CN117036760A - Multi-view clustering model implementation method based on graph comparison learning - Google Patents
Multi-view clustering model implementation method based on graph comparison learning Download PDFInfo
- Publication number
- CN117036760A CN117036760A CN202310957418.0A CN202310957418A CN117036760A CN 117036760 A CN117036760 A CN 117036760A CN 202310957418 A CN202310957418 A CN 202310957418A CN 117036760 A CN117036760 A CN 117036760A
- Authority
- CN
- China
- Prior art keywords
- node
- graph
- clustering
- view
- matrix
- 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
- 238000000034 method Methods 0.000 title claims abstract description 38
- 230000006870 function Effects 0.000 claims abstract description 42
- 238000012549 training Methods 0.000 claims abstract description 18
- 230000008569 process Effects 0.000 claims abstract description 11
- 239000011159 matrix material Substances 0.000 claims description 59
- 238000004364 calculation method Methods 0.000 claims description 9
- 230000007246 mechanism Effects 0.000 claims description 9
- 238000010586 diagram Methods 0.000 claims description 8
- 230000004913 activation Effects 0.000 claims description 7
- 238000001228 spectrum Methods 0.000 claims description 6
- 238000005259 measurement Methods 0.000 claims description 4
- 238000010606 normalization Methods 0.000 claims description 4
- 230000009466 transformation Effects 0.000 claims description 4
- 238000012512 characterization method Methods 0.000 claims description 3
- 238000005457 optimization Methods 0.000 claims description 3
- 230000003595 spectral effect Effects 0.000 claims description 3
- 230000001105 regulatory effect Effects 0.000 claims description 2
- 230000000694 effects Effects 0.000 description 4
- 230000006978 adaptation Effects 0.000 description 2
- 238000013528 artificial neural network Methods 0.000 description 2
- 230000000295 complement effect Effects 0.000 description 2
- 238000012986 modification Methods 0.000 description 2
- 230000004048 modification Effects 0.000 description 2
- 238000011524 similarity measure Methods 0.000 description 2
- 238000004458 analytical method Methods 0.000 description 1
- 230000006399 behavior Effects 0.000 description 1
- 238000004891 communication Methods 0.000 description 1
- 238000001514 detection method Methods 0.000 description 1
- 238000005516 engineering process Methods 0.000 description 1
- 230000002708 enhancing effect Effects 0.000 description 1
- 230000004927 fusion Effects 0.000 description 1
- 238000012545 processing Methods 0.000 description 1
- 230000003014 reinforcing effect Effects 0.000 description 1
- 238000011160 research Methods 0.000 description 1
- 238000005096 rolling process Methods 0.000 description 1
- 230000011218 segmentation Effects 0.000 description 1
Classifications
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06V—IMAGE OR VIDEO RECOGNITION OR UNDERSTANDING
- G06V10/00—Arrangements for image or video recognition or understanding
- G06V10/70—Arrangements for image or video recognition or understanding using pattern recognition or machine learning
- G06V10/762—Arrangements for image or video recognition or understanding using pattern recognition or machine learning using clustering, e.g. of similar faces in social 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
- G06N3/04—Architecture, e.g. interconnection topology
- G06N3/0464—Convolutional networks [CNN, ConvNet]
-
- 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
- G06N3/048—Activation functions
-
- 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
- G06N3/0895—Weakly supervised learning, e.g. semi-supervised or self-supervised learning
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06V—IMAGE OR VIDEO RECOGNITION OR UNDERSTANDING
- G06V10/00—Arrangements for image or video recognition or understanding
- G06V10/70—Arrangements for image or video recognition or understanding using pattern recognition or machine learning
- G06V10/74—Image or video pattern matching; Proximity measures in feature spaces
- G06V10/761—Proximity, similarity or dissimilarity measures
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06V—IMAGE OR VIDEO RECOGNITION OR UNDERSTANDING
- G06V10/00—Arrangements for image or video recognition or understanding
- G06V10/70—Arrangements for image or video recognition or understanding using pattern recognition or machine learning
- G06V10/77—Processing image or video features in feature spaces; using data integration or data reduction, e.g. principal component analysis [PCA] or independent component analysis [ICA] or self-organising maps [SOM]; Blind source separation
