CN111581297A - Symbolic network symbol prediction method based on graph convolution network - Google Patents
Symbolic network symbol prediction method based on graph convolution network Download PDFInfo
- Publication number
- CN111581297A CN111581297A CN202010325465.XA CN202010325465A CN111581297A CN 111581297 A CN111581297 A CN 111581297A CN 202010325465 A CN202010325465 A CN 202010325465A CN 111581297 A CN111581297 A CN 111581297A
- Authority
- CN
- China
- Prior art keywords
- network
- sub
- edge
- negative
- positive
- 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
- G06F16/00—Information retrieval; Database structures therefor; File system structures therefor
- G06F16/20—Information retrieval; Database structures therefor; File system structures therefor of structured data, e.g. relational data
- G06F16/28—Databases characterised by their database models, e.g. relational or object models
- G06F16/284—Relational databases
- G06F16/288—Entity relationship models
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
- G06N3/00—Computing arrangements based on biological models
- G06N3/02—Neural networks
- G06N3/04—Architecture, e.g. interconnection topology
- G06N3/045—Combinations of networks
Landscapes
- Engineering & Computer Science (AREA)
- Theoretical Computer Science (AREA)
- Databases & Information Systems (AREA)
- Physics & Mathematics (AREA)
- Data Mining & Analysis (AREA)
- General Physics & Mathematics (AREA)
- General Engineering & Computer Science (AREA)
- General Health & Medical Sciences (AREA)
- Computational Linguistics (AREA)
- Evolutionary Computation (AREA)
- Biophysics (AREA)
- Molecular Biology (AREA)
- Computing Systems (AREA)
- Biomedical Technology (AREA)
- Artificial Intelligence (AREA)
- Mathematical Physics (AREA)
- Software Systems (AREA)
- Life Sciences & Earth Sciences (AREA)
- Health & Medical Sciences (AREA)
- Management, Administration, Business Operations System, And Electronic Commerce (AREA)
Abstract
The invention discloses a prediction method of a symbol network symbol based on a graph convolution network, which is characterized in that based on a GCN model in a graph neural network, the symbol network is divided into two sub-networks only comprising positive edges and negative edges by combining two conditions of dividing the symbol network into the positive edges and the negative edges, and the two sub-networks are preprocessed to obtain an adjacent matrix and a characteristic matrix of the two sub-networks. And training the adjacency matrix and the feature matrix of the two sub-networks by using the designed graph convolution network model to obtain a network representation containing the two sub-networks, and splicing the network representation to obtain a node representation of the original symbol network. In a symbol prediction test experiment of a symbol network, after a test set is input, positive probability and negative probability of an edge can be respectively obtained in network representation after model training, the symbol tendency of the edge is judged by comparing the positive probability and the negative probability, if the negative probability is higher than the positive probability, the symbol of the edge tends to be more negative, otherwise, the symbol tends to be more positive.
Description
Technical Field
The invention belongs to the field of complex networks, and particularly relates to a symbolic network based on a graph convolution network.
Background
Various relationships in the real world can be abstracted into a complex network to a certain extent, including human-to-human relationships, human-to-object relationships and object-to-object relationships, and a heterogeneous information network containing multiple individuals and multiple relationships can be formed by integrating the relationships among the human-to-human relationships, the human-to-object relationships and the object-to-object relationships. Individuals in the real world are referred to as nodes in the network, and relationships between individuals in the real world are abstracted as edges in the network. The relationship between people includes positive relationship such as good and favorite, and negative relationship such as dislike, and these relationships are abstracted into edges in the network, so that the network includes two structural relationships, the network including the positive and negative relationships is called symbol network, and the symbol network is a special heterogeneous information network. Therefore, in a recommendation system, identifying the type of relationship between people is an important task.
With the wide application of the convolutional neural network in image processing, the application of the convolutional neural network to the image processing direction becomes a new research idea, signals can be converted from a time domain space to a frequency domain space through image Fourier transform, and the view angle of the frequency domain space can be helpful for processing the signals, so that the convolution operation in the image has a theoretical basis, and the image convolution layer can be generated by designing an image filter, so that an image convolution network model is generated.
