Spectrum recommendation method based on knowledge graph

Abstract

The invention relates to a frequency spectrum recommendation method based on a knowledge graph, and belongs to the field of wireless communication. The method comprises the following steps: S1, defining a knowledge system of a satellite-ground spectrum shared knowledge graph; s2, importing the satellite parameter library and the ground frequency spectrum use condition into a knowledge graph to form a satellite-ground frequency spectrum shared knowledge graph; s3, defining parameters of a Node2vec algorithm, and calculating an embedded vector; s4, carrying out embedded vector PCA dimension reduction; s5, constructing a similarity matrix S among the satellite nodes, wherein the similarity measure is Euclidean distance; s6, screening out node pairs of which the similarity S is greater than 0.9, and writing the node pairs back to the knowledge graph to create a recovery relationship; and S7, updating the knowledge graph and visualizing the recommendation list. According to the method, the frequency spectrum recommendation method is provided based on the knowledge graph and is applied to the satellite-ground integrated system, the satellite-ground frequency spectrum shares the knowledge graph to serve as graph structure input, similar satellites are found, ground frequency bands are recommended for the satellites, and therefore the frequency spectrum resource utilization rate is increased, and the frequency spectrum scarcity problem is solved.

Description

Technical Field

[0001] The present invention belongs to the field of wireless communications and relates to a spectrum recommendation method based on a knowledge graph. Background Art

[0002] With the rapid development of wireless communication technology, the scarcity of spectrum resources has become increasingly prominent. Especially in satellite communications and terrestrial cellular networks, efficient use of spectrum resources has become a key technical challenge. Traditional spectrum allocation methods cannot adapt to the growing communication needs, resulting in low spectrum utilization and waste of spectrum resources. In addition, the co-channel interference problem between terrestrial cellular networks and satellite networks is becoming increasingly serious, affecting the communication quality.

[0003] Satellite-ground spectrum sharing technology can effectively alleviate resource shortages. Through the satellite-ground fusion network, broadband access services can be realized worldwide, providing users with ubiquitous communication services. In addition, satellite-ground spectrum sharing can also use the idle spectrum resources of the satellite-ground fusion network based on the spectrum situation awareness results through intelligent spectrum situation awareness and utilization, effectively reduce the co-frequency interference between the two systems, and greatly improve resource utilization efficiency. The development of satellite-ground spectrum sharing technology is of great significance for the construction of future 6G networks, and can provide higher quality communication services for massive users.

[0004] As a structured semantic knowledge base, knowledge graph can store and represent complex entity relationships and attribute information. In the field of satellite-ground spectrum sharing, building a knowledge graph can effectively integrate and analyze spectrum usage data of satellites and ground networks, providing deeper insights and decision support. The advantage of knowledge graph is that it can process and understand a large amount of heterogeneous data. Through graph embedding technology such as Node2Vec, complex graph structure data can be converted into low-dimensional vectors, which is convenient for calculation and analysis. In addition, knowledge graph can also support complex queries and reasoning, providing strong data support for the formulation and optimization of spectrum sharing strategies.

[0005] In summary, it is of great significance to use knowledge graphs to propose a recommendation plan for the frequency utilization strategy of satellite-ground spectrum sharing. Summary of the invention

[0006] In view of this, the present invention provides a spectrum recommendation method based on knowledge graph. In the present invention, a spectrum recommendation method is proposed based on knowledge graph, the satellite-ground spectrum sharing knowledge graph is used as the graph structure input, the node embedding vector is calculated through the Node2vec model and the Euclidean distance is used as the similarity metric to construct the similarity matrix, PCA is introduced to reduce the computational complexity, and finally the knowledge graph is updated and the ground frequency band is recommended for similar satellites, thereby improving the utilization rate of spectrum resources and solving the problem of spectrum scarcity.