- G06V10/774—Generating sets of training patterns; Bootstrap methods, e.g. bagging or boosting
-
- Y—GENERAL TAGGING OF NEW TECHNOLOGICAL DEVELOPMENTS; GENERAL TAGGING OF CROSS-SECTIONAL TECHNOLOGIES SPANNING OVER SEVERAL SECTIONS OF THE IPC; TECHNICAL SUBJECTS COVERED BY FORMER USPC CROSS-REFERENCE ART COLLECTIONS [XRACs] AND DIGESTS
- Y02—TECHNOLOGIES OR APPLICATIONS FOR MITIGATION OR ADAPTATION AGAINST CLIMATE CHANGE
- Y02T—CLIMATE CHANGE MITIGATION TECHNOLOGIES RELATED TO TRANSPORTATION
- Y02T10/00—Road transport of goods or passengers
- Y02T10/10—Internal combustion engine [ICE] based vehicles
- Y02T10/40—Engine management systems
Landscapes
- Engineering & Computer Science (AREA)
- Theoretical Computer Science (AREA)
- Physics & Mathematics (AREA)
- Artificial Intelligence (AREA)
- Computing Systems (AREA)
- Health & Medical Sciences (AREA)
- Evolutionary Computation (AREA)
- General Health & Medical Sciences (AREA)
- General Physics & Mathematics (AREA)
- Software Systems (AREA)
- Computer Vision & Pattern Recognition (AREA)
- Databases & Information Systems (AREA)
- Medical Informatics (AREA)
- Multimedia (AREA)
- Life Sciences & Earth Sciences (AREA)
- Biomedical Technology (AREA)
- Biophysics (AREA)
- Computational Linguistics (AREA)
- Data Mining & Analysis (AREA)
- Molecular Biology (AREA)
- General Engineering & Computer Science (AREA)
- Mathematical Physics (AREA)
- Image Analysis (AREA)
Abstract
The invention discloses a multi-view clustering model implementation method based on graph comparison learning, which comprises the following steps: the system comprises a graph learning module, a graph comparison module and a self-supervision module; the method comprises the following steps: s1, establishing a unified multi-view model, and defining the input and output of the multi-view model; s2, selecting a public data set as initial data, and initializing the initial data through a constructed graph learning module; s3, performing contrast learning on the initialized data through a constructed graph contrast module to obtain comprehensive and accurate data structure information; and S4, clustering the acquired data structure information through a self-supervision module, constructing an objective function for training, stopping training when the set condition is met, and taking the obtained predicted clustering label as the output of the multi-view model. The invention introduces a self-supervision module to assist in clustering tasks, and uses soft labels to supervise the graph clustering process, thereby improving the generalization capability of the model.
Description
Technical Field
The invention relates to the field of contrast learning and graph convolution neural networks, in particular to a method for realizing a multi-view clustering model based on graph contrast learning.
Background
Graph clustering is a fundamental task in the area of graph network research, aimed at dividing a graph into several closely connected groups. Graph clustering techniques are widely used in practice, such as group segmentation, communication network structure analysis, community detection in social networks, and the like. However, most of the existing graph clustering algorithms can only handle single views, and the graph data in reality is more complex, multiple views are usually needed to better represent the real graph data instead of single views, and multiple views can better represent multiple relationships of nodes, such as co-authors and co-conferences in academic networks.
The sources of real-life data are typically different and may be represented by different features or views. Noise and imperfections may exist for each view, but important factors (e.g., geometry and semantics) exist among all views sharing information. The features and data of the different views are complementary, and integrating the features of the different views is critical to improving the performance of the clustering task.
However, most of the current clustering methods focus on single-view clustering, and the effect of processing complex multi-view clustering by using the single-view method is poor, and complementary information between different views cannot be focused.
Disclosure of Invention
The invention aims to: the invention aims to provide a multi-view clustering model realization method based on graph contrast learning, which improves the generalization capability of a model.
