Wireless communication network knowledge graph representation learning method based on heterogeneous graph neural network

Abstract

The application discloses a wireless communication network knowledge graph representation learning method based on a heterogeneous graph neural network and belongs to the technical field of artificial intelligence assisted wireless communication. The application converts a wireless communication network knowledge graph into a heterogeneous graph containing multiple source heterogeneous nodes, constructs a wireless communication network knowledge graph representation learning model based on the heterogeneous graph neural network, trains the constructed heterogeneous graph neural network model, verifies the effect of the model by using a link prediction downstream task, solves the positive and negative edge imbalance problem in the link prediction by using a negative sampling technology, and the constructed heterogeneous graph neural network model can effectively learn and mine the wireless communication network knowledge graph, optimize and complete the wireless communication network knowledge graph, and understand the correlation characteristics of various communication indexes in the wireless communication network, which has important research significance for realizing the endogenous intelligence of the wireless communication network.

Description

TECHNICAL FIELD

[0001] The application relates to a wireless communication network knowledge graph representation learning method based on a heterogeneous graph neural network and belongs to the technical field of artificial intelligence assisted wireless communication. BACKGROUND

[0002] With the rapid development of digitization and large-scale Internet of Things, wireless communication networks pay more attention to the needs of specific application scenarios and the performance and architecture of the overall communication network. The current wireless communication network cannot effectively accumulate experience to continuously improve, and the wireless communication network needs to have intelligent capabilities to solve this problem. With the rapid development of artificial intelligence, the intelligence level and network quality of the wireless communication network will gradually improve. Artificial intelligence empowers 6G wireless communication networks to optimize network configuration and improve network performance by learning wireless network infrastructure and sensor device features in big data. Knowledge graphs can store knowledge in graph structures, can reason and discover new knowledge. Using a knowledge graph to depict a wireless communication network can effectively explain and reason about the wireless communication network, convert wireless communication data into usable wireless communication knowledge, and build an endogenous intelligent data model, which is a technical basis for realizing endogenous intelligence of a 6G network. At present, there are few related researches on constructing a wireless communication network knowledge graph to visualize the correlation characteristics of wireless communication data, so it is of great research significance to study the representation learning method of the wireless communication network knowledge graph. Graph neural networks are a new research hotspot and can process graph structure data and adapt to different field tasks. Considering the multi-source heterogeneous characteristics of wireless communication network data, using a homogeneous graph neural network for wireless communication network knowledge graph representation learning will sacrifice a lot of semantic information, which is not appropriate. The application constructs a wireless communication network knowledge graph representation learning method based on a heterogeneous graph neural network to mine the correlation characteristics of various communication indicators in the wireless communication network knowledge graph, optimize and complete the wireless communication network knowledge graph, provide help for subsequent network optimization tasks, and have important significance for realizing endogenous intelligence of a 6G network. SUMMARY

[0003] Inventive purpose: In view of the mining and learning of the wireless communication network knowledge graph, the application constructs a wireless communication network knowledge graph representation learning method based on a heterogeneous graph neural network, adopts the wireless communication network knowledge graph to depict the visual wireless communication data, considers the multi-source heterogeneous characteristics of the wireless communication network data, constructs a heterogeneous graph neural network model based on the wireless communication network knowledge graph to perform heterogeneous representation learning on the wireless communication network knowledge graph, maps the correlation between the key performance indicator nodes and the factor indicator nodes in the wireless communication network knowledge graph to a continuous vector space, learns the node representation vectors of the nodes, then calculates the correlation between the nodes, and verifies the test model by using the downstream task of link prediction, so as to optimize and complete the wireless communication network knowledge graph, provide help for subsequent network optimization tasks, and help realize the endogenous intelligence of the 6G network.

