CN114372506A - Network structure node category prediction method and system - Google Patents
Network structure node category prediction method and system Download PDFInfo
- Publication number
- CN114372506A CN114372506A CN202111502872.4A CN202111502872A CN114372506A CN 114372506 A CN114372506 A CN 114372506A CN 202111502872 A CN202111502872 A CN 202111502872A CN 114372506 A CN114372506 A CN 114372506A
- Authority
- CN
- China
- Prior art keywords
- vector
- source network
- loss function
- network
- representation vector
- 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
-
- 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
- G06F18/2411—Classification techniques relating to the classification model, e.g. parametric or non-parametric approaches based on the proximity to a decision surface, e.g. support vector machines
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
- G06N20/00—Machine learning
- G06N20/10—Machine learning using kernel methods, e.g. support vector machines [SVM]
-
- 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
- G06N3/084—Backpropagation, e.g. using gradient descent
Landscapes
- Engineering & Computer Science (AREA)
- Theoretical Computer Science (AREA)
- Data Mining & Analysis (AREA)
- Physics & Mathematics (AREA)
- Evolutionary Computation (AREA)
- Artificial Intelligence (AREA)
- General Physics & Mathematics (AREA)
- General Engineering & Computer Science (AREA)
- Computer Vision & Pattern Recognition (AREA)
- Life Sciences & Earth Sciences (AREA)
- Software Systems (AREA)
- Computing Systems (AREA)
- Bioinformatics & Cheminformatics (AREA)
- Bioinformatics & Computational Biology (AREA)
- Evolutionary Biology (AREA)
- Mathematical Physics (AREA)
- Medical Informatics (AREA)
- Health & Medical Sciences (AREA)
- Biomedical Technology (AREA)
- Biophysics (AREA)
- Computational Linguistics (AREA)
- General Health & Medical Sciences (AREA)
- Molecular Biology (AREA)
- Management, Administration, Business Operations System, And Electronic Commerce (AREA)
Abstract
The application discloses a method and a system for predicting the class of a network structure node, wherein the method for predicting the class of the network structure node comprises the following steps: the classifier generates a cross entropy loss function by using a source network structure, calculates the sinkhorn distance between a source network expression vector and a target network expression vector through a sinkhorn distance formula, and adds the cross entropy loss function and the sinkhorn distance calculation result to obtain a GAT overall loss function; classifying the source network representation vectors through an attention network according to the GAT overall loss function to obtain high-dimensional source network representation vectors; mapping the high-dimensional source network representation vector through TLDA (transport layer object) to obtain a source network representation vector after dimension reduction and a target network representation vector after dimension reduction; and after the dividing line of the linear mapping matrix is obtained by a maximum interval method, the support vector is obtained according to the dividing line, and the target network representation vector after dimension reduction is classified and predicted through the dividing line and the support vector to obtain a node class prediction result of the target network structure.
Description
Technical Field
The present application relates to the field of network analysis technologies, and in particular, to a method and a system for predicting a class of a network structure node.
Background
A network is a data structure that is ubiquitous in the real world. Accurate and efficient network analysis is the key to realizing multiple intelligent network services. However, most of the existing network analysis methods are developed for a single network, a large amount of labeled data is needed, and the data acquisition cost is high and the time consumption is long. Representative classifiers include Support Vector Machines (SVMs), neural networks, and logistic regression. In recent years, the development of graph embedding technology has greatly promoted the completion of a node classification task that represents nodes with low-dimensional vectors while maintaining the close proximity of structures and attributes. In addition, some approaches utilize a unified framework to learn class-specific embedding of nodes through joint optimization of the embedding module and the prediction module. However, most existing node classification approaches learn and predict node classes on a single network and require large amounts of labeled data, which is both expensive and difficult to acquire.
In recent years, transfer learning has been widely accepted as an effective method and successfully applied to solve problems of medical image classification, cross-domain retrieval, object instance segmentation, and the like. But since most existing migration learning approaches require independent co-distributed (IID) data in both the source and target domains, it solves this problem by reusing models trained in supervised tasks, whereas data of non-euclidean graph structures do not. Most existing cross-network analysis studies are based on a large number of anchor links transmitting knowledge across the network, but this is not the case in most cases. Therefore, how to accurately classify the knowledge into nodes in network analysis becomes an urgent problem to be solved.