The technical scheme is as follows: the invention discloses a multi-view clustering model implementation method, which comprises the following steps: the system comprises a graph learning module, a graph comparison module and a self-supervision module;
the diagram learning module carries out convolution calculation on multi-view data by adopting a spectrum convolution function, carries out initialization representation on each node in the diagram by utilizing a diagram embedding method, and strengthens node embedding on the multi-view data by using an attention mechanism to obtain initialization data;
the graph comparison module learns similarity measurement between initialization data through a cluster comparison method, and the similarity measurement is used for explaining the structure of multi-view data; the comprehensive and accurate data structure information is obtained by comparing and learning the multiple views;
after the self-supervision module obtains the optimal view coding and node embedding, the process of graph clustering is supervised by adopting a highly-trusted node q as a soft label, the similarity between each node and a central point is measured by using t distribution as a core, and an auxiliary target matrix P is introduced to optimize the clustering task;
the method comprises the following steps:
s1, establishing a unified multi-view model, and defining the input and output of the multi-view model;
s2, selecting a public data set as initial data, and initializing the initial data through a constructed graph learning module;
s3, performing contrast learning on the initialized data through a constructed graph contrast module to obtain comprehensive and accurate data structure information;
and S4, clustering the acquired data structure information through a self-supervision module, constructing an objective function for training, stopping training when the set condition is met, and taking the obtained predicted clustering label as the output of the multi-view model.
Further, in step S1, the input of the multi-view model is as follows:
define the multi-layer graph data as g= { V, E 1 ，…，E m ，X 1 ，…，X m -wherein V represents a set of N nodes; e (E) 1 ，…，E m Respectively connecting edges in m graph views, wherein the topological structure of the graph is represented by an adjacent matrix A; in a certain layer, if A ij =1, indicating that there is a border between the i-th node and the j-th node in the layer; if A ij =0, indicating that there is a border between the i-th node and the j-th node in the layer; { X 1 ，…，X m -representing a matrix of attributes of the node at different levels;a reconstructed adjacency matrix +.>Reconstructing an attribute matrix;
model output: clustering result y r 。
Further, in step S2, the implementation steps of the graph learning module are as follows:
s21, inputting an adjacent matrix and an attribute matrix in initial data into an encoder, carrying out convolution calculation on a graph by using a spectrum convolution function, and carrying out nonlinear transformation on a convolved result by using an activation function to obtain node representation of a node in a current layer, wherein the coded input of a next layer is representation of a node in a previous layer and a corresponding adjacent matrix; for the first embedding of the mth layer, the expression is as follows:
wherein,is the first-1 embedding of the m-th layer, f () is a spectral convolution function, W (l) Is a trainable parameter matrix;wherein I is an identity matrix,>an adjacency matrix for the m-th layer; d (D) m A degree matrix, σ () is an activation function;
s22, embedding the multi-view data by using an attention mechanism to strengthen nodes, so as to obtain initialization data; correlation of node i with node j in layer lThe expression is:
wherein sigmoid () is a heuristic function,respectively trainable parameter matrix +.>And->For matrix->I and j rows of (a);
for a pair ofNormalization was performed with the following expression:
finally obtaining the embedding of the node i:
wherein,representing the association between the ith node and the kth neighbor node of the first layer,/and->The set of neighbor nodes representing the node, exp () represents the representation base on the natural constant eAn exponential function.
Further, in step S3, the implementation steps of the graph comparison module are as follows:
s31, taking the embedded representation of each node as a feature vector, dividing N nodes into G clusters by adopting a k-means algorithm, and simultaneously solving the central node of each cluster, wherein the central node is represented as { c } 1 ，…，c G }；
S32, for each node i, if the central node of the cluster to which the node i belongs is c i Then the embedding of the node with the central node of the cluster to which it belongs characterizes z' i Forming a positive sample, forming a negative sample by the embedded characterization of the node and the central nodes of other clusters, and calculating cluster contrast loss L con Gradient updating encoder, cluster contrast loss L con The expression of (2) is as follows:
wherein z is i Representing the embedding of node i, z' i Embedding, z 'for the node of the belonging central cluster' t Representing the embedding of central nodes of clusters to which other clusters belong, wherein sim () is a cosine similarity formula, τ is a temperature parameter, G is the number of the divided central clusters, e is a natural constant, and N is the total number of nodes;
s33, training the self-encoder by minimizing the sum of reconstruction errors of each view data, and predicting whether a link exists between two nodes, wherein the expression is as follows:
total reconstruction loss function L r ：
Wherein,representing a reconstructed view m, Z being a learned representation of the graph data nodes; w (W) m Is a training weight matrix for projecting Z into view m to predict the connection probability between nodes; z is Z T Is a transpose of Z, sigmoid () is a heuristic function, A m Is the original data of view m; l (L) r Loss, representing M views, constitutes the total reconstruction loss.
