CN114821119B - Method and device for training graph neural network model aiming at graph data invariant features - Google Patents

Method and device for training graph neural network model aiming at graph data invariant features Download PDF

Info

Publication number
CN114821119B
CN114821119B CN202210714507.8A CN202210714507A CN114821119B CN 114821119 B CN114821119 B CN 114821119B CN 202210714507 A CN202210714507 A CN 202210714507A CN 114821119 B CN114821119 B CN 114821119B
Authority
CN
China
Prior art keywords
invariant
vector
environmental noise
intrinsic feature
graph
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
Application number
CN202210714507.8A
Other languages
Chinese (zh)
Other versions
CN114821119A (en
Inventor
王翔
李思杭
何向南
张岸
Current Assignee (The listed assignees may be inaccurate. Google has not performed a legal analysis and makes no representation or warranty as to the accuracy of the list.)
University of Science and Technology of China USTC
Original Assignee
University of Science and Technology of China USTC
Priority date (The priority date 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 date listed.)
Filing date
Publication date
Application filed by University of Science and Technology of China USTC filed Critical University of Science and Technology of China USTC
Priority to CN202210714507.8A priority Critical patent/CN114821119B/en
Publication of CN114821119A publication Critical patent/CN114821119A/en
Application granted granted Critical
Publication of CN114821119B publication Critical patent/CN114821119B/en
Active legal-status Critical Current
Anticipated expiration legal-status Critical

Links

Images

Classifications

    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06F—ELECTRIC DIGITAL DATA PROCESSING
    • G06F18/00—Pattern recognition
    • G06F18/20—Analysing
    • G06F18/21—Design or setup of recognition systems or techniques; Extraction of features in feature space; Blind source separation
    • G06F18/214—Generating training patterns; Bootstrap methods, e.g. bagging or boosting
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06F—ELECTRIC DIGITAL DATA PROCESSING
    • G06F18/00—Pattern recognition
    • G06F18/20—Analysing
    • G06F18/21—Design or setup of recognition systems or techniques; Extraction of features in feature space; Blind source separation
    • G06F18/217—Validation; Performance evaluation; Active pattern learning techniques
    • G06F18/2193—Validation; Performance evaluation; Active pattern learning techniques based on specific statistical tests
    • 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

Landscapes

  • Engineering & Computer Science (AREA)
  • Theoretical Computer Science (AREA)
  • Physics & Mathematics (AREA)
  • Data Mining & Analysis (AREA)
  • Evolutionary Computation (AREA)
  • Life Sciences & Earth Sciences (AREA)
  • Artificial Intelligence (AREA)
  • General Physics & Mathematics (AREA)
  • General Engineering & Computer Science (AREA)
  • Bioinformatics & Cheminformatics (AREA)
  • Molecular Biology (AREA)
  • Computer Vision & Pattern Recognition (AREA)
  • Bioinformatics & Computational Biology (AREA)
  • Software Systems (AREA)
  • Health & Medical Sciences (AREA)
  • Biomedical Technology (AREA)
  • Biophysics (AREA)
  • Computational Linguistics (AREA)
  • General Health & Medical Sciences (AREA)
  • Evolutionary Biology (AREA)
  • Computing Systems (AREA)
  • Mathematical Physics (AREA)
  • Probability & Statistics with Applications (AREA)
  • Image Analysis (AREA)

Abstract

The invention discloses a method and a device for training a graph neural network model aiming at invariant features of graph data. The method comprises the following steps: performing feature extraction on the original image by using an interpretation generator to obtain probability distribution of invariant intrinsic feature subgraphs and probability distribution of environmental noise subgraphs; obtaining an invariant intrinsic characteristic enhancement view and an environmental noise enhancement view according to a preset sampling proportion; processing by using a backbone graph neural network encoder to obtain a first invariant intrinsic feature graph vector, a second invariant intrinsic feature graph vector and an environmental noise graph vector; obtaining a first invariant intrinsic feature hidden space vector, a second invariant intrinsic feature hidden space vector and an environmental noise hidden space vector by using a multi-layer perceptron projector; optimizing the backbone graph neural network encoder using a loss value; and (4) iterating to perform feature extraction operation, sampling operation, processing operation and optimization operation until the loss value is converged to a preset condition, so as to obtain the trained backbone graph neural network encoder.