Disclosure of Invention
The embodiment of the application provides a method and a system for predicting the class of a network structure node, and at least solves the problems that in the process of predicting the class of a network node, accurate node classification of knowledge learned in a label source network cannot be realized, the effect of distinguishing labels of cross-network node classification tasks is low, cross-network analysis needs to transmit knowledge across networks based on a large number of anchor links, and the like.
The invention provides a network structure node category prediction method, which comprises the following steps:
GAT overall loss function acquisition step: the classifier generates a cross entropy loss function by using a source network structure, calculates the sinkhorn distance between a source network representation vector and a target network representation vector through a sinkhorn distance formula, and adds the cross entropy loss function and the sinkhorn distance calculation result to obtain a GAT overall loss function;
a representative vector acquisition step: classifying the source network representation vectors through an attention network according to the GAT overall loss function to obtain high-dimensional source network representation vectors;
and a representation vector dimensionality reduction step: mapping the high-dimensional source network representation vector through TLDA to obtain a source network representation vector after dimension reduction and a target network representation vector after dimension reduction;
a prediction result obtaining step: and after the parting line of the linear mapping matrix is obtained by a method of maximizing the interval, a support vector is obtained according to the parting line, and the target network representation vector after dimensionality reduction is classified and predicted through the parting line and the support vector to obtain a node class prediction result of the target network structure.
In the above method for predicting a node class of a network structure, after the source network structure and the target network structure are extracted through the GAT machine learning network, the first classifier (relu) and the second classifier (softmax) use the source network structure to generate the cross entropy loss function.
In the foregoing method for predicting a category of a network structure node, the step of obtaining the GAT total loss function further includes:
obtaining the source network representation vector of the source network structure and the target network representation vector of the target network structure by an attention-seeking neural network;
and calculating the sinkhorn distance between the source network representation vector and the target network representation vector through the sinkhorn distance formula to obtain a sinkhorn distance calculation result.
In the foregoing method for predicting a category of a network structure node, the step of obtaining the GAT total loss function further includes:
and adding the cross entropy loss function and the sinkhorn distance calculation result to obtain the GAT overall loss function.
In the method for predicting a type of a network structure node, the obtaining step of the expression vector includes:
constructing an attention network by the gat model through stacking graph attention layers;
and classifying the source network representation vector through the attention network according to the GAT overall loss function to obtain the high-dimensional source network representation vector.
In the above method for predicting a class of a node in a network structure, the step of reducing the dimension of the expression vector includes:
and mapping the high-dimensional source network representation vector to the same plane as the low-dimensional representation vector of the target network structure through the TLDA to obtain a mapping result.
In the above method for predicting a class of a network structure node, the step of reducing the dimension of the expression vector further includes:
and clustering the mapping result to obtain the source network representation vector after dimensionality reduction and the target network representation vector after dimensionality reduction.
In the method for predicting a type of a network structure node, the step of obtaining a prediction result includes:
pre-training the third classifier through the source network structure to obtain a pre-trained final classifier;
and training the source network representation vector after dimension reduction through the final classifier to obtain a source network projection representation vector.
In the method for predicting a type of a network structure node, the step of obtaining a prediction result further includes:
converting the source network projection representation vector by a kernel function method to obtain the linear projection matrix;
after the maximum partition line of the linear projection matrix is obtained by the maximum interval method, the support vector is obtained according to the partition line;
and performing classified prediction on the reduced-dimension target network representation vector through the dividing line and the support vector to obtain the node class prediction result of the target network structure.
