussian mixture model corresponding to the candidate graph class.

6. The method according to any one of claims 1 to 5, wherein for each node, determining the class of the node from the candidate graph classes according to the similarity between the node feature of the node and the graph class feature corresponding to each candidate graph class comprises:

for each node, determining the highest similarity from the similarity between the node characteristics of the node and the graph category characteristics corresponding to each candidate graph category;

and determining the candidate graph category corresponding to the highest similarity as the category of the node.

7. The method according to claim 1, wherein said performing the corresponding processing on at least one node in the graph to be processed based on the category of each node comprises:

acquiring the graph category of the graph to be processed;

and according to the category of each node, correspondingly processing the nodes with the categories of the nodes in each node different from the graph category of the graph to be processed.

8. A node feature extraction network training method is characterized by comprising the following steps:

acquiring an initial graph classification network, wherein the initial graph classification network comprises a node feature extraction module, a graph feature extraction module and a graph classification module which are sequentially cascaded;

acquiring training data, wherein each sample icon in the training data is marked with a sample label, and the sample label represents the real image category of the sample image;

inputting each sample graph into the node feature extraction module to obtain the node feature of each node of each sample graph;

inputting the node characteristics of each node into the graph characteristic extraction module to obtain the graph characteristics of each sample graph;

inputting the graph characteristics of each sample graph into the graph classification module to obtain the prediction graph category of the sample graph;

determining a total training loss value according to the prediction graph type of each sample graph and the sample label of each sample graph;

and performing iterative training on the initial graph classification network according to the total training loss value and the training data until the total training loss value meets the training end condition, and determining a node feature extraction module in the initial graph classification network at the training end as a node feature extraction network.

9. The method of claim 8, wherein determining a total training loss value based on the prediction graph class for each of the sample graphs and the sample label for each of the sample graphs comprises:

determining a first training loss value according to the prediction graph type of each sample graph and the sample label of each sample graph;

determining a feature distance between each graph feature and a graph category feature corresponding to each candidate graph category;

determining a second training loss value according to the feature distance corresponding to each graph feature and each candidate graph category;

and determining a total training loss value according to the first training loss value and the second training loss value.

10. The method of claim 8, wherein inputting the node feature of each node into the graph feature extraction module to obtain the graph feature of each sample graph comprises:

inputting the node characteristics of each node into the graph characteristic extraction module, and executing the following operations through the graph characteristic extraction module:

for each sample graph, determining attention distribution corresponding to each node in the sample graph according to the node characteristics of each node in the sample graph;

and determining the graph characteristics of the sample graph according to the node characteristics of each node in the sample graph and the attention distribution.

11. The method of claim 9, wherein determining a feature distance between each graph feature and a graph class feature corresponding to each candidate graph class comprises:

for each graph feature, determining a first feature distance between the graph feature and a graph category feature corresponding to each candidate graph category;

determining a target candidate graph category which is consistent with the real graph category corresponding to the graph feature from all the candidate graph categories;

and determining a second feature distance between the graph feature and the graph category feature corresponding to the target graph category, and determining a feature distance between the graph feature and the graph category feature corresponding to each candidate graph category according to the first feature distance and the second feature distance.

12. The method of any of claims 9 to 11, wherein determining a total training loss value based on the first training loss value and the second training loss value comprises:

acquiring a first weight corresponding to the first training loss value and a second weight corresponding to the second training loss value;

determining a total training loss value according to the first training loss value, the second training loss value, the first weight, and the second weight.

13. A graph processing apparatus, characterized in that the apparatus comprises:

the graph acquisition module is used for acquiring a graph to be processed;

the characteristic acquisition module is used for acquiring the node characteristics of each node in the graph to be processed;

a classification module, configured to, for each node, determine a class of the node from each candidate graph class according to a similarity between a node feature of the node and a graph class feature corresponding to each candidate graph class, where each candidate graph class includes a graph class of the to-be-processed graph;

and the graph processing module is used for correspondingly processing at least one node in the graph to be processed based on the category of each node.

14. An electronic device comprising a processor and a memory, the processor and the memory being interconnected;

the memory is used for storing a computer program;

the processor is configured to perform the method of any of claims 1 to 7 or the method of any of claims 8 to 12 when the computer program is invoked.

15. A computer-readable storage medium, characterized in that the computer-readable storage medium stores a computer program which is executed by a processor to implement the method of any of claims 1 to 7 or the method of any of claims 8 to 12.

Images