CN114610950B - Graph network node representation method - Google Patents
Graph network node representation method Download PDFInfo
- Publication number
- CN114610950B CN114610950B CN202011416368.8A CN202011416368A CN114610950B CN 114610950 B CN114610950 B CN 114610950B CN 202011416368 A CN202011416368 A CN 202011416368A CN 114610950 B CN114610950 B CN 114610950B
- Authority
- CN
- China
- Prior art keywords
- node
- vector
- attribute
- network
- nodes
- Prior art date
- Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
- Active
Links
- 238000000034 method Methods 0.000 title claims abstract description 52
- 230000003993 interaction Effects 0.000 claims abstract description 51
- 230000002776 aggregation Effects 0.000 claims abstract description 32
- 238000004220 aggregation Methods 0.000 claims abstract description 32
- 239000013604 expression vector Substances 0.000 claims abstract description 29
- 239000013598 vector Substances 0.000 claims description 84
- 239000011159 matrix material Substances 0.000 claims description 30
- 238000013528 artificial neural network Methods 0.000 claims description 13
- 208000006011 Stroke Diseases 0.000 claims description 11
- 208000015181 infectious disease Diseases 0.000 claims description 8
- 238000011176 pooling Methods 0.000 claims description 8
- 230000002685 pulmonary effect Effects 0.000 claims description 8
- 206010008190 Cerebrovascular accident Diseases 0.000 claims description 6
- 230000002490 cerebral effect Effects 0.000 claims description 6
- 230000004913 activation Effects 0.000 claims description 5
- 238000010606 normalization Methods 0.000 claims description 4
- 230000017105 transposition Effects 0.000 claims 1
- 230000007246 mechanism Effects 0.000 abstract description 13
- 230000004069 differentiation Effects 0.000 abstract description 3
- 239000013589 supplement Substances 0.000 abstract description 3
- 230000006870 function Effects 0.000 description 14
- 238000012549 training Methods 0.000 description 6
- 201000010099 disease Diseases 0.000 description 5
- 208000037265 diseases, disorders, signs and symptoms Diseases 0.000 description 5
- 238000009826 distribution Methods 0.000 description 4
- 238000003012 network analysis Methods 0.000 description 4
- 239000000047 product Substances 0.000 description 4
- 238000004364 calculation method Methods 0.000 description 3
- 238000013527 convolutional neural network Methods 0.000 description 3
- 238000013507 mapping Methods 0.000 description 3
- 230000009286 beneficial effect Effects 0.000 description 2
- 230000009466 transformation Effects 0.000 description 2
- 239000000654 additive Substances 0.000 description 1
- 230000000996 additive effect Effects 0.000 description 1
- 238000013459 approach Methods 0.000 description 1
- 230000008033 biological extinction Effects 0.000 description 1
- 238000012512 characterization method Methods 0.000 description 1
- 238000000354 decomposition reaction Methods 0.000 description 1
- 238000003745 diagnosis Methods 0.000 description 1
- 238000010586 diagram Methods 0.000 description 1
- 230000000694 effects Effects 0.000 description 1
- 238000002474 experimental method Methods 0.000 description 1
- 238000004880 explosion Methods 0.000 description 1
- 230000001939 inductive effect Effects 0.000 description 1
- 230000002452 interceptive effect Effects 0.000 description 1
- 239000000463 material Substances 0.000 description 1
- 238000012986 modification Methods 0.000 description 1
- 230000004048 modification Effects 0.000 description 1
- 238000007781 pre-processing Methods 0.000 description 1
- 238000012545 processing Methods 0.000 description 1
- 238000011160 research Methods 0.000 description 1
- 238000005096 rolling process Methods 0.000 description 1
- 235000002020 sage Nutrition 0.000 description 1
- 238000003860 storage Methods 0.000 description 1
- 238000006467 substitution reaction Methods 0.000 description 1
- 208000024891 symptom Diseases 0.000 description 1
- 230000001131 transforming effect Effects 0.000 description 1
Classifications
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F16/00—Information retrieval; Database structures therefor; File system structures therefor
- G06F16/90—Details of database functions independent of the retrieved data types
- G06F16/901—Indexing; Data structures therefor; Storage structures
- G06F16/9024—Graphs; Linked lists
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
- G06N3/00—Computing arrangements based on biological models
- G06N3/02—Neural networks
- G06N3/04—Architecture, e.g. interconnection topology
- G06N3/045—Combinations of networks
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
- G06N3/00—Computing arrangements based on biological models
- G06N3/02—Neural networks
- G06N3/08—Learning methods
-
- H—ELECTRICITY
- H04—ELECTRIC COMMUNICATION TECHNIQUE
- H04L—TRANSMISSION OF DIGITAL INFORMATION, e.g. TELEGRAPHIC COMMUNICATION
- H04L41/00—Arrangements for maintenance, administration or management of data switching networks, e.g. of packet switching networks
- H04L41/14—Network analysis or design
Landscapes
- Engineering & Computer Science (AREA)
- Theoretical Computer Science (AREA)
- Physics & Mathematics (AREA)
- Software Systems (AREA)
- Data Mining & Analysis (AREA)
- General Engineering & Computer Science (AREA)
- General Physics & Mathematics (AREA)
- Artificial Intelligence (AREA)
- Evolutionary Computation (AREA)
- Life Sciences & Earth Sciences (AREA)
- Databases & Information Systems (AREA)
- Biomedical Technology (AREA)
- Biophysics (AREA)
- Computational Linguistics (AREA)
- Health & Medical Sciences (AREA)
- General Health & Medical Sciences (AREA)
- Molecular Biology (AREA)
- Computing Systems (AREA)
- Mathematical Physics (AREA)
- Computer Networks & Wireless Communication (AREA)
- Signal Processing (AREA)
- Management, Administration, Business Operations System, And Electronic Commerce (AREA)
- Information Retrieval, Db Structures And Fs Structures Therefor (AREA)
Abstract
The application relates to the technical field of graph networks, in particular to a graph network node representation method, which comprises the following steps: respectively obtaining attribute interaction expression vectors and neighborhood aggregation expression vectors of target nodes in a graph network; and splicing the attribute interaction expression vector and the neighborhood aggregation expression vector to obtain a final expression vector of the target node in the graph network. In the embodiment, the neighborhood aggregation mechanism of the graph network is considered to acquire the information of the additional information supplement node from the neighbor nodes, so that the learned target node representation has richer information and stronger differentiation.