[0004] To achieve the above object, the application adopts the following technical scheme:

[0005] A wireless communication network knowledge graph representation learning method based on a heterogeneous graph neural network, the method comprising the following steps:

[0006] Step 1: The wireless communication network knowledge graph constructed according to expert knowledge contains the correlation between the communication performance indicators of different categories of attributes in the wireless communication network and the relationship between the data of the communication performance indicators of different categories of attributes in the wireless communication network; the data in the wireless communication network has the multi-source heterogeneous characteristics, and the wireless communication network knowledge graph is converted into a heterogeneous graph; in the heterogeneous graph, the entity set representing the communication performance indicators in the wireless communication network is mapped to each node in the heterogeneous graph, and the relationship set representing the relationship between the communication performance indicators in the wireless communication network is mapped to the edge in the heterogeneous graph.

[0007] Step 2: Construct a heterogeneous graph neural network model suitable for the wireless communication network knowledge graph, map the entities and relationships in the wireless communication network knowledge graph into nodes and edges in the heterogeneous graph neural network; the heterogeneous graph neural network model can automatically select a meta-path, and can not need to divide the meta-path in advance according to prior knowledge, so as to obtain the representation vectors of the key performance indicator nodes and the factor indicator nodes in the wireless communication network knowledge graph, and mine and learn the wireless communication network knowledge graph.

[0008] Step 3: Realize the representation learning of the wireless communication network knowledge graph based on the constructed heterogeneous graph neural network model and perform model training; the trained model verifies the test model effect by using the downstream task of link prediction, so as to construct the effective knowledge representation of the related communication performance indicators in the wireless communication network.

[0009] Step 4: based on the trained heterogeneous graph neural network model suitable for the wireless communication network knowledge graph, to optimize and complete the wireless communication network knowledge graph; the trained heterogeneous graph neural network model can effectively understand the correlation characteristics of the related communication performance indicators in the wireless communication network, and has important significance for realizing the endogenous intelligence of the wireless communication network.

[0010] Specifically, the communication performance indicators of each different category attribute in the wireless communication network in step 1 include:

[0011] According to the specific key performance indicators, the wireless communication network knowledge graph constructed includes physical layer key performance indicators, medium access control layer key performance indicators, wireless link control layer key performance indicators, upper and lower limit indicators of each layer key performance indicators, modulation coding and resource block control feedback process factor indicators, initial non-retransmission factor indicators, factors affecting uplink and downlink error rate indicators, factors affecting uplink and downlink resource block number indicators, other factors affecting each layer key performance indicators, and other user level evaluation indicators in wireless communication network, general non-adjustable data parameter indicators and adjustable data parameter indicators and the like.

[0012] Specifically, the wireless communication network knowledge graph is converted into a heterogeneous graph in step 1, which specifically includes:

[0013] Based on the expert knowledge in the field of wireless communication, the wireless communication network is constructed into a wireless communication network knowledge graph u i represents the i-th head node, r j represents the j-th relationship, v k represents the k-th tail node, the set ε is an entity set representing communication performance indicators of each different category attribute in the wireless communication network, and the set is a relationship set representing the relationship between the communication performance indicators of each different category attribute in the wireless communication network; the wireless communication network knowledge graph κ is converted into a heterogeneous graph The set V is a node set of communication performance indicators of each different category attribute in the wireless communication network, and the set E is a set of edges between nodes in the set V, the node type is the edge type is.

[0014] Further, the heterogeneous graph neural network model constructed in step 2 specifically includes the following steps:

[0015] Step 2.1: the present application constructs a heterogeneous graph V is a node set, and E is an edge set. The initial input feature matrix is the node feature matrix, where N represents the number of nodes, and M is the dimension of the feature vector. The communication performance indicators of nodes of different categories in the wireless communication network constitute the heterogeneous nodes in the heterogeneous graph, and the definition is is the node type mapping function, and for each node v i Each node v V (i = 1,...,N) has a different category, and the nodes are divided into different categories according to the category of the node, that is, Similarly, in order to construct a heterogeneous graph suitable for the knowledge graph of the wireless communication network, the semantic relationship between the nodes in the wireless communication network is fully considered, and the present application defines a type of node to another type of node as the heterogeneous edge required by the model constructed by the present application, is the edge type mapping function, where

[0016] Step 2.2: Constructing the candidate adjacency matrix tensor of the heterogeneous graph HetG A piece of the candidate adjacency matrix tensor represents the heterogeneous edge type adjacency matrix.