The method is characterized in that the edge sign prediction between the sign networks is evaluated in various ways, prediction learning analysis is performed essentially based on positive and negative edge relations of the sign networks, the main properties of the sign networks comprise a structure balance theory, a similarity theory and a node state theory, each theory is performed on the basis of the positive and negative edge relations of the nodes, the trend relations of the positive and negative edges of the nodes are obtained through the theory, for example, the structure balance theory, and the sign tendencies of the edges are identified through analyzing the balance relations between the nodes of the triad.
Disclosure of Invention
Aiming at the prior art, the invention provides a symbol network symbol prediction method based on a Graph Convolution Network (GCN), which is mainly used for predicting the symbol tendency of nodes in the network based on the positive-negative relation of edges between the nodes in the GCN symbol network, outputting vector representation of the positive tendency and the negative tendency of the nodes according to the flexible application of the GCN, and predicting the tendency between the nodes by predicting the probability of the positive tendency and the probability of the negative tendency of the nodes and splicing the vector representation to obtain the result with the maximum probability. The method of the invention has great breakthrough in both theoretical innovation and experimental effect.
In order to solve the above technical problem, the present invention provides a symbol network symbol prediction method based on a graph convolution network, which comprises the following steps:
the method comprises the following steps that firstly, based on a graph convolution network model in a graph neural network, the graph convolution network model is combined with two situations that the edge of the graph network is only divided into a positive edge and a negative edge, the graph network is divided into two sub-networks, wherein one sub-network is a positive edge sub-network only containing the positive edge, the other sub-network is a negative edge sub-network only containing the negative edge, the positive edge sub-network and the negative edge sub-network are preprocessed respectively, and an adjacent matrix and a feature matrix of the two sub-networks are obtained;
training an adjacency matrix and a feature matrix of the two sub-networks by using the constructed two-layer graph convolution network model to obtain positive edge sub-network representation and negative edge sub-network representation, and splicing the positive edge sub-network representation and the negative edge sub-network representation to obtain node representation of the symbol network;
and thirdly, symbol prediction of a symbol network is carried out, a wiki-vote test data set is selected, the test data set is tested in the two-layer graph convolution network model, node representation generated based on the two-layer graph convolution network model is obtained, the symbol tendency of the edge of the test data set is evaluated by utilizing the generated node representation, after the node representation is obtained, the probability of the edge is obtained by using a logistic regression model, the probability of the edge comprises the positive probability of the edge and the negative probability of the edge, and the symbol tendency of the edge is judged by comparing the positive probability of the edge with the negative probability of the edge.
Further, in the step one, the specific process of respectively preprocessing the positive edge sub-network and the negative edge sub-network to obtain the adjacency matrix and the feature matrix of the two sub-networks is as follows:
obtaining symbolic network use case data, wherein the use case data is a wiki-vote data set, and the data structure of the data set is as follows: the name is Wiki-vote, the number of the nodes is 7118, the number of the edges is 103747, the proportion of positive edges is 78.8%, and the proportion of negative edges is 21.2%; reading wiki-vote data of a data set, separating positive edges and negative edges of nodes, namely performing iterative judgment on the edges of the nodes, wherein all the positive edges form a set, all the negative edges form a set, and constructing an adjacent matrix of a positive edge sub-network only containing the positive edges and an adjacent matrix of a negative edge sub-network only containing the negative edges based on the sets of the two edges and in combination with the nodes in original data; since the default data weight in the data set wiki-vote is 1, the adjacency matrix of the positive side subnetwork and the adjacency matrix of the negative side subnetwork are constructed and obtained according to the degree distribution of the nodes and the degree-of-entry correlation attributes of the nodes based on the degree distribution of the nodes.