Further, in step S4, the implementation process of the self-supervision module is as follows:
s41, measuring the similarity between each node and the clustering center point by adopting t distribution as a core, and q kr In order to assign the node k to the probability of the cluster r, the specific calculation method is as follows:
wherein mu r For embedding cluster centers, z k Embedding for node k; mu (mu) r′ Representing the embedding of the r' th cluster center;
s42, introducing an auxiliary target matrix P to optimize clustering, and defining an element P in the auxiliary target matrix P by normalizing the probability between each node and the clustering center kr The expression is as follows:
wherein f r The sum of probabilities assigned to clusters r for all nodes; f (f) r′ The sum of probabilities assigned to the cluster centers r' for all nodes;
from q kr Obtaining soft distribution matrix Q, p kr Obtaining the element P in the auxiliary target matrix P kr Then the loss function in self-supervision clusteringL clu The method comprises the following steps:
the overall objective function is:
L＝L r +αL con +βL clu
wherein alpha and beta are respectively super parameters, and the contribution degree of different losses to a final optimization target is regulated;
performing network training by adopting a total objective function L, and deducing a predicted clustering label y when a network training result meets a set condition r ：
Wherein argmax represents an index corresponding to the maximum value.
Compared with the prior art, the invention has the following remarkable effects:
1. the invention adopts the node representation learning method based on the attention mechanism, can capture the relevance between the nodes and generate high-quality node embedding; then, node embedding is further enhanced by using a contrast learning technology, and similar nodes are enabled to be more close in a low-dimensional space, so that a subsequent clustering task is more accurate;
2. the self-supervision module is introduced to assist the clustering task, and the soft label is used for supervising the graph clustering process, so that the generalization capability of the model is improved, and a good effect is achieved on the clustering task.
Drawings
Fig. 1 is a general frame diagram of the present invention.
Detailed Description
The invention is described in further detail below with reference to the drawings and the detailed description.
The present invention employs a multi-layer graph clustering auto-encoder framework that employs a multi-layer graph rolling network (GCN) to encode multiple networks to learn the representation of nodes. Through the attention mechanism and contrast learning, the framework can better capture the relevance between nodes and enhance the embedding of the nodes. In addition, a self-supervision module is used for optimizing the cluster centroid, and a view fusion module is introduced for integrating information of different views. FIG. 1 is a diagram of a Multi-view neural network clustering model based on graph contrast learning-MCAN (Multi-view Contrative Attention Network) model overall framework, the overall model framework comprising 3 modules:
and a graph learning module: in order to learn the representation of each node, the adjacent matrix and the attribute matrix in the initial data are firstly required to be input into an encoder, a spectrum convolution function is used for carrying out convolution calculation on the graph, an activation function is used for carrying out nonlinear transformation on the convolved result, the node representation of the node in the current layer is obtained, and the encoding input of the next layer is the representation of the node in the previous layer and the corresponding adjacent matrix.
And (5) a contrast learning module: similarity measures between the data may be learned using contrast learning, which may be used to visualize and interpret the structure of the multi-view data. By contrast learning of multiple views, more comprehensive and accurate data structure information can be obtained, so that the interpretability of the multi-view data is improved.
And a self-supervision module: after the optimal view coding and node embedding are obtained, in order to be applied to a clustering task, because the views are coded independently, a distance-based clustering algorithm can be tried, and because the process can be an unsupervised process, under the condition of no label information, the generated embedded vector can have certain deviation or error, so that the clustering effect is inferior to that of a supervised method. In order to improve the performance of the model and train the model with more generalization capability, the invention uses a self-supervision algorithm to perform clustering tasks, and uses a highly-trusted node as a soft label to supervise the graph clustering process.
The specific implementation steps of the invention are as follows:
step 1, a unified multi-view model is established, and input and output of the multi-view model are defined.
Model input: define the multi-layer graph data as g= { V, E 1 ,…,E m X 1 ,…,X m V represents N node sets, E 1 ,…,E m Respectively connecting edges in m graphs, wherein the topological structure of the graphs is represented by an adjacent matrix A; in a certain layer, if A ij =1, indicating that there is a border between the i-th node and the j-th node in the layer; if A ij =0, indicating that there is a border between the i-th node and the j-th node in the layer; { X 1 ,…,X m The attribute matrix of the node at different layers is represented,a reconstructed adjacency matrix +.>Reconstructing the attribute matrix.