Description

Method and device for training graph neural network model aiming at graph data invariant features
Technical Field
The invention relates to the field of image processing and artificial intelligence, in particular to a method and a device for training a graph neural network model aiming at invariant features of graph data, electronic equipment and a storage medium.
Background
Although deep learning has made great progress in many research fields in recent years, it is still a data-driven method, and in practical application, there are often challenges that the acquisition of labeled training samples is difficult and limited in number. In the field of graph neural networks, graph data enhancement technology and contrast learning technology in the prior art also face the problems of high training cost, poor model generalization capability and the like.
Disclosure of Invention
In view of the above problems, the present invention provides a method and an apparatus for training a graph neural network model for graph data invariant features, an electronic device, and a storage medium, so as to solve at least one of the above problems.
According to a first aspect of the present invention, there is provided a training method of a graph neural network model for graph data invariant features, comprising:
performing feature extraction on an original image by using an interpretation Generator (RG) to obtain probability distribution of invariant intrinsic feature subgraphs and probability distribution of environment noise subgraphs;
sampling the probability distribution of the invariant intrinsic feature subgraph according to a preset sampling proportion to obtain a first invariant intrinsic feature enhanced view and a second invariant intrinsic feature enhanced view, and sampling the probability distribution of the environmental noise subgraph according to the preset sampling proportion to obtain an environmental noise enhanced view;
processing the first invariant intrinsic feature enhanced view, the second invariant intrinsic feature enhanced view and the environmental noise enhanced view by using a backbone diagram neural network encoder to obtain a first invariant intrinsic feature diagram vector, a second invariant intrinsic feature diagram vector and an environmental noise diagram vector;
processing the first invariant intrinsic feature map vector, the second invariant intrinsic feature map vector and the ambient noise map vector by using a Multi-Layer Perceptron (MLP) Projection Head (PH) to obtain a first invariant intrinsic feature latent space vector, a second invariant intrinsic feature latent space vector and an ambient noise latent space vector;
inputting the first invariant intrinsic characteristic hidden space vector, the second invariant intrinsic characteristic hidden space vector and the environmental noise hidden space vector into a loss function to obtain a loss value, and optimizing a backbone graph neural network encoder according to the loss value;
and (4) iterating to perform feature extraction operation, sampling operation, processing operation and optimization operation until the loss value is converged to a preset condition, so as to obtain the trained backbone graph neural network encoder.
According to an embodiment of the invention, the interpretation generator comprises a neural network encoder and a multi-layered perceptron.
According to the embodiment of the invention, the extracting the feature of the original image by using the interpretation generator to obtain the probability distribution of the invariant intrinsic feature subgraph and the probability distribution of the environmental noise subgraph comprises the following steps:
processing the original image by using the graph neural network encoder to obtain vector expressions of all nodes of the original image;
calculating the vector expression of each node in the original image by using a multilayer perception computer to obtain the importance score of each node in the original image;
normalizing the importance score of each node in the original graph to obtain the sampling probability of each node in the original graph;
and obtaining the probability distribution of the unchanged intrinsic characteristic subgraph and the probability distribution of the environmental noise subgraph according to the sampling probability of each node in the original graph.
According to the embodiment of the invention, the probability distribution of the invariant intrinsic feature subgraph is determined by formula (1):
Figure 962820DEST_PATH_IMAGE001
（1），
wherein the ambient noise sub-graph probability distribution is determined by equation (2):
Figure 434122DEST_PATH_IMAGE002
（2），
wherein,
Figure 315490DEST_PATH_IMAGE003
a sub-graph of the invariant intrinsic features is represented,
Figure 812330DEST_PATH_IMAGE004
a sub-graph of the environmental noise is represented,
Figure 712022DEST_PATH_IMAGE005
the nodes are represented as a list of nodes,
Figure 370537DEST_PATH_IMAGE006
representing the artwork
Figure 739201DEST_PATH_IMAGE007
The set of nodes of (a) is,
Figure 774153DEST_PATH_IMAGE008
a set of nodes representing an invariant intrinsic feature subgraph,
Figure 793931DEST_PATH_IMAGE009
a set of nodes representing an ambient noise sub-graph,
Figure 623346DEST_PATH_IMAGE010
representing original drawings
Figure 213728DEST_PATH_IMAGE007
Node (a) of
Figure 52371DEST_PATH_IMAGE005
The probability of (c).
According to an embodiment of the present invention, the processing the first invariant intrinsic feature enhancement view, the second invariant intrinsic feature enhancement view, and the environmental noise enhancement view by using the backbone neural network encoder to obtain the first invariant intrinsic feature map vector, the second invariant intrinsic feature map vector, and the environmental noise map vector includes:
respectively extracting vector expressions of each node in a first invariant intrinsic characteristic enhanced view, a second invariant intrinsic characteristic enhanced view and an environmental noise enhanced view by using a backbone diagram neural network encoder;
performing point multiplication on the vector expression of each node in the first invariant intrinsic characteristic enhancement view and the probability distribution of the invariant intrinsic characteristic subgraph to obtain a first point multiplication result;
performing point multiplication on the vector expression of each node in the second invariant intrinsic characteristic enhancement view and the probability distribution of the invariant intrinsic characteristic subgraph to obtain a second point multiplication result;
carrying out point multiplication on the vector expression of each node in the environmental noise enhancement view and the probability distribution of the environmental noise subgraph to obtain an environmental noise point multiplication result;
and performing pooling operation on the first point multiplication result, the second point multiplication result and the environmental noise point multiplication result respectively to obtain a first invariant intrinsic feature map vector, a second invariant intrinsic feature map vector and an environmental noise map vector.
According to an embodiment of the present invention, the above-mentioned loss function includes a sufficiency loss function and an independence loss function;
wherein the loss function is determined by equation (3):
Figure 684513DEST_PATH_IMAGE011
（3），
wherein,
Figure 950409DEST_PATH_IMAGE012
is a hyper-parameter that balances the adequacy loss function and the independence loss function,
Figure 824824DEST_PATH_IMAGE013
an interpretation generator is represented that generates a set of interpretation data,
Figure 467158DEST_PATH_IMAGE014
a neural network encoder of a backbone diagram is shown,
Figure 946681DEST_PATH_IMAGE015
a multi-layer perceptron projector head is shown,
Figure 367167DEST_PATH_IMAGE016
a function representing the loss of adequacy is expressed,
Figure 932141DEST_PATH_IMAGE017
the loss of independence function is expressed as,
Figure 112586DEST_PATH_IMAGE018
the image of the original image is represented,
Figure 243353DEST_PATH_IMAGE019
a collection of the original pictures is represented,
Figure 851052DEST_PATH_IMAGE020
representing a mathematical expectation.
According to an embodiment of the present invention, the above-mentioned sufficiency loss function is determined by equation (4):
Figure 152589DEST_PATH_IMAGE021
（4），
wherein the independence loss function is determined by equation (5):
Figure 871147DEST_PATH_IMAGE022
（5），
wherein,
Figure 59683DEST_PATH_IMAGE023
representing a transpose of the first invariant intrinsic feature hidden space vector,
Figure 838283DEST_PATH_IMAGE024
representing a second invariant intrinsic feature hidden space vector,
Figure 627116DEST_PATH_IMAGE025
representing an ambient noise-hidden spatial vector,
Figure 946102DEST_PATH_IMAGE026
which is indicative of a temperature over-parameter,
Figure 254723DEST_PATH_IMAGE027
representing a set of ambient noise subgraphs generated from each raw graph data in the same set of graph training data,
Figure 204225DEST_PATH_IMAGE028
is to divide the training data of the same group of graphs
Figure 699928DEST_PATH_IMAGE029
