CN113792753A - Dynamic hypergraph neural network classification method and system - Google Patents
Dynamic hypergraph neural network classification method and system Download PDFInfo
- Publication number
- CN113792753A CN113792753A CN202110921299.4A CN202110921299A CN113792753A CN 113792753 A CN113792753 A CN 113792753A CN 202110921299 A CN202110921299 A CN 202110921299A CN 113792753 A CN113792753 A CN 113792753A
- Authority
- CN
- China
- Prior art keywords
- hypergraph
- node
- dynamic
- static
- data
- 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.)
- Pending
Links
Images
Classifications
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F18/00—Pattern recognition
- G06F18/20—Analysing
- G06F18/24—Classification techniques
- G06F18/241—Classification techniques relating to the classification model, e.g. parametric or non-parametric approaches
- G06F18/2413—Classification techniques relating to the classification model, e.g. parametric or non-parametric approaches based on distances to training or reference patterns
- G06F18/24147—Distances to closest patterns, e.g. nearest neighbour classification
-
- 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
-
- 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)
- Life Sciences & Earth Sciences (AREA)
- Artificial Intelligence (AREA)
- General Physics & Mathematics (AREA)
- General Engineering & Computer Science (AREA)
- Evolutionary Computation (AREA)
- Biophysics (AREA)
- Computational Linguistics (AREA)
- Software Systems (AREA)
- Mathematical Physics (AREA)
- Health & Medical Sciences (AREA)
- Biomedical Technology (AREA)
- Computing Systems (AREA)
- Molecular Biology (AREA)
- General Health & Medical Sciences (AREA)
- Evolutionary Biology (AREA)
- Bioinformatics & Cheminformatics (AREA)
- Bioinformatics & Computational Biology (AREA)
- Computer Vision & Pattern Recognition (AREA)
- Information Retrieval, Db Structures And Fs Structures Therefor (AREA)
Abstract
The application provides a dynamic hypergraph neural network classification method, a device and a medium, wherein the method comprises the following steps: acquiring original data to be classified, wherein the original data comprises known label data and unknown label data, and constructing a static hypergraph structure according to known association between the original data; converting the embedded characteristics of the original data nodes into corresponding initial dynamic overcarriages through a preset algorithm for mutual conversion of the node characteristics and the associated structures; selecting a plurality of known label data to form a training set of the dynamic hypergraph neural network, and performing iterative training on the dynamic hypergraph neural network; and constructing a static association hypergraph of the unknown label data and the known label data according to the original node characteristics of the unknown label data, and extracting high-order characteristics of the nodes of the unknown label data through a trained dynamic hypergraph neural network to obtain a classification result. The method models node semantic information of different depths by combining a static hypergraph structure and a dynamic hypergraph structure, and has better classification performance.
Description
Technical Field
The application relates to the technical field of complex network classification, in particular to a dynamic hypergraph neural network classification method and system.
Background
At present, complex network analysis has wide application in real life, such as: brain disease diagnosis, recommendation systems, social network recommendations, and the like. The development of the graph/hypergraph neural network also brings great improvement to the performance of the existing complex network classification problem. Graph structures are widely used in the modeling of various complex networks due to their simplicity and ease of operation. And the hypergraph structure gradually becomes a new growing point in the field of complex network analysis due to the inherent high-order correlation modeling capability of the hypergraph structure.
In the related technology, the traditional hypergraph learning algorithm is slow in modeling and reasoning of a complex network and is difficult to be used in actual large-scale network reasoning, and the dynamic hypergraph learning algorithm is difficult to capture the dynamic change of characteristic parameters in the model learning process. Therefore, the main challenges for the complex network classification method are:
(1) it is difficult to model high-order complex association data.
(2) Modeling has a poor ability to capture dynamic changes in features.
(3) The learned model parameters cannot adapt to tasks of different feature spaces.
(4) The classification performance of the network nodes is crossed, and under the condition of sufficient known information, the direction of the optimal parameter gradient is difficult to find, so that the overfitting of the model parameters and the performance are reduced.
Disclosure of Invention
The present application is directed to solving, at least to some extent, one of the technical problems in the related art.
To this end, the first objective of the present application is to propose a dynamic hypergraph neural network classification method, which uses the known correlation between the original data and uses the first-order neighbor transformation and other ways to construct static hypergraph correlation as the initial hypergraph structure information. Then designing a function for interconversion of the node characteristics and the association structure, converting the embedded characteristics of the nodes into corresponding hypergraph associations in real time by using the function, and modeling the original data from another angle. And then constructing a dynamic hypergraph neural network, combining trainable parameters with different node characteristics and an associated structure conversion algorithm to transform the characteristics, and designing node convolution, an attention layer and hypergraph convolution to embed static and dynamic hypergraph associated structures into high-order characteristics of nodes in real time, so that association among data is fully mined, and more accurate node description is generated. And finally, constructing a static association hypergraph of the node data and the original characteristics of the unknown label and the known nodes, and simultaneously extracting high-order characteristics of the unknown label nodes by using a trained dynamic hypergraph neural network so as to obtain a better node classification prediction result. The method and the device can more conveniently model high-order complex associated data; the capability of capturing the dynamic change of the characteristics is stronger; the learned model parameters can be adapted to different feature spaces; under the condition of sufficient known information, the direction of the optimal parameter gradient can be easily found, and the model parameter overfitting and the performance are enhanced.
A second objective of the present application is to provide a dynamic hypergraph neural network classification system;
a third object of the present application is to propose a non-transitory computer-readable storage medium.