In the second step, a two-layer graph convolution network model is constructed, and the method comprises the following steps:
the following iterative formula of the graph convolution network was used:
in formula (1), σ represents a nonlinear transformation; h represents a feature representation result; l represents the number of layers of the neural network of the designed graph; w represents a weight, w is 1; c represents a normalization factor;
substituting the characteristic matrix and the adjacency matrix into the iterative formula, and substituting the characteristic expression result therein to obtain:
Hl+1＝σ(AHlWl) (2)
in the formula (2), H represents a feature matrix, A represents an adjacent matrix, and W represents a weight matrix, and the input feature matrix H and the adjacent matrix are usedArray A is with sign tendencies, with A+And H+Respectively representing the adjacency matrix of the positive edge sub-network and the feature matrix of the positive edge sub-network, and using A-And H-Respectively representing an adjacency matrix of the negative-side subnetwork and a feature matrix of the negative-side subnetwork; thus, the result of the calculation
And
respectively representing the feature matrix of the positive edge sub-network and the feature matrix of the negative edge sub-network, and splicing the feature matrices to obtain the final result:
and obtaining the network representation of the original symbol network according to the constructed two-layer graph convolution network model.
The invention relates to a symbol network symbol prediction method based on a graph convolution network, wherein user related data are obtained from a database, and the method comprises the steps of table screening and field screening in a table; preprocessing and cleaning data; entities and relationships are abstracted from the data to construct a symbolic network.
Compared with the prior art, the invention has the beneficial effects that:
the traditional symbolic network symbolic prediction method and the graph convolution network application method are more, but the symbolic network symbolic prediction calculation method based on the graph convolution network has a lot of values and is not explored by combining the two methods. The method of the invention mainly has the following beneficial effects:
firstly, the method combines a theoretical symbol network symbol prediction method with a graph convolution network in the field of practical application, and theoretical practice is combined with each other, so that the value of theoretical research can be embodied in practice.
And secondly, the method is different from most machine learning methods, does not need to label data in advance, and can be better fused with the data environment in actual production.
Finally, the method has good expansibility, can add a full connection layer and other models (such as a transit model, a graph attention mechanism and the like) in the graph convolution network, and improves the symbol network symbol prediction effect.
Drawings
FIG. 1 is a block flow diagram of the method of the present invention;
FIG. 2 is an example of a designed graph convolution network model;
fig. 3 is an example of result analysis.
Detailed Description
The invention will be further described with reference to the following figures and specific examples, which are not intended to limit the invention in any way.
The design idea of the invention is that based on a graph convolution Network (GCN grapHgConvolitional Network) model in a graph neural Network, the symbol Network is divided into two sub-networks by combining two situations that the self edge of the symbol Network is only divided into a positive edge and a negative edge, the first sub-Network only comprises the positive edge, the second sub-Network only comprises the negative edge, the two sub-networks are respectively preprocessed, and an adjacent matrix and a feature matrix of the two sub-networks can be obtained. And training the adjacency matrix and the feature matrix of the two sub-networks by using the designed graph convolution network model to obtain a network representation containing a positive edge and a network representation containing a negative edge, and splicing the network representations to obtain a node representation of the original symbol network. In a symbol prediction test experiment of a symbol network, after a test set is input, positive probabilities and negative probabilities of edges can be respectively obtained in network representations obtained on the basis of node representations of the symbol network, and the positive probabilities and the negative probabilities of the edges are compared, so that the symbol tendentiousness of the edges is judged, if the negative probabilities are higher than the positive probabilities, the symbols of the edges tend to be more negative, and if not, the symbols tend to be more positive.
As shown in fig. 1, the symbolic network symbol prediction method based on the graph convolution network proposed by the present invention is that: acquiring user related data from a database, wherein the user related data comprises table screening and field screening in a table; preprocessing and cleaning data; entities and relationships are abstracted from the data to construct a symbolic network.
The method for predicting the symbolic network symbols based on the graph convolution network comprises the following steps:
the method comprises the steps that firstly, based on a graph convolution network model in a graph neural network, the graph convolution network model is combined with two situations that the edge of the graph network is only divided into a positive edge and a negative edge, the graph network is divided into two sub-networks, one sub-network is a positive edge sub-network only containing the positive edge, the other sub-network is a negative edge sub-network only containing the negative edge, the positive edge sub-network and the negative edge sub-network are preprocessed respectively, and an adjacent matrix and a feature matrix of the two sub-networks are obtained.
