A GCN-based network edge user clustering method

Abstract

The application discloses a network edge user clustering method based on GCN, comprising the following steps: constructing a user graph structure according to a user node set and a non-directional edge existing between users due to interference, obtaining an interference matrix between users, constructing a node feature matrix, calculating an adjacency matrix of the user graph structure of the network edge according to the user graph structure, selecting a clustering number F, finding out F users with the largest interference between users, making each user correspond to a cluster, marking the F users, obtaining a labeled user set and a user label represented by a F-dimensional vector, building a two-layer graph neural network GCN, training the graph neural network GCN by using the user label of a small amount of the labeled user set, constructing a final graph neural network GCN, inputting the adjacency matrix into the final graph neural network GCN, outputting F user clusters, and allocating the same subchannel to each user cluster. The method effectively reduces the interference of the system and improves the total throughput of the system by clustering the edge users.

Description

Technical Field

[0001] This invention relates to the field of network communication technology, and more specifically to a method for clustering network edge users based on GCN. Background Technology

[0002] With the development of 5G communication technology, the explosive growth of mobile user devices has led to an exponential increase in network traffic. To address the high deployment costs of traditional base stations and the difficulty in guaranteeing service quality for edge users, Ultra-Dense Networks (UDNs) technology has emerged. UDNs typically consist of densely deployed micro base stations, improving spectrum utilization, traffic capacity, and data transmission rates. However, high-density UDN deployment results in edge users receiving almost identical signal strengths from surrounding base stations, placing them in a state of interference and being interfered with. For these edge users, clustering is necessary to reduce inter-user interference and improve user performance. This paper proposes a GCN-based edge user clustering algorithm. By aggregating connectivity information between users, it provides more detailed information for user clustering. Through a semi-supervised classification learning method, users are grouped, reducing system interference and effectively improving system throughput. Summary of the Invention

[0003] This invention provides a network edge user clustering method based on GCN, comprising:

[0004] Construct the user graph structure based on the user node set and the undirected edges that exist due to interference between users;

[0005] Obtain the interference matrix between users and construct the node feature matrix;

[0006] Calculate the adjacency matrix of the user graph structure at the network edge based on the user graph structure;

[0007] Choose the number of clusters F, find the F users with the greatest inter-user interference, make each user correspond to a cluster, label the F users, and obtain the labeled user set and the user labels represented by F-dimensional vectors;

[0008] Using user labels from a small set of labeled users, train a graph neural network (GCN) to construct the final graph neural network (GCN);

[0009] The final graph neural network (GCN) takes an adjacency matrix as input, outputs F user clusters, and assigns the same sub-channel to each user cluster.

[0010] Preferably, constructing a node feature matrix includes:

[0011] Let V = {V1, V2, V3…V} N} represents the set of user nodes, where V NThis represents the Nth user node;

[0012] Obtain the interference matrix between users:

[0013]

[0014] Among them, W ab (i,j) represents user node V i With V j Interference between them;

[0015] Set the interference matrix to be equal to the node feature matrix.

[0016] Preferably, obtaining the set of labeled users and the user labels represented by F-dimensional vectors includes:

[0017] The set of labeled users is K l The label uses an F-dimensional vector Y l The label vector of a labeled user i is determined by its cluster f. When a labeled user is located in the f-th cluster, its F-dimensional label vector Y is... i Only the f-th bit is 1, and the rest are 0.

[0018] Preferably, the model expression of the graph neural network (GCN) includes:

[0019]

[0020] in, I is the identity matrix, and A is the adjacency matrix;

[0021] X is N×N dimensional, representing the input node feature matrix, i.e., X = W ab ;

[0022] W (0) ∈R N×H , where H is the weight matrix from the input layer to the hidden layer, and H is manually set, with F < H < N;

[0023] W (1) ∈R H×F , is the weight matrix from the hidden layer to the output layer;

[0024] The Softmax function is used to normalize input values.