Model output: clustering result y r 。
And 2, constructing a graph learning module of the multi-view data, wherein the module comprises two parts, firstly, multi-view node data learning is carried out, and then, a attention mechanism is used for reinforcing node embedding so as to obtain more comprehensive and accurate data structure information.
Specifically, the implementation process of the multi-view graph learning module includes the following steps:
step 21, selecting a public data set as initial data, firstly, inputting an adjacent matrix and an attribute matrix in the initial data into an encoder, carrying out convolution calculation on a graph by using a spectrum convolution function f (), and carrying out nonlinear transformation on a convolved result by using an activation function to obtain node representation of a node in a current layer, wherein the coding input of a next layer is the representation of the node in the previous layer and the corresponding adjacent matrix; for the first embedding of the mth layer, the expression is as follows:
wherein the method comprises the steps of，Is the first-1 embedding of the m-th layer, f () is a spectral convolution function, W (l) Is a trainable parameter matrix;wherein I is an identity matrix>An adjacency matrix for the m-th layer; d (D) m For the degree matrix, σ () is the activation function.
Step 22, using an attention mechanism to strengthen node embedding on the multi-view data to obtain initialization data: in order to better acquire the correlation between the nodes in each layer, the invention refers to an attention mechanism to better acquire the embedding of the nodes. Relationships between adjacent nodes, i.e. correlation of nodes i and j in layer iThe expression is:
wherein sigmoid () is a heuristic function,are trainable parameter matrices, +.>And->For matrix->I and j rows of (a).
Next, forNormalization is carried out to make the correlation coefficient of the neighbor set of the node i have comparability, and the normalization expression is as follows:
wherein, the method comprises the steps of,representing the association between the ith node and the kth neighbor node of the first layer,/and->The set of neighboring nodes representing the node, exp () represents an exponential function that bases on the natural constant e.
Finally, the embedding of the node i can be obtained:
and 3, performing contrast learning by using a graph contrast module, and learning a consistent similarity measure, thereby improving the performance of multi-view clustering.
The graph contrast module learns similarity metrics between the data by using cluster contrast, which can be used to visualize and interpret the structure of the multi-view data. By contrast learning of multiple views, more comprehensive and accurate data structure information can be obtained, so that the performance of multi-view clustering is improved. The specific implementation steps are as follows:
step 31, after the multi-view node data is encoded by the encoder, the invention uses a k-means algorithm to acquire the center node. First, an embedded token of each node is taken as a feature vector, and an Nxd feature matrix X is constructed, wherein d is the dimension of an embedded space. Then, the present invention uses k-means algorithm to divide N nodes into G clusters, and simultaneously calculate the center node of each cluster, denoted as { c } 1 ,…,c G }。
Step 32, using contrast module learning, enhancing node learning: for each node i, if the central node of the cluster to which it belongs is c i Then the embedding of the node with the central node of the cluster to which it belongs characterizes z' i Forming a positive sample, forming a negative sample by the embedded characterization of the node and the central nodes of other clusters, and calculating cluster contrast loss L con Gradient update encoder coding, cluster contrast loss L con Is defined as follows:
wherein z is i Representing the embedding of node i, z' i Embedding, z 'for the node of the belonging central cluster' t The embedding of the central nodes of the clusters to which other clusters belong is represented, sim () is a cosine similarity formula, τ is a temperature parameter, G is the number of the divided central clusters, e is a natural constant, and N is the total number of nodes.
Step 33, reconstructing the loss, training the self-encoder by minimizing the sum of the reconstruction errors for each view data: the decoder is designed to handle multiple views of the input map data, while the encoder is trained to learn a shared representation that captures common features between all views.
If two nodes are closer together in the low dimensional space, then the two nodes may also have similar attributes or similar behavior patterns in the original graph. To better understand the relationships between nodes, reconstruct all views, it is necessary to predict whether there is a link between two nodes, expressed as follows:
wherein,representing the reconstructed view m, Z is the graph data node learning representation, W m Is a specific weight matrix for projecting Z into view m to predict the probability of connection between nodes, Z T Is a transpose of Z.
Total reconstruction loss function L r The expression of (2) is as follows:
wherein A is m Is the original data of view m and,representing a reconstructed view m; l (L) r The loss representing the M views constitutes an overall reconstruction loss.