The present invention also provides a network structure node type prediction system, which is applicable to the network structure node type prediction method described above, and the network structure node type prediction system includes:
GAT total loss function acquisition unit: the classifier generates a cross entropy loss function by using a source network structure, calculates the sinkhorn distance between a source network representation vector and a target network representation vector through a sinkhorn distance formula, and adds the cross entropy loss function and the sinkhorn distance calculation result to obtain a GAT overall loss function;
a representative vector acquisition unit: classifying the source network representation vectors through an attention network according to the GAT overall loss function to obtain high-dimensional source network representation vectors;
representing a vector dimension reduction unit: mapping the high-dimensional source network representation vector through TLDA to obtain a source network representation vector after dimension reduction and a target network representation vector after dimension reduction;
a prediction result acquisition unit: and after the parting line of the linear mapping matrix is obtained by a method of maximizing the interval, a support vector is obtained according to the parting line, and the target network representation vector after dimensionality reduction is classified and predicted through the parting line and the support vector to obtain a node class prediction result of the target network structure.
Compared with the related technology, the network structure node type prediction method and the network structure node type prediction system provided by the invention provide a learning method of network invariant representation and label discriminant representation based on graph embedding and linear discriminant analysis. The method solves the challenging task of cross-network node classification without anchoring links through an unsupervised network transfer learning model; learning network invariant embedding under supervision of Sinkhorn distance; the nodes are projected to the label distinguishing subspace by adopting a Transferable Linear Discriminant Analysis (TLDA) method, so that the separability of the node representation is improved, and compared with the traditional LDA method, the projection using the TLDA method has higher label distinguishability and is beneficial to the cross-network node classification task; meanwhile, the distribution divergence loss plays an important role in the proposed method, and due to the supervision of the Sinkhorn distance, the GAT can learn network invariant embedding, and the classifier trained on the source network can be effectively applied to the target network. The invention transfers knowledge from a source network to a target network through the invariant embedding of a learning network, and maps the embedding to an identification subspace, thereby improving the classification performance of the nodes.
The details of one or more embodiments of the application are set forth in the accompanying drawings and the description below to provide a more thorough understanding of the application.
Drawings
The accompanying drawings, which are included to provide a further understanding of the application and are incorporated in and constitute a part of this application, illustrate embodiment(s) of the application and together with the description serve to explain the application and not to limit the application in a non-limiting sense. In the drawings:
fig. 1 is a flow chart of a network structure node class prediction method according to an embodiment of the present application;
FIG. 2 is a framework diagram of a network fabric node class prediction flow according to an embodiment of the present application;
fig. 3 is a schematic structural diagram of a network structure node type prediction system according to the present invention.
Wherein the reference numerals are:
GAT total loss function acquisition unit: 51;
a representative vector acquisition unit: 52;
representing a vector dimension reduction unit: 53;
a prediction result acquisition unit: 54.
Detailed Description
In order to make the objects, technical solutions and advantages of the present application more apparent, the present application will be described and illustrated below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative of the present application and are not intended to limit the present application. All other embodiments obtained by a person of ordinary skill in the art based on the embodiments provided in the present application without any inventive step are within the scope of protection of the present application.
It is obvious that the drawings in the following description are only examples or embodiments of the present application, and that it is also possible for a person skilled in the art to apply the present application to other similar contexts on the basis of these drawings without inventive effort. Moreover, it should be appreciated that such a development effort might be complex and tedious, but would nevertheless be a routine undertaking of design, fabrication, and manufacture for those of ordinary skill having the benefit of this disclosure, and is not intended to limit the scope of this disclosure.
Reference in the specification to "an embodiment" means that a particular feature, structure, or characteristic described in connection with the embodiment can be included in at least one embodiment of the specification. The appearances of the phrase in various places in the specification are not necessarily all referring to the same embodiment, nor are separate or alternative embodiments mutually exclusive of other embodiments. Those of ordinary skill in the art will explicitly and implicitly appreciate that the embodiments described herein may be combined with other embodiments without conflict.