Obtaining hidden space vectors of enhanced views of other samples through a backbone diagram neural network encoder and a multi-layer perceptron projector head
Figure 275135DEST_PATH_IMAGE030
A set is formed.
According to a second aspect of the present invention, there is provided a training apparatus for a graph neural network model for graph data invariant features, comprising:
the feature extraction module is used for extracting features of the original image by using the interpretation generator to obtain probability distribution of invariant intrinsic feature subgraphs and probability distribution of environmental noise subgraphs;
the sampling module is used for sampling the probability distribution of the invariant intrinsic characteristic subgraph according to a preset sampling proportion to obtain a first invariant intrinsic characteristic enhanced view and a second invariant intrinsic characteristic enhanced view, and sampling the probability distribution of the environmental noise subgraph according to the preset sampling proportion to obtain an environmental noise enhanced view;
the first processing module is used for processing the first invariant intrinsic feature enhancement view, the second invariant intrinsic feature enhancement view and the environmental noise enhancement view by using the backbone map neural network encoder to obtain a first invariant intrinsic feature map vector, a second invariant intrinsic feature map vector and an environmental noise map vector;
the second processing module is used for processing the first invariant intrinsic feature map vector, the second invariant intrinsic feature map vector and the environmental noise map vector by using the multi-layer perceptron projector head to obtain a first invariant intrinsic feature hidden space vector, a second invariant intrinsic feature hidden space vector and an environmental noise hidden space vector;
the optimization module is used for inputting the first invariant intrinsic characteristic hidden space vector, the second invariant intrinsic characteristic hidden space vector and the environmental noise hidden space vector into a loss function to obtain a loss value and optimizing the backbone graph neural network encoder according to the loss value;
and the iterative training module is used for iteratively performing feature extraction operation, sampling operation, processing operation and optimization operation until the loss value is converged to a preset condition to obtain the trained backbone graph neural network encoder.
According to a third aspect of the present invention, there is provided an electronic apparatus comprising:
one or more processors;
a storage device for storing one or more programs,
wherein the one or more programs, when executed by the one or more processors, cause the one or more processors to perform a method of training a graph neural network model for graph data invariant features.
According to a fourth aspect of the present invention, there is provided a computer readable storage medium having stored thereon executable instructions that, when executed by a processor, cause the processor to perform a method of training a graph neural network model for graph data invariant features.
According to the method and the device for training the graph neural network model aiming at the invariant features of the graph data, the nodes containing important semantic information in the graph data are captured by the interpretation generator, the ability of the graph neural network model for interpreting the input data is given, and the performance of a backbone graph neural network encoder in the graph neural network model is improved, so that the training cost of the model is reduced, the generalization ability of the model is improved, and the application scene of the trained model is expanded.
Drawings
FIG. 1 is a flow diagram of a method of training a graph neural network model for graph data invariant features, according to an embodiment of the present invention;
FIG. 2 is a flow chart for obtaining probability distribution of an invariant intrinsic feature sub-graph and probability distribution of an environmental noise sub-graph according to an embodiment of the present invention;
FIG. 3 is a flow chart of obtaining a first invariant intrinsic feature map vector, a second invariant intrinsic feature map vector, and an ambient noise map vector according to an embodiment of the present invention;
FIG. 4 is a diagram of a comparative learning framework for graph data invariant features, according to an embodiment of the present invention;
FIG. 5 is a block diagram of a training apparatus for a graph neural network model for graph data invariant features, in accordance with an embodiment of the present invention;
fig. 6 is a diagram of MNIST visualization according to an embodiment of the present invention;
FIG. 7 is a diagram illustrating a MUTAG visualization according to an embodiment of the present invention;
FIG. 8 schematically illustrates a block diagram of an electronic device suitable for implementing a method of training a graph neural network model for graph data invariant features in accordance with an embodiment of the present invention.
Detailed Description
In order that the objects, technical solutions and advantages of the present invention will become more apparent, the present invention will be further described in detail with reference to the accompanying drawings in conjunction with the following specific embodiments.
At present, a method based on self-supervision contrast learning utilizes a large amount of non-labeled data to perform pre-training, and then utilizes a small amount of labeled data to perform model parameter fine tuning, so that the performance of a model in a downstream task is remarkably improved, and the method is widely applied to the field of graph neural networks.
In general, the graph contrast learning framework can be divided into two parts: the image data enhancement module creates a plurality of enhanced views of original image data through various image data enhancement methods; and secondly, the contrast learning module enables the projections of the multiple enhanced views expressed by the same sample on the hypersphere to be as close as possible, and meanwhile forces the projections of the enhanced views of different samples to be dispersed as much as possible. Obviously, the graph data enhancement module plays an important role in characterizing the essential features of the graph data.
Current graph data enhancement techniques can be broadly divided into two categories: one is a method based on a random mechanism. The graph data often contains abundant structural information, and random sampling of nodes, edges or attributes of the graph data can cause changes of semantic information of the graph data, such as discarding cyano groups of cyanide molecules to change the properties of the molecules from being virulent to non-toxic, thereby misleading a subsequent contrast learning module. The second is a method based on domain knowledge. The most important substructures of the graph data are identified through the guidance of expert knowledge, and the substructures are reserved as much as possible when data enhancement is carried out, so that the essential characteristics and semantic information of the graph data are maintained. Expert knowledge is often expensive, difficult to obtain even in some scenarios, and models that have been pre-trained using domain-specific expert knowledge are difficult to generalize to new domains. Therefore, domain knowledge based methods are still very limited in application.
One of the purposes of the invention is to explore the influence of a data enhancement module on the performance of a pre-training model in graph contrast learning, and design an automatic graph data enhancement method which does not depend on expert knowledge and simultaneously retains the essential characteristics and semantic information of original graph data, thereby improving the performance of the existing graph contrast learning framework.
In the interpretable domain of research, the invariant intrinsic properties (or invariant intrinsic features) of graph data training samples are generally defined as: some of the input features that determine their distinction from other samples. According to its definition, graph data
Figure 438263DEST_PATH_IMAGE018
Invariant intrinsic characteristic sub-graph (or invariant intrinsic characteristic sub-graph)
Figure 355404DEST_PATH_IMAGE031
Two conditions should be met: one is adequacy condition, i.e. conditions of sufficiency
Figure 869561DEST_PATH_IMAGE031
Retaining graph data
Figure 733612DEST_PATH_IMAGE018
Neutralizing possible predicted results
Figure 734935DEST_PATH_IMAGE032
Relevant semantic information:
Figure 291818DEST_PATH_IMAGE033
represents a given input
Figure 27693DEST_PATH_IMAGE034
Conditions of (2)A probability density function; the second is independence condition, namely an invariant intrinsic characteristic subgraph
Figure 695435DEST_PATH_IMAGE031
Data on the graph
Figure 816844DEST_PATH_IMAGE018
Complement of
Figure 544628DEST_PATH_IMAGE035
Is defined as the ambient noise that is,
Figure 767799DEST_PATH_IMAGE035
should be associated with possible predicted outcomes
Figure 973653DEST_PATH_IMAGE032