[0025] Preferably, a graph neural network (GCN) model is trained using user labels from a small set of labeled users, including:

[0026] Set the initial learning rate l of the Graph Neural Network (GCN) model. r Initialize the weight matrix

[0027] X containing user graph structure dataN×N and as well as Substituting the model expression of the graph neural network (GCN), we obtain the clustering matrix Z. N×F ;

[0028] Get the set of labeled users K l The output matrix Z corresponding to the marked node N×F The vector Z in l , where l∈K l , indicating K l Centralized users;

[0029] Obtain the vectors and user labels of the labeled user nodes from the classification matrix corresponding to the labeled user nodes in the labeled user set;

[0030] According to the loss function Calculate the cross-entropy loss function for the labeled user set;

[0031] The parameters of the graph neural network (GCN) are updated using gradient descent with the cross-entropy loss function.

[0032] Set a training period m. When the training period is reached, stop training and output the preliminary clustering results, including those with a number of users greater than m. Users within the cluster are sorted in descending order of their softmax output results, and then... Individual users are ranked according to their number of users. The probability of clusters is redistributed until the number of users in all clusters is equal or differs by only one user.

[0033] This invention proposes a user clustering algorithm based on GCN. This algorithm effectively reduces system interference and improves the overall system throughput by clustering edge users. Attached Figure Description

[0034] Figure 1 A flowchart of a network edge user clustering method based on GCN provided in this embodiment of the invention;

[0035] Figure 2 This is a performance comparison chart of a network edge user clustering method based on GCN and a comparative method provided in an embodiment of the present invention. Detailed Implementation

[0036] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0037] See Figure 1-2 This invention provides a network edge user clustering method based on GCN, the method comprising:

[0038] 1) Construct an edge user communication network graph, with users as nodes, and the node set (users) represented by V, V = {V1, V2, V3…V}. N}; with the inter-user interference matrix W ab This is the node feature matrix.

[0039] W ab (i,j) represents user V i With V j Interference between them.

[0040] Then, the features of node V1 are [0, W] ab (1,2), W ab (1,3)…W ab (1,N)]. When there is interference between users, the users are considered to be connected, that is, there is an undirected edge between the two nodes.

[0041] Find the adjacency matrix A of the communication network graph. N×N When two nodes V i With V j If there is an edge between them, then a ij =1, otherwise a ij =0.

[0042] Based on the cluster number F, select F users with the greatest interference among them, assigning each user to a cluster, and label these F users. The set of labeled users is K. l The label uses an F-dimensional vector Y l This means that the label vector of the labeled user i is determined by the cluster f in which it belongs. When the labeled user is located in the f-th cluster, its F-dimensional label vector Y is... i Only the f-th bit is 1, and the rest are 0.

[0043] Taking a binary cluster as an example, for user V a Marked as the first cluster, user V b If labeled as the second cluster, then the user set K is labeled. l ={V a V b}, V a The label is Y a =[1 0], V b The label is Y b =[0 1].

[0044] 4) A two-layer GCN is used for forward propagation of node and structural information, ensuring that each node possesses information about its neighboring nodes and the graph structure. The GCN model formula is expressed as follows:

[0045]

[0046] in, I is the identity matrix, and A is the adjacency matrix.

[0047] X (N×N dimensional) is the feature matrix of the input nodes, i.e., X = W ab .

[0048] W (0) ∈R N×H H is the weight matrix from the input layer to the hidden layer. H can be set manually. Generally, F < H < N.

[0049] W (1) ∈R H×F , where is the weight matrix from the hidden layer to the output layer.

[0050] Softmax is used to normalize input values.

[0051] 5) Set the initial learning rate l of the model. r Initialize the weight matrix

[0052] 6) Calculate the labeled user set K l The cross-entropy loss.

[0053] a. Transfer the above communication network diagram data X N×N and as well as Substituting into equation (1), we obtain the output matrix Z. N×F .

[0054] b. Obtain the tag user set K l The output matrix Z corresponding to the marked node N×F The vector Z in l , where l∈K l , indicating K l Centralized users.

[0055] Taking a binary cluster as an example, K l ={V a V b}, then user V a Corresponding vector Z a That is, the output matrix Z N×F The a-th row vector; user V b Corresponding vector Z b That is, the output matrix Z N×F The b-th row vector.

[0056] c. Based on the loss function: Calculate the label set K l The cross-entropy loss.

[0057] Where, l∈K l , indicating K l Centralized node users, Y l Z is the real label for node user l. l The predicted probability of node user l is represented by the output matrix Z corresponding to node l. N×F The vector Z in l .