[0017] Step 2.3: Constructing the 1x1 non-negative selection weight of the candidate adjacency matrix tensor required for the lth layer of a single channel,

[0018] Step 2.4: Convolution of the 1x1 non-negative selection weight in the candidate adjacency matrix tensor and a softmax(W conv ), the lth layer graph structure P (l) of a single channel is selected by the weight of the lth layer of the heterogeneous edge type , and can be represented as:

[0019]

[0020] Step 2.5: Automatically learning the meta path of the constructed heterogeneous graph neural network model, for a class of meta path p, the adjacency matrix of the meta path p is obtained by multiplying the matrices of the selected first l-1 layer graph structure and the lth layer graph structure, and the adjacency matrix of the meta path p is the weighted sum of the candidate adjacency matrix obtained by 1x1 non-negative weight convolution in softmax(W conv ):

[0021]

[0022] Step 2.6: In order to automatically learn multiple types of meta-paths at the same time, multi-channel sampling is adopted, and the output channel of the 1x1 filter is set to C, so as to select the adjacency matrix tensor of multiple types of meta-paths The graph convolution network is applied to each output channel of the C meta-paths adjacency matrix, so as to obtain the node representation under different meta-paths, and the node representations under different meta-paths are spliced together to obtain the representation vector of the KPI node and the factor index node in the wireless communication network knowledge graph:

[0023]

[0024] wherein || represents splicing operation, represents the adjacency matrix of the cth channel, D c represents the degree matrix of, and W is a channel-shared trainable weight matrix.

[0025] Further, the step 3 realizes representation learning of the wireless communication network knowledge graph and model training based on the constructed heterogeneous graph neural network model, and specifically includes:

[0026] Step 3.1: Real-time real acquisition of the data of the communication performance indicators in the wireless communication network in the B5G / 6G test network, and pre-processing using interpolation algorithm and normalization algorithm to obtain the initial node feature matrix X of the KPI node and the factor index node in the wireless communication network knowledge graph.

[0027] Step 3.2: Designing a heterogeneous graph neural network model to embed the KPI node and the factor index node in the wireless communication network knowledge graph into a continuous vector space to obtain the representation vector of the KPI node and the factor index node in the wireless communication network knowledge graph.

[0028] Step 3.3: Training the constructed heterogeneous graph neural network model, and optimizing the model weight by minimizing the cross-entropy through back propagation and gradient descent.

[0029] Step 3.4: Testing and verifying the constructed heterogeneous graph neural network model, testing the trained model on the test set in the communication performance indicator data set of the wireless communication network, and verifying the performance of the constructed heterogeneous graph neural network model by using the downstream task of link prediction.

[0030] Step 3.5: After the trained model, the effective knowledge representation vector of the communication performance indicators of the wireless communication network is obtained.

[0031] Specifically, the data of the communication performance indicators in the wireless communication network in step 3 can be divided into a training data set D train And a test data set D test .

[0032] Specifically, the downstream task of link prediction in step 3 includes the following specific steps:

[0033] Since all the communication performance indicators in the actual wireless communication network do not have the same category features, each communication performance indicator has different category information, and therefore it is not appropriate to use a homogeneous graph neural network model to perform representation learning on the wireless communication network knowledge graph, and the application considers the different category information of the nodes in the wireless communication network, and constructs a heterogeneous graph neural network model to embed the key performance indicator nodes and factor indicator nodes in the wireless communication network knowledge graph into a continuous vector space; for the link prediction task, the nodes in the training set are represented using the heterogeneous graph neural network model constructed by the application to obtain the representation vectors of the nodes, and then the representation vectors are used for supervised learning on the positive and negative samples in the training set (the negative samples are resampled at each training round) to predict whether there is a connection between two nodes and to infer or predict the unobserved connection relationship in the graph, wherein the negative sampling technique is used to balance the imbalance between positive and negative edges in link prediction.