Description
Technical Field
The application relates to the technical field of graph networks, in particular to a graph network node representation method.
Background
The efficient representation of network nodes is a key issue in performing network analysis tasks. In recent years, researchers have proposed network representation learning aimed at learning low-dimensional dense vector representations for all nodes in a high-dimensional sparse network, preserving rich network information in order to perform subsequent network analysis tasks.
In the medical field, the traditional adjacent matrix graph network analysis method only considers the adjacent relation between nodes, and the high-dimensional characteristics of the adjacent matrix graph network analysis method bring huge calculation and storage cost, so that the clinical large-scale network containing mass facts is difficult to deal with. The existing network representation learning method is used for exploring neighborhood aggregation information, and most of the network representation learning method belongs to direct push learning rather than inductive learning, and node signals which do not exist cannot be captured. Clinically, however, negative symptoms often help to rule out possible diseases. The limited representation of clinical path characteristics results in bottlenecks in practical use.
At present, a plurality of network representation learning methods not only utilize the topology structure of a network, but also explore attribute information reflecting certain characteristics of the network. In a real-world network, node attributes often have the characteristic of high-dimensional sparsity, face the problem of dimension disasters, and limit the capacity of node representation.
Disclosure of Invention
The application mainly solves the technical problem that the node representation capability in the prior art is limited.
A graph network node representation method, comprising:
respectively obtaining attribute interaction expression vectors and neighborhood aggregation expression vectors of target nodes in a graph network;
and splicing the attribute interaction expression vector and the neighborhood aggregation expression vector to obtain a final expression vector of the target node in the graph network.
The obtaining the attribute interaction expression vector of the target node in the graph network comprises the following steps:
in the embedded layer, searching and calculating an embedded vector set corresponding to the non-zero attribute of the target node through an embedded table;
calculating second-order attribute interaction by adopting second-order interaction pooling operation, and compressing an embedded vector set of the target node into a single vector;
and inputting the single vector into a full-connection layer, stacking a plurality of full-connection layers, and taking the output of the last layer as the attribute interaction representation vector of the target node.
The searching and calculating the embedded vector set corresponding to the non-zero attribute of the target node through the embedded table in the embedded layer comprises the following steps:
for graph network g= (V, E, X), where V represents the set of all nodes of the network, E represents the set of edges in the network, the number of nodes and edges in the network are |v| and |e|, respectively; the attribute matrix of all nodes is: x= [ X ] 1 ,x 2 ,...,x |V| ]E|V|xF, and searching through an embedded table to obtain a target node V i Corresponding embedded vectors of all non-zero attributes of the set S are obtained to form an embedded vector set S i =[x i1 v 1 ,x i2 v 2 ,...,x ij v j ]Wherein x is ij Representing a target node v i Is the j-th dimensional attribute value of (c).
The calculating the second-order attribute interaction by adopting the second-order interaction pooling operation, and compressing the embedded vector set of the target node into a single vector comprises:
calculating second-order attribute interaction in a dot product manner by adopting second-order interaction pooling operation, and compressing the embedded vector set into a single vector t i :
Wherein,representing attribute->And->Second order interactions between.
1. The graph network node representation method of claim 4, wherein inputting the single vector into the fully-connected layer, stacking a plurality of fully-connected layers, and taking the output of the last layer as the attribute interaction representation vector of the target node comprises:
the full-connection layer is as follows:
wherein the method comprises the steps ofRespectively a first layer weight matrix and an offset vector, wherein sigma (·) is a nonlinear activation function, and L' is represented;
after stacking a plurality of fully connected layers, the output of the last layer is taken as a target node v i Attribute interaction representation vector f i 。
The obtaining the neighborhood aggregation expression vector of the target node in the graph network comprises the following steps:
calculating a target node v i With neighbor node v i Is related to (a)
Wherein,
all nodes v with target by softmax function i Adjacent node v j Is related to (a)Normalization is performed to obtain the attention weighting coefficient +.>
By taking the attentionWeighting all neighbor nodes of the target node by the weight coefficient to obtain a target node v i Is a neighborhood aggregate representation vector h i ;
Wherein N is i Representation and v i A set of neighboring neighbor nodes, whereinRepresents the layer I output of the neural network, +.>Representing a weight matrix.