Independent of each other:
Figure 215147DEST_PATH_IMAGE036
，
Figure 113833DEST_PATH_IMAGE037
indicating probability independence. Sufficiency and independence conditions guarantee
Figure 355458DEST_PATH_IMAGE031
Capturing image data
Figure 365003DEST_PATH_IMAGE018
All features that are distinguishable from the rest of the sample, while preventing them from capturing only a very small number of background features
Figure 477315DEST_PATH_IMAGE031
And (4) degrading.
FIG. 1 is a flow chart of a method of training a graph neural network model for graph data invariant features, according to an embodiment of the present invention.
As shown in FIG. 1, the method includes operations S110 to S160.
In operation S110, feature extraction is performed on the original image by using the interpretation generator to obtain invariant intrinsic feature sub-image probability distribution and environmental noise sub-image probability distribution.
The interpretation Generator (RG) is actually an intrinsic feature extraction neural network, and is intended to extract features of an original image. The probability distribution of the unchanged intrinsic characteristic subgraph and the probability distribution of the environmental noise subgraph are complementary sets relative to the original graph.
In operation S120, the invariant intrinsic feature sub-graph probability distribution is sampled according to a preset sampling ratio to obtain a first invariant intrinsic feature enhanced view and a second invariant intrinsic feature enhanced view, and the environmental noise sub-graph probability distribution is sampled according to the preset sampling ratio to obtain an environmental noise enhanced view.
The first invariant intrinsic feature enhanced view, the second invariant intrinsic feature enhanced view and the environmental noise enhanced view meet the requirements of sufficiency and independence.
In operation S130, the first invariant intrinsic feature enhanced view, the second invariant intrinsic feature enhanced view, and the ambient noise enhanced view are processed by using the backbone map neural network encoder to obtain a first invariant intrinsic feature map vector, a second invariant intrinsic feature map vector, and an ambient noise map vector.
In operation S140, the first invariant intrinsic feature map vector, the second invariant intrinsic feature map vector, and the ambient noise map vector are processed by a Multi-Layer Perceptron (MLP) Projection Head (PH) to obtain a first invariant intrinsic feature hidden space vector, a second invariant intrinsic feature hidden space vector, and an ambient noise hidden space vector.
In operation S150, the first invariant intrinsic feature hidden space vector, the second invariant intrinsic feature hidden space vector, and the environmental noise hidden space vector are input into a loss function to obtain a loss value, and the backbone neural network encoder is optimized according to the loss value.
In operation S160, the feature extraction operation, the sampling operation, the processing operation, and the optimization operation are performed iteratively until the loss value converges to the preset condition, so as to obtain the trained backbone graph neural network encoder.
The preset conditions include, but are not limited to: the loss value converges to a fixed value, the loss value oscillates in a certain interval or the descending amplitude of the loss value reaches the expected effect.
According to the method and the device for training the graph neural network model aiming at the invariant features of the graph data, the nodes containing important semantic information in the graph data are captured by the interpretation generator, the ability of the graph neural network model for interpreting the input data is given, and the performance of a backbone graph neural network encoder in the graph neural network model is improved, so that the training cost of the model is reduced, the generalization ability of the model is improved, and the application scene of the trained model is expanded.
According to an embodiment of the invention, the interpretation generator comprises a neural network encoder and a multi-layer perceptron.
FIG. 2 is a flow chart for obtaining probability distribution of an invariant intrinsic feature sub-graph and probability distribution of an environmental noise sub-graph according to an embodiment of the present invention.
As shown in fig. 2, the feature extraction of the original image by using the interpretation generator to obtain the probability distribution of the invariant intrinsic feature subgraph and the probability distribution of the environmental noise subgraph includes operations S210 to S240.
In operation S210, the original image is processed by the neural network encoder to obtain vector representations of all nodes of the original image.
In operation S220, a vector representation of each node in the original image is calculated using the multi-layer perceptron, and an importance score of each node in the original image is obtained.
In operation S230, the importance score of each node in the original image is normalized to obtain a sampling probability of each node in the original image.
In operation S240, according to the sampling probability of each node in the original graph, a probability distribution of the invariant intrinsic feature subgraph and a probability distribution of the environmental noise subgraph are obtained.
According to the embodiment of the invention, the probability distribution of the invariant intrinsic feature subgraph is determined by formula (1):
Figure 530591DEST_PATH_IMAGE001
（1），
wherein the ambient noise sub-graph probability distribution is determined by equation (2):
Figure 462775DEST_PATH_IMAGE002
（2），
wherein,
Figure 10431DEST_PATH_IMAGE003
a sub-graph of the invariant intrinsic features is represented,
Figure 977250DEST_PATH_IMAGE004
a sub-graph of the ambient noise is represented,
Figure 748896DEST_PATH_IMAGE005
the nodes are represented as a list of nodes,
Figure 683223DEST_PATH_IMAGE006
representing the artwork
Figure 768991DEST_PATH_IMAGE007
The set of nodes of (a) is,
Figure 590316DEST_PATH_IMAGE008
a set of nodes representing an invariant intrinsic feature subgraph,
Figure 1706DEST_PATH_IMAGE009
a set of nodes representing an ambient noise sub-graph,
Figure 157750DEST_PATH_IMAGE010
representing original drawings
Figure 843946DEST_PATH_IMAGE007
Node (a) of
Figure 785357DEST_PATH_IMAGE005
The probability of (c).
Fig. 3 is a flowchart for obtaining a first invariant intrinsic feature map vector, a second invariant intrinsic feature map vector, and an ambient noise map vector according to an embodiment of the present invention.
As shown in fig. 3, the processing the first invariant intrinsic feature enhancement view, the second invariant intrinsic feature enhancement view and the environmental noise enhancement view by using the backbone neural network encoder to obtain the first invariant intrinsic feature map vector, the second invariant intrinsic feature map vector and the environmental noise map vector includes operations S310 to S350.
In operation S310, vector expressions of each node in the first invariant intrinsic feature enhanced view, the second invariant intrinsic feature enhanced view and the environmental noise enhanced view are respectively extracted by using a backbone graph neural network encoder.
In operation S320, the vector expression of each node in the first invariant intrinsic feature enhancement view is point-multiplied with the probability distribution of the invariant intrinsic feature subgraph to obtain a first point-multiplied result.
In operation S330, the vector expression of each node in the second invariant intrinsic feature enhanced view is point-multiplied with the probability distribution of the invariant intrinsic feature subgraph to obtain a second point-multiplied result.
In operation S340, the vector expression of each node in the environmental noise enhanced view is point-multiplied with the environmental noise subgraph probability distribution to obtain an environmental noise point-multiplied result.
In operation S350, pooling operations are performed on the first point multiplication result, the second point multiplication result, and the ambient noise point multiplication result, respectively, to obtain a first invariant intrinsic feature map vector, a second invariant intrinsic feature map vector, and an ambient noise map vector.
According to an embodiment of the present invention, the above-mentioned loss function includes a sufficiency loss function and an independence loss function;
wherein the loss function is determined by equation (3):
Figure 102069DEST_PATH_IMAGE011
（3），
wherein,
Figure 496141DEST_PATH_IMAGE012
is a hyper-parameter that balances the adequacy loss function and the independence loss function,
Figure 172979DEST_PATH_IMAGE013
an interpretation generator is represented that generates a set of interpretation data,
Figure 968897DEST_PATH_IMAGE014
a neural network encoder of a backbone diagram is shown,
Figure 722089DEST_PATH_IMAGE015
a multi-layer perceptron projector head is shown,
Figure 400195DEST_PATH_IMAGE016
a function representing the loss of adequacy is expressed,
Figure 631457DEST_PATH_IMAGE017