And 4, acquiring optimal view codes and node embedding by a self-supervision module, training a model with more generalization capability for being applied to clustering tasks to improve the performance of the model, adopting 'highly-trusted nodes' q as a soft label to supervise the process of graph clustering, taking t distribution as a core to measure the similarity between each node and a central point, and introducing an auxiliary target matrix P to refine clustering for linking the embedding by using the clustering tasks.
The realization of the self-supervision module comprises the following steps:
and step 41, the embedding and clustering tasks are mutually supplemented and iteratively updated under a unified framework, so that the embedded representation and the clustering result can be mutually influenced, and the consistency of the clustering quality and the embedded representation is improved. The similarity between each node and the cluster center is measured using the t distribution as a core. q kr To assign node k to the probability of cluster r, the expression is as follows:
wherein mu r For embedding cluster centers, z k For the embedded representation of node k, μ r′ Representing the embedding of the r' th cluster center;
step 42, introducing an auxiliary target matrix P to optimize clustering by determining the probability q between each sample (i.e., node) and the cluster center kr Normalizing to define an element P in the auxiliary target matrix P kr The expression is as follows:
wherein f r The sum of the probabilities assigned to clusters r (i.e., cluster frequencies) for all samples; f (f) r′ The sum of the probabilities assigned to the cluster centers r' for all samples (i.e., nodes) is used to normalize the loss contribution for each cluster center.
The element Q in the soft distribution matrix Q can be obtained through calculation kr Element P in auxiliary target matrix P kr And then using the Frobenius norm to compare the distance or similarity between the two matrixes, the loss function in self-supervision clustering is as follows:
finally, a general objective function is obtained:
L＝L r +αL c on +βL clu (12)
wherein L is r To reconstruct the loss function as a whole, L con For cluster contrast loss function, L clu And (3) as a loss function in self-supervision clustering, alpha and beta are super parameters respectively, and the contribution degree of different losses to a final optimization target is adjusted.
The total objective function is adopted for network training, and when the network training result is good, the predicted clustering label y can be deduced r ：
Wherein argmax represents the index corresponding to the maximum value, and the position of the maximum value in the vector or matrix is found.
The foregoing is only a preferred embodiment of the invention, it being noted that: it will be apparent to those skilled in the art that various modifications and adaptations can be made without departing from the principles of the present invention, and such modifications and adaptations are intended to be comprehended within the scope of the invention.
Claims (5)
1. A multi-view clustering model implementation method based on graph contrast learning, the multi-view clustering model comprising: the system comprises a graph learning module, a graph comparison module and a self-supervision module;
the diagram learning module carries out convolution calculation on multi-view data by adopting a spectrum convolution function, carries out initialization representation on each node in the diagram by utilizing a diagram embedding method, and strengthens node embedding on the multi-view data by using an attention mechanism to obtain initialization data;
the graph comparison module learns similarity measurement between initialization data through a cluster comparison method, and the similarity measurement is used for explaining the structure of multi-view data; the comprehensive and accurate data structure information is obtained by comparing and learning the multiple views;
after the self-supervision module obtains the optimal view coding and node embedding, the process of graph clustering is supervised by adopting a highly-trusted node q as a soft label, the similarity between each node and a central point is measured by using t distribution as a core, and an auxiliary target matrix P is introduced to optimize the clustering task;
the method is characterized by comprising the following steps of:
s1, establishing a unified multi-view model, and defining the input and output of the multi-view model;
s2, selecting a public data set as initial data, and initializing the initial data through a constructed graph learning module;
s3, performing contrast learning on the initialized data through a constructed graph contrast module to obtain comprehensive and accurate data structure information;
and S4, clustering the acquired data structure information through a self-supervision module, constructing an objective function for training, stopping training when the set condition is met, and taking the obtained predicted clustering label as the output of the multi-view model.