The graph network node representation method further comprises the following steps: calculating the attribute smoothness lambda of the current graph network according to the final representation vector a Smoothness lambda according to the attribute a Judging attribute similarity between the target node and the neighbor nodes;
wherein,
further, the method further comprises: calculating label smoothness lambda of graph network node l According to the label smoothness lambda l Judging the validity of neighbor information;
wherein,
further, the method further comprises the following steps: smoothness lambda according to the attribute a And label smoothness lambda l Calculating the noise S of all nodes in the graph network;
wherein sigma 2 Representation squareAnd (3) difference.
Further, the method further comprises: for node classification tasks, calculating a loss function L by the following formula, and updating the weight by minimizing the loss function L;
wherein, C is the number of labels,to mark the number of the node, Y lc =softmax(z l ) The predicted probability vector after the input softmax classifier of the vector is represented for the node.
The graph network node representation method according to the above embodiment includes: respectively obtaining attribute interaction expression vectors and neighborhood aggregation expression vectors of target nodes in a graph network; and splicing the attribute interaction expression vector and the neighborhood aggregation expression vector to obtain a final expression vector of the target node in the graph network. In the embodiment, the neighborhood aggregation mechanism of the graph network is considered to acquire the information of the additional information supplement node from the neighbor nodes, so that the learned target node representation has richer information and stronger differentiation.
Drawings
FIG. 1 is a flow chart of a method for representing a graph network node in an embodiment of the application;
fig. 2 is a schematic diagram of a processing procedure of a graph network node representation method according to an embodiment of the present application.
Detailed Description
The application will be described in further detail below with reference to the drawings by means of specific embodiments. Wherein like elements in different embodiments are numbered alike in association. In the following embodiments, numerous specific details are set forth in order to provide a better understanding of the present application. However, one skilled in the art will readily recognize that some of the features may be omitted, or replaced by other elements, materials, or methods in different situations. In some instances, related operations of the present application have not been shown or described in the specification in order to avoid obscuring the core portions of the present application, and may be unnecessary to persons skilled in the art from a detailed description of the related operations, which may be presented in the description and general knowledge of one skilled in the art.
Furthermore, the described features, operations, or characteristics of the description may be combined in any suitable manner in various embodiments. Also, various steps or acts in the method descriptions may be interchanged or modified in a manner apparent to those of ordinary skill in the art. Thus, the various orders in the description and drawings are for clarity of description of only certain embodiments, and are not meant to be required orders unless otherwise indicated.
According to the method, the neighborhood aggregation mechanism of the graph neural network is considered to acquire additional information from the neighbor nodes to supplement node self information, and the neighborhood aggregation network is designed to fuse network structure and node attribute information by utilizing attribute smoothness and label smoothness, so that the learned node representation has richer information and stronger differentiation.
According to the application, node classification task experiments are carried out on a plurality of different types of network data sets, and the node representation method is verified to be capable of improving the utilization degree of network structure and attribute information, adapting to the characteristics of different networks, keeping good performance and improving the stability of a model.
In the embodiment of the application, the representation of the nodes in the graph network is thinned to two layers of attribute interaction state feature dimension and neighborhood aggregation state feature dimension for modeling so as to acquire more node information and improve the accuracy of the node information representation.
Referring to fig. 1 and 2, the present application provides a graph network node representation method, which includes:
step 101: respectively obtaining attribute interaction expression vectors and neighborhood aggregation expression vectors of target nodes in a graph network;
and 102, splicing the attribute interaction representation vector and the neighborhood aggregation representation vector to obtain a final representation vector of the target node in the graph network.
Given graph network g= (V, E, X) in this embodiment, where V represents the set of all nodes of the network, E represents the set of edges in the network, and the number of nodes and edges in the network are |v| and |e|, respectively.
The application uses X= [ X ] 1 ,x 2 ,...,x |V| ]The E|V|x F is expressed as an attribute matrix of all nodes, and the dimension of the node attribute is F. Study node v i Attribute x of (2) i Mapping to an embedding vector set S by an embedding layer i =[x i1 v 1 ,x i2 v 2 ,...,x ij v j ]Wherein x is ij Representing node v i Is the j-th dimensional attribute value of (c).
The topology information of the graph is then represented by the adjacency matrix a= {0,1} |V|×|V| And (3) representing. If node v i And node v ij There is a connecting edge e between ij E, then a ij =1; conversely, a ij =0. Node v i N for neighbor set of (2) i Representation, N i ={v j |e ij ∈E}。
Wherein in step 101, obtaining the attribute interaction expression vector of the target node in the graph network includes:
(1) In the embedded layer, the application maps each dimension feature of the high-dimension sparse attribute of the graph network to a low-dimension dense vector through embedded table lookup. Because of the sparseness of the graph network attributes, most elements in the graph network are 0. In order to improve the calculation efficiency, the application only calculates the embedded vector corresponding to the non-zero attribute, namely S i =[x ij v j |x ij ≠0]Wherein S is i < F. The target node may be any node in the graph network.