And secondly, training the adjacent matrix and the feature matrix of the two sub-networks by using the constructed two-layer graph convolution network model to obtain positive edge sub-network representation and negative edge sub-network representation, and splicing the positive edge sub-network representation and the negative edge sub-network representation to obtain node representation of the symbol network.
And thirdly, symbol prediction of a symbol network is carried out, a wiki-vote test data set is selected, the test data set is tested in the two-layer graph convolution network model, node representation generated based on the two-layer graph convolution network model is obtained, the symbol tendency of the edge of the test data set is evaluated by utilizing the generated node representation, after the node representation is obtained, the probability of the edge is obtained by using a logistic regression model, the probability of the edge comprises the positive probability of the edge and the negative probability of the edge, and the symbol tendency of the edge is judged by comparing the positive probability of the edge with the negative probability of the edge.
In the first step of the method of the present invention, the specific process of preprocessing the positive edge sub-network and the negative edge sub-network respectively to obtain the adjacency matrix and the feature matrix of the two sub-networks is as follows:
obtaining symbolic network use case data, wherein the use case data is a wiki-vote data set, and the data structure of the data set is as follows: the name is Wiki-vote, the number of the nodes is 7118, the number of the edges is 103747, the proportion of positive edges is 78.8%, and the proportion of negative edges is 21.2%;
reading data set wiki-vote data, separating positive edges and negative edges of nodes through data processing, namely performing iterative judgment on the edges of the nodes, wherein all the positive edges form a set, all the negative edges form a set, and constructing an adjacent matrix of a positive edge sub-network only containing the positive edges and an adjacent matrix of a negative edge sub-network only containing the negative edges based on the sets of the two edges and in combination with the nodes in the original data;
since the default data weight in the data set wiki-vote is 1, the adjacency matrix of the positive side subnetwork and the adjacency matrix of the negative side subnetwork are constructed and obtained according to the degree distribution of the nodes and the degree-of-entry correlation attributes of the nodes based on the degree distribution of the nodes.
In the second step of the invention, a two-layer graph convolution network model is constructed, as shown in fig. 2, the method comprises the following steps:
the following iterative formula of the graph convolution network was used:
in formula (1), σ represents a nonlinear transformation; h represents a feature representation result; l represents the number of layers of the neural network of the designed graph; w represents a weight, w is 1; c represents a normalization factor;
substituting the characteristic matrix and the adjacency matrix into the iterative formula, and substituting the characteristic expression result therein to obtain:
Hl+1＝σ(AHlWl) (2)
in the formula (2), H represents a feature matrix, A represents an adjacent matrix, W represents a weight matrix, and A is used for the input feature matrix H and the adjacent matrix A with sign tendencies+And H+Respectively representing the adjacency matrix of the positive edge sub-network and the feature matrix of the positive edge sub-network, and using A-And H-Respectively representing an adjacency matrix of the negative-side subnetwork and a feature matrix of the negative-side subnetwork; thus, the result of the calculation
And
respectively representing the feature matrix of the positive edge sub-network and the feature matrix of the negative edge sub-network, and splicing the feature matrices to obtain the final result:
and obtaining the network representation of the symbol network according to the constructed two-layer graph convolution network model.
In the third step, symbol prediction of a symbol network is carried out, and in a test data experiment, wiki-vote test data is used for carrying out the first step in the invention to obtain preprocessed data;
inputting the data into a trained two-layer gcn model, and after the data is expressed, predicting the probability of an edge by using a logistic regression model to predict a link, wherein the specific method is as follows: and (4) testing by using a logistic regression model, and judging the sign tendency of the edge by comparing the positive probability of the edge with the negative probability of the edge. Data validation was also performed using the evaluation function, which was F1, and the results are shown in fig. 3: wherein f1 is a common evaluation function, f1-micro is a micro average value, and f1-macro is a macro average value.
