WO2020114122A1

WO2020114122A1 - Neural network system and method for analyzing relationship network graph

Info

Publication number: WO2020114122A1
Application number: PCT/CN2019/112564
Authority: WO
Inventors: 常晓夫; 宋乐
Original assignee: 阿里巴巴集团控股有限公司
Priority date: 2018-12-07
Filing date: 2019-10-22
Publication date: 2020-06-11
Also published as: TW202032422A; CN110009093B; CN110009093A; TWI709086B

Abstract

Disclosed are a neural network system and method executed by means of a computer and used for analyzing a relationship network graph. The neural network system comprises: a feature extraction layer (21) for extracting a feature vector of a node in a relationship network graph; a deep neural network (22) for carrying out first processing on the feature vector, so as to obtain a first output; a graph neural network (23) for carrying out second processing on the feature vector in combination with adjacency information of the relationship network graph, so as to obtain a second output, wherein the adjacency information is used for representing connection relationships between various nodes included in the relationship network graph; and a fusion layer (24) for fusing the first output and the second output, and outputting, based on a fusion result, a prediction result regarding the node.

Description

Neural network system and method for analyzing relational network graph

Technical field

One or more embodiments of this specification relate to a neural network system executed by a computer, and particularly to a neural network system and method for analyzing a relational network graph.

Background technique

The graph is a powerful tool for modeling relational data. Therefore, at present, the data in which the relationship exists is often expressed and modeled in the form of a graph. On the other hand, a graph-based neural network using deep learning methods, Graph Neural Network (Graph Neural Network, Graph NN or GNN) is proposed to learn graph information. The graph neural network GNN can effectively use the information transfer on the graph and fuse the feature information of the nodes or edges to complete the machine learning tasks such as the classification or regression of the nodes or edges on the graph.

However, in real business scenarios, especially at the initial stage of the business, such as inviting new users, the relational data is incomplete. From the perspective of the graph, there are many isolated nodes. Under such circumstances, the graph neural network GNN cannot achieve the expected analysis results.

Therefore, it is hoped that there can be an improved solution to learn, analyze and predict the relationship network graph more effectively.

Summary of the invention

One or more embodiments of this specification describe a neural network system and method executed by a computer for analyzing a relationship network graph, which can learn, analyze, and predict the relationship network graph more effectively.

According to a first aspect, a neural network system for analyzing a relationship network graph executed by a computer is provided, including:

The feature extraction layer is used to extract the feature vectors of the nodes in the relationship network graph;

A deep neural network, which is used to perform first processing on the feature vector to obtain a first output;

A graph neural network, used to combine the adjacency information of the relational network graph to perform a second processing on the feature vector to obtain a second output; wherein the adjacency information is used to represent each of the nodes included in the relational network graph Connection between

A fusion layer is used to fuse the first output and the second output, and output a prediction result for the node based on the fusion result.

In one embodiment, each node included in the relationship network diagram corresponds to each user, and the connection relationship between the various nodes includes one or more of the following: social relationship, media relationship, and fund relationship between users .

According to a possible implementation manner, the relationship network graph is a directed graph, and accordingly, the adjacency information includes an adjacency list or a cross-linked list corresponding to the directed graph.

According to an embodiment, the adjacency information includes the adjacency matrix of the relationship network graph.

According to an embodiment, the graph neural network is a graph convolutional network, and the graph convolutional network includes multiple network layers to perform the second processing, and the second processing includes at least using the adjacency matrix. The element is a weighting factor, which performs a weighted sum operation on the feature vectors of the node and its neighbors.

According to an embodiment, the above fusion layer is specifically used for weighted summation of the first output and the second output, where the first output corresponds to a first weighting factor and the second output corresponds to a second Weighting factor.

Further, in one embodiment, the first weighting factor is a function of the first output, and the second weighting factor is a function of the second output.

In another embodiment, the sum of the first weighting factor and the second weighting factor is 1, and:

The first weighting factor is a function of the first output; or,

The second weighting factor is a function of the second output.