[0007] In order to achieve the above object, the present invention provides the following technical solutions:

[0008] A spectrum recommendation method based on knowledge graph, characterized in that the method comprises the following steps:

[0009] S1: A knowledge system that defines the knowledge graph of satellite-ground spectrum sharing;

[0010] S2: Import the satellite parameter library and ground spectrum usage into the knowledge graph to form satellite-ground spectrum sharing knowledge

[0011] Atlas;

[0012] S3: Define the parameters of the Node2vec algorithm and calculate the embedding vector;

[0013] S4: Embedding vector PCA dimensionality reduction;

[0014] S5: construct the similarity matrix S between satellite nodes, and the similarity metric is Euclidean distance;

[0015] S6: Filter out node pairs with similarity S>0.9 and write them back to the knowledge graph to create a recommend relationship;

[0016] S7: Update the knowledge graph and visualize the recommendation list.

[0017] Further, in step S1, entities satellite, application type, cell, contractor, country, launch site, operator, orbit class, purpose, and user are defined in the knowledge system; the satellite-ground spectrum sharing knowledge graph describes instance relationships in the form of triples, namely, "head entity-relationship-tail entity", denoted as "hrt"; the head entity h is the defined entities such as satellite, cell, and user, the relationship r is defined as share, communicate, interfere, and the tail entity t is the defined application type, launch site, and other entities or attributes, and the complete triple is called a fact, denoted as F = (h, r, t); all entities in the satellite-ground spectrum sharing knowledge graph correspond to nodes in the graph model, and the relationships between entities correspond to edges in the graph, forming a mesh data model. The relationships between the nodes are intricate, and the knowledge graph organizes various scattered factual relationships tightly together.

[0018] Furthermore, in step S2, after defining the knowledge system of the knowledge graph, it is necessary to find spectrum data, which comes from professional practical experience, spectrum management policies, frequency usage equipment information, spectrum management logs, spectrum monitoring data tables, spectrum situation images, etc. The present invention selects Neo4j graph database as the repository of the satellite-ground spectrum sharing knowledge graph, and uses Python language and its py2neo function library for data processing and graph database input, and uses Cypher language for querying, modifying, and deleting the graph database.

[0019] Further, in step S3, the input of the Node2vec algorithm is a graph structure, the node name must be an integer or a string, and the knowledge graph is used as the input of the algorithm; in the feature learning framework of the Node2vec algorithm, it is assumed that G = (V, E) is a given network, V represents the vertex set, E represents the edge set, f: →R d is the mapping function from node to feature representation that we want to learn for the experiment, R represents a set of real numbers, d is the dimension parameter of the feature representation, and f is a parameter matrix of |V|*d dimension; for each source node μ∈V, define For the network neighbors generated by node μ through the neighbor sampling strategy S, the optimized objective function is:

[0020]

[0021] Where f(μ) is the mapping function that maps vertex μ to embedding vector; in order to make this optimization problem easier to handle, two standard assumptions are made:

[0022] Conditional independence assumption: given the feature representation of a source node, the probability of one of its neighbor nodes appearing is independent of any other neighbor nodes:

[0023]

[0024] Symmetric in feature space: The source node and its neighbor nodes have symmetrical influence on each other in the feature space, by taking the conditional probability of each "source node-neighbor node" pair as a softmax unit parameterized by the dot product between their features:

[0025]

[0026] In summary, the objective function is:

[0027]

[0028] Since the allocation function of each node

[0029] In order to solve the computational cost problem, the negative sampling method is adopted; thus, Node2vec optimizes the objective function by using the stochastic gradient descent method on the model parameters. Since the network is not linear, it is not possible to define the concept of neighbors using a sliding window on continuous words like text. Node2vec proposes a random process to sample multiple different neighbor nodes of a given source node μ. The node μ generates the network neighbor nodes N through the neighbor sampling strategy S. s (μ) is not limited to direct neighbors, but has many different structures according to different sampling strategies S, which is determined by the p and q parameters; in the Node2vec algorithm, it is necessary to define iterations, that is, the number of iterations when the Word2vec model is trained. This parameter determines the number of times the Word2vec model will completely traverse the data set during the training process. The present invention defines iterations = 10, and it is necessary to define Embedding dimension, that is, the embedding vector dimension. This parameter determines the dimension of the vector to which each node will be mapped. The present invention defines Embedding dimension = 4, and it is necessary to define Walklength, that is, the length of the random walk. This parameter determines the length of the node sequence continuously visited from a node when performing a random walk in the graph. The present invention defines Walk length = 80, and it is necessary to define Retum factor, in out factor, that is, parameters p and q. Parameters p and q control the tendency of random walks, which are specifically manifested as search parameters α, t represents the current node, and x represents the next node, which are defined as follows:

[0030]

[0031] If the p value is relatively small, the sampling will linger near the node, showing the characteristics of BFS, i.e., breadth first. If the q value is relatively small, the sampling will gradually move away from the node, showing the characteristics of DFS, i.e., depth first. The present invention defines p=1, q=1, and needs to define Window size, i.e., window size. This parameter determines the range of context words considered around each target word when training the Word2vec model. The present invention defines Window size=10.