[0034] Specifically, the negative sampling technique used in step 3 to balance the imbalance between positive and negative edges in link prediction specifically includes:

[0035] The wireless communication network knowledge graph is converted into a heterogeneous graph, which is sparse, and link prediction is performed in a supervised manner, that is, it is regarded as a binary classification task, and all existing edges in the network are regarded as positive samples and non-existing edges are regarded as negative samples; perform segmentation on edge_index to ensure that there is no target leakage in node embedding when predicting test data, add two new attributes edge_label and edge_label_index to each segmented data, which are the edge label and edge index corresponding to each segmentation, edge_label_index is used to calculate the error, and edge_label will be used for model evaluation, add the same number of negative edges as positive edges to the test set, which are added to the edge_label and edge_label_index attributes, but not to the edge_index, the application uses 30% of the edges as positive samples of the test set, randomly breaks 70% of the edges, and the remaining edges are used as training samples; construct node pairs according to the edges.

[0036] Specifically, the model training of the constructed heterogeneous graph neural network model in step 3 specifically includes:

[0037] After importing the heterogeneous graph data including node information and edge information, the graph data is divided into a training set and a test set, the heterogeneous graph neural network model constructed by the application obtains a representation vector of a node, defines a proper loss function to measure the difference between the prediction result of the heterogeneous graph neural network model constructed by the application and the real label, and continuously updates the parameters of the model through a back propagation algorithm to reduce the value of the loss function, the loss function is a binary cross-entropy loss function, is the value of the loss function, η is a correction factor (η>1), y is a real label, is a prediction probability value of the model:

[0038]

[0039] Further, after being sufficiently trained and optimized in step 4, the trained heterogeneous graph neural network model is used to predict new edges in the graph according to specific application requirements, effectively understand the correlation characteristics of various communication indicators in the wireless communication network, optimize and complete the wireless communication network knowledge graph, and has important significance for realizing the endogenous intelligence of the wireless communication network.

[0040] Advantages:

[0041] The wireless communication network knowledge graph representation learning method based on the heterogeneous graph neural network constructed by the application considers the multi-source heterogeneous characteristics of wireless communication data, does not use a homogeneous graph neural network, but uses a heterogeneous graph neural network to perform heterogeneous representation learning on the wireless communication network knowledge graph, automatically learns a meta path through a candidate adjacency matrix tensor, obtains effective representation vectors of key performance indicator nodes and factor indicator nodes in the wireless communication network knowledge graph, and the heterogeneous graph neural network model trained is used for the downstream task of link prediction, and good classification effects are obtained. The application first proposes the wireless communication network knowledge graph representation learning method based on the heterogeneous graph neural network, good knowledge graph knowledge representation effects are obtained, the wireless communication network knowledge graph can be optimized and completed, the subsequent network optimization task can be facilitated, and the endogenous intelligence of the 6G network has important research significance. BRIEF DESCRIPTION OF DRAWINGS

[0042] Figure 1 The figure is a step diagram of the wireless communication network knowledge graph representation learning method based on the heterogeneous graph neural network in the embodiment of the application.

[0043] Figure 2 The figure is a simple schematic diagram of the correlation relationship between the key performance indicators and the factor indicators in the wireless communication network knowledge graph in the embodiment of the application.

[0044] Figure 3A schematic diagram of a knowledge graph of an uplink throughput of a key performance indicator of a wireless communication network in an embodiment of the present application.

[0045] Figure 4 A heterogeneous graph neural network model diagram suitable for a knowledge graph of a wireless communication network in an embodiment of the present application.

[0046] Figure 5 A F1 score result diagram of a representation learning method of a knowledge graph of an uplink throughput of a key performance indicator of a wireless communication network in an embodiment of the present application on a link prediction task.

[0047] Figure 6 An AUC score result diagram of a representation learning method of a knowledge graph of an uplink throughput of a key performance indicator of a wireless communication network in an embodiment of the present application on a link prediction task. DETAILED DESCRIPTION

[0048] The technical solutions provided by the present application will be described in detail below with reference to specific embodiments, and it should be understood that the following specific embodiments are only used to illustrate the present application and not to limit the scope of the present application.