In yet another embodiment, the sum of the first weighting factor and the second weighting factor is 1, and:

The first weighting factor is a function of the first output and the second output; or,

The second weighting factor is a function of the first output and the second output.

According to one implementation, the neural network system is trained in an end-to-end manner.

According to a second aspect, there is provided a computer-implemented method for analyzing a relationship network graph, including:

Extract the feature vectors of the nodes in the relationship network graph;

A deep neural network is used to perform the first processing on the feature vector to obtain a first output;

A graph neural network is used, combined with the adjacency information of the relational network graph, to perform a second processing on the feature vector to obtain a second output; wherein the adjacency information is used to represent the relationship between each node included in the relational network graph Connection

The first output and the second output are fused, and a prediction result for the node is output based on the fusion result.

According to a third aspect, a computer-readable storage medium is provided on which a computer program is stored, and when the computer program is executed in a computer, the computer is caused to perform the method of the second aspect.

According to a fourth aspect, a computing device is provided, including a memory and a processor, wherein the memory stores executable code, and when the processor executes the executable code, the method of the second aspect is implemented .

Through the neural network system and method provided by the embodiments of the present specification, a deep neural network DNN and a graph neural network GNN are combined, and the single-node feature processing capability of DNN and the relational feature processing capability of GNN are fused to make the combined neural network system Can effectively analyze and learn a variety of relationship network diagrams. In the case where the relationship features in the relationship network graph are perfect and effective, the graph neural network GNN can play a major role, and the deep neural network DNN supplements the single node analysis; and if the relationship features are missing or the effect is limited, the deep neural network can still be used The branch of DNN effectively analyzes and processes the nodes in the graph to give a more ideal prediction result.

BRIEF DESCRIPTION

In order to more clearly explain the technical solutions of the embodiments of the present application, the following will briefly introduce the drawings used in the description of the embodiments. Obviously, the drawings in the following description are only some embodiments of the present application. Those of ordinary skill in the art can obtain other drawings based on these drawings without creative efforts.

FIG. 1 is a schematic diagram of a relationship network diagram disclosed in this specification;

2 shows a schematic diagram of a neural network system according to an embodiment;

FIG. 3 shows a schematic diagram of a deep neural network DNN according to an embodiment;

4 shows a schematic diagram of a graph convolutional network GCN according to an embodiment;

FIG. 5 shows a flowchart of a method for analyzing a relationship network diagram according to an embodiment.

detailed description

The solution provided in this specification will be described below in conjunction with the drawings.

According to one or more embodiments of the present specification, a neural network system for processing relational data is proposed. The neural network system can be used for learning and predicting relational network graphs.

The following first describes the relationship network diagram. FIG. 1 is a schematic diagram of a relationship network diagram disclosed in this specification. It can be seen that the relational network graph includes multiple nodes, and the nodes with associated relationships are connected to each other by connecting edges. Nodes that are not associated with other nodes form isolated nodes, such as nodes A, B, and C in Figure 1.

In one embodiment, the type and/or strength of the association relationship can also be distinguished, so as to assign certain attributes or weights to the connected edges. For example, in FIG. 1, thick lines indicate strong connections, and thin lines indicate weak connections. However, this is not necessary.

According to the different entities represented by the nodes, the relationship network graph can reflect the association relationship between various entities. For example, in one embodiment, the nodes may represent sellers or buyers, and the edges between the nodes may represent that a transaction has occurred, thereby reflecting the transaction relationship between entities through a relationship network graph.

In another embodiment, the nodes represent individual users, and the edges between the nodes represent the association between users. More specifically, in different examples, a connection relationship may be established for nodes based on different types of association relationships between users.

In one example, the association relationship between users may include a social relationship between users. In a relationship network formed based on social relationships, if two users have common followers (such as Weibo accounts following the same person together), or they have previous contacts, or join a common group (such as QQ group, WeChat group) Etc.), or interactions in activities such as red envelopes and lottery tickets, then it can be considered that there is a social relationship between the two nodes, and an edge can be established to connect.