a function representing the loss of independence is expressed,
Figure 531148DEST_PATH_IMAGE018
the image of the original image is represented,
Figure 189663DEST_PATH_IMAGE019
a collection of the original pictures is represented,
Figure 558327DEST_PATH_IMAGE020
representing a mathematical expectation.
The loss function gives consideration to sufficiency and independence, and can better reflect important semantic information of the graph data.
According to an embodiment of the present invention, the above-mentioned sufficiency loss function is determined by equation (4):
Figure 593279DEST_PATH_IMAGE038
（4），
wherein the independence loss function is determined by equation (5):
Figure 347478DEST_PATH_IMAGE039
（5），
wherein,
Figure 239210DEST_PATH_IMAGE040
representing a transpose of the first invariant intrinsic feature hidden space vector,
Figure 829592DEST_PATH_IMAGE041
representing a second invariant intrinsic feature latent space vector,
Figure 668235DEST_PATH_IMAGE042
representing an ambient noise-hidden spatial vector,
Figure 27672DEST_PATH_IMAGE043
which is indicative of a temperature over-parameter,
Figure 542836DEST_PATH_IMAGE044
representing a set of ambient noise subgraphs generated from each raw graph data in the same set of graph training data,
Figure 620513DEST_PATH_IMAGE045
is to divide the training data of the same group of graphs
Figure 59585DEST_PATH_IMAGE046
Obtaining hidden space vectors of enhanced views of other samples through a backbone diagram neural network encoder and a multi-layer perceptron projector head
Figure 539108DEST_PATH_IMAGE047
The formed set.
The sufficiency loss function and the independence loss function are improved on the basis of the InfonCE function; wherein the sufficiency loss function is optimized
Figure 975905DEST_PATH_IMAGE048
The essential characteristic estimation network can extract the key information in the original image and capture the expression
Figure 790146DEST_PATH_IMAGE018
The essential characteristic important nodes ensure the sufficiency condition; optimizing independence loss function
Figure 705013DEST_PATH_IMAGE049
The essential characteristic estimation network is more stable, the influence of environmental noise nodes except the key information in the input data is ignored, and the independence condition is ensured.
FIG. 4 is a comparative learning framework diagram for graph data invariant features according to an embodiment of the present invention.
The above-described method provided by the present invention is described in further detail below with reference to fig. 4.
The new graph contrast learning framework shown in fig. 4 allows the graph data reinforcement part to generate a reinforced view that satisfies the sufficiency and independence conditions.
Firstly, the following components are mixed
Figure 304621DEST_PATH_IMAGE031
The conditional probability distribution of the original image is approximate to the condition that each node in the original image node set is sampled to enter
Figure 646741DEST_PATH_IMAGE031
The probability of (c), i.e. as shown in equation (1):
Figure 948278DEST_PATH_IMAGE050
（1），
Figure 666835DEST_PATH_IMAGE051
is an original drawing
Figure 917688DEST_PATH_IMAGE018
Of the plurality of nodes in the network,
Figure 696288DEST_PATH_IMAGE052
represents a node
Figure 235854DEST_PATH_IMAGE053
Is sampled into
Figure 7370DEST_PATH_IMAGE031
Is a probability of reflecting a node
Figure 50412DEST_PATH_IMAGE053
For expression
Figure 999914DEST_PATH_IMAGE018
Importance of the essential features. Similarly, will
Figure 761196DEST_PATH_IMAGE035
Is approximated to the form shown in equation (2):
Figure 883873DEST_PATH_IMAGE054
（2）。
by an interpretation generator
Figure 296269DEST_PATH_IMAGE055
Will make the above probability distribution function
Figure 416671DEST_PATH_IMAGE056
The parameters are parameterized,
Figure 930829DEST_PATH_IMAGE055
the method comprises two sub-modules, namely a graph neural network encoder and a multilayer perceptron, as shown in formula (6) to formula (8):
Figure 794880DEST_PATH_IMAGE057
（6），
Figure 61782DEST_PATH_IMAGE058
（7），
Figure 353086DEST_PATH_IMAGE059
（8），
Figure 88961DEST_PATH_IMAGE060
is a neural network encoder for encoding the original image
Figure 287861DEST_PATH_IMAGE018
As input, and output
Figure 425582DEST_PATH_IMAGE018
All node expressions of
Figure 402634DEST_PATH_IMAGE061
，
Figure 94646DEST_PATH_IMAGE062
Is a simple multi-layer perceptron based on nodes
Figure 566079DEST_PATH_IMAGE053
Expression of
Figure 558306DEST_PATH_IMAGE063
Derive its importance score, i.e. is
Figure 988150DEST_PATH_IMAGE052
Parameterization of (2). Based on
Figure 947885DEST_PATH_IMAGE052
A constant intrinsic feature subgraph can be obtained by probabilistically sampling a fixed proportion (e.g., 80%) of the nodes and preserving the interconnections between them
Figure 957429DEST_PATH_IMAGE031
And ambient noise figure
Figure 804162DEST_PATH_IMAGE035
。
Inputting the enhanced view obtained by sampling into a backbone diagram neural network encoder
Figure 608170DEST_PATH_IMAGE064
(i.e., the pre-trained target model) to obtain a graph-level representation thereof, as shown in equation (9):
Figure 55201DEST_PATH_IMAGE065
（9），
unlike the general framework, before the pooling operation, the probability vector obtained by the node expression and interpretation generator extracted by the backbone neural network is expressed and interpreted
Figure 602857DEST_PATH_IMAGE066
A dot product is performed so that subsequent gradients can be passed back to the interpretation generator. Then will be
Figure 835255DEST_PATH_IMAGE067
Inputting a multi-layer perceptron projector head
Figure 810164DEST_PATH_IMAGE068
And obtaining the final implicit space expression.
Finally, the implicit space expression is input into a loss function to train the model, and after the pre-training is completed, only the backbone graph neural network encoder is reserved
Figure 744491DEST_PATH_IMAGE064
And used as a feature encoder for downstream tasks. It should be noted that the pre-training framework is model independent, and an appropriate backbone neural network encoder may be selected according to task requirements
Figure 830259DEST_PATH_IMAGE064
。
FIG. 5 is a block diagram of a training apparatus for a graph neural network model for graph data invariant features, according to an embodiment of the present invention.
As shown in fig. 5, the training apparatus 500 for the graph neural network model for the graph data invariant features includes a feature extraction module 510, a sampling module 520, a first processing module 530, a second processing module 540, an optimization module 550, and an iterative training module 560.
And the feature extraction module 510 is configured to perform feature extraction on the original image by using the interpretation generator to obtain probability distribution of invariant intrinsic feature subgraph and probability distribution of environmental noise subgraph.
The sampling module 520 is configured to sample the probability distribution of the invariant intrinsic feature subgraph according to a preset sampling ratio to obtain a first invariant intrinsic feature enhanced view and a second invariant intrinsic feature enhanced view, and sample the probability distribution of the environmental noise subgraph according to the preset sampling ratio to obtain an environmental noise enhanced view.
The first processing module 530 is configured to process the first invariant intrinsic feature enhancement view, the second invariant intrinsic feature enhancement view, and the ambient noise enhancement view by using the backbone neural network encoder to obtain a first invariant intrinsic feature map vector, a second invariant intrinsic feature map vector, and an ambient noise map vector.
The second processing module 540 is configured to process the first invariant intrinsic feature map vector, the second invariant intrinsic feature map vector, and the environmental noise map vector by using the multi-layer perceptron projector to obtain a first invariant intrinsic feature hidden space vector, a second invariant intrinsic feature hidden space vector, and an environmental noise hidden space vector.
And an optimizing module 550, configured to input the first invariant intrinsic feature hidden space vector, the second invariant intrinsic feature hidden space vector, and the environmental noise hidden space vector into a loss function, obtain a loss value, and optimize the backbone graph neural network encoder according to the loss value.
And the iterative training module 560 is used for iteratively performing feature extraction operation, sampling operation, processing operation and optimization operation until the loss value is converged to a preset condition, so as to obtain the trained backbone graph neural network encoder.
The device provided by the invention can improve the generalization capability of the graph neural network model, expand the application scene of the device and greatly reduce the training cost because no expert marking is needed in the training process.
Fig. 6 is a diagram illustrating MNIST visualization according to an embodiment of the present invention.
Fig. 7 is a schematic diagram of a MUTAG visualization result according to an embodiment of the present invention.