Unless defined otherwise, technical or scientific terms referred to herein shall have the ordinary meaning as understood by those of ordinary skill in the art to which this application belongs. Reference to "a," "an," "the," and similar words throughout this application are not to be construed as limiting in number, and may refer to the singular or the plural. The use of the terms "including," "comprising," "having," and any variations thereof herein, is intended to cover a non-exclusive inclusion; for example, a process, method, system, article, or apparatus that comprises a list of steps or modules (elements) is not limited to the listed steps or elements, but may include other steps or elements not expressly listed or inherent to such process, method, article, or apparatus. Reference to "connected," "coupled," and the like in this application is not intended to be limited to physical or mechanical connections, but may include electrical connections, whether direct or indirect. The term "plurality" as referred to herein means two or more. "and/or" describes an association relationship of associated objects, meaning that three relationships may exist, for example, "A and/or B" may mean: a exists alone, A and B exist simultaneously, and B exists alone. The character "/" generally indicates that the former and latter associated objects are in an "or" relationship. Reference herein to the terms "first," "second," "third," and the like, are merely to distinguish similar objects and do not denote a particular ordering for the objects.
The invention aims at two node networks, which are divided into a source network and a target network, and the node category of the target network is necessarily the same as that of the source network. In addition, there is no shared common node between the source network and the target network, nor edges connecting the source network and the target network nodes. The invention utilizes two parameter shared graph attention network gat to learn network invariant embedding by minimizing Sinkhorn distance, and then adopts TLDA to map the embedding into the label identification subspace.
The present invention will be described with reference to specific examples.
Example one
The embodiment provides a network structure node category prediction method. Referring to fig. 1 to 2, fig. 1 is a flowchart illustrating a method for predicting a network node type according to an embodiment of the present disclosure; fig. 2 is a framework diagram of a network structure node class prediction process according to an embodiment of the present application, and as shown in fig. 1 to fig. 2, the network structure node class prediction method includes the following steps:
GAT total loss function acquisition step S1: the classifier generates a cross entropy loss function by using a source network structure, calculates the sinkhorn distance between a source network expression vector and a target network expression vector through a sinkhorn distance formula, and adds the cross entropy loss function and the sinkhorn distance calculation result to obtain a GAT overall loss function;
representative vector acquisition step S2: classifying the source network representation vectors through an attention network according to the GAT overall loss function to obtain high-dimensional source network representation vectors;
representation vector dimension reduction step S3: mapping the high-dimensional source network representation vector through TLDA to obtain a source network representation vector after dimension reduction and a target network representation vector after dimension reduction;
prediction result acquisition step S4: and after the parting line of the linear mapping matrix is obtained by a method of maximizing the interval, the support vector is obtained according to the parting line, and the target network representation vector after dimension reduction is classified and predicted through the parting line and the support vector to obtain a node class prediction result of the target network structure.
In an embodiment, the GAT total loss function obtaining step S1 includes:
after extracting a source network structure and a target network structure through a GAT machine learning network, a first classifier (relu) and a second classifier (softmax) generate a cross entropy loss function by using the source network structure;
obtaining a source network representation vector of a source network structure and a target network representation vector of a target network structure through an attention-seeking neural network;
calculating the sinkhorn distance between the source network expression vector and the target network expression vector through a sinkhorn distance formula to obtain a sinkhorn distance calculation result;
and adding the cross entropy loss function and the sinkhorn distance calculation result to obtain a GAT overall loss function.
In the specific implementation, after a source network structure and a target network structure are extracted through a GAT machine learning network, a first classifier (relu) and a second classifier (softmax) generate a cross entropy loss function by using the source network structure, a source network representation vector of the source network structure and a target network representation vector of the target network structure are obtained through an attention-seeking neural network, a sinkhorn distance between the source network representation vector and the target network representation vector is calculated through a sinkhorn distance formula, and after a sinkhorn distance calculation result is obtained, the cross entropy loss function and the sinkhorn distance calculation result are added to obtain a GAT total loss function.
A graph embedding method based on the GAT model is employed given a source network and a target network. In general, GAT updates the representation of a node by focusing on the neighborhood characteristics of the node, which may be expressed as:
wherein
Represents a hidden representation of the ith node of the l-th layer, σ (·) represents an activation function ReLU (·) max (0,), W(l)Is a learnable linear transformation matrix, N, shared by all nodesiIs a neighbor set, α, of the ith node containing itselfijAre the weights of the ith and jth nodes,
in the formula, is
The attention coefficient between the ith and jth nodes is calculated by the following equation:
is shown inTThe operation of the transfer is carried out,
representing a vector concatenation. Transforming the matrix W(l)And limiting the matrix to be a diagonal matrix so as to reduce the number of parameters and improve generalization capability.