In another example, the association relationship between users may include a media relationship between users. In the relationship network formed based on the media relationship, if two users have used the same media, such as encrypted bank card, ID card, mailbox, account number, mobile phone number, physical address (such as MAC address), terminal device number ( For example, UMID, TID, UTDID, etc., then there is a media relationship between these two users, you can establish an edge to connect.

In yet another example, the association relationship between users may include the financial relationship between users. In the relationship network formed based on the fund relationship, if at least one of the fund transactions exists between two users, then there is an edge between the two nodes. Fund transactions can include charge, charge, bar code collection, bar code payment, AA collection, C2C mobile phone face-to-face payment, gift money, rent, red envelope, credit card repayment, purchase, intimate payment, subscription service, etc.

In other embodiments, the nodes in the relationship network graph may also represent other entities, and the connections between the nodes may be based on various types of association relationships.

For the relational network graph of FIG. 1, the graph neural network GNN can generally be used for learning and prediction. The learning process generally corresponds to the training process of graph neural network GNN.

When training the graph neural network GNN, it is necessary to add labels to at least some nodes in FIG. 1 according to the purpose of the prediction business, so as to perform supervised learning. For example, assuming that the nodes in the relational network graph in FIG. 1 correspond to individual users, if you want to learn and predict the credit risk of each user, at least some nodes need to add "high risk users" (users suspected of fraud) and "normal users" ", provide these labeled nodes, together with the connection relationship of these nodes in the graph, to the graph neural network GNN for GNN to train and learn. After the training is completed, the unknown user can be input to the graph neural network GNN, so that the graph neural network GNN uses the trained network parameters to predict the unknown user.

However, in many cases, the effect of the above scheme is not ideal.

On the one hand, the relationship network graph is constructed based on the association relationship. As mentioned earlier, there are various association relationships between nodes. If the relationship between the selected association relationship and the business purpose is not high enough when constructing the relationship network graph, then the relationship network graph constructed based on such association relationship Learning, the results are often not ideal. For example, in the case where the business purpose is to learn the user’s purchase intention of a certain type of product, if the association relationship is selected as whether it has a social relationship with a common object of interest, such relationship data may not be effective for determining the business purpose of the purchase intention .

On the other hand, in some business scenarios, such as pull-in, promotion, and other business stages, a large number of new users are added. At this time, the relationship data of the new users is very incomplete, forming a large number of isolated nodes in the graph. There are certain difficulties in predicting these isolated nodes. Because in the training phase, the graph neural network GNN is usually learned based on the relationship graph with a relatively perfect connection relationship, the sample distribution between the training phase and the prediction phase is quite different, which will affect the prediction effect of the graph neural network GNN. Moreover, the lack of characteristic data about the connection relationship of the isolated node itself makes the graph neural network GNN unable to achieve its optimal effect.

Based on the above analysis and research, in the embodiments of the present specification, a combined neural network system is provided, which can learn the relationship network graph more effectively.

FIG. 2 shows a schematic diagram of a neural network system according to an embodiment. The neural network system is executed by a computer and is used to learn a relationship network graph and process relationship data. As shown in FIG. 2, the neural network system includes a feature extraction layer 21 for extracting feature vectors of nodes in the relational network graph; a deep neural network 22 for performing first processing on the feature vectors to obtain An output; graph neural network 23, used to combine the adjacency information of the relational network graph, and perform a second processing on the feature vector to obtain a second output; a fusion layer 24, used for the first output and the The second output is fused, and a prediction result for the node is output based on the fusion result. The following specifically describes the implementation of each of the above network parts.

The feature extraction layer 21 is used to extract feature vectors of nodes in the relationship network graph. The relationship network diagram is, for example, the relationship network diagram shown in FIG. 1, which includes multiple nodes, and there is a connection relationship between nodes having an association relationship. The feature extraction layer 21 performs feature extraction on the nodes in the relational network graph, and the extracted features are the features of the relevant nodes themselves, so the extracted features constitute a feature vector.