[0058] Taking a binary cluster as an example, K l ={V a V b}, assuming V a This is the first cluster, and its label is Y. a =[1 0], V b It is the second cluster, and its label is Y. b =[0 1]; Z a =[a1 a2], Z b =[b1 b2], substitute into formula (2)K l Cross-entropy loss:

[0059] Γ=-(1×Ina1+0×Ina2+0×Inb1+1×Inb2)=-Ina1b2.

[0060] 7) Using the loss function to... Take the derivative and use gradient descent to update the parameters.

[0061] 8) Set the training period m. When the training period is reached, stop training and output the preliminary clustering results, including those for users with a larger number of users. Users within the cluster are sorted in descending order of their softmax output results, and then... Individual users are ranked according to their number of users. The probability of clusters is redistributed until the number of users in all clusters is equal or differs by only one user.

[0062] 9) F user clusters are obtained, and the same sub-channel is allocated to each user cluster, which reduces system interference and improves system throughput.

[0063] The above-disclosed embodiments are merely a few specific examples of the present invention. Those skilled in the art can make various modifications and variations to the embodiments of the present invention without departing from the spirit and scope of the present invention. However, the embodiments of the present invention are not limited thereto, and any changes that those skilled in the art can conceive of should fall within the protection scope of the present invention.

Claims

1. A network edge user clustering method based on GCN, characterized in that, include: Construct the user graph structure based on the user node set and the undirected edges that exist due to interference between users; Obtain the interference matrix between users and construct the node feature matrix; Calculate the adjacency matrix of the user graph structure at the network edge based on the user graph structure; Choose the number of clusters F, find the F users with the greatest inter-user interference, make each user correspond to a cluster, label the F users, and obtain the labeled user set and the user labels represented by F-dimensional vectors; Using user labels from a small set of labeled users, train a graph neural network (GCN) to construct the final graph neural network (GCN); The final graph neural network (GCN) takes an adjacency matrix as input, outputs F user clusters, and assigns the same sub-channel to each user cluster.

2. The network edge user clustering method based on GCN as described in claim 1, characterized in that, The construction of the node feature matrix includes: Use V={ } represents the set of user nodes, where V N This represents the Nth user node; Obtain the interference matrix between users: in, (i,j) represents the user node and Interference between them; Set the interference matrix to be equal to the node feature matrix.

3. The network edge user clustering method based on GCN as described in claim 2, characterized in that, The process of obtaining the set of labeled users and the user labels represented by F-dimensional vectors includes: The set of marked users is The label is an F-dimensional vector. The label vector of a labeled user i is determined by its cluster f. When a labeled user is in the f-th cluster, its F-dimensional label vector... Only the f-th bit is 1, and the rest are 0.

4. The network edge user clustering method based on GCN as described in claim 3, characterized in that, The model expression of the graph neural network (GCN) includes: ⑴, in, I is the identity matrix, and A is the adjacency matrix; X is N×N dimensional, representing the input node feature matrix, i.e., X = ; , where H is the weight matrix from the input layer to the hidden layer, and H is set manually. ; , is the weight matrix from the hidden layer to the output layer; The Softmax function is used to normalize input values.

5. The network edge user clustering method based on GCN as described in claim 4, characterized in that, The method of training a graph neural network (GCN) model using user labels from a small set of labeled users includes: Set the initial learning rate for the Graph Neural Network (GCN) model. Initialize the weight matrix , ; Will contain user graph structure data and ,as well as , Substituting the model expression of the Graph Neural Network (GCN) into the formula, we obtain the clustering matrix. ; Get Tag User Set The output matrix corresponding to the marked nodes vectors in ,in ,express Centralized users; Obtain the vectors and user labels of the labeled user nodes from the classification matrix corresponding to the labeled user nodes in the labeled user set; According to the loss function Calculate the cross-entropy loss function for the labeled user set; The parameters of the graph neural network (GCN) are updated using gradient descent with the cross-entropy loss function. Set a training period m. When the training period is reached, stop training and output the preliminary clustering results, including those with a number of users greater than a certain threshold. Users within the cluster are sorted in descending order of their softmax output results, and then... Individual users are ranked according to their number of users. The probability of clusters is redistributed until the number of users in all clusters is equal or differs by only one user.

Images