While the present invention has been described with reference to the accompanying drawings, the present invention is not limited to the above-described embodiments, which are illustrative only and not restrictive, and various modifications which do not depart from the spirit of the present invention and which are intended to be covered by the claims of the present invention may be made by those skilled in the art.
Claims (4)
1. A method for predicting symbolic network symbols based on a graph convolution network is characterized by comprising the following steps:
the method comprises the following steps that firstly, based on a graph convolution network model in a graph neural network, the graph convolution network model is combined with two situations that the edge of the graph network is only divided into a positive edge and a negative edge, the graph network is divided into two sub-networks, wherein one sub-network is a positive edge sub-network only containing the positive edge, the other sub-network is a negative edge sub-network only containing the negative edge, the positive edge sub-network and the negative edge sub-network are preprocessed respectively, and an adjacent matrix and a feature matrix of the two sub-networks are obtained;
training an adjacency matrix and a feature matrix of the two sub-networks by using the constructed two-layer graph convolution network model to obtain positive edge sub-network representation and negative edge sub-network representation, and splicing the positive edge sub-network representation and the negative edge sub-network representation to obtain node representation of the symbol network;
and thirdly, symbol prediction of a symbol network is carried out, a wiki-vote test data set is selected, the test data set is tested in the two-layer graph convolution network model, node representation generated based on the two-layer graph convolution network model is obtained, the symbol tendency of the edge of the test data set is evaluated by utilizing the generated node representation, after the node representation is obtained, the probability of the edge is obtained by using a logistic regression model, the probability of the edge comprises the positive probability of the edge and the negative probability of the edge, and the symbol tendency of the edge is judged by comparing the positive probability of the edge with the negative probability of the edge.
2. The method for predicting symbol network symbols based on graph convolution network of claim 1, wherein in the first step, the specific process of preprocessing the positive sub-network and the negative sub-network respectively to obtain the adjacency matrix and the feature matrix of the two sub-networks is as follows:
obtaining symbolic network use case data, wherein the use case data is a wiki-vote data set, and the data structure of the data set is as follows: the name is Wiki-vote, the number of the nodes is 7118, the number of the edges is 103747, the proportion of positive edges is 78.8%, and the proportion of negative edges is 21.2%;
reading wiki-vote data of a data set, separating positive edges and negative edges of nodes, namely performing iterative judgment on the edges of the nodes, wherein all the positive edges form a set, all the negative edges form a set, and constructing an adjacent matrix of a positive edge sub-network only containing the positive edges and an adjacent matrix of a negative edge sub-network only containing the negative edges based on the sets of the two edges and in combination with the nodes in original data;
since the default data weight in the data set wiki-vote is 1, the adjacency matrix of the positive side subnetwork and the adjacency matrix of the negative side subnetwork are constructed and obtained according to the degree distribution of the nodes and the degree-of-entry correlation attributes of the nodes based on the degree distribution of the nodes.
3. The method for predicting symbolic network symbols based on a graph and volume network according to claim 1, wherein in the second step, constructing a two-layer graph and volume network model comprises the following steps:
the following iterative formula of the graph convolution network was used:
in formula (1), σ represents a nonlinear transformation; h represents a feature representation result; l represents the number of layers of the neural network of the designed graph; w represents a weight, w is 1; c represents a normalization factor;
substituting the characteristic matrix and the adjacency matrix into the iterative formula, and substituting the characteristic expression result therein to obtain:
Hl+1＝σ(AHlWl) (2)
in the formula (2), H represents a feature matrix, A represents an adjacent matrix, W represents a weight matrix, and A is used for the input feature matrix H and the adjacent matrix A with sign tendencies+And H+Respectively representing the adjacency matrix of the positive edge sub-network and the feature matrix of the positive edge sub-network, and using A-And H-Respectively representing an adjacency matrix of the negative-side subnetwork and a feature matrix of the negative-side subnetwork; thus, the result of the calculation
And
respectively representing the feature matrix of the positive edge sub-network and the feature matrix of the negative edge sub-network, and splicing the feature matrices to obtain the final result:
and obtaining the network representation of the original symbol network according to the constructed two-layer graph convolution network model.