(2) The goal of the interaction layer is to capture all attribute interaction relations of the nodes, the application uses a factor decomposition machine, adopts a second-order interaction pooling operation to calculate second-order attribute interaction in a dot product mode, and compresses an embedded vector set into a single vector through the following formula (1);
wherein +.,representing attribute->And->Second order interactions between. By summation pooling, a k-dimensional vector can be output that encodes the second order interactions between all attribute embeddings. The k-dimensional vector encodes the second order interactions between all attribute embeddings. The second-order interactive pooling operation does not introduce any additional parameters, and at the same time, the computational complexity can be reduced to linear complexity through mathematical transformation, the k-dimensional vector is weighted by the following formula (2), and the weight formula is obtained by the following formula (3).
(3) In order to better clarify interaction among different features, the method introduces a attention-aware pooling layer in an attribute interaction network, and performs weighted summation on each attribute interaction through a formula (2):
wherein the method comprises the steps ofIs node v i Attribute interaction item->Weight coefficient of>Representing node attribute vector dot product. For learning the weight parameter->The application introduces a multi-layer perceptron to build an attention network. The input of the multi-layer perceptron is two attribute embedded vectors v i And v j :
Wherein W is f ∈d f ×k,b j ∈d f ,r f ∈d f Weight parameters for model training, d f Indicating the size of the hidden layer of the attention network. To learn weight parametersOne conventional idea is to solve it directly as a parameter to be learned by gradient descent, but this approach has a significant drawback: for attribute interaction items which never appear in the training set, the weight coefficient of the attribute interaction items cannot be learned, and the generalization capability is lacked. Thus, to solve the generalization problem, a Multi-Layer Perceptron (MLP) is introduced herein to construct an attention network (Attention Network), input as two attribute-embedded vectors v i And v j 。
(4) In order to learn the higher-order interaction relation between the attributes, the application pools the second-order interaction vector t i Inputting a full connection layer, stacking a plurality of full connection layers, and taking the output of the last layer as an attribute interaction expression vector f of the node i . The definition of the full connection layer in this embodiment is as follows:
wherein the method comprises the steps ofThe first layer weight matrix and the offset vector are respectively, sigma (·) is a nonlinear activation function, and L' represents the output of the last layer after stacking a plurality of fully connected layers. After stacking multiple fully connected layers, the output of the last layer can be regarded as the target node v i Attribute interaction representation vector f i 。
Further, the following focuses on the acquisition of the neighborhood aggregation expression vector in the research layer of the neighborhood aggregation module.
(1) The present embodiment uses a graph convolutional neural network to learn a mapping function f (X, a) that takes as input the adjacency matrix and the attribute matrix. The propagation formula between layers of the graph roll-up neural network is as follows:
wherein the method comprises the steps ofI is an identity matrix; a is an adjacent matrix, which itself represents the link relation between each node and the adjacent node, and the matrix is obtained after I is added>The graph rolling operation object comprises information of own nodes and adjacent nodes; />Is the degree matrix of the network,/>The element value on the diagonal is perThe degree of each vertex, the remaining elements are 0. To prevent gradient extinction and gradient explosion, a symmetric normalization operation was introduced: />So that the sum of each row of the matrix is 1; w (W) (l) Is a trainable layer-1 weight matrix; h (l) Is an implicit state of the first layer, for input layer H (()) =x; σ (·) represents a nonlinear activation function, e.g., RELU (·) =max (0, ·). By stacking multiple graph convolutional neural network layers, the relationship between a node and neighboring nodes that are multi-hop away can be captured.
(2) Based on a two-layer graph convolution neural network, the embodiment performs symmetrical normalization on the adjacency matrix according to the following formula (6) in a data preprocessing stage.
(3) Then the processed adjacency matrixAnd the attribute matrix X is input into a two-layer graph roll-up neural network. After two layers of graph convolution, obtaining an hidden layer matrix H= [ H ] coded by a graph convolution network 1 ,h 2 ,...h |V| ]. Wherein the forward propagation model is as shown in the following formula (7):
wherein W is (0) E F x d is a first layer weight matrix for mapping the node attribute to hidden state, W (1) ∈d 0 ×d 1 Is the weight matrix of the second hidden layer.
(4) The application provides a smoothness-aware graph annotation meaning network, which adopts a weighted sum as an aggregation function, adopts a splicing mode as a combination function, and learns weights among nodes by using a self-attention mechanism, so that the performance of different graph networks and node representations is improved.
The core of the network is a graph attention layer, and the input node vector passes through the graph attention layer to generate a new node vector. Specifically, at layer l, the target node v i The input of (2) isThe new target node vector of its output is +.>Its neighbor node v j Input of +.>The new neighbor node vector output by it is +.>Specifically, the vector is obtained by linearly transforming the input vector by the following formula (8).
Wherein the method comprises the steps ofAnd->For the weight matrix to be trained, d k Is vector quantityAnd->Is a dimension of (c). />Is composed of->And->The nodes with more different attributes between adjacent nodes are obtained by subtraction calculation after linear transformation, namely, the nodes with more different attributes contain more information and higher weight is obtained.