Training of GATs involves supervision of two aspects, the first being to make the node representation labels distinguishable based on known source labels, and the second being to learn distributed network invariant embedding, enabling the knowledge learned from the source network to be migrated to the target network.
First, a node classifier is added on top of the embedding module to incorporate node labels from the source network. Specifically, the output of the classifier is calculated by the following formula:
yi∈RCrepresents the prediction probability of the ith node over all classes, phi (-) is the ReLu function, WcAnd
are trainable parameters of the classifier. Training the classifier using cross entropy loss, which is defined as:
wherein | VsAnd | is the size of the source network node set. y isikIs the basic true label of the ith node, if the ith node belongs to the kth class, then yik＝1；Otherwise, yik＝0。
Representing the prediction probability of the ith node belonging to the kth class. The loss function is also used to optimize the embedding.
The Sinkhorn distance is used to measure the closeness between the source network and target network embedding. The output of the GAT is a normalized embedding. Target embeddings similar in distribution to the source embeddings are also adjacent to the source embeddings in vector space. Embedding H1And H2The Sinkhorn distance between is defined as:
wherein <, > represents a Frobenius dot product. M is a transition cost matrix defined as:
where e (T.) is an entropy function. r and c are weights from nodes of the source and target networks, which are set in proportion to their degrees.
For a batch of samples with n nodes in the source and target networks, the Sinkhorn distance between their embeddings is calculated by algorithm 1, which is the entropy constrained lagrangian multiplier. For each batch, the Sinkhorn distance was calculated for T replicates.
Algorithm 1: calculating the Sinhoen distance
Inputting: m, r, c, μ, T
And (3) outputting: d (H)s,Ht)
K＝e-μM
When T is 1 → T, run
u＝r./Kv
v＝c./KTu
The dispersion loss is then defined as the Sinkhorn loss, expressed as:
the total loss of the embedded module optimization is
Wherein λ1The relative weight of the distribution divergence loss is adopted, and the total loss achieves the purpose of learning the embedded vector which has good classification separability and minimized distribution difference among different networks through the combination of different loss functions.
In an embodiment, the representative vector obtaining step S2 includes:
constructing an attention network by the gat model through stacking graph attention layers;
and classifying the source network representation vectors through the attention network according to the GAT overall loss function to obtain high-dimensional source network representation vectors.
In specific implementation, after the gat model constructs the attention network through the stacked graph attention layer, the attention network parameters are set, wherein the main parameters are as follows:
aij is the attention coefficient;
h', h is node characteristics;
f', F is a characteristic dimension;
w is a change weight matrix;
k is the number of attention mechanisms;
after the parameters are set, the attention layer adopts an 8-head attention mechanism to stably learn, K independent attention mechanisms are applied to calculate the hidden states and the attention coefficients, a back propagation mechanism adopts a gradient descent method through a loss function, source network representation vectors are classified through feature vector calculation in the attention mechanism, and high-dimensional source network representation vectors are obtained.
In an embodiment, the representative vector dimension reduction step S3 includes:
mapping the high-dimensional source network representation vector to a plane which is the same as the low-dimensional representation vector of the target network structure through TLDA (transport layer object) to obtain a mapping result;
and clustering the mapping result to obtain a source network representation vector after dimension reduction and a target network representation vector after dimension reduction.
In specific implementation, the high-dimensional source network representation vector is mapped to the same plane as the low-dimensional representation vector of the target network structure through TLDA, after the mapping result is obtained, the mapping result is clustered, the unlabeled network converges to the clustering of the labeled network, and then the source network representation vector after dimension reduction and the target network representation vector after dimension reduction are obtained.