4. The method of claim 1, wherein the user-related data is obtained from a database, including table screening and table field screening; preprocessing and cleaning data; entities and relationships are abstracted from the data to construct a symbolic network.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202010325465.XA CN111581297A (en) | 2020-04-23 | 2020-04-23 | Symbolic network symbol prediction method based on graph convolution network |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202010325465.XA CN111581297A (en) | 2020-04-23 | 2020-04-23 | Symbolic network symbol prediction method based on graph convolution network |
Publications (1)
| Publication Number | Publication Date |
|---|---|
| CN111581297A true CN111581297A (en) | 2020-08-25 |
Family
ID=72114991
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN202010325465.XA Pending CN111581297A (en) | 2020-04-23 | 2020-04-23 | Symbolic network symbol prediction method based on graph convolution network |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN111581297A (en) |
Cited By (1)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN117557118A (en) * | 2023-11-13 | 2024-02-13 | 国网江苏省电力有限公司镇江供电分公司 | UPS system power supply topological graph generation method based on machine learning |
-
2020
- 2020-04-23 CN CN202010325465.XA patent/CN111581297A/en active Pending
Cited By (1)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN117557118A (en) * | 2023-11-13 | 2024-02-13 | 国网江苏省电力有限公司镇江供电分公司 | UPS system power supply topological graph generation method based on machine learning |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CN110462604B (en) | Data processing system and method based on device use associated internet device | |
| CN111061856B (en) | Knowledge perception-based news recommendation method | |
| CN112215837B (en) | Multi-attribute image semantic analysis method and device | |
| Gao et al. | A k-core decomposition-based opinion leaders identifying method and clustering-based consensus model for large-scale group decision making | |
| Gharehchopogh et al. | Data mining application for cyber space users tendency in blog writing: a case study | |
| CN116127190B (en) | Digital earth resource recommendation system and method | |
| CN113407784A (en) | Social network-based community dividing method, system and storage medium | |
| Alhamdani et al. | Recommender system for global terrorist database based on deep learning | |
| CN110910235A (en) | Method for detecting abnormal behavior in credit based on user relationship network | |
| CN113239143B (en) | Power transmission and transformation equipment fault processing method and system fusing power grid fault case base | |
| CN111581297A (en) | Symbolic network symbol prediction method based on graph convolution network | |
| CN116541792A (en) | Method for carrying out group partner identification based on graph neural network node classification | |
| CN114265954B (en) | Graph representation learning method based on position and structure information | |
| CN116318925A (en) | Multi-CNN fusion intrusion detection method, system, medium, equipment and terminal | |
| CN112463974A (en) | Method and device for establishing knowledge graph | |
| CN113159976B (en) | Identification method for important users of microblog network | |
| CN114462707B (en) | Web service multidimensional QoS joint prediction method based on feature depth fusion | |
| De La Hoz‐M et al. | Trends and topics in geographically weighted regression research from 1996 to 2019 | |
| CN115169521A (en) | Graph neural network interpretation method for keeping prediction sequence and structure dependency relationship | |
| CN113378075A (en) | Community discovery method for adaptively fusing network topology and node content | |
| CN114757581A (en) | Financial transaction risk assessment method and device, electronic equipment and computer readable medium | |
| CN114329099A (en) | Overlapping community identification method, device, equipment, storage medium and program product | |
| CN114387005A (en) | Arbitrage group identification method based on graph classification | |
| CN114529096A (en) | Social network link prediction method and system based on ternary closure graph embedding | |
| CN114417958A (en) | Unbalanced financial data credit evaluation method based on improved graph convolution neural network |
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 | ||
| RJ01 | Rejection of invention patent application after publication |
Application publication date: 20200825 |
|
| RJ01 | Rejection of invention patent application after publication |