Wherein, in step 102, obtaining a neighborhood aggregate representation vector of a target node in the graph network includes:
the target node v is calculated by the following formula (9) i With neighbor node v j Is related to (a)
Wherein,
wherein,the target node vector passes through the attention layer of the graph and then outputs a new target node vector, wherein +_>Is a new neighbor node vector output after the neighbor node vector passes through the attention layer of the graph.
The application introduces the masked attention into the graph structure and distributes the attention to the target node v i Adjacent neighbor node set N i And pair all and v by softmax function i Adjacent node v j Normalized to obtain the attention coefficient
The present embodiment utilizes a trainable weight matrixAnd splicing operation, namely weighting neighbor nodes by adopting attention weight coefficients, and providing an innovative graph neural network neighborhood aggregation mechanism to learn and obtain a target node v i Is a neighborhood aggregate representation vector h i ;
Wherein,represents the layer I output of the neural network, +.>Also representing the weight matrix, ">And->And the weight matrixes are all represented, and are randomly initialized, and the most suitable weight matrix is automatically learned by continuous iteration of training, so that the model is optimal, and an optimal result is obtained. This is an important part of neural network training. Wherein (1)>And->The outputs of the (l-1) th layers of the i-th and j-th neural network nodes, respectively.
Equation (11) is a graph neural network neighborhood aggregation mechanism provided by the application. In contrast to prior art models, in the aggregation operation, the graph nervesNetwork (GNN)Graph convolutional neural network (GCN)>Andno attention mechanism is used. Whereas the existing graphical attention network (GAT)/(GAT)>Unlike the attention mechanism used by the proposed method of the present application: the attention mechanism used by the application utilizes the network smoothness information, thereby optimizing the introduction of neighbor information; secondly, GAT uses an additive attention model, while the dot product model used by the present application is more beneficial to learning interactions between attributes. The relationship between the current node and its neighborhood nodes is learned through an aggregation mechanism.
Finally, the present application represents the attribute interaction vector f i And neighborhood aggregate representation vector h i Splicing to obtain final representation vector of graph network node
Wherein the attribute interactions represent vectors: taking the medical field as an example, one of complications of cerebral apoplexy is pulmonary infection, and the attribute coincidence degree between cerebral apoplexy (central node) and pulmonary infection (neighbor node) is very low, so that the introduction of the node information of pulmonary infection is beneficial to better learning the clinical characterization of cerebral apoplexy. f (f) i Vector information representing stroke and pulmonary infection is presented herein. Neighborhood aggregation represents a vector: still taking the medical field as an example, the label of cerebral stroke is "disease", and a patient in the neighborhood with a label of "politics" suffers from this disease. The information of "politics" is not helpful in understanding the disease and may introduce noise-like information about major events in the relevant country. In this example, h i Only stroke vectors, and not the political figure vectors.The attribute interaction representation vector and the overall vector representation of the neighborhood aggregate representation vector are considered together. It should be noted that the illustration is given here only by way of example in the medical field, and does not represent that the method of the application can only be applied in the medical field, and that the method of the application is applied in the medical field only by way of representation of network nodes, so that it has a richer information for providing reference information to a doctor, and is not directly used for diagnosis and treatment of diseases.
In the combination operation, the GCN and the GAT combine the aggregation operation and the combination operation into a whole, and aggregate the neighbor node set and the self node in a weighted aggregation mode to finally obtain the node representation. The graph SAGE and the method of the application combine the representation of the upper layer of the node with the representation of the neighbor node in a splicing mode, so that the updated final representation vector of the node is obtained, and the operation can better keep the information of the node and the relation between the input of the current neural network and the output of the upper layer of the neural network. The obtained graph network node of the embodiment is used for tasks such as downstream task link prediction, node classification and the like. Compared with the existing method, the method has the advantages that:
1) Node information of the graph network data is better characterized;
2) A new domain aggregation mechanism is proposed;
3) A joint training method for mashup tag smoothness and attribute smoothness is provided.
Furthermore, the application provides a training method of the model provided by the embodiment, and the method introduces the attribute smoothness and the label smoothness of the graph network, so that the isomorphic effect of the neighborhood network is utilized to more fully explore the interaction relationship between the heterogeneous topological structure and the node attribute information in the clinical attribute graph network.
Further, the method of the application further comprises: calculating the attribute smoothness lambda of the current graph network according to the final representation vector a According to the attributeSmoothness lambda a And judging attribute similarity between the target node and the neighbor nodes thereof. The method specifically comprises the following steps:
definition 1. Aggregation operation of the graph network should capture neighbor node information dissimilar to the attribute of the central node as much as possible so as to promote the introduced information quantity and enhance the representation capability of the nodes.
Given a clinical attribute map network G and its attribute matrix X, the present application defines the attribute smoothness as follows:
Z i representing attribute interaction representation vectors, z j Representing a neighborhood aggregate representation vector, in general, the neighbor node attributes and the attributes of the center node follow the same distribution. Lambda (lambda) a The larger the attribute signal, the higher the frequency of the attribute signal, the more dissimilar the attribute between the nodes and the neighbor nodes in the graph network, and the larger the amount of information introduced.
Further, the method of the present application further comprises calculating the label smoothness lambda of the graph network node l According to the label smoothness lambda l And judging the validity of the neighbor information. The method specifically comprises the following steps:
and 2, retaining the information of the nodes in a neighborhood aggregation mechanism, and preventing excessive interference information from being introduced due to overlarge smoothness of the labels.