The core idea of TLDA (transferable linear discriminant analysis) is to reduce the variation distributed within the same subspace, to extend the distance between different subspaces, while recovering the low-rank structure of different domains during linear discriminant analysis. Thus, the loss function that projects the embedding into different subspaces for cross-network node classification is defined as:
s.t.‖P‖2＝1
wherein P represents a linear projection matrix, SwAnd SbRepresenting intra-class and inter-class scattering matrices, H represents the concatenation embedding of the source and target networks, | · |)*Denotes the nuclear norm, λ2Is a weight override parameter. Pt+1＝Pt-ρΔPtWhere Pt is P at the tth iteration and ρ is the step size. P is renormalized after each iteration. The minor gradient is calculated by the following formula:
wherein Δ | · |)*Is the sub-differential of the kernel specification.
In an embodiment, the prediction result acquisition step S4 includes:
pre-training the third classifier through a source network structure to obtain a final classifier after pre-training;
training the source network representation vector after dimension reduction through a final classifier to obtain a source network projection representation vector;
converting the source network projection expression vector by a kernel function method to obtain a linear projection matrix;
obtaining a maximum partition line of the linear projection matrix by a maximum partition method, and then obtaining a support vector according to the partition line;
and carrying out classified prediction on the reduced-dimension target network representation vector through the partition line and the support vector to obtain a node class prediction result of the target network structure.
In the specific implementation, a classifier SVM (or any classifier) is pre-trained through a source network structure to obtain a pre-trained final classifier, a reduced-dimension source network representation vector is trained through the final classifier to obtain a source network projection representation vector, the source network projection representation vector is converted through a kernel function method to obtain a linear projection matrix, a maximum partition line of the linear projection matrix is obtained through a maximum partition method, simple scaling operation is carried out on the linear projection matrix according to the partition line, specifically, after a kernel function is set, optimal cross-validation parameters C and g are set, optimal parameters C and g are called out to train the whole training set to obtain a support vector machine model, and the reduced-dimension target network representation vector is classified and predicted through the partition line and the support vector model, and obtaining a node type prediction result of the target network structure.
Referring to fig. 3, fig. 3 is a schematic structural diagram of a network structure node type prediction system according to a second embodiment of the present invention. As shown in fig. 3, the network structure node type prediction system according to the present invention is applied to the above network structure node type prediction method, and includes:
the GAT total loss function acquisition unit 51: the classifier generates a cross entropy loss function by using a source network structure, calculates the sinkhorn distance between a source network expression vector and a target network expression vector through a sinkhorn distance formula, and adds the cross entropy loss function and the sinkhorn distance calculation result to obtain a GAT overall loss function;
the representative vector acquisition unit 52: classifying the source network representation vectors through an attention network according to the GAT overall loss function to obtain high-dimensional source network representation vectors;
representation vector dimension reduction unit 53: mapping the high-dimensional source network representation vector through TLDA to obtain a source network representation vector after dimension reduction and a target network representation vector after dimension reduction;
prediction result acquisition unit 54: and after the parting line of the linear mapping matrix is obtained by a method of maximizing the interval, the support vector is obtained according to the parting line, and the target network representation vector after dimension reduction is classified and predicted through the parting line and the support vector to obtain a node class prediction result of the target network structure.
In summary, the present invention represents nodes in two networks by a pair of graph attention networks (GAT) with parameter sharing, minimizes classification errors on a source network and Sinkhorn distribution distance between source domain and target domain embeddings, uses Transferable Linear Discriminant Analysis (TLDA), learns label discriminant projection by recovering low rank structures of different distribution data and expanding distances between different subspaces, trains an edge-maximized SVM classifier on the source network, and predicts class labels of target nodes by using the classifier.
The above-mentioned embodiments only express several embodiments of the present application, and the description thereof is specific and detailed, but not construed as limiting the scope of the invention. It should be noted that, for a person skilled in the art, several variations and modifications can be made without departing from the concept of the present application, which falls within the scope of protection of the present application. Therefore, the protection scope of the present patent application shall be subject to the protection scope of the appended claims.