To achieve the above object, a first embodiment of the present application is directed to a dynamic hypergraph neural network classification method, including the following steps:
acquiring original data to be classified, and constructing a static hypergraph structure according to known association among the original data, wherein the original data comprises known label data and unknown label data;
converting the embedded characteristics of the nodes of the original data into corresponding initial dynamic overcurrents through a preset algorithm for interconversion between the node characteristics and the associated structures;
constructing a dynamic hypergraph neural network, selecting a plurality of known label data to form a training set of the dynamic hypergraph neural network, and performing iterative training on the dynamic hypergraph neural network by combining the static hypergraph structure and the initial dynamic hypergraph edge;
and constructing a static association hypergraph of the unknown label data and the known label data according to the original node characteristics of the unknown label data, and extracting high-order characteristics of the nodes of the unknown label data through a trained dynamic hypergraph neural network to obtain a node classification prediction result.
Optionally, in one embodiment of the present applicationIn the above, constructing a static hypergraph structure according to the known association between the original data includes: constructing a simple graph structure G ' ═ V, E ' } through the known association between the original data, wherein V represents a set of nodes of the known label data and the unknown label data, and E ' represents a pair of associated edges abstracted through the known association between the original data; converting the simple graph structure G' into a static hypergraph structure G through first-order neighbor conversions＝{V,Es,WsIn which EsIs the set of the super-edges in the static hypergraph structure obtained by the transformation of E', WsRepresenting each of said super edges EsThe corresponding weight.
Optionally, in an embodiment of the present application, the converting the embedded features of the nodes of the original data into corresponding initial dynamic hyper-edges includes: acquiring an embedded feature set of the nodes of the original data; according to the embedded characteristics of each node, calculating characteristic vectors between any node in the node set of the original data and nodes except the any node; selecting k nodes closest to any node, and constructing an initial dynamic super edge E formed by connecting any node and the k nodes through the following formulai：
Ei＝knnk(Xi)
Wherein, knnk() Is a neighbor function, k is a positive integer, XiIs an embedded feature of any node.
Optionally, in an embodiment of the present application, iteratively training the dynamic hypergraph neural network includes: s1: constructing an initial static hypergraph incidence matrix, and collecting E for the hypergraphsAdding a corresponding data column in the initial static hypergraph incidence matrix according to the judgment result of each hyperedge, wherein each bit in the data column corresponds to each node in the training set, setting the value of the bit in the data column corresponding to the node connected by each hyperedge to be 1 according to the judgment result, and setting the value of the bit in the data column corresponding to the node not connected to be 0 to construct a static hyperedgeGraph correlation matrix Hs(ii) a S2: constructing an initial dynamic hypergraph structure according to original node characteristics through a dynamic hypergraph structure construction algorithm, converting the initial dynamic hypergraph structure into a corresponding initial dynamic hypergraph incidence matrix, and setting values in the initial dynamic hypergraph incidence matrix through the initial dynamic hypergraph edges; s3: fusing the static hypergraph incidence matrix and the initial dynamic hypergraph incidence matrix to construct an initial static and dynamic combined incidence matrix; s4: convolving the node characteristics of the original data through a preset characteristic conversion matrix to obtain modified first node characteristics, performing residual connection with the original node characteristics to obtain updated second node characteristics, and aggregating the second node characteristics into the super-edge characteristics through the transposition of the dynamic and static combined hypergraph association matrix; s5: performing super-edge convolution on the super-edge feature, and obtaining an updated third node feature by combining an attention mechanism S6: repeating steps S1-S5 for each node feature-tag pair in the training set to iteratively train network parameters until the dynamic hypergraph neural network converges.
Optionally, in an embodiment of the present application, the second node features are aggregated into a super-edge feature by the following formula:
wherein, Yi+1Is a super-edge feature that is characterized in that,
is the transpose of a hyper-map associative matrix of a dynamic and static union, Xi′Is a feature of the second node that,
is a diagonal matrix made up of the inverse of each node degree.
Optionally, in an embodiment of the present application, the updated third node characteristic is obtained by the following formula:
Xi+1＝atten(Yi+1)Yi+1
wherein, atten (x) x + mlp (x),
wherein, atten () is the embedded attention mechanism function, MLP () is the learnable full link layer for automatically learning the weight of each node feature, Xi+1Is a third node feature.
In order to achieve the above object, a second embodiment of the present application further provides a dynamic hypergraph neural network classification system, including the following modules:
the system comprises an acquisition module, a classification module and a classification module, wherein the acquisition module is used for acquiring original data to be classified and constructing a static hypergraph structure according to known association among the original data, and the original data comprises known label data and unknown label data;
the conversion module is used for converting the embedded characteristics of the nodes of the original data into corresponding initial dynamic super edges through a preset algorithm of mutual conversion between the node characteristics and the associated structures;
the construction module is used for constructing a dynamic hypergraph neural network, selecting a plurality of known label data to form a training set of the dynamic hypergraph neural network, and performing iterative training on the dynamic hypergraph neural network by combining the static hypergraph structure and the initial dynamic hypergraph edge;
and the extraction module is used for constructing a static association hypergraph of the unknown label data and the known label data according to the original node characteristics of the unknown label data, and extracting the high-order characteristics of the nodes of the unknown label data through a trained dynamic hypergraph neural network so as to obtain a node classification prediction result.