The label smoothness designed by the application is used for analyzing the validity of neighbor information. In the node classification task, each node v i The label corresponding to E V is y i . In the neighborhood aggregation mechanism, attribute input of a neighbor node can be decomposed into the following two parts according to labels:
wherein I (·) represents an indication function,neighbor node representing the same label as the current node, +.>Representing neighbor nodes of different labels than the current node. The application is summarized as follows: the neighbor nodes of the same label provide useful information, while the neighbor nodes of different labels provide interference information. Thus, given a network of attributes, the present application obtains its label smoothness by the following equation (14):
wherein lambda is l The smaller the label the more similar the neighbor node is to the central node, the more valid information is obtained from the neighbor node.
Further, the method of the application further comprises: according to the following formula (15), the smoothness lambda is obtained according to the attribute a And label smoothness lambda l Calculating the noise S of all nodes in the graph network;
wherein sigma 2 Representing the variance.
Specifically, the exploration thought of the noise S in the application is as follows: it is assumed that the noise of all nodes in the graph network obeys the same distribution, and that the noise distribution and the signal distribution are independent of each other. If the noise is random, the energy is defined as the variance σ 2 Node v i The first layer neighborhood attribute input isThe noise energy of the neighborhood aggregation attribute input is +.>
Given a given(true signal)>(noise) and c i (node v) i Information of itself), since noise is random noise, the average value is 0, then:
expected formula from varianceWeighted aggregated inputThe noise of (2) is:
further, considering the tag smoothness and the node smoothness, the noise can be expressed as:
further, in terms of the loss function, for the node classification task, the cross entropy of all labeled nodes is used as the loss function, see equation (16).
Wherein C is the number of tags to be attached,to mark the number of the node, Y lc =softmax(z l ) Representing the input of a vector for a nodePredictive probability vectors after entering the softmax classifier. The weight update is performed by minimizing the above loss function.
The foregoing description of the application has been presented for purposes of illustration and description, and is not intended to be limiting. Several simple deductions, modifications or substitutions may also be made by a person skilled in the art to which the application pertains, based on the idea of the application.
Claims (5)
1. A graph network node representation method, comprising:
acquiring complications of cerebral apoplexy, wherein the complications are pulmonary infection, the cerebral apoplexy is taken as a central node, and the pulmonary infection is taken as a neighbor node to determine attribute interaction expression vectors of the cerebral apoplexy and the pulmonary infection;
respectively obtaining attribute interaction expression vectors of cerebral stroke and pulmonary infection of target nodes in a graph network and neighborhood aggregation expression vectors of cerebral stroke;
for graph networksThe method comprises the steps of carrying out a first treatment on the surface of the V represents the set of all nodes of the network, E represents the set of edges in the network, the number of nodes and edges in the network are +.>And->;
The attribute matrix of all nodes is:the method comprises the steps of carrying out a first treatment on the surface of the F represents the dimension of node attributes, +.>Representing the attributes of the nodes;
target node is obtained through embedded table lookupObtaining all embedded vectors corresponding to all non-zero attributes to form an embedded vector set +.>;/>Representing the target node +.>Is>A dimension attribute value;
computing second order attribute interactions in a dot product manner using a second order interaction pooling operation, compressing the set of embedded vectors into a single vector:
;
Wherein,representing attribute->And->A second order interaction relationship between the two;
inputting the single vector into the full connection layer, stacking a plurality of full connection layers, and taking the output of the last layer as a target nodeAttribute interaction representation vector +.>;
The full-connection layer is as follows:
;
wherein,is->Layer weight matrix,/->Is->Layer offset vector->As a function of the non-linear activation,representing the output of the last layer after stacking a plurality of fully connected layers, < >>Indicate->The value of a single vector of layers,indicate->The value of a single vector of layers,/>Representing a single vector of layer 0Value of->Values representing individual vectors of the last layer after a plurality of fully connected layers; computing target node->And neighbor node->Correlation of->;
;
Wherein, ELU represents the activation function,a new target node vector which is output after the target node passes through the attention layer of the graph is represented, T represents transposition,/->Representing a new neighbor node vector output by a neighbor node through a graph attention layer;
all nodes with target node through softmax functionAdjacent node->Correlation of->Normalization is performed to obtain the attention weighting coefficient +.>:
;
Weighting all neighbor nodes of the target node by adopting the attention weight coefficient to obtain the target nodeNeighborhood aggregate representation vector +.>,/>Representation and->A set of neighboring neighbor nodes; r represents->Any node in the network;
;
wherein,represents the->Layer output->Representing a weight matrix, +.>Representation and->Neighboring neighbor nodesGather (S)>And->The (th) of the (th) and (th) neural network nodes respectively>Layer output;
and splicing the attribute interaction expression vector and the neighborhood aggregation expression vector to obtain a final expression vector of the target node in the graph network, so as to display the final expression vector to a doctor.