2. The method for implementing a multi-view clustering model based on graph contrast learning according to claim 1, wherein in step S1, the inputs of the multi-view model are as follows:
define the multi-layer graph data as g= { V, E 1 ，…，E m ，X 1 ，…，X m -wherein V represents a set of N nodes; e (E) 1 ，…，E m Respectively connecting edges in m graph views, wherein the topological structure of the graph is represented by an adjacent matrix A; in a certain layer, if A ij =1, indicating that there is a border between the i-th node and the j-th node in the layer; if A ij =0, indicating that there is a border between the i-th node and the j-th node in the layer; { X 1 ，…，X m -representing a matrix of attributes of the node at different levels;a reconstructed adjacency matrix +.>Reconstructing an attribute matrix;
model output: clustering result y r 。
3. The method for implementing a multi-view clustering model based on graph contrast learning according to claim 1, wherein in step S2, the implementation steps of the graph learning module are as follows:
s21, inputting an adjacent matrix and an attribute matrix in initial data into an encoder, carrying out convolution calculation on a graph by using a spectrum convolution function, and carrying out nonlinear transformation on a convolved result by using an activation function to obtain node representation of a node in a current layer, wherein the coded input of a next layer is representation of a node in a previous layer and a corresponding adjacent matrix; for the first embedding of the mth layer, the expression is as follows:
wherein,is the first-1 embedding of the m-th layer, f () is a spectral convolution function, W (l) Is a trainable parameter matrix;wherein I is an identity matrix,>an adjacency matrix for the m-th layer; d (D) m A degree matrix, σ () is an activation function;
s22, embedding the multi-view data by using an attention mechanism to strengthen nodes, so as to obtain initialization data; correlation of node i with node j in layer lThe expression is:
wherein sigmoid () is a heuristic function,respectively trainable parameter matrix +.>And->Is a matrixI and j rows of (a);
for a pair ofNormalization was performed with the following expression:
finally obtaining the embedding of the node i:
wherein,representing the association between the ith node and the kth neighbor node of the first layer,/and->The set of neighboring nodes representing the node, exp () represents an exponential function that bases on the natural constant e.
4. The method for implementing a multi-view clustering model based on graph comparison learning according to claim 1, wherein in step S3, the implementing steps of the graph comparison module are as follows:
s31, taking the embedded representation of each node as a feature vector, dividing N nodes into G clusters by adopting a k-means algorithm, and simultaneously solving the central node of each cluster, wherein the central node is represented as { c } 1 ，…，c G }；
S32, for each node i, if the central node of the cluster to which the node i belongs is c i Then the embedding of the node with the central node of the cluster to which it belongs characterizes z' i Forming a positive sample, forming a negative sample by the embedded characterization of the node and the central nodes of other clusters, and calculating cluster contrast loss L con Gradient updating encoder, cluster contrast loss L con The expression of (2) is as follows:
wherein z is i Representing the embedding of node i, z' i Embedding, z 'for the node of the belonging central cluster' t Representing the embedding of central nodes of clusters to which other clusters belong, wherein sim () is a cosine similarity formula, τ is a temperature parameter, G is the number of the divided central clusters, e is a natural constant, and N is the total number of nodes;
s33, training the self-encoder by minimizing the sum of reconstruction errors of each view data, and predicting whether a link exists between two nodes, wherein the expression is as follows:
total reconstruction loss function L r ：
Wherein,representing a reconstructed view m, Z being a learned representation of the graph data nodes; w (W) m Is a training weight matrix for projecting Z into view m to predict the connection probability between nodes; z is Z T Is a transpose of Z, sigmoid () is a heuristic function, A m Is the original data of view m; l (L) r Loss, representing M views, constitutes the total reconstruction loss.