Optionally, in an embodiment of the present application, the first obtaining module is specifically configured to: constructing a simple graph structure G ' ═ V, E ' } through the known association between the original data, wherein V represents a set of nodes of the known label data and the unknown label data, and E ' represents a pair of associated edges abstracted through the known association between the original data; converting the simple graph structure G' into a static hypergraph structure G through first-order neighbor conversions＝{V,Es,WsAnd (c) the step of (c) in which,Esis the set of the super-edges in the static hypergraph structure obtained by the transformation of E', WsRepresenting each of said super edges EsThe corresponding weight.
Optionally, in an embodiment of the present application, the conversion module is specifically configured to: acquiring an embedded feature set of the nodes of the original data; according to the embedded characteristics of each node, calculating characteristic vectors between any node in the node set of the original data and nodes except the any node; selecting k nodes closest to any node, and constructing an initial dynamic super edge E formed by connecting any node and the k nodes through the following formulai：
Ei＝knnk(Xi)
Wherein, knnk() Is a neighbor function, k is a positive integer, XiIs an embedded feature of any node.
The technical scheme provided by the embodiment of the application at least has the following beneficial effects: the method utilizes the known correlation among original data, and uses a first-order neighbor conversion and other modes to construct static hypergraph correlation as initial hypergraph structure information. Then designing a function for interconversion of the node characteristics and the association structure, converting the embedded characteristics of the nodes into corresponding hypergraph associations in real time by using the function, and modeling the original data from another angle. And then constructing a dynamic hypergraph neural network, combining trainable parameters with different node characteristics and an associated structure conversion algorithm to transform the characteristics, and designing node convolution, an attention layer and hypergraph convolution to embed static and dynamic hypergraph associated structures into high-order characteristics of nodes in real time, so that association among data is fully mined, and more accurate node description is generated. And finally, constructing a static association hypergraph of the node data and the original characteristics of the unknown label and the known nodes, and simultaneously extracting high-order characteristics of the unknown label nodes by using a trained dynamic hypergraph neural network so as to obtain a better node classification prediction result. The method and the device can more conveniently model high-order complex associated data; the capability of capturing the dynamic change of the characteristics is stronger; the learned model parameters can be adapted to different feature spaces; under the condition of sufficient known information, the direction of the optimal parameter gradient can be easily found, and the model parameter overfitting and the performance are enhanced.
In order to implement the foregoing embodiments, the third aspect of the present application further provides a non-transitory computer-readable storage medium, on which a computer program is stored, and when the computer program is executed by a processor, the computer program implements the classification method of the dynamic hypergraph neural network in the foregoing embodiments.
Additional aspects and advantages of the invention will be set forth in part in the description which follows and, in part, will be obvious from the description, or may be learned by practice of the invention.
Drawings
The foregoing and/or additional aspects and advantages of the present application will become apparent and readily appreciated from the following description of the embodiments, taken in conjunction with the accompanying drawings of which:
FIG. 1 is a flow chart of a dynamic hypergraph neural network classification method according to an embodiment of the present application;
FIG. 2 is a schematic diagram of a convolutional layer of a dynamic hypergraph neural network according to an embodiment of the present application;
FIG. 3 is an architecture diagram of a dynamic hypergraph neural network classification method and system according to an embodiment of the present application;
fig. 4 is a schematic structural diagram of a dynamic hypergraph neural network classification system according to an embodiment of the present application.
Detailed Description
Reference will now be made in detail to embodiments of the present invention, examples of which are illustrated in the accompanying drawings, wherein like or similar reference numerals refer to the same or similar elements or elements having the same or similar function throughout. The embodiments described below with reference to the drawings are illustrative and intended to be illustrative of the invention and are not to be construed as limiting the invention.
A method and apparatus for block coding according to an embodiment of the present invention will be described with reference to the accompanying drawings.
Fig. 1 is a flowchart of a dynamic hypergraph neural network classification method according to an embodiment of the present application, and as shown in fig. 1, the method includes the following steps:
step 101, obtaining original data to be classified, and constructing a static hypergraph structure according to known association among the original data, wherein the original data comprises known label data and unknown label data.
In one embodiment of the present application, a simple graph structure G ' ═ V, E ' is first constructed using the associations between the raw data, where V denotes a set of nodes of known label data and unknown label data, and E ' denotes paired association edges abstracted by the raw associations. Wherein, if there is an element E in the set EijIt means that there is some association between node i and node j.
Then, the constructed graph structure G' is converted into a static hypergraph structure G by adopting a first-order neighbor conversion mode and the likesEach hypergraph structure uses Gs＝{V,Es,WsV is that the node set is consistent with V in the simple graph structure. EsIs a set of super edges in a static hypergraph, which is converted from an edge set E' in a simple hypergraph, WsThe weight corresponding to each super edge is included in the table.
And 102, converting the embedded characteristics of the nodes of the original data into corresponding initial dynamic super edges through a preset algorithm for mutual conversion of the node characteristics and the associated structures.
In one embodiment of the present application, transforming the embedded features of the nodes of the original data into corresponding initial dynamic hyper-edges comprises: acquiring an embedded feature set of a node of original data; calculating a feature vector between any node in the node set of the original data and nodes except any node according to the embedded feature of each node; selecting k nodes closest to any node, and constructing an initial dynamic super edge E formed by connecting any node and the k nodes through the following formulai：
Ei＝knnk(Xi)
Wherein, knnk() Is a neighbor function, k is a positive integer, XiIs an embedded feature of any node.