To better assist those skilled in the art to understand the present invention, the improvement and advantages of the above method provided by the present invention over the prior art are further described in detail by the specific embodiments in conjunction with fig. 6-7 and table 1.
Compared with the existing Graph contrast Learning framework, the Graph neural network model general pre-training framework Invariant feature Graph contrast Learning (GCL-IR) aiming at the Graph data Invariant feature has the following advantages: 1) In the absence of expert knowledge, the interpretation generator can accurately capture nodes containing important semantic information, and a certain ability of the model for interpreting input data is given. 2) The backbone model pre-trained under the GCL-IR framework performs significantly better in downstream tasks.
The advantages described above are demonstrated by exhaustive experiments on various data sets in various fields.
Table 1 shows data statistics of data sets used in the experiments
Figure 651584DEST_PATH_IMAGE069
(1) And estimating the importance of the node. To verify the ability of the interpretation generator to capture nodes containing important semantic information, without the constraints of expert knowledge, experiments were conducted on the MNIST-Superpixel dataset, first using the IGNORE of all labels in the training set and pre-training using the GCL-IR framework, and then visualizing the interpretation generator's estimate of the importance of a portion of the data nodes as shown in FIG. 6. The three lines from top to bottom are the raw MNIST picture data, the MNIST superpixel graph data and the evaluation of the interpretation generator on the importance of the nodes. The darker the color, the higher the score, and the more likely it is to be retained in the enhanced view. From experimental results, interpretation generators pre-trained using GCL-IR have the ability to accurately capture nodes containing important semantic information.
Verification was then also performed on the real world biochemical molecular dataset MUTAG. Firstly, model pre-training is carried out on a ZINC-2M data set containing 2,000,000 unlabeled biochemical molecules by using a GCL-IR framework, and then the estimation result of the node importance of a part of MUTAG data set is visualized as shown in FIG. 7. After six randomly selected molecular samples, two chemists marked groups with mutation properties to the chemical molecules, and the results are shown as yellow marks on the first row. The interpretation generator from the GCL-IR pre-training marks the equivalent number of nodes with the highest significance, and the results are shown in the second row with green marks. If the professional annotation is regarded as a true value, the accuracy of the GCL-IR interpretation generator reaches 83.3%. Experimental results demonstrate the ability of the GCL-IR pre-trained interpretation generator to accurately capture important nodes on a real biochemical molecular dataset.
(2) And the performance of the downstream tasks of the backbone network is improved. In the GCL-IR framework, a backbone graph neural network encoder is pre-trained by using a ZINC-2M biochemical molecular data set, and then supervised learning is carried out on 8 different downstream biochemical molecular data sets by using the pre-trained backbone network. The GCL-IR performance was compared fairly with the most advanced pre-training frameworks Attramking, contextPred, graphCL, graphLoG and AD-GCL at present using the same backbone neural network. The results of the experiment are shown in table 2. Compared with an untrained model, the GCL-IR remarkably improves the ROC-AUC index of the model in a downstream classification task. The GCL-IR also achieves the optimal effect compared with other pre-training methods.
Table 2: GCL-IR backbone network migration learning ROC-AUC index
Figure 797395DEST_PATH_IMAGE070
The GCL-IR is a general pre-training method of a graph neural network model aiming at the invariant characteristic of graph data and independent of a model, and can be applied to the graph neural network model under various self-supervision pre-training-fine tuning training paradigms. The interpretation generator of GCL-IR has the ability of accurately capturing important nodes, so that the interpretability of the model is increased, and meanwhile, the reinforced view of essential features retained in GCL-IR more accurately guides the subsequent backbone graph neural network comparison learning process, so that the feature expression ability of the GCL-IR in downstream tasks is obviously improved.
FIG. 8 schematically illustrates a block diagram of an electronic device suitable for implementing a method of training a graph neural network model for graph data invariant features in accordance with an embodiment of the present invention.
As shown in fig. 8, an electronic device 800 according to an embodiment of the present invention includes a processor 801 which can perform various appropriate actions and processes according to a program stored in a Read Only Memory (ROM) 802 or a program loaded from a storage section 808 into a Random Access Memory (RAM) 803. The processor 801 may include, for example, a general purpose microprocessor (e.g., CPU), an instruction set processor and/or associated chipset, and/or a special purpose microprocessor (e.g., application Specific Integrated Circuit (ASIC)), among others. The processor 801 may also include onboard memory for caching purposes. The processor 801 may comprise a single processing unit or a plurality of processing units for performing the different actions of the method flows according to embodiments of the present invention.
In the RAM 803, various programs and data necessary for the operation of the electronic apparatus 800 are stored. The processor 801, the ROM 802, and the RAM 803 are connected to each other by a bus 804. The processor 801 performs various operations of the method flow according to the embodiment of the present invention by executing programs in the ROM 802 and/or the RAM 803. Note that the programs may also be stored in one or more memories other than the ROM 802 and the RAM 803. The processor 801 may also perform various operations of method flows according to embodiments of the present invention by executing programs stored in the one or more memories.
Electronic device 800 may also include input/output (I/O) interface 805, input/output (I/O) interface 805 also connected to bus 804, according to an embodiment of the invention. Electronic device 800 may also include one or more of the following components connected to I/O interface 805: an input portion 806 including a keyboard, a mouse, and the like; an output section 807 including a signal such as a Cathode Ray Tube (CRT), a Liquid Crystal Display (LCD), and the like, and a speaker; a storage portion 808 including a hard disk and the like; and a communication section 809 including a network interface card such as a LAN card, a modem, or the like. The communication section 809 performs communication processing via a network such as the internet. A drive 810 is also connected to the I/O interface 805 as necessary. A removable medium 811 such as a magnetic disk, an optical disk, a magneto-optical disk, a semiconductor memory, or the like is mounted on the drive 810 as necessary, so that a computer program read out therefrom is mounted on the storage section 808 as necessary.
The present invention also provides a computer-readable storage medium, which may be contained in the apparatus/device/system described in the above embodiments; or may exist separately and not be assembled into the device/apparatus/system. The computer-readable storage medium carries one or more programs which, when executed, implement the method according to an embodiment of the present invention.
According to embodiments of the present invention, the computer readable storage medium may be a non-volatile computer readable storage medium, which may include, for example but is not limited to: a portable computer diskette, a hard disk, a Random Access Memory (RAM), a read-only memory (ROM), an erasable programmable read-only memory (EPROM or flash memory), a portable compact disc read-only memory (CD-ROM), an optical storage device, a magnetic storage device, or any suitable combination of the foregoing. In the present invention, a computer readable storage medium may be any tangible medium that can contain, or store a program for use by or in connection with an instruction execution system, apparatus, or device. For example, according to embodiments of the present invention, a computer-readable storage medium may include the ROM 802 and/or the RAM 803 described above and/or one or more memories other than the ROM 802 and the RAM 803.
The above embodiments are provided to further explain the objects, technical solutions and advantages of the present invention in detail, and it should be understood that the above embodiments are only examples of the present invention and are not intended to limit the present invention, and any modifications, equivalents, improvements and the like made within the spirit and principle of the present invention should be included in the protection scope of the present invention.