In one embodiment, each node in the above relational network diagram corresponds to each user, for example, Alipay user. For such a node, the feature extraction layer 21 may extract the user's basic attribute features to form a feature vector. The basic attribute features include, for example, user ID, registration duration, gender, age, and so on.

In one embodiment, the feature extraction layer 21 also extracts features related to the business purpose according to the business purpose. For example, in the case where the business purpose is to predict the user's purchase intention, the feature extraction layer 21 also obtains the user's purchase records, and performs feature extraction based on the purchase records, such as extracting multiple features such as the number of purchases, purchase categories, and purchase amount. For another example, when the business purpose is to predict the borrowing risk of the user, the feature extraction layer 21 also obtains the borrowing record of the user, and extracts the features based on the borrowing record, such as extracting the number of loans, the amount of loan, the amount of repayment, and the number of trustworthiness , Overdue times and other characteristics.

After the feature extraction layer 21 extracts the feature vectors of the nodes, the feature vectors are input to the deep neural network 22 and the graph neural network 23 in parallel for processing.

Deep Neural Network (DNN) is a multi-layer fully connected forward-structured artificial neural network. Figure 3 shows a schematic diagram of a deep neural network according to one embodiment. As shown in Figure 3, the deep neural network DNN contains multiple network layers. These network layers can be divided into an input layer, a hidden layer, and an output layer. Between the input layer and the output layer are hidden layers. Generally, when there are many hidden layers in a neural network, we call it a deep neural network. Each network layer in the DNN contains several neurons, and all neurons except the input layer operate on the input data through the activation function. In DNN, the network layers are fully connected, that is, any neuron in layer i is connected to any neuron in layer i+1.

Deep neural network DNN can be designed and trained for analysis and prediction of various business scenarios.

According to an embodiment of the present specification, the deep neural network DNN is taken as a branch part of the neural network system of an example. After the feature extraction layer 21 extracts the feature vector of the node, the feature vector is provided to the input layer of the deep neural network DNN 22, which is processed through the hidden layer, and the processing result is output through the output layer of the DNN. For simplicity of description, the process of processing the feature vectors by the hidden layer of the DNN is called first processing, and the processing result output by the DNN output layer is called the first output.

It should be noted that the descriptions of "first" and "second" in this article are only for distinguishing similar concepts, and for simplicity and clarity of description, they do not have a limiting effect on order and other aspects.

On the other hand, the node feature vector extracted by the feature extraction layer 21 is also input to the graph neural network GNN23. It can be understood that the graph neural network GNN is used to analyze and learn the relational network graph. Similar to the conventional neural network, the graph neural network GNN also has a multi-layer network structure, which operates and processes input data through the function mapping of neurons. In particular, as a neural network dedicated to processing relational network graphs, GNN will process the feature vectors of nodes in the process of combining the connection information between the nodes in the relational network graph. The connection relationship information between the nodes in the above relationship network graph is also called adjacency information. For convenience of description, the process of processing the feature vectors of the nodes in the GNN in conjunction with the adjacency information is called second processing, and the result of the GNN processing is called the second output.

As shown in FIG. 2, for the needs of GNN processing, the adjacency information of the relationship network graph is obtained in advance and provided to the graph neural network GNN23.

The aforementioned adjacency information can be embodied in various forms. Typically, the connection relationship between nodes in the relational network graph can be represented by an adjacency matrix. Assuming that the relational network graph contains n nodes, the adjacency matrix is an n*n-dimensional matrix. In the simplest case (in the case where the connecting edges between nodes have equal weights), if node i and node j There is a connection relationship, then the matrix element A _ij =1, otherwise, the matrix element A _ij =0.

In addition, you can also use the degree matrix, Laplace matrix and other forms to represent the connection relationship between nodes in the relationship network graph.

In one embodiment, the above relational network graph is a directed graph, that is, the connection between nodes is directed. For directed graphs, in some cases, an adjacency table is also used to record the adjacency information of the relationship network graph, where the adjacency table may further include a forward adjacency table and a reverse adjacency table. In another example, a cross-linked list is generated based on the positive adjacency list and the reverse adjacency list, and the cross-linked list is used to record the connection relationship between the nodes in the directed graph.