That is to sayTo say, firstly, a node feature set of original data is acquired: x ═ X1,X2,…,Xn}. Selecting any node characteristic X from the node characteristicsiAnd X isiCalculating the distance between the node and all other node characteristics to obtain a distance vector DijWhere j requires enumeration of all nodes in the point set. Then obtaining the initial dynamic hyper-edge epsilon according to the neighbor functiond＝{E1,E2,…,EnThe meaning of the neighbor function is that any node i selects k nodes nearest to the node to connect with to construct an initial dynamic hyper-edge EiWherein the vector D can be obtained by calculationijAnd constructing the super edge.
And 103, constructing a dynamic hypergraph neural network, selecting a plurality of known label data to form a training set of the dynamic hypergraph neural network, and performing iterative training on the dynamic hypergraph neural network by combining a static hypergraph structure and an initial dynamic hypergraph edge.
In the embodiment of the application, when the dynamic hypergraph neural network is constructed, trainable parameters and different node characteristics and associated structure conversion algorithms are combined to transform the characteristics, and meanwhile, node convolution, attention layer and hypergraph convolution are designed to embed static and dynamic hypergraph associated structures into high-order characteristics of nodes in real time, so that association among data is fully mined, and more accurate node description is generated. As a possible implementation manner, as shown in fig. 2, the embodiment provides an architecture diagram of a dynamic hypergraph convolutional layer, and the specific construction process includes the following steps:
s1: constructing an initial static hypergraph incidence matrix and setting a hyperedge set EsEach hyper-edge in the initial static hyper-graph correlation matrix is judged, a corresponding data column is added in the initial static hyper-graph correlation matrix according to the judgment result of each hyper-edge, wherein each bit in the data column corresponds to each node in the training set, the value on the bit in the data column corresponding to the node connected with each hyper-edge is set to be 1 according to the judgment result, the value on the bit in the data column corresponding to the node not connected is set to be 0, so that the static hyper-graph correlation matrix H is constructeds
In one embodiment of the present applicationFor the super edge set EsEach of the super edges in (1) is judged. I.e. set E of hyper-edges in the static hypergraph in step 101sThe static excess edge in (1) is judged, and the following formula can be used for judging in this embodiment:
setting the nodes associated with the static super edge to be 1 and keeping the nodes not associated with the static super edge to be 0, thereby constructing a static super graph association matrix Hs。
S2: and constructing an initial dynamic hypergraph structure according to the original node characteristics by a dynamic hypergraph structure construction algorithm, converting the initial dynamic hypergraph structure into a corresponding initial dynamic hypergraph incidence matrix, and setting values in the initial dynamic hypergraph incidence matrix through the initial dynamic hypergraph edges.
In one embodiment of the present application, the specific steps of creating the initial dynamic hypergraph correlation matrix are: constructing an initial dynamic hypergraph structure from the original node characteristics through a dynamic hypergraph structure construction algorithm, and converting the initial dynamic hypergraph structure into a corresponding initial dynamic hypergraph incidence matrix
Further, the dynamic hyper-edge set epsilon in step 102 is processedd＝{E1,E2,…,EnJudging the dynamic excess edge in the matrix, and generating an initial dynamic excess edge incidence matrix through the initial dynamic excess edge
This embodiment can be judged by the following formula:
setting the value in the initial dynamic hypergraph incidence matrix through the initial dynamic hypergraph, namely setting the node associated with the dynamic hypergraph to be 1 and not being related to the dynamic hypergraphThe nodes of the link are kept to be 0, so that a dynamic hypergraph incidence matrix is constructed
S3: and fusing the static hypergraph incidence matrix and the initial dynamic hypergraph incidence matrix to construct an initial static and dynamic combined incidence matrix.
In one embodiment of the present application, a static hypergraph is associated with a matrix HsAnd dynamic hypergraph correlation matrix
Performing combination to construct an initial static and dynamic combined incidence matrix
The formula is expressed as follows:
where | | |, refers to the splicing operation of the matrix.
S4: and after obtaining the modified first node characteristic by convolution of a preset characteristic conversion matrix and the node characteristic of the original data, carrying out residual connection with the original node characteristic to obtain an updated second node characteristic, and aggregating the second node characteristic into a super-edge characteristic by transposition of a dynamic and static combined hypergraph incidence matrix.
In one embodiment of the present application, a predetermined feature transformation matrix is convolved with node features of original data, i.e. a defined feature transformation matrix is defined
And node characteristics
The elements in the data are convoluted to obtain a first node characteristic after being corrected, and then the original node characteristic is subjected to
And performing residual connection to obtain a second node characteristic, wherein the calculation formula is as follows:
Xi′＝Xiθi+Xi
further, a hypergraph incidence matrix combining dynamic state and static state
Transposing as a guide to information aggregation of the second node features Xi′Converging to a super-edge feature Yi+1The formula is as follows:
wherein, Yi+1Is a super-edge feature that is characterized in that,
is the transpose of a hyper-map associative matrix of a dynamic and static union, Xi′Is a feature of the second node that,
is a diagonal matrix made up of the inverse of each node degree.
S5: and performing super-edge convolution on the super-edge characteristics, and combining an attention mechanism to obtain updated third node characteristics.
In an embodiment of the present application, the super-edge convolution is performed on the super-edge feature, that is, the super-edge feature is input into a preset super-edge convolution, and meanwhile, a third node feature is obtained by combining an attention mechanism, where a formula is expressed as:
wherein Xi+1Is a third node characteristic, atten (-) is an embedded attention mechanism function, MLP (-) is a learnable full-connection layer for automatically learning each node characteristicThe weight of the token.