Claims (7)

1. A training method of a graph neural network model aiming at graph data invariant features comprises the following steps:
performing feature extraction on an original image by using an interpretation generator to obtain invariant intrinsic feature sub-image probability distribution and environmental noise sub-image probability distribution, wherein the interpretation generator comprises a graph neural network encoder and a multilayer perceptron;
sampling the probability distribution of the invariant intrinsic feature subgraph according to a preset sampling proportion to obtain a first invariant intrinsic feature enhanced view and a second invariant intrinsic feature enhanced view, and sampling the probability distribution of the environmental noise subgraph according to the preset sampling proportion to obtain an environmental noise enhanced view;
processing the first invariant intrinsic feature enhancement view, the second invariant intrinsic feature enhancement view and the environmental noise enhancement view by using a backbone map neural network encoder to obtain a first invariant intrinsic feature map vector, a second invariant intrinsic feature map vector and an environmental noise map vector;
processing the first invariant intrinsic feature map vector, the second invariant intrinsic feature map vector and the environmental noise map vector by using a multi-layer perceptron projector to obtain a first invariant intrinsic feature hidden space vector, a second invariant intrinsic feature hidden space vector and an environmental noise hidden space vector;
inputting the first invariant intrinsic feature hidden space vector, the second invariant intrinsic feature hidden space vector and the environmental noise hidden space vector into a loss function to obtain a loss value, and optimizing the backbone graph neural network encoder according to the loss value;
iteratively performing feature extraction operation, sampling operation, processing operation and optimization operation until the loss value is converged to a preset condition to obtain a trained backbone graph neural network encoder;
the method for extracting the features of the original image by using the interpretation generator to obtain the probability distribution of the invariant intrinsic feature subgraph and the probability distribution of the environmental noise subgraph comprises the following steps:
processing the original image by using the image neural network encoder to obtain vector expressions of all nodes of the original image;
calculating the vector expression of each node in the original image by using the multilayer perceptron to obtain the importance score of each node in the original image;
normalizing the importance score of each node in the original graph to obtain the sampling probability of each node in the original graph;
obtaining the probability distribution of the unchanged intrinsic characteristic subgraph and the probability distribution of the environmental noise subgraph according to the sampling probability of each node in the original graph;
wherein the processing the first invariant intrinsic feature enhanced view, the second invariant intrinsic feature enhanced view, and the environmental noise enhanced view by using the backbone graph neural network encoder to obtain a first invariant intrinsic feature map vector, a second invariant intrinsic feature map vector, and an environmental noise map vector includes:
respectively extracting vector expressions of each node in the first invariant intrinsic feature enhanced view, the second invariant intrinsic feature enhanced view and the environmental noise enhanced view by utilizing the backbone graph neural network encoder;
performing point multiplication on the vector expression of each node in the first invariant intrinsic characteristic enhancement view and the probability distribution of the invariant intrinsic characteristic subgraph to obtain a first point multiplication result;
performing point multiplication on the vector expression of each node in the second invariant intrinsic characteristic enhancement view and the probability distribution of the invariant intrinsic characteristic subgraph to obtain a second point multiplication result;
performing point multiplication on the vector expression of each node in the environmental noise enhancement view and the probability distribution of the environmental noise subgraph to obtain an environmental noise point multiplication result;
and performing pooling operation on the first point multiplication result, the second point multiplication result and the environmental noise point multiplication result respectively to obtain the first invariant intrinsic feature map vector, the second invariant intrinsic feature map vector and the environmental noise map vector.
2. The method of claim 1, wherein the invariant intrinsic feature subgraph probability distribution is determined by equation (1):
Figure DEST_PATH_IMAGE002
（1），
wherein the ambient noise sub-graph probability distribution is determined by equation (2):
Figure DEST_PATH_IMAGE004
（2），
wherein,
Figure DEST_PATH_IMAGE006
a sub-graph of the invariant intrinsic features is represented,
Figure DEST_PATH_IMAGE008
a sub-graph of the environmental noise is represented,
Figure DEST_PATH_IMAGE010
the nodes are represented as a list of nodes,
Figure DEST_PATH_IMAGE012
representing the artwork
Figure DEST_PATH_IMAGE014
The set of nodes of (a) is,
Figure DEST_PATH_IMAGE016
a set of nodes representing an invariant intrinsic feature subgraph,
Figure DEST_PATH_IMAGE018
a set of nodes representing an ambient noise sub-graph,
Figure DEST_PATH_IMAGE020
representing the artwork
Figure 812649DEST_PATH_IMAGE014
Node (a) of
Figure 729789DEST_PATH_IMAGE010
The probability of (c).
3. The method of claim 1, wherein the loss function comprises a sufficiency loss function and an independence loss function;
wherein the loss function is determined by equation (3):
Figure DEST_PATH_IMAGE022
（3），
wherein,
Figure DEST_PATH_IMAGE024
is a hyper-parameter that balances the adequacy loss function and the independence loss function,
Figure DEST_PATH_IMAGE026
the interpretation generator is represented and the interpretation generator is represented,
Figure DEST_PATH_IMAGE028
represents the backbone graph neural network encoder,
Figure DEST_PATH_IMAGE030
representing the multi-layer perceptron projection head,
Figure DEST_PATH_IMAGE032
a function representing the loss of adequacy function,
Figure DEST_PATH_IMAGE034
representing the function of the loss of independence,
Figure DEST_PATH_IMAGE036
the original image is represented by a representation of the original image,
Figure DEST_PATH_IMAGE038
the set of artwork is represented, representing a mathematical expectation.
4. The method of claim 3, wherein the sufficiency loss function is determined by equation (4):
Figure DEST_PATH_IMAGE040
（4），
wherein the independence loss function is determined by equation (5):
Figure DEST_PATH_IMAGE042
（5），
wherein,
Figure DEST_PATH_IMAGE044
representing a transpose of the first invariant intrinsic feature hidden space vector,
Figure DEST_PATH_IMAGE046
representing the second invariant intrinsic feature latent space vector,
Figure DEST_PATH_IMAGE048
representing the ambient noise-hidden spatial vector,
Figure DEST_PATH_IMAGE050
indicating the temperatureThe super-parameter is set to be,
Figure DEST_PATH_IMAGE052
representing a set of ambient noise subgraphs generated from each raw graph data in the same set of graph training data,
Figure DEST_PATH_IMAGE054
is to divide the training data of the same group of graphs
Figure DEST_PATH_IMAGE056
Obtaining hidden space vectors of enhanced views of other samples through a backbone diagram neural network encoder and a multi-layer perceptron projector head
Figure DEST_PATH_IMAGE058
The formed set.
5. A training apparatus for a graph neural network model for graph data invariant features, comprising:
the characteristic extraction module is used for extracting the characteristics of the original image by using the interpretation generator to obtain the probability distribution of the unchanged intrinsic characteristic subgraph and the probability distribution of the environmental noise subgraph, wherein the interpretation generator comprises a graph neural network encoder and a multilayer perceptron;
the sampling module is used for sampling the probability distribution of the invariant intrinsic characteristic subgraph according to a preset sampling proportion to obtain a first invariant intrinsic characteristic enhanced view and a second invariant intrinsic characteristic enhanced view, and sampling the probability distribution of the environmental noise subgraph according to the preset sampling proportion to obtain an environmental noise enhanced view;
a first processing module, configured to process the first invariant intrinsic feature enhancement view, the second invariant intrinsic feature enhancement view, and the ambient noise enhancement view by using a backbone graph neural network encoder to obtain a first invariant intrinsic feature map vector, a second invariant intrinsic feature map vector, and an ambient noise map vector;
the second processing module is used for processing the first invariant intrinsic feature map vector, the second invariant intrinsic feature map vector and the environmental noise map vector by using a multi-layer perceptron projector head to obtain a first invariant intrinsic feature hidden space vector, a second invariant intrinsic feature hidden space vector and an environmental noise hidden space vector;
the optimization module is used for inputting the first invariant intrinsic feature hidden space vector, the second invariant intrinsic feature hidden space vector and the environmental noise hidden space vector into a loss function to obtain a loss value, and optimizing the backbone graph neural network encoder according to the loss value;
an iterative training module for iteratively performing feature extraction operation, sampling operation, processing operation and optimization operation until the loss value converges to a preset condition to obtain a trained backbone graph neural network encoder
The method for extracting the features of the original image by using the interpretation generator to obtain the probability distribution of the invariant intrinsic feature subgraph and the probability distribution of the environmental noise subgraph comprises the following steps:
processing the original image by using the image neural network encoder to obtain vector expressions of all nodes of the original image;
calculating the vector expression of each node in the original image by using the multilayer perceptron to obtain the importance score of each node in the original image;
normalizing the importance score of each node in the original graph to obtain the sampling probability of each node in the original graph;
obtaining the probability distribution of the unchanged intrinsic characteristic subgraph and the probability distribution of the environmental noise subgraph according to the sampling probability of each node in the original graph;
wherein the processing the first invariant intrinsic feature enhanced view, the second invariant intrinsic feature enhanced view, and the environmental noise enhanced view by using the backbone graph neural network encoder to obtain a first invariant intrinsic feature map vector, a second invariant intrinsic feature map vector, and an environmental noise map vector includes:
respectively extracting vector expressions of each node in the first invariant intrinsic feature enhanced view, the second invariant intrinsic feature enhanced view and the environmental noise enhanced view by utilizing the backbone graph neural network encoder;
performing point multiplication on the vector expression of each node in the first invariant intrinsic characteristic enhancement view and the probability distribution of the invariant intrinsic characteristic subgraph to obtain a first point multiplication result;
performing point multiplication on the vector expression of each node in the second invariant intrinsic characteristic enhancement view and the probability distribution of the invariant intrinsic characteristic subgraph to obtain a second point multiplication result;
performing point multiplication on the vector expression of each node in the environmental noise enhancement view and the probability distribution of the environmental noise subgraph to obtain an environmental noise point multiplication result;
and performing pooling operation on the first point multiplication result, the second point multiplication result and the environmental noise point multiplication result respectively to obtain the first invariant intrinsic feature map vector, the second invariant intrinsic feature map vector and the environmental noise map vector.
6. An electronic device, comprising:
one or more processors;
a storage device for storing one or more programs,
wherein the one or more programs, when executed by the one or more processors, cause the one or more processors to perform the method of any of claims 1~4.
7. A computer readable storage medium having stored thereon executable instructions that, when executed by a processor, cause the processor to perform the method of any of claims 1~4.
CN202210714507.8A 2022-06-22 2022-06-22 Method and device for training graph neural network model aiming at graph data invariant features Active CN114821119B (en)