5. The method for implementing a multi-view clustering model based on graph comparison learning according to claim 1, wherein in step S4, the implementation process of the self-supervision module is as follows:
s41, measuring the similarity between each node and the clustering center point by adopting t distribution as a core, and q kr In order to assign the node k to the probability of the cluster r, the specific calculation method is as follows:
wherein mu r For embedding cluster centers, z k Embedding for node k; mu (mu) r′ Representing the embedding of the r' th cluster center;
s42, introducing an auxiliary target matrix P to optimize clustering, and defining an element P in the auxiliary target matrix P by normalizing the probability between each node and the clustering center kr The expression is as follows:
wherein f r The sum of probabilities assigned to clusters r for all nodes; f (f) r′ The sum of probabilities assigned to the cluster centers r' for all nodes;
from q kr Obtaining soft distribution matrix Q, p kr Obtaining the element P in the auxiliary target matrix P kr Then the loss function L in self-supervision clustering clu The method comprises the following steps:
the overall objective function is:
L＝L r +αL con +βL clu
wherein alpha and beta are respectively super parameters, and the contribution degree of different losses to a final optimization target is regulated;
performing network training by adopting a total objective function L, and deducing a predicted clustering label y when a network training result meets a set condition r ：
Wherein argmax represents an index corresponding to the maximum value.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202310957418.0A CN117036760A (en) | 2023-07-31 | 2023-07-31 | Multi-view clustering model implementation method based on graph comparison learning |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202310957418.0A CN117036760A (en) | 2023-07-31 | 2023-07-31 | Multi-view clustering model implementation method based on graph comparison learning |
Publications (1)
| Publication Number | Publication Date |
|---|---|
| CN117036760A true CN117036760A (en) | 2023-11-10 |
Family
ID=88638373
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN202310957418.0A Pending CN117036760A (en) | 2023-07-31 | 2023-07-31 | Multi-view clustering model implementation method based on graph comparison learning |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN117036760A (en) |
Cited By (2)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN117292162A (en) * | 2023-11-27 | 2023-12-26 | 烟台大学 | Target tracking method, system, equipment and medium for multi-view image clustering |
| CN117934891A (en) * | 2024-03-25 | 2024-04-26 | 南京信息工程大学 | Image contrast clustering method and system based on graph structure |
-
2023
- 2023-07-31 CN CN202310957418.0A patent/CN117036760A/en active Pending
Cited By (4)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN117292162A (en) * | 2023-11-27 | 2023-12-26 | 烟台大学 | Target tracking method, system, equipment and medium for multi-view image clustering |
| CN117292162B (en) * | 2023-11-27 | 2024-03-08 | 烟台大学 | Target tracking method, system, equipment and medium for multi-view image clustering |
| CN117934891A (en) * | 2024-03-25 | 2024-04-26 | 南京信息工程大学 | Image contrast clustering method and system based on graph structure |
| CN117934891B (en) * | 2024-03-25 | 2024-06-07 | 南京信息工程大学 | Image contrast clustering method and system based on graph structure |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CN109086805B (en) | Clustering method based on deep neural network and pairwise constraints | |
| CN112966114B (en) | Literature classification method and device based on symmetrical graph convolutional neural network | |
| CN110347932B (en) | Cross-network user alignment method based on deep learning | |
| CN117036760A (en) | Multi-view clustering model implementation method based on graph comparison learning | |
| CN114418954A (en) | Mutual learning-based semi-supervised medical image segmentation method and system | |
| CN113157957A (en) | Attribute graph document clustering method based on graph convolution neural network | |
| CN112861976B (en) | Sensitive image identification method based on twin graph convolution hash network | |
| Cai et al. | A robust interclass and intraclass loss function for deep learning based tongue segmentation | |
| CN111488498A (en) | Node-graph cross-layer graph matching method and system based on graph neural network | |
| Zhou et al. | Improved cross-label suppression dictionary learning for face recognition | |
| CN115841607A (en) | Brain network structure and similarity joint learning method based on graph attention network | |
| CN112883756A (en) | Generation method of age-transformed face image and generation countermeasure network model | |
| CN111582506A (en) | Multi-label learning method based on global and local label relation | |
| CN114743037A (en) | Deep medical image clustering method based on multi-scale structure learning | |
| CN117152459A (en) | Image detection method, device, computer readable medium and electronic equipment | |
| CN114266321A (en) | Weak supervision fuzzy clustering algorithm based on unconstrained prior information mode | |
| Lu et al. | Soft-orthogonal constrained dual-stream encoder with self-supervised clustering network for brain functional connectivity data | |
| CN116452241B (en) | User loss probability calculation method based on multi-mode fusion neural network | |
| CN117349494A (en) | Graph classification method, system, medium and equipment for space graph convolution neural network | |
| CN115359297B (en) | Classification method, system, electronic equipment and medium based on higher-order brain network | |
| CN116188428A (en) | Bridging multi-source domain self-adaptive cross-domain histopathological image recognition method | |
| CN115588487A (en) | Medical image data set making method based on federal learning and generation countermeasure network | |
| CN115240782A (en) | Drug attribute prediction method, device, electronic device and storage medium | |
| CN108304546B (en) | Medical image retrieval method based on content similarity and Softmax classifier | |
| Wei et al. | Negatives make a positive: An embarrassingly simple approach to semi-supervised few-shot learning |
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 |