S6: the steps S1 to S5 are repeatedly performed for each node feature-tag pair in the training set to iteratively train the network parameters until the dynamic hypergraph neural network converges.
And 104, constructing a static association hypergraph of the unknown label data and the known label data according to the original node characteristics of the unknown label data, and extracting high-order characteristics of the nodes of the unknown label data through a trained dynamic hypergraph neural network to obtain a node classification prediction result.
In an embodiment of the present application, the known label data is data participating in training of the network model, and as shown in fig. 3, according to the original node features of the unknown label data, the high-order features of the unknown label data nodes are extracted through a trained dynamic hypergraph neural network, so that the classification prediction result of the unknown nodes can be obtained.
In summary, the dynamic hypergraph neural network classification method in the embodiment of the present application uses the known correlation between the original data and uses the first-order neighbor transformation and other ways to construct the static hypergraph correlation as the initial hypergraph structure information. Then, a function for mutual conversion of the node characteristics and the associated structures is designed, the embedded characteristics of the nodes are converted into corresponding hypergraph associations in real time by utilizing the function, and modeling is carried out on the original data. And then constructing a dynamic hypergraph neural network, combining trainable parameters with different node characteristics and an associated structure conversion algorithm to transform the characteristics, and designing node convolution, an attention layer and hypergraph convolution to embed static and dynamic hypergraph associated structures into high-order characteristics of nodes in real time, so that association among data is fully mined, and more accurate node description is generated. And finally, constructing a static association hypergraph of the node data and the original characteristics of the unknown label and the known nodes, and simultaneously extracting high-order characteristics of the unknown label nodes by using a trained dynamic hypergraph neural network so as to obtain a better node classification prediction result. The method and the device can more conveniently model high-order complex associated data; the capability of capturing the dynamic change of the characteristics is stronger; the learned model parameters can be adapted to different feature spaces; under the condition of sufficient known information, the direction of the optimal parameter gradient can be easily found, and the model parameter overfitting and the performance are enhanced.
In order to implement the foregoing embodiment, the present application further provides a dynamic hypergraph neural network classification system, and fig. 4 is a schematic structural diagram of the dynamic hypergraph neural network classification system provided in the embodiment of the present application, and as shown in fig. 4, the classification system includes a first obtaining module 100, a transforming module 200, a constructing module 300, and an extracting module 400.
The first obtaining module 100 is configured to obtain original data to be classified, and construct a static hypergraph structure according to a known association between the original data, where the original data includes known tag data and unknown tag data.
The conversion module 200 is configured to convert the embedded features of the nodes of the original data into corresponding initial dynamic super edges through a preset algorithm for mutual conversion between the node features and the associated structures.
The building module 300 is configured to build a dynamic hypergraph neural network, select a plurality of known label data to form a training set of the dynamic hypergraph neural network, and perform iterative training on the dynamic hypergraph neural network by combining a static hypergraph structure and an initial dynamic hypergraph edge
And the extraction module 400 is configured to construct a static association hypergraph of the unknown label data and the known label data according to the original node features of the unknown label data, and extract high-order features of the nodes of the unknown label data through a trained dynamic hypergraph neural network to obtain a node classification prediction result.
Optionally, in an embodiment of the present application, the first obtaining module 100 is specifically configured to: constructing a simple graph structure G ' ═ V, E ' through the known association between the original data, wherein V represents a set of nodes of the known label data and the unknown label data, and E ' represents paired association edges abstracted through the known association between the original data; converting a simple graph structure G' into a static hypergraph structure G through first-order neighbor conversions＝{V,Es,WsIn which EsIs a set of super-edges in the static hypergraph structure obtained by converting E', WsEach of which is shownThe above-mentioned excess edge EsThe corresponding weight.
Optionally, in an embodiment of the present application, the conversion module 200 is specifically configured to: acquiring an embedded feature set of a node of original data; calculating a feature vector between any node in the node set of the original data and nodes except any node according to the embedded feature of each node; selecting k nodes closest to any node, and constructing an initial dynamic super edge E formed by connecting any node and the k nodes through the following formulai：
Ei＝knnk(Xi)
Wherein, knnk() Is a neighbor function, k is a positive integer, XiIs an embedded feature of any node.
Optionally, in an embodiment of the present application, the building module 300 is specifically configured to: constructing an initial static hypergraph incidence matrix and setting a hyperedge set EsEach hyper-edge in the initial static hyper-graph correlation matrix is judged, a corresponding data column is added in the initial static hyper-graph correlation matrix according to the judgment result of each hyper-edge, wherein each bit in the data column corresponds to each node in the training set, the value on the bit in the data column corresponding to the node connected with each hyper-edge is set to be 1 according to the judgment result, the value on the bit in the data column corresponding to the node not connected is set to be 0, so that the static hyper-graph correlation matrix H is constructeds(ii) a Constructing an initial dynamic hypergraph structure according to original node characteristics through a dynamic hypergraph structure construction algorithm, converting the initial dynamic hypergraph structure into a corresponding initial dynamic hypergraph incidence matrix, and setting values in the initial dynamic hypergraph incidence matrix through initial dynamic hypergraph edges; fusing the static hypergraph incidence matrix and the initial dynamic hypergraph incidence matrix to construct an initial static and dynamic combined incidence matrix; performing convolution on a preset feature transformation matrix and node features of original data to obtain modified first node features, performing residual connection on the modified first node features and the original node features to obtain updated second node features, and aggregating the second node features into super-edge features through transposition of a dynamic and static combined hypergraph incidence matrix; performing super-edge convolution on the super-edge characteristics, and combiningAnd repeatedly executing the steps on each node feature-label pair in the training set by the attention mechanism after the updated third node feature pair is obtained so as to iteratively train the network parameters until the dynamic hypergraph neural network converges.