In other embodiments, the adjacency information may also be recorded in other forms.

Therefore, the graph neural network GNN 23 can determine the neighbor nodes of the current node and the connection edge information between the current node and these neighbor nodes based on the adjacency information, and extract the node information of these neighbor nodes and the edge information and features of the connection edges. The feature vectors of the current node input at layer 21 are integrated, and a second output is obtained.

Specifically, in one embodiment, the graph neural network GNN 23 is implemented using a graph convolutional network GCN (Graph Convolutional Network).

Fig. 4 shows a schematic diagram of a graph convolution network GCN according to one embodiment. In one embodiment, the graph convolutional network GCN contains multiple network layers, and each network layer defines a neural network model f(X,A) through the neurons therein, where X is the input feature vector, that is, the aforementioned features The feature vector of the current node input to the GCN by the extraction layer 21, A is the adjacency matrix of the relational network graph, and the neural network model f(X, A) can be more specifically expressed as:

among them,

Is a degree matrix of A+λI, and λ is a hyperparameter, used to control the weight of a node relative to its neighbors, and is set to 1 in the original model.

H ^(l+1) represents the output of each network layer. When l=0, H ⁰ =X, that is, the input layer receives the feature vector X of the node. The d*d dimension vector W ^(l) and the d*1 dimension parameter b are both trainable network layer parameters, and σ is a nonlinear function. In different examples, the σ function may be a Relu function, a sigmoid function, a tanh function, a softmax function, and so on.

It can be seen from the above equation that the first layer of the network uses the elements in the adjacency matrix as the weighting factors to sum the feature vectors of the current node and its neighbors (labeled or unlabeled), and then use W ^(l) and b Perform a linear transformation operation, and then apply a nonlinear activation function σ. The subsequent operations of each network layer also include at least, using the elements in the adjacency matrix as a weighting factor, and performing a weighted sum operation on the node vector output by the previous network layer and its neighbor node vectors; in addition, using W ^{(l )} And b linear transformation operation, and the applied nonlinear activation function σ operation.

In other words, for node u _i , the output of layer l is calculated by the following formula:

Where W ^l and b ^l are the trainable network layer parameters of layer l, and nhood(i) represents the neighbor node of node i.

In this way, in the graph convolutional network GCN, the feature vectors of the nodes are processed by combining the adjacency information expressed by the adjacency matrix.

In an embodiment, the above adjacency matrix may be a matrix after normalization, so as to avoid the situation that the distribution of elements in some adjacency matrices is too large. For example, some relational network graphs contain some super nodes, which are connected to almost every node in the graph; on the other hand, some nodes are very isolated and have few connections, which will cause different nodes in the adjacency matrix The number of corresponding connecting edges (for example, the sum of elements corresponding to a row or a column in the matrix) is very different. To this end, the adjacency matrix can be normalized. The use of normalized adjacency matrix for weighted summation in GCN is equivalent to the average pooling operation of the current node and the adjacent nodes.

In other embodiments, the graph neural network GNN 23 may also use other network structures and hidden layer algorithms. However, the common point is that the second processing performed by the graph neural network GNN 23 needs to combine the adjacency information of the relational network graph to comprehensively process the feature vector of the current node to obtain the second output.

Comparing the processing of the deep neural network DNN 22 and the graph neural network GNN 23, it can be understood that the first processing performed by the deep neural network DNN 22 is only for the feature vector of the current node, focusing on the analysis of the attribute characteristics of the node itself, that is, the single point feature The second processing performed by the graph neural network GNN 23 needs to incorporate the adjacency information of the relational network graph and introduce the relationship characteristics between the current node and other nodes.

Next, the first output of the deep neural network DNN 22 and the second output of the graph neural network 23 are fused through the fusion layer 24, and the prediction result for the current node is output based on the fusion result. For convenience of expression, let the first output be H1 and the second output H2. In different embodiments, the fusion layer 24 may fuse H1 and H2 in various ways to obtain a fusion result H.