[0032] Further, in step S4, since S3 defines Embedding dimension=4 through the Node2vec algorithm, that is, the node embedding vector output by the algorithm is 4-dimensional, in order to construct the similarity matrix, the 4-dimensional embedding vector needs to be reduced to 2 dimensions and most of the features of the original 4-dimensional vector are retained. The present invention uses the PCA method. PCA first subtracts the mean of each feature (dimension) so that the mean of the data set is zero:

[0033] X oentered=X-μ

[0034] Where X is the original data and μ is the mean of the feature; then the covariance matrix is ​​calculated. The covariance matrix is ​​a square matrix whose elements are the covariances between each pair of features in the original data set:

[0035]

[0036] Where ∑ is the covariance matrix and n is the number of samples; then perform eigendecomposition on the covariance matrix to obtain the eigenvalues ​​and corresponding eigenvectors:

[0037] ∑v=λv

[0038] Where v is the eigenvector, λ is the corresponding eigenvalue, the eigenvalue represents the variance of the data in the direction of each eigenvector, and the eigenvector represents the new direction of the data; finally, according to the size of the eigenvalue, the eigenvectors corresponding to the two largest eigenvalues ​​are selected. These two eigenvectors represent the largest variance in the data and are orthogonal. Then the original data is projected onto the two selected eigenvectors:

[0039] Y=X centerted P

[0040] Where Y is the data after dimensionality reduction, and P is a matrix composed of the eigenvectors corresponding to the selected principal components, thereby obtaining new 2D data.

[0041] Further, in step S5, whether the original nodes are similar can be judged according to the similarity of the embedded vectors. This is because the Node2vec algorithm can retain the local domain information of the node, that is, the embedded vector of a node contains the information of its neighboring nodes. Therefore, if the embedded vectors of two nodes are similar, it means that they have similar local domain structures in the graph structure, so the embedded vector obtained by the Node2vec algorithm can effectively represent the similarity of the nodes; after obtaining the reduced-dimensional graph embedding vector, a suitable similarity metric can be selected to construct the similarity matrix between nodes. This method selects Euclidean distance as the metric of similarity:

[0042]

[0043] where u i , v i are the coordinates of point u and point v in the i-th dimension, n is the number of dimensions after dimensionality reduction. The similarity of nodes in the graph structure can be determined by calculating the Euclidean distance of nodes in the two-dimensional vector space; construct a similarity matrix and normalize it:

[0044]

[0045] Where max((d)) is the maximum distance between all point pairs. It is verified by the scatter plot that close points in the graph will get a higher similarity coefficient, while separated points in the graph will get a lower similarity coefficient.

[0046] Furthermore, in step S6, according to the similarity matrix, the corresponding vectors with S>0.9 are screened out, and the threshold value of S depends on the number of recommendations required. The node pairs are then stored as new graph structure relationships to be established, and a recommend relationship is created for similar nodes of the original node and nodes shared by the original node through the Cypher language.

[0047] Furthermore, in step S7, after the new relationship is created, updating the satellite-ground spectrum sharing knowledge graph will add a new relationship named recommend to the knowledge graph, and the recommendation list can be visualized by querying the nodes associated with recommend through the Cypher language.