Optionally, in an embodiment of the present application, the building module 300 is further configured to aggregate the second node feature into the super-edge feature by the following formula:
wherein, Yi+1Is a super-edge feature that is characterized in that,
is the transpose of a hyper-map associative matrix of a dynamic and static union, Xi′Is a feature of the second node that,
is a diagonal matrix made up of the inverse of each node degree.
Optionally, in an embodiment of the present application, the building module 300 is further configured to obtain an updated third node characteristic by the following formula:
Xi+1＝atten(Yi+1)Yi+1
wherein, atten (x) x + mlp (x),
wherein, atten () is the embedded attention mechanism function, MLP () is the learnable full link layer for automatically learning the weight of each node feature, Xi+1Is a third node feature.
In summary, the dynamic hypergraph neural network classification system of the embodiment of the present application uses the known correlation between the original data and uses the first-order neighbor transformation and other ways to construct the static hypergraph correlation as the initial hypergraph structure information. Then, a function for mutual conversion of the node characteristics and the associated structures is designed, the embedded characteristics of the nodes are converted into corresponding hypergraph associations in real time by utilizing the function, and modeling is carried out on the original data. And then constructing a dynamic hypergraph neural network, combining trainable parameters with different node characteristics and an associated structure conversion algorithm to transform the characteristics, and designing node convolution, an attention layer and hypergraph convolution to embed static and dynamic hypergraph associated structures into high-order characteristics of nodes in real time, so that association among data is fully mined, and more accurate node description is generated. And finally, constructing a static association hypergraph of the node data and the original characteristics of the unknown label and the known nodes, and simultaneously extracting high-order characteristics of the unknown label nodes by using a trained dynamic hypergraph neural network so as to obtain a better node classification prediction result. The method and the device can more conveniently model high-order complex associated data; the capability of capturing the dynamic change of the characteristics is stronger; the learned model parameters can be adapted to different feature spaces; under the condition of sufficient known information, the direction of the optimal parameter gradient can be easily found, and the model parameter overfitting and the performance are enhanced.
In the description herein, reference to the description of the term "one embodiment," "some embodiments," "an example," "a specific example," or "some examples," etc., means that a particular feature, structure, material, or characteristic described in connection with the embodiment or example is included in at least one embodiment or example of the application. In this specification, the schematic representations of the terms used above are not necessarily intended to refer to the same embodiment or example. Furthermore, the particular features, structures, materials, or characteristics described may be combined in any suitable manner in any one or more embodiments or examples. Furthermore, various embodiments or examples and features of different embodiments or examples described in this specification can be combined and combined by one skilled in the art without contradiction.
Furthermore, the terms "first", "second" and "first" are used for descriptive purposes only and are not to be construed as indicating or implying relative importance or implicitly indicating the number of technical features indicated. Thus, a feature defined as "first" or "second" may explicitly or implicitly include at least one such feature. In the description of the present application, "plurality" means at least two, e.g., two, three, etc., unless specifically limited otherwise.
Any process or method descriptions in flow charts or otherwise described herein may be understood as representing modules, segments, or portions of code which include one or more executable instructions for implementing steps of a custom logic function or process, and alternate implementations are included within the scope of the preferred embodiment of the present application in which functions may be executed out of order from that shown or discussed, including substantially concurrently or in reverse order, depending on the functionality involved, as would be understood by those reasonably skilled in the art of the present application.
The logic and/or steps represented in the flowcharts or otherwise described herein, e.g., an ordered listing of executable instructions that can be considered to implement logical functions, can be embodied in any computer-readable medium for use by or in connection with an instruction execution system, apparatus, or device, such as a computer-based system, processor-containing system, or other system that can fetch the instructions from the instruction execution system, apparatus, or device and execute the instructions. For the purposes of this description, a "computer-readable medium" can be any means that can contain, store, communicate, propagate, or transport the program for use by or in connection with the instruction execution system, apparatus, or device. More specific examples (a non-exhaustive list) of the computer-readable medium would include the following: an electrical connection (electronic device) having one or more wires, a portable computer diskette (magnetic device), a Random Access Memory (RAM), a read-only memory (ROM), an erasable programmable read-only memory (EPROM or flash memory), an optical fiber device, and a portable compact disc read-only memory (CDROM). Additionally, the computer-readable medium could even be paper or another suitable medium upon which the program is printed, as the program can be electronically captured, via for instance optical scanning of the paper or other medium, then compiled, interpreted or otherwise processed in a suitable manner if necessary, and then stored in a computer memory.
It should be understood that portions of the present application may be implemented in hardware, software, firmware, or a combination thereof. In the above embodiments, the various steps or methods may be implemented in software or firmware stored in memory and executed by a suitable instruction execution system. If implemented in hardware, as in another embodiment, any one or combination of the following techniques, which are known in the art, may be used: a discrete logic circuit having a logic gate circuit for implementing a logic function on a data signal, an application specific integrated circuit having an appropriate combinational logic gate circuit, a Programmable Gate Array (PGA), a Field Programmable Gate Array (FPGA), or the like.
It will be understood by those skilled in the art that all or part of the steps carried by the method for implementing the above embodiments may be implemented by hardware related to instructions of a program, which may be stored in a computer readable storage medium, and when the program is executed, the program includes one or a combination of the steps of the method embodiments.