In one embodiment, the fusion layer 24 fuses the first output H1 and the second output H2 through a fusion function F:

H=F(H1, H2)

The fusion function F can be various linear or nonlinear functions.

In one embodiment, the fusion layer 24 performs weighted summation on the first output H1 and the second output H2 (corresponding to the case where the fusion function is linear summation), that is:

H＝w1*H1+w2*H2

Where w1 is the first weighting factor corresponding to the first output, and w2 is the second weighting factor corresponding to the second output.

It can be understood that the first output H1 and the second output H2 are both in the form of output vectors; and the weighting factors w1 and w2 may be scalars, vectors, or even matrices. The values of the weighting factors w1 and w2 are optimized and determined through the training process of the neural network system.

In one embodiment, further, the above weight factor is set to a function corresponding to the output, for example, the first weight factor w1 is set to a function of the first output H1, and the second weight factor w2 is set to the second output H2 function:

w1＝f1(H1)

w2=f2(H2)

More specifically, the specific form of the above function may be:

w1＝g(u1*H1+b1)

w2＝g(u2*H2+b2)

The function g is preferably a non-linear function, such as sigmoid function and tanh function.

In this way, the weighting factors w1 and w2 are trained and determined, that is, the training and determination parameters u1, b1, u2, b2.

In the above manner, the weighting factors w1 and w2 are trained independently of each other, and the value range of the final result H is not guaranteed.

In another embodiment, first the sum of the first weighting factor and the second weighting factor is set to 1, and then only one of them is set and adjusted. For example, set the fusion result H as:

H＝α*H1+(1-α)*H2

In this way, only the first weighting factor α needs to be set and adjusted, and the second weighting factor is determined accordingly.

In an embodiment, the first weighting factor α may be set as a function of the first output, or a function of the first output and the second output, namely:

α=g(H1)=g(u*H1+b), or

α=g(H1, H2)=g(u1*H1+u2*H2+b)

Of course, the second weighting factor β can also be set and adjusted so that the first weighting factor is (1-β), that is:

H＝(1-β)*H1+β*H2

Further, the second weighting factor may also be set as a function of the second output, or a function of the first output and the second output, namely:

β=g(H2)=g(u*H2+b), or

β=g(H1, H2)=g(u1*H1+u2*H2+b)

The above examples illustrate several typical fusion methods. On the basis of the above examples, those skilled in the art can also think of other similar fusion methods, which should be covered by the concept of this specification.

Through various fusion methods, the fusion layer 24 obtains a fusion result H, and outputs a prediction result for the current node based on the fusion result H. The prediction result is for the predicted value of the labeled node in the training phase; in the use phase, it is the final classification prediction for the unknown result. The following describes the execution process of the neural network system shown in FIG. 2 in the training phase and the use phase.

For the neural network system shown in FIG. 2 combining the deep neural network DNN and the graph neural network GNN, the end-to-end training can be used. Specifically, in the training stage, labeled node information is input on the input side of the entire neural network system, that is, the feature extraction layer 21 extracts feature vectors of several labeled nodes. As mentioned above, according to different business purposes, the labels may be various types of labels, for example, labels used to represent risk levels, such as 1 for high-risk users, 0 for ordinary users, and so on. Then, the prediction results for each node are obtained on the output side of the entire neural network system. Specifically, the prediction result is output by the fusion layer 24 according to the fusion result, and may be embodied as a prediction value for each node. Compare the predicted value of each node with its label, according to the comparison result and the preset loss function, get the error of this batch of samples, and then through the error back propagation, adjust the network parameters of the entire neural network system, and finally Identify the network parameters that minimize the error. Once the optimal network parameters are determined, it can be considered that the training of the neural network system is completed, and the neural network system can be used for the prediction of unknown nodes.