2. The graph network node representation method of claim 1, wherein the graph network node representation method further comprises: calculating the attribute smoothness of the current graph network according to the final representation vectorSmoothness according to the attribute->Judging attribute similarity between the target node and the neighbor nodes;
wherein,。
3. the graph network node representation method of claim 2, further comprising: calculating label smoothness for graph network nodesAccording to the tag smoothness +.>Judging the validity of neighbor information;
wherein,representing the target node +.>Corresponding tag->Representing the target node +.>Adjacent node->A corresponding tag.
4. A graph network node representation method in accordance with claim 3, further comprising: smoothness according to the attributeAnd tag smoothness +.>Calculating noise for all nodes in a graph networkS;
;
Wherein,representing the variance.
5. The graph network node representation method of claim 4, further comprising: for node classification tasks, calculating a loss function L by the following formula, and updating the weight by minimizing the loss function L;
;
wherein, C is the number of labels,for marking the number of the node>The predicted probability vector after the input softmax classifier of the vector is represented for the node.
Priority Applications (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN202011416368.8A CN114610950B (en) | 2020-12-04 | 2020-12-04 | Graph network node representation method |
Applications Claiming Priority (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN202011416368.8A CN114610950B (en) | 2020-12-04 | 2020-12-04 | Graph network node representation method |
Publications (2)
Publication Number | Publication Date |
---|---|
CN114610950A CN114610950A (en) | 2022-06-10 |
CN114610950B true CN114610950B (en) | 2023-11-07 |
Family
ID=81856169
Family Applications (1)
Application Number | Title | Priority Date | Filing Date |
---|---|---|---|
CN202011416368.8A Active CN114610950B (en) | 2020-12-04 | 2020-12-04 | Graph network node representation method |
Country Status (1)
Country | Link |
---|---|
CN (1) | CN114610950B (en) |
Citations (11)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN104850613A (en) * | 2015-05-14 | 2015-08-19 | 苗琦超 | Multi-form map marks-based social aggregating map interaction technology |
CN105808931A (en) * | 2016-03-03 | 2016-07-27 | 北京大学深圳研究生院 | Knowledge graph based acupuncture and moxibustion decision support method and apparatus |
CN108039206A (en) * | 2017-12-21 | 2018-05-15 | 北京大学深圳研究生院 | A kind of diagnosis and treatment method and system of infectious disease |
WO2019183191A1 (en) * | 2018-03-22 | 2019-09-26 | Michael Bronstein | Method of news evaluation in social media networks |
CN110575178A (en) * | 2019-09-10 | 2019-12-17 | 贾英 | Diagnosis and monitoring integrated medical system for judging motion state and judging method thereof |
CN110600109A (en) * | 2019-09-10 | 2019-12-20 | 秦学玉 | Diagnosis and monitoring comprehensive medical system with color image fusion and fusion method thereof |
CN110826700A (en) * | 2019-11-13 | 2020-02-21 | 中国科学技术大学 | Method for realizing and classifying bilinear graph neural network model for modeling neighbor interaction |
CN110991483A (en) * | 2019-11-01 | 2020-04-10 | 北京邮电大学 | High-order neighborhood mixed network representation learning method and device |
CN111339415A (en) * | 2020-02-25 | 2020-06-26 | 中国科学技术大学 | Click rate prediction method and device based on multi-interactive attention network |
CN111582443A (en) * | 2020-04-22 | 2020-08-25 | 成都信息工程大学 | Recommendation method based on Mask mechanism and level attention mechanism |
CN111797321A (en) * | 2020-07-07 | 2020-10-20 | 山东大学 | Personalized knowledge recommendation method and system for different scenes |
Family Cites Families (9)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
US10691743B2 (en) * | 2014-08-05 | 2020-06-23 | Sri International | Multi-dimensional realization of visual content of an image collection |
US10416999B2 (en) * | 2016-12-30 | 2019-09-17 | Intel Corporation | Processors, methods, and systems with a configurable spatial accelerator |
US11797838B2 (en) * | 2018-03-13 | 2023-10-24 | Pinterest, Inc. | Efficient convolutional network for recommender systems |
US10628708B2 (en) * | 2018-05-18 | 2020-04-21 | Adobe Inc. | Utilizing a deep neural network-based model to identify visually similar digital images based on user-selected visual attributes |
US11526765B2 (en) * | 2019-01-10 | 2022-12-13 | Arizona Board Of Regents On Behalf Of Arizona State University | Systems and methods for a supra-fusion graph attention model for multi-layered embeddings and deep learning applications |
CN112085124B (en) * | 2020-09-27 | 2022-08-09 | 西安交通大学 | Complex network node classification method based on graph attention network |
CN113035350B (en) * | 2021-03-29 | 2022-09-27 | 北京大学深圳研究生院 | Medical decision support method and system based on knowledge extraction |
CN114691936A (en) * | 2022-03-23 | 2022-07-01 | 支付宝(杭州)信息技术有限公司 | Prediction method and device for first attribute of first node in graph network |
CN115221413B (en) * | 2022-08-03 | 2023-04-14 | 湖北工业大学 | Sequence recommendation method and system based on interactive graph attention network |
-
2020
- 2020-12-04 CN CN202011416368.8A patent/CN114610950B/en active Active