[0048] The effective effect of the present invention is: a spectrum recommendation method based on knowledge graph, in view of the complex graph structure relationship of the knowledge graph, uses the Node2vec algorithm to learn the complex structure of the satellite node to obtain an embedded vector, the embedded vector contains the structural information of the node, and finds similar satellite nodes through the vector. It can recommend frequency bands used on the ground for satellites, improve the utilization rate of spectrum resources, and solve the problem of spectrum scarcity. BRIEF DESCRIPTION OF THE DRAWINGS

[0049] In order to make the purpose, technical solution and beneficial effects of the present invention more clear, the present invention provides the following drawings for illustration:

[0050] Figure 1 This is a flow chart of the recommendation algorithm described in an embodiment of the present invention;

[0051] Figure 2 This is a schematic diagram of a graph embedding algorithm according to an embodiment of the present invention;

[0052] Figure 3 It is a low-dimensional space scatter plot of the embedded vector according to an embodiment of the present invention;

[0053] Figure 4 The satellite node similarity matrix described in the embodiment of the present invention;

[0054] Figure 5 This is a knowledge graph of satellite-ground spectrum sharing according to an embodiment of the present invention. DETAILED DESCRIPTION

[0055] The preferred embodiments of the present invention will be described in detail below with reference to the accompanying drawings.

[0056] Aiming at the problem of scarce spectrum resources, the present invention proposes a spectrum recommendation method based on knowledge graph. Different from the traditional fixed spectrum allocation mode, in the present invention, it is necessary to first build a satellite-ground spectrum sharing knowledge graph, which consists of two parts: satellite parameter library and historical ground spectrum usage data. Both parts of data come from professional practical experience, spectrum management policies, frequency equipment information, spectrum management logs, spectrum monitoring data tables, spectrum situation images, etc. After building the satellite-ground spectrum sharing knowledge graph, since the knowledge graph is a typical graph structure network, and the Node2vec algorithm can calculate the embedding vector of the graph structure, the knowledge graph is used as the input of the Node2vec algorithm, and the output is the embedding vector of the corresponding node. Since the defined vector dimension is 4-dimensional, the vector needs to be reduced in dimension. PCA is used for vector dimension reduction. PCA can retain the characteristics of the original data as much as possible. Then, the similarity matrix S is constructed according to the size of the Euclidean distance of the nodes in the 2-dimensional space. The node pairs with S>0.9 are selected and written back to the satellite-ground spectrum sharing knowledge graph, a new recommend relationship is created and the knowledge graph is updated. Finally, the recommendation list is visualized in the knowledge graph. The process is as follows: Figure 1 As shown. The Node2vec algorithm uses the graph structure as input. The difference from the Word2vec model is that Node2vec adds p and q parameters to control the tendency of random walks starting from the source node. A small p value will cause the random walk sequence to be a breadth-first traversal, and a small q value will cause the random walk sequence to be a depth-first traversal. Reasonable control of the p and q values ​​can make the random walk sequence present content similarity and structural similarity with different weights. Sampling large-scale graphs is extremely time-consuming, so some nodes in the network can be sampled to save sampling time. In partial sampling, we need to control the probability of a node being sampled, the frequency of node appearance Z (v i ), the higher the probability of sampling it as the starting point, that is, nodes that have been repeatedly sampled do not need to be resampled. After obtaining the random walk sequence, the sequence is input to the SkipGram model, and the model will output the embedding vector of the sequence. Figure 2 shown.

[0057] After PCA reduces the 4-dimensional vector of the satellite node to 2-dimensional, some nodes are represented in the scatter diagram, such as Figure 3 As shown. In the scatter plot, some nodes are close to each other, and some nodes are far away from each other. From this 2D vector space, we can roughly observe which satellites are similar in structure and content. Now we need to define a function to determine which satellites are considered similar. The present invention uses the Euclidean distance, that is, the distance between two nodes in the 2D vector space is used as a measure of similarity. By calculating the Euclidean distance between each node and other nodes and normalizing them, we get a similarity matrix, as shown in Figure 4 shown.

[0058] The similarity matrix can be used to filter out satellite nodes that are considered similar. Then, the same ground frequency band can be recommended for sharing based on similar satellite nodes. The filtered satellite nodes are written back to the satellite-ground spectrum sharing knowledge graph, and a recommend relationship is created and the knowledge graph is updated. The recommendation list can be obtained by querying the established recommend relationship in the knowledge graph using the Cypher language. The complete satellite-ground spectrum sharing knowledge graph is shown in the figure. Figure 5 shown.

[0059] Finally, it should be noted that the above preferred embodiments are only used to illustrate the technical solutions of the present invention rather than to limit it. Although the present invention has been described in detail through the above preferred embodiments, those skilled in the art should understand that various changes can be made in form and details without departing from the scope defined by the claims of the present invention.