Claims (10)
1. A network structure node class prediction method is characterized in that the network structure node class prediction method comprises the following steps:
GAT overall loss function acquisition step: the classifier generates a cross entropy loss function by using a source network structure, calculates the sinkhorn distance between a source network expression vector and a target network expression vector through a sinkhorn distance formula, and adds the cross entropy loss function and the sinkhorn distance calculation result to obtain a GAT overall loss function;
a representative vector acquisition step: classifying the source network representation vectors through an attention network according to the GAT overall loss function to obtain high-dimensional source network representation vectors;
and a representation vector dimensionality reduction step: mapping the high-dimensional source network representation vector through TLDA to obtain a source network representation vector after dimension reduction and a target network representation vector after dimension reduction;
a prediction result obtaining step: and after the parting line of the linear mapping matrix is obtained by a method of maximizing the interval, a support vector is obtained according to the parting line, and the target network representation vector after dimensionality reduction is classified and predicted through the parting line and the support vector to obtain a node class prediction result of the target network structure.
2. The method according to claim 1, wherein the GAT global loss function obtaining step comprises:
upon extracting the source and target network structures through a GAT machine learning network, a first classifier (relu) and a second classifier (softmax) generate the cross entropy loss function using the source network structure.
3. The method according to claim 2, wherein the GAT global loss function obtaining step further comprises:
obtaining the source network representation vector of the source network structure and the target network representation vector of the target network structure by an attention-seeking neural network;
and calculating the sinkhorn distance between the source network representation vector and the target network representation vector through the sinkhorn distance formula to obtain a sinkhorn distance calculation result.
4. The method according to claim 1, wherein the GAT global loss function obtaining step further comprises:
and adding the cross entropy loss function and the sinkhorn distance calculation result to obtain the GAT overall loss function.
5. The method of claim 1, wherein the step of obtaining the representation vector comprises:
constructing an attention network by the gat model through stacking graph attention layers;
and classifying the source network representation vector through the attention network according to the GAT overall loss function to obtain the high-dimensional source network representation vector.
6. The method of claim 2, wherein the step of reducing the dimension of the representation vector comprises:
and mapping the high-dimensional source network representation vector to the same plane as the low-dimensional representation vector of the target network structure through the TLDA to obtain a mapping result.
7. The method of claim 6, wherein the step of reducing the dimension of the representation vector further comprises:
and clustering the mapping result to obtain the source network representation vector after dimensionality reduction and the target network representation vector after dimensionality reduction.
8. The method according to claim 1, wherein the prediction result obtaining step includes:
pre-training the third classifier through the source network structure to obtain a pre-trained final classifier;
and training the source network representation vector after dimension reduction through the final classifier to obtain a source network projection representation vector.
9. The method according to claim 7, wherein the step of obtaining the prediction result further comprises:
converting the source network projection representation vector by a kernel function method to obtain the linear projection matrix;
after the maximum partition line of the linear projection matrix is obtained by the maximum interval method, the support vector is obtained according to the partition line;
and performing classified prediction on the reduced-dimension target network representation vector through the dividing line and the support vector to obtain the node class prediction result of the target network structure.