Priority Applications (1)

Application Number Priority Date Filing Date Title
CN202210714507.8A CN114821119B (en) 2022-06-22 2022-06-22 Method and device for training graph neural network model aiming at graph data invariant features

Applications Claiming Priority (1)

Application Number Priority Date Filing Date Title
CN202210714507.8A CN114821119B (en) 2022-06-22 2022-06-22 Method and device for training graph neural network model aiming at graph data invariant features

Publications (2)

Publication Number Publication Date
CN114821119A CN114821119A (en) 2022-07-29
CN114821119B true CN114821119B (en) 2022-10-28

Family

ID=82521832

Family Applications (1)

Application Number Title Priority Date Filing Date
CN202210714507.8A Active CN114821119B (en) 2022-06-22 2022-06-22 Method and device for training graph neural network model aiming at graph data invariant features

Country Status (1)

Country Link
CN (1) CN114821119B (en)

Family Cites Families (6)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN111428874B (en) * 2020-02-29 2024-07-12 平安科技（深圳）有限公司 Wind control method, electronic device and computer readable storage medium
US11354778B2 (en) * 2020-04-13 2022-06-07 Google Llc Systems and methods for contrastive learning of visual representations
CN111930992B (en) * 2020-08-14 2022-10-28 腾讯科技（深圳）有限公司 Neural network training method and device and electronic equipment
CN112862093B (en) * 2021-01-29 2024-01-12 北京邮电大学 Graphic neural network training method and device
CN113314188B (en) * 2021-06-16 2022-07-15 中国科学技术大学 Graph structure enhanced small sample learning method, system, equipment and storage medium
CN114020928A (en) * 2021-11-02 2022-02-08 上海交通大学 False news identification method based on heterogeneous graph comparison learning

Also Published As

Publication number Publication date
CN114821119A (en) 2022-07-29

Similar Documents

Publication Publication Date Title
Liang et al. Explaining the black-box model: A survey of local interpretation methods for deep neural networks
Eslami et al. Attend, infer, repeat: Fast scene understanding with generative models
Samek et al. Evaluating the visualization of what a deep neural network has learned
Wang et al. Rethinking the learning paradigm for dynamic facial expression recognition
CN107784288A (en) A kind of iteration positioning formula method for detecting human face based on deep neural network
CN115861462B (en) Training method and device for image generation model, electronic equipment and storage medium
Lin et al. Saliency detection via multi-scale global cues
US20210056353A1 (en) Joint representation learning from images and text
CN115661246A (en) Attitude estimation method based on self-supervision learning
CN115147607A (en) Anti-noise zero-sample image classification method based on convex optimization theory
CN113743474A (en) Digital picture classification method and system based on cooperative semi-supervised convolutional neural network
CN117253044A (en) Farmland remote sensing image segmentation method based on semi-supervised interactive learning
CN113807425B (en) Tissue pathology image classification method based on self-adaptive regular depth clustering
CN117690098A (en) Multi-label identification method based on dynamic graph convolution under open driving scene
CN114821119B (en) Method and device for training graph neural network model aiming at graph data invariant features
CN111950582A (en) Determining a perturbation mask for a classification model
CN116597275A (en) High-speed moving target recognition method based on data enhancement
CN115810109A (en) Tongue image feature extraction method, system and storage medium
CN111797732B (en) Video motion identification anti-attack method insensitive to sampling
CN114170485A (en) Deep learning interpretable method and apparatus, storage medium, and program product
CN113920291A (en) Error correction method and device based on picture recognition result, electronic equipment and medium
CN112861601A (en) Method for generating confrontation sample and related equipment
CN112446345A (en) Low-quality three-dimensional face recognition method, system, equipment and storage medium
Zhang et al. A modified image processing method for deblurring based on GAN networks
CN114998990B (en) Method and device for identifying safety behaviors of personnel on construction site

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