Claims

1. A spectrum recommendation method based on knowledge graph, characterized in that: The method comprises the following steps: S1: A knowledge system that defines the knowledge graph of satellite-ground spectrum sharing; S2: Import the satellite parameter library and ground spectrum usage into the knowledge graph to form a satellite-ground spectrum sharing knowledge graph; S3: Define the parameters of the Node2vec algorithm and calculate the embedding vector; S4: Embedding vector PCA dimensionality reduction; S5: construct the similarity matrix S between satellite nodes, and the similarity metric is Euclidean distance; S6: Filter out node pairs with similarity S>0.9 and write them back to the knowledge graph to create a recommend relationship; S7: Update the knowledge graph and visualize the recommendation list.

2. The spectrum recommendation method based on knowledge graph according to claim 1, characterized in that: In step S1, entities satellite, application type, cell, contractor, country, launch site, operator, orbit class, purpose, and user are defined in the knowledge system; the satellite-ground spectrum sharing knowledge graph describes instance relationships in the form of triples, namely "head entity-relationship-tail entity", denoted as "hrt"; the head entity h is the defined entities such as satellite, cell, and user, the relationship r is defined as share, communicate, interfere, etc., and the tail entity t is the defined application type, launch site, and other entities or attributes. The complete triple is called a fact, denoted as F = (hr, t); all entities in the satellite-ground spectrum sharing knowledge graph correspond to nodes in the graph model, and the relationships between entities correspond to edges in the graph, forming a mesh data model. The relationships between nodes are intricate, and the knowledge graph organizes various scattered factual relationships tightly together.

3. The spectrum recommendation method based on knowledge graph according to claim 1, characterized in that: In step S2, after defining the knowledge system of the knowledge graph, it is necessary to find spectrum data. The spectrum data comes from professional practical experience, spectrum management policies, frequency usage equipment information, spectrum management logs, spectrum monitoring data tables, spectrum situation images, etc. The present invention selects Neo4j graph database as the repository of the satellite-ground spectrum sharing knowledge graph, and uses Python language and its py2neo function library for data processing and graph database input. The query, modification, and deletion of the graph database use Cypher language.

4. The spectrum recommendation method based on knowledge graph according to claim 1, characterized in that: In step S3, the input of the Node2vec algorithm is a graph structure, the node name must be an integer or a string, and the knowledge graph is used as the input of the algorithm; in the feature learning framework of the Node2vec algorithm, it is assumed that G = (V, E) is a given network, V represents the vertex set, E represents the edge set, f: →R d is the mapping function from node to feature representation that we want to learn for the experiment, R represents a set of real numbers, d is the dimension parameter of the feature representation, and f is a parameter matrix of |V|*d dimension; for each source node μ∈V, define For the network neighbors generated by node μ through the neighbor sampling strategy S, the optimized objective function is: Where f(μ) is the mapping function that maps vertex μ to embedding vector; in order to make this optimization problem easier to handle, two standard assumptions are made: Conditional independence assumption: given the feature representation of a source node, the probability of one of its neighbor nodes appearing is independent of any other neighbor nodes: Symmetric in feature space: The source node and its neighbor nodes have symmetrical influence on each other in the feature space, by taking the conditional probability of each "source node-neighbor node" pair as a softmax unit parameterized by the dot product between their features: In summary, the objective function is: Since the allocation function of each node In order to solve the computational cost problem, the negative sampling method is adopted; thus, Node2vec optimizes the objective function by using the stochastic gradient descent method on the model parameters. Since the network is not linear, it is not possible to define the concept of neighbors using a sliding window on continuous words like text. Node2vec proposes a random process to sample multiple different neighbor nodes of a given source node μ. The node μ generates the network neighbor nodes N through the neighbor sampling strategy S. s (μ) is not limited to direct neighbors, but has many different structures according to different sampling strategies S, which is determined by the p and q parameters; in the Node2vec algorithm, it is necessary to define iterations, that is, the number of iterations when the Word2vec model is trained. This parameter determines the number of times the Word2vec model will completely traverse the data set during the training process. The present invention defines iterations = 10, and it is necessary to define Embeddingdimension, that is, the embedding vector dimension. This parameter determines the dimension of the vector to which each node will be mapped. The present invention defines Embedding dimension = 4, and it is necessary to define Walklength, that is, the length of the random walk. This parameter determines the length of the node sequence continuously visited from a node when performing a random walk in the graph. The present invention defines Walk length = 80, and it is necessary to define Retum factor, in out factor, that is, parameters p and q. Parameters p and q control the tendency of random walks, which are specifically manifested as search parameters α, t represents the current node, and x represents the next node, which are defined as follows: If the p value is relatively small, the sampling will linger near the node, showing the characteristics of BFS, i.e., breadth first. If the q value is relatively small, the sampling will gradually move away from the node, showing the characteristics of DFS, i.e., depth first. The present invention defines p=1, q=1, and needs to define Window size, i.e., window size. This parameter determines the range of context words considered around each target word when training the Word2vec model. The present invention defines Window size=10.