10. A network fabric node class prediction system, comprising:
GAT total loss function acquisition unit: the classifier generates a cross entropy loss function by using a source network structure, calculates the sinkhorn distance between a source network expression vector and a target network expression vector through a sinkhorn distance formula, and adds the cross entropy loss function and the sinkhorn distance calculation result to obtain a GAT overall loss function;
a representative vector acquisition unit: classifying the source network representation vectors through an attention network according to the GAT overall loss function to obtain high-dimensional source network representation vectors;
representing a vector dimension reduction unit: mapping the high-dimensional source network representation vector through TLDA to obtain a source network representation vector after dimension reduction and a target network representation vector after dimension reduction;
a prediction result acquisition unit: and after the parting line of the linear mapping matrix is obtained by a method of maximizing the interval, a support vector is obtained according to the parting line, and the target network representation vector after dimensionality reduction is classified and predicted through the parting line and the support vector to obtain a node class prediction result of the target network structure.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202111502872.4A CN114372506A (en) | 2021-12-09 | 2021-12-09 | Network structure node category prediction method and system |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202111502872.4A CN114372506A (en) | 2021-12-09 | 2021-12-09 | Network structure node category prediction method and system |
Publications (1)
| Publication Number | Publication Date |
|---|---|
| CN114372506A true CN114372506A (en) | 2022-04-19 |
Family
ID=81139888
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN202111502872.4A Pending CN114372506A (en) | 2021-12-09 | 2021-12-09 | Network structure node category prediction method and system |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN114372506A (en) |
Cited By (1)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN115100491A (en) * | 2022-08-25 | 2022-09-23 | 山东省凯麟环保设备股份有限公司 | Abnormal robust segmentation method and system for complex automatic driving scene |
-
2021
- 2021-12-09 CN CN202111502872.4A patent/CN114372506A/en active Pending
Cited By (3)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN115100491A (en) * | 2022-08-25 | 2022-09-23 | 山东省凯麟环保设备股份有限公司 | Abnormal robust segmentation method and system for complex automatic driving scene |
| CN115100491B (en) * | 2022-08-25 | 2022-11-18 | 山东省凯麟环保设备股份有限公司 | Abnormal robust segmentation method and system for complex automatic driving scene |
| US11954917B2 (en) | 2022-08-25 | 2024-04-09 | Shandong Kailin Environmental Protection Equipment Co., Ltd. | Method of segmenting abnormal robust for complex autonomous driving scenes and system thereof |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| Chi et al. | Infogcn: Representation learning for human skeleton-based action recognition | |
| US11586905B2 (en) | Systems and methods for customizing kernel machines with deep neural networks | |
| Liao et al. | Learning deep parsimonious representations | |
| Bouchachia et al. | Enhancement of fuzzy clustering by mechanisms of partial supervision | |
| Peng et al. | Active transfer learning | |
| CN107944410B (en) | Cross-domain facial feature analysis method based on convolutional neural network | |
| Li et al. | Max-margin zero-shot learning for multi-class classification | |
| CN109784405A (en) | Cross-module state search method and system based on pseudo label study and semantic consistency | |
| Das et al. | NAS-SGAN: a semi-supervised generative adversarial network model for atypia scoring of breast cancer histopathological images | |
| Zhao et al. | Few-shot class-incremental learning via class-aware bilateral distillation | |
| Liu et al. | High‐Performance Machine Learning for Large‐Scale Data Classification considering Class Imbalance | |
| CN116188941A (en) | Manifold regularized width learning method and system based on relaxation annotation | |
| CN117494051A (en) | Classification processing method, model training method and related device | |
| CN114372506A (en) | Network structure node category prediction method and system | |
| Ren et al. | Deep metric learning via subtype fuzzy clustering | |
| Rastin et al. | A stacking weighted k-nearest neighbour with thresholding | |
| Zhang et al. | Projected-prototype based classifier for text categorization | |
| CN112668633A (en) | Adaptive graph migration learning method based on fine granularity field | |
| Wang et al. | Distance correlation autoencoder | |
| CN114818979A (en) | Noise-containing multi-label classification method based on maximum interval mechanism | |
| Zhao et al. | Distance metric learning based on the class center and nearest neighbor relationship | |
| Dornaika et al. | Simultaneous label inference and discriminant projection estimation through adaptive self-taught graphs | |
| Vijaya Saradhi et al. | Incremental maximum margin clustering | |
| Trentin et al. | Unsupervised nonparametric density estimation: A neural network approach | |
| Acharya et al. | Active multitask learning using supervised and shared latent topics |
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 | ||
| TA01 | Transfer of patent application right |
Effective date of registration: 20230317 Address after: 266000 26F, block B, Chuangye building, high tech Zone, Qingdao, Shandong Province Applicant after: QINGDAO ACADEMY OF INTELLIGENT INDUSTRIES Applicant after: STATE GRID ZHEJIANG ELECTRIC POWER Co.,Ltd. Address before: 266000 26F, block B, Chuangye building, high tech Zone, Qingdao, Shandong Province Applicant before: QINGDAO ACADEMY OF INTELLIGENT INDUSTRIES |
|
| TA01 | Transfer of patent application right |