In addition, functional units in the embodiments of the present application may be integrated into one processing module, or each unit may exist alone physically, or two or more units are integrated into one module. The integrated module can be realized in a hardware mode, and can also be realized in a software functional module mode. The integrated module, if implemented in the form of a software functional module and sold or used as a stand-alone product, may also be stored in a computer readable storage medium.
The storage medium mentioned above may be a read-only memory, a magnetic or optical disk, etc. Although embodiments of the present application have been shown and described above, it is understood that the above embodiments are exemplary and should not be construed as limiting the present application, and that variations, modifications, substitutions and alterations may be made to the above embodiments by those of ordinary skill in the art within the scope of the present application.
Claims (10)
1. A dynamic hypergraph neural network classification method is characterized by comprising the following steps:
acquiring original data to be classified, and constructing a static hypergraph structure according to known association among the original data, wherein the original data comprises known label data and unknown label data;
converting the embedded characteristics of the nodes of the original data into corresponding initial dynamic overcurrents through a preset algorithm for interconversion between the node characteristics and the associated structures;
constructing a dynamic hypergraph neural network, selecting a plurality of known label data to form a training set of the dynamic hypergraph neural network, and performing iterative training on the dynamic hypergraph neural network by combining the static hypergraph structure and the initial dynamic hypergraph edge;
and constructing a static association hypergraph of the unknown label data and the known label data according to the original node characteristics of the unknown label data, and extracting high-order characteristics of the nodes of the unknown label data through a trained dynamic hypergraph neural network to obtain a node classification prediction result.
2. The classification method according to claim 1, wherein the constructing a static hypergraph structure from known associations between the raw data comprises:
constructing a simple graph structure G ' ═ V, E ' } through the known association between the original data, wherein V represents a set of nodes of the known label data and the unknown label data, and E ' represents a pair of associated edges abstracted through the known association between the original data;
converting the simple graph structure G' into a static hypergraph structure G through first-order neighbor conversions＝{V，Es，WsIn which EsIs the set of the super-edges in the static hypergraph structure obtained by the transformation of E', WsRepresenting each of said super edges EsThe corresponding weight.
3. The classification method according to claim 1 or 2, wherein the converting of the embedded features of the nodes of the original data into corresponding initial dynamic hyper-edges comprises:
acquiring an embedded feature set of the nodes of the original data;
according to the embedded characteristics of each node, calculating characteristic vectors between any node in the node set of the original data and nodes except the any node;
selecting k nodes closest to any node, and constructing an initial dynamic super edge E formed by connecting any node and the k nodes through the following formulai：
Ei＝knnk(Xi)
Wherein, knnk() Is a neighbor function, k is a positive integer, XiIs an embedded feature of any node.
4. The classification method according to claim 1, wherein the iteratively training the dynamic hypergraph neural network comprises:
s1: constructing an initial static hypergraph incidence matrix, and collecting E for the hypergraphsAdding a corresponding data column in the initial static hypergraph correlation matrix according to the judgment result of each hyperedge, wherein each bit in the data column corresponds to each node in the training set, setting the value of the bit in the data column corresponding to the node connected by each hyperedge to be 1 according to the judgment result, and setting the value of the bit in the data column corresponding to the node not connected to be 0 to construct a static hypergraph correlation matrix Hs；
S2: constructing an initial dynamic hypergraph structure according to original node characteristics through a dynamic hypergraph structure construction algorithm, converting the initial dynamic hypergraph structure into a corresponding initial dynamic hypergraph incidence matrix, and setting values in the initial dynamic hypergraph incidence matrix through the initial dynamic hypergraph edges;
s3: fusing the static hypergraph incidence matrix and the initial dynamic hypergraph incidence matrix to construct an initial static and dynamic combined incidence matrix;
s4: convolving the node characteristics of the original data through a preset characteristic conversion matrix to obtain modified first node characteristics, performing residual connection with the original node characteristics to obtain updated second node characteristics, and aggregating the second node characteristics into the super-edge characteristics through the transposition of the dynamic and static combined hypergraph association matrix;
s5: performing super-edge convolution on the super-edge characteristics, and obtaining updated third node characteristics by combining an attention mechanism
S6: repeating steps S1-S5 for each node feature-tag pair in the training set to iteratively train network parameters until the dynamic hypergraph neural network converges.
5. The classification method according to claim 4, wherein the second node features are aggregated into super-edge features by the following formula:
wherein, Yi+1Is a super-edge feature that is characterized in that,
is the transpose of a hyper-map associative matrix of a dynamic and static union, Xi′Is a feature of the second node that,
is a diagonal matrix made up of the inverse of each node degree.
6. The classification method according to claim 4, wherein the updated third node characteristic is obtained by the following formula:
Xi+1＝atten(Yi+1)Yi+1
wherein, atten (x) x + mlp (x),
wherein, atten () is the embedded attention mechanism function, MLP () is the learnable full link layer for automatically learning the weight of each node feature, Xi+1Is a third node feature.