In the use phase, the node information of the unknown node is input to the input side of the neural network system, that is, the feature extraction layer 21 extracts the feature vector of the unknown node. Then, the network parameters determined by the training stage in the neural network system are used, and the feature vector is processed by the parallel deep neural network DNN and the graph neural network GNN. The fusion layer 24 outputs a prediction result based on the fusion result. The prediction result is The output result of business prediction for the unknown node.

It can be seen from the above that the neural network system of FIG. 2 combines the deep neural network DNN and the graph neural network GNN, which fuses the single-node feature processing ability of the DNN and the relational feature processing ability of the GNN, making the combined neural network The system can effectively analyze and learn all kinds of relational network graphs. Graph neural network GNN can play a major role in the case where the relationship features in the relationship network graph are perfect and effective. The deep neural network DNN supplements the analysis of single nodes; and if the relationship features are missing or have limited effects, such as the existence of the relationship network graph A large number of isolated nodes, or the relationship based on the construction of the relationship network graph is not very effective for the business. In such a case, the nodes in the graph can be effectively analyzed and processed through the branch of the deep neural network DNN to give More ideal prediction results.

According to an embodiment of another aspect, a method for analyzing a relationship network graph executed by a computer is also provided. FIG. 5 shows a flowchart of a method for analyzing a relationship network diagram according to an embodiment. It can be understood that this method can be performed by any device, device, computing platform, or computing cluster with computing and processing capabilities. As shown in Figure 5, the method includes:

Step 51: Extract the feature vectors of the nodes in the relationship network graph;

Step 52: Use a deep neural network to perform first processing on the feature vector to obtain a first output;

Step 53: A graph neural network is used to combine the adjacency information of the relational network graph to perform a second processing on the feature vector to obtain a second output; wherein the adjacency information is used to represent each of the relations network graph The connection relationship between nodes;

Step 54: Fusion the first output and the second output, and output a prediction result for the node based on the fusion result.

It should be noted that step 52 and step 53 may be executed in any order, or in parallel, and are not limited herein.

In a possible implementation manner, the relationship network graph is a directed graph, and the adjacency information includes an adjacency list or a cross-linked list corresponding to the directed graph.

In one embodiment, the adjacency information includes the adjacency matrix of the relationship network graph.

According to an embodiment, the graph neural network described above is a graph convolutional network. The graph convolutional network includes a plurality of network layers to perform the second processing. The second processing includes at least using the adjacency matrix. The element is a weighting factor, which performs a weighted sum operation on the feature vectors of the node and its neighbors.

In one embodiment, the fusing the first output and the second output in step 54 specifically includes weighting and summing the first output and the second output, wherein the first output Corresponding to the first weighting factor, the second output corresponds to the second weighting factor.

Further, in an embodiment, the first weighting factor is a function of the first output, and the second weighting factor is a function of the second output.

The first weighting factor is a function of the first output; or,

The second weighting factor is a function of the second output.

Or, in yet another embodiment, the sum of the first weighting factor and the second weighting factor is 1, and:

Through the above methods, combined with the deep neural network processing of single-node features and the graph neural network processing of node relationship features, the relationship network graph is comprehensively analyzed.

According to an embodiment of another aspect, there is also provided a computer-readable storage medium on which a computer program is stored, and when the computer program is executed in a computer, the computer is caused to perform the method described in conjunction with FIG. 5.

According to an embodiment of still another aspect, a computing device is further provided, including a memory and a processor, where executable code is stored in the memory, and when the processor executes the executable code, the implementation described in conjunction with FIG. 5 is implemented method.

Those skilled in the art should be aware that in one or more of the above examples, the functions described in this application may be implemented by hardware, software, firmware, or any combination thereof. When implemented in software, these functions can be stored in a computer-readable medium or transmitted as one or more instructions or code on a computer-readable medium.

The specific embodiments described above further describe the purpose, technical solutions, and beneficial effects of this application in detail. It should be understood that the above descriptions are only specific implementations of this application, and are not intended to limit the scope of this application. The scope of protection, any modifications, equivalent replacements, improvements, etc. made on the basis of the technical solutions of this application, shall be included in the scope of protection of this application.