Patent Citations (11)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN104850613A (en) * | 2015-05-14 | 2015-08-19 | 苗琦超 | Multi-form map marks-based social aggregating map interaction technology |
CN105808931A (en) * | 2016-03-03 | 2016-07-27 | 北京大学深圳研究生院 | Knowledge graph based acupuncture and moxibustion decision support method and apparatus |
CN108039206A (en) * | 2017-12-21 | 2018-05-15 | 北京大学深圳研究生院 | A kind of diagnosis and treatment method and system of infectious disease |
WO2019183191A1 (en) * | 2018-03-22 | 2019-09-26 | Michael Bronstein | Method of news evaluation in social media networks |
CN110575178A (en) * | 2019-09-10 | 2019-12-17 | 贾英 | Diagnosis and monitoring integrated medical system for judging motion state and judging method thereof |
CN110600109A (en) * | 2019-09-10 | 2019-12-20 | 秦学玉 | Diagnosis and monitoring comprehensive medical system with color image fusion and fusion method thereof |
CN110991483A (en) * | 2019-11-01 | 2020-04-10 | 北京邮电大学 | High-order neighborhood mixed network representation learning method and device |
CN110826700A (en) * | 2019-11-13 | 2020-02-21 | 中国科学技术大学 | Method for realizing and classifying bilinear graph neural network model for modeling neighbor interaction |
CN111339415A (en) * | 2020-02-25 | 2020-06-26 | 中国科学技术大学 | Click rate prediction method and device based on multi-interactive attention network |
CN111582443A (en) * | 2020-04-22 | 2020-08-25 | 成都信息工程大学 | Recommendation method based on Mask mechanism and level attention mechanism |
CN111797321A (en) * | 2020-07-07 | 2020-10-20 | 山东大学 | Personalized knowledge recommendation method and system for different scenes |
Non-Patent Citations (4)
Title |
---|
Kai Lei1,3 • Jin Zhang1 • Yuexiang Xie1 • Desi Wen1 • Daoyuan Chen1 • Min Yang2 • Ying Shen1.Path-based reasoning with constrained type attention for knowledge graph completion.Neural Computing and Applications.2019,全文. * |
KGDDS: A System for Drug-Drug Similarity Measure in Therapeutic Substitution based on Knowledge Graph Curation;Ying Shen;Journal of Medical Systems;全文 * |
Knowledge-aware Textual Entailment with Graph Attention Network;Daoyuan Chen1, Yaliang Li2, Min Yang3, Hai-Tao Zheng4, Ying Shen1;The 28th ACM International Conference on Information and Knowledge Management (CIKM ’19);全文 * |
基于知识图谱嵌入与多神经网络的序列推荐算法;沈冬东;汪海涛;姜瑛;陈星;;计算机工程与科学(第09期);1-5 * |
Also Published As
Publication number | Publication date |
---|---|
CN114610950A (en) | 2022-06-10 |
Similar Documents
Publication | Publication Date | Title |
---|---|---|
CN112966127B (en) | Cross-modal retrieval method based on multilayer semantic alignment | |
Fu et al. | HesGCN: Hessian graph convolutional networks for semi-supervised classification | |
CN111310707B (en) | Bone-based graph annotation meaning network action recognition method and system | |
CN113157957A (en) | Attribute graph document clustering method based on graph convolution neural network | |
CN112765370B (en) | Entity alignment method and device of knowledge graph, computer equipment and storage medium | |
Cai et al. | A robust interclass and intraclass loss function for deep learning based tongue segmentation | |
Zhang et al. | Dual-constrained deep semi-supervised coupled factorization network with enriched prior | |
CN113688253B (en) | Hierarchical perception temporal knowledge graph representation learning method | |
CN115841607A (en) | Brain network structure and similarity joint learning method based on graph attention network | |
CN112069825B (en) | Entity relation joint extraction method for alert condition record data | |
Golzari Oskouei et al. | EDCWRN: efficient deep clustering with the weight of representations and the help of neighbors | |
Shu et al. | A SVM multi-class image classification method based on DE and KNN in smart city management | |
Du et al. | Multi-feature based network for multivariate time series classification | |
CN115761275A (en) | Unsupervised community discovery method and system based on graph neural network | |
Li et al. | Few-shot image classification via contrastive self-supervised learning | |
CN117036760A (en) | Multi-view clustering model implementation method based on graph comparison learning | |
CN116469561A (en) | Breast cancer survival prediction method based on deep learning | |
CN116883723A (en) | Combined zero sample image classification method based on parallel semantic embedding | |
CN115983341A (en) | Node classification method based on relation aggregation hypergraph | |
CN114780767A (en) | Large-scale image retrieval method and system based on deep convolutional neural network | |
CN114093507A (en) | Skin disease intelligent classification method based on contrast learning in edge computing network | |
CN117349494A (en) | Graph classification method, system, medium and equipment for space graph convolution neural network | |
CN114610950B (en) | Graph network node representation method | |
CN112905599B (en) | Distributed deep hash retrieval method based on end-to-end | |
CN112861882B (en) | Image-text matching method and system based on frequency self-adaption |
Legal Events
Date | Code | Title | Description |
---|---|---|---|
PB01 | Publication | ||
PB01 | Publication | ||
SE01 | Entry into force of request for substantive examination | ||
SE01 | Entry into force of request for substantive examination | ||
GR01 | Patent grant | ||
GR01 | Patent grant |