7. A dynamic hypergraph neural network classification system, comprising:
the system comprises an acquisition module, a classification module and a classification module, wherein the acquisition module is used for acquiring original data to be classified and constructing a static hypergraph structure according to known association among the original data, and the original data comprises known label data and unknown label data;
the conversion module is used for converting the embedded characteristics of the nodes of the original data into corresponding initial dynamic super edges through a preset algorithm of mutual conversion between the node characteristics and the associated structures;
the construction module is used for constructing a dynamic hypergraph neural network, selecting a plurality of known label data to form a training set of the dynamic hypergraph neural network, and performing iterative training on the dynamic hypergraph neural network by combining the static hypergraph structure and the initial dynamic hypergraph edge;
and the extraction module is used for constructing a static association hypergraph of the unknown label data and the known label data according to the original node characteristics of the unknown label data, and extracting the high-order characteristics of the nodes of the unknown label data through a trained dynamic hypergraph neural network so as to obtain a node classification prediction result.
8. The classification system for a dynamic hypergraph neural network of claim 7, wherein the first obtaining module is specifically configured to:
constructing a simple graph structure G ' ═ V, E ' } through the known association between the original data, wherein V represents a set of nodes of the known label data and the unknown label data, and E ' represents a pair of associated edges abstracted through the known association between the original data;
converting the simple graph structure G' into a static hypergraph structure G through first-order neighbor conversions＝{V，Es，WsIn which EsIs the set of the super-edges in the static hypergraph structure obtained by the transformation of E', WsRepresenting each of said super edges EsThe corresponding weight.
9. The classification system according to claims 7 and 8, wherein the conversion module is specifically configured to:
acquiring an embedded feature set of the nodes of the original data;
according to the embedded characteristics of each node, calculating characteristic vectors between any node in the node set of the original data and nodes except the any node;
selecting k nodes closest to any node, and constructing an initial dynamic super edge E formed by connecting any node and the k nodes through the following formulai：
Ei＝knnk(Xi)
Wherein, knnk() Is a neighbor function, k is a positive integer, XiIs an embedded feature of any node.
10. A non-transitory computer readable storage medium having stored thereon a computer program, wherein the computer program, when executed by a processor, implements the dynamic hypergraph neural network classification method of any one of claims 1-6.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202110921299.4A CN113792753A (en) | 2021-08-11 | 2021-08-11 | Dynamic hypergraph neural network classification method and system |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202110921299.4A CN113792753A (en) | 2021-08-11 | 2021-08-11 | Dynamic hypergraph neural network classification method and system |
Publications (1)
| Publication Number | Publication Date |
|---|---|
| CN113792753A true CN113792753A (en) | 2021-12-14 |
Family
ID=78875892
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN202110921299.4A Pending CN113792753A (en) | 2021-08-11 | 2021-08-11 | Dynamic hypergraph neural network classification method and system |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN113792753A (en) |
Cited By (2)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN114372573A (en) * | 2022-01-07 | 2022-04-19 | 中国人民解放军国防科技大学 | User portrait information recognition method and device, computer equipment and storage medium |
| CN114463687A (en) * | 2022-04-12 | 2022-05-10 | 北京云恒科技研究院有限公司 | Movement track prediction method based on big data |
-
2021
- 2021-08-11 CN CN202110921299.4A patent/CN113792753A/en active Pending
Cited By (2)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN114372573A (en) * | 2022-01-07 | 2022-04-19 | 中国人民解放军国防科技大学 | User portrait information recognition method and device, computer equipment and storage medium |
| CN114463687A (en) * | 2022-04-12 | 2022-05-10 | 北京云恒科技研究院有限公司 | Movement track prediction method based on big data |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| Feng et al. | Domain knowledge-based deep-broad learning framework for fault diagnosis | |
| CN109685135B (en) | Few-sample image classification method based on improved metric learning | |
| CN113360673B (en) | Entity alignment method, device and storage medium of multi-mode knowledge graph | |
| CN115456160A (en) | Data processing method and data processing equipment | |
| CN109192298B (en) | Deep brain disease diagnosis algorithm based on brain network | |
| CN113792753A (en) | Dynamic hypergraph neural network classification method and system | |
| US11574198B2 (en) | Apparatus and method with neural network implementation of domain adaptation | |
| CN113780292A (en) | Semantic segmentation network model uncertainty quantification method based on evidence reasoning | |
| CN113792768A (en) | Hypergraph neural network classification method and device | |
| Cheng et al. | Wind turbine blade icing detection: A federated learning approach | |
| CN112733933B (en) | Data classification method and device based on unified optimization target frame graph neural network | |
| CN115358487A (en) | Federal learning aggregation optimization system and method for power data sharing | |
| Luciano et al. | Deep similarity network fusion for 3D shape classification | |
| CN116403730A (en) | Medicine interaction prediction method and system based on graph neural network | |
| CN111027610B (en) | Image feature fusion method, apparatus, and medium | |
| Seo et al. | Graph neural networks and implicit neural representation for near-optimal topology prediction over irregular design domains | |
| CN113592769B (en) | Abnormal image detection and model training method, device, equipment and medium | |
| CN113688882A (en) | Training method and device of memory-enhanced continuous learning neural network model | |
| CN116432053A (en) | Multi-mode data representation method based on modal interaction deep hypergraph neural network | |
| CN112733724A (en) | Relativity relationship verification method and device based on discrimination sample meta-digger | |
| CN116958613A (en) | Depth multi-view clustering method and device, electronic equipment and readable storage medium | |
| CN116028891A (en) | Industrial anomaly detection model training method and device based on multi-model fusion | |
| CN113208641B (en) | Auxiliary diagnosis method for lung nodule based on three-dimensional multi-resolution attention capsule network | |
| Ha et al. | Comparing deep learning strategies and attention mechanisms of discrete registration for multimodal image-guided interventions | |
| Msuya et al. | Deep Learning Model Compression Techniques: Advances, Opportunities, and Perspective |
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 |