5. The spectrum recommendation method based on knowledge graph according to claim 1, characterized in that: In step S4, since S3 defines Embedding dimension=4 through the Node2vec algorithm, that is, the node embedding vector output by the algorithm is 4-dimensional, in order to construct the similarity matrix, the 4-dimensional embedding vector needs to be reduced to 2 dimensions and most of the features of the original 4-dimensional vector are retained. The present invention uses the PCA method. PCA first subtracts the mean of each feature (dimension) so that the mean of the data set is zero: X centered =X-μ Where X is the original data and μ is the mean of the feature; Then calculate the covariance matrix, which is a square matrix whose elements are the covariances between each pair of features in the original dataset: Where ∑ is the covariance matrix and n is the number of samples; then perform eigendecomposition on the covariance matrix to obtain the eigenvalues ​​and corresponding eigenvectors: ∑υ=λυ Among them, υ is the eigenvector, λ is the corresponding eigenvalue, the eigenvalue represents the variance of the data in the direction of each eigenvector, and the eigenvector represents the new direction of the data; finally, according to the size of the eigenvalue, the eigenvectors corresponding to the two largest eigenvalues ​​are selected. These two eigenvectors represent the largest variance in the data and are orthogonal. Then the original data is projected onto the two selected eigenvectors: Y=X centerted P Where Y is the data after dimensionality reduction, and P is a matrix composed of the eigenvectors corresponding to the selected principal components, thereby obtaining new 2D data.

6. The spectrum recommendation method based on knowledge graph according to claim 1, characterized in that: In step S5, whether the original nodes are similar can be determined based on the similarity of the embedded vectors. This is because the Node2vec algorithm can retain the local domain information of the nodes, that is, the embedded vector of a node contains the information of its neighboring nodes. Therefore, if the embedded vectors of two nodes are similar, it means that they have similar local domain structures in the graph structure, so the embedded vector obtained by the Node2vec algorithm can effectively represent the similarity of the nodes. After obtaining the reduced-dimensional graph embedding vector, a suitable similarity metric can be selected to construct the similarity matrix between nodes. This method selects Euclidean distance as the metric of similarity: where u i , v i are the coordinates of point u and point v in the i-th dimension, n is the number of dimensions after dimensionality reduction. The similarity of nodes in the graph structure can be determined by calculating the Euclidean distance of nodes in the two-dimensional vector space; construct a similarity matrix and normalize it: Among them, max(d) is the maximum distance between all point pairs. It is verified by the scatter plot that close points in the graph will get a higher similarity coefficient, while separated points in the graph will get a lower similarity coefficient.

7. The spectrum recommendation method based on knowledge graph according to claim 1, characterized in that: In step S6, according to the similarity matrix, the corresponding vectors with S>0.9 are screened out. The threshold value of S depends on the number of recommendations required. The node pairs are then stored as new graph structure relationships to be established. The Cypher language is used to create a recommend relationship between similar nodes of the original node and nodes shared by the original node.

8. The spectrum recommendation method based on knowledge graph according to claim 1, characterized in that: In step S7, after the new relationship is created, updating the satellite-ground spectrum sharing knowledge graph will add a new relationship named recommend to the knowledge graph. The recommendation list can be visualized by querying the nodes associated with recommend through the Cypher language.

Images