Claims

A neural network system executed by a computer and used to analyze a relational network graph, including:

The feature extraction layer is used to extract the feature vectors of the nodes in the relationship network graph;

A deep neural network, which is used to perform first processing on the feature vector to obtain a first output;

A graph neural network, used to combine the adjacency information of the relational network graph to perform a second processing on the feature vector to obtain a second output; wherein the adjacency information is used to represent each of the nodes included in the relational network graph Connection between

A fusion layer is used to fuse the first output and the second output, and output a prediction result for the node based on the fusion result.
The neural network system according to claim 1, wherein each node included in the relational network graph corresponds to each user, and the connection relationship between the various nodes includes one or more of the following: between users Social relations, media relations and financial relations.
The neural network system according to claim 1, wherein the relational network graph is a directed graph, and the adjacency information includes an adjacency list or a cross list corresponding to the directed graph.
The neural network system according to claim 1, wherein the adjacency information includes an adjacency matrix of the relational network graph.
The neural network system according to claim 4, wherein the graph neural network is a graph convolutional network, the graph convolutional network includes a plurality of network layers to perform the second processing, the second processing includes at least , Using the elements in the adjacency matrix as a weighting factor to perform a weighted sum operation on the feature vectors of the node and its neighbors.
The neural network system according to claim 1, the fusion layer is specifically used for weighted summation of the first output and the second output, wherein the first output corresponds to a first weighting factor, the The second output corresponds to the second weighting factor.
The neural network system of claim 6, wherein the first weighting factor is a function of the first output and the second weighting factor is a function of the second output.
The neural network system according to claim 6, wherein the sum of the first weighting factor and the second weighting factor is 1, and:

The first weighting factor is a function of the first output; or,

The second weighting factor is a function of the second output.
The neural network system according to claim 6, wherein the sum of the first weighting factor and the second weighting factor is 1, and:

The first weighting factor is a function of the first output and the second output; or,

The second weighting factor is a function of the first output and the second output.
The neural network system according to claim 1, the neural network system being trained in an end-to-end manner.
A computer-implemented method for analyzing relational network diagrams, including:

Extract the feature vectors of the nodes in the relationship network graph;

A deep neural network is used to perform the first processing on the feature vector to obtain a first output;

A graph neural network is used, combined with the adjacency information of the relational network graph, to perform a second processing on the feature vector to obtain a second output; wherein the adjacency information is used to represent the relationship between each node included in the relational network graph Connection

The first output and the second output are fused, and a prediction result for the node is output based on the fusion result.
The method according to claim 11, wherein each node included in the relationship network graph corresponds to each user, and the connection relationship between the various nodes includes one or more of the following: social between users Relations, media relations and funding relations.
The method according to claim 11, wherein the relationship network graph is a directed graph, and the adjacency information includes an adjacency list or a cross-linked list corresponding to the directed graph.
The method according to claim 11, wherein the adjacency information includes an adjacency matrix of the relational network graph.
The method according to claim 14, wherein the graph neural network is a graph convolutional network, the graph convolutional network includes a plurality of network layers to perform the second processing, the second processing includes at least, using The elements in the adjacency matrix are weighting factors, and a weighted sum operation is performed on the feature vectors of the node and its neighbors.
The method of claim 11, wherein the fusing the first output and the second output includes weighted summing the first output and the second output, wherein the first One output corresponds to the first weighting factor, and the second output corresponds to the second weighting factor.
The method of claim 16, wherein the first weighting factor is a function of the first output and the second weighting factor is a function of the second output.
The method according to claim 16, wherein the sum of the first weighting factor and the second weighting factor is 1, and:

The first weighting factor is a function of the first output; or,

The second weighting factor is a function of the second output.
The method according to claim 16, wherein the sum of the first weighting factor and the second weighting factor is 1, and:

The first weighting factor is a function of the first output and the second output; or,

The second weighting factor is a function of the first output and the second output.
A computing device, including a memory and a processor, characterized in that executable code is stored in the memory, and when the processor executes the executable code, it implements any one of claims 11-19 method.