[0049] A heterogeneous graph neural network-based representation learning method of a knowledge graph of a wireless communication network, a step diagram of the method is shown in Figure 1 The method specifically includes the following steps:

[0050] Step 1: a knowledge graph of an uplink throughput of a key performance indicator of a wireless communication network constructed according to expert knowledge, containing the correlation between the uplink throughput of the key performance indicator and the factor indicators in the wireless communication network and the relationship between the data of the uplink throughput of the key performance indicator and the factor indicators in the wireless communication network; the data in the wireless communication network has a multi-source heterogeneous characteristic, and the knowledge graph of the wireless communication network is converted into a heterogeneous graph; in the heterogeneous graph, the entity set representing the uplink throughput of the key performance indicator and the factor indicators in the wireless communication network is mapped to each node in the heterogeneous graph, and the relationship set representing the relationship between the uplink throughput of the key performance indicator and the factor indicators in the wireless communication network is mapped to the edge in the heterogeneous graph; a simple schematic diagram of the correlation between the key performance indicators and the factor indicators in the knowledge graph of the wireless communication network is shown in Figure 2 , wherein the key performance indicators and the factor indicators belong to different categories.

[0051] Step 2: Construct a heterogeneous graph neural network model suitable for a wireless communication network knowledge graph, and map entities and relationships in the wireless communication network knowledge graph into nodes and edges in the heterogeneous graph neural network; the heterogeneous graph neural network model can automatically select a meta-path, without the need to pre-classify the meta-path according to prior knowledge, to obtain a representation vector of a key performance indicator node and a factor indicator node in the wireless communication network knowledge graph, for mining and learning the wireless communication network knowledge graph.

[0052] Step 3: Implement representation learning of the wireless communication network knowledge graph based on the constructed heterogeneous graph neural network model and perform model training; the trained model is verified by a downstream task of link prediction to test the model effect, to construct an effective knowledge representation of related communication performance indicators in the wireless communication network.

[0053] Step 4: Based on the trained heterogeneous graph neural network model suitable for the wireless communication network knowledge graph, the wireless communication network knowledge graph is optimized and completed; the trained heterogeneous graph neural network model can effectively understand the correlation characteristics of related communication performance indicators in the wireless communication network, and has important significance for realizing the endogenous intelligence of the wireless communication network.

[0054] Specifically, the key performance indicator uplink throughput and the factor indicator in the wireless communication network in step 1 include:

[0055] According to the specific key performance indicator uplink throughput, the constructed wireless communication network knowledge graph is a wireless communication network key performance indicator uplink throughput knowledge graph, and the entity set includes physical layer uplink throughput, medium access control layer uplink throughput, radio link control layer uplink throughput, upper and lower limit indicators of each layer uplink throughput, factor indicators of modulation and coding and resource block control feedback processes, initial non-retransmission factor indicators, factor indicators affecting uplink error rate, factor indicators affecting uplink resource block number, remaining factor indicators affecting each layer key performance indicator, and other user level evaluation indicators in the wireless communication network, general non-tunable data parameter indicators, and tunable data parameter indicators, and the like.

[0056] Specifically, the wireless communication network knowledge graph is converted into a heterogeneous graph in step 1, which specifically includes:

[0057] Based on the expert knowledge in the field of wireless communication, the wireless communication network is constructed as a wireless communication network knowledge graph u i represents the i-th head node, r j represents the j-th relationship, v k represents the k-th tail node, the set ε represents an entity set of the key performance indicator uplink throughput and the factor indicator in the wireless communication network, and the set is a set of relationships representing the relationship between the key performance indicator uplink throughput and the factor index in the wireless communication network; a part of the schematic diagram of the wireless communication network key performance indicator uplink throughput knowledge graph is shown in Figure 3 ; the wireless communication network knowledge graph K is converted into a heterogeneous graph The set V is a set of nodes of the key performance indicator uplink throughput and the factor index in the wireless communication network, and the set E is a set of edges between the nodes in the set V, is a node type, is an edge type.

[0058] Further, the heterogeneous graph neural network model suitable for the wireless communication network knowledge graph constructed in step 2 has a model diagram as shown in Figure 4 , and the specific construction steps include:

[0059] Step 2.1: The present application constructs a heterogeneous graph V is a node set, and E is an edge set. The initial input feature matrix is a node feature matrix, where N represents the number of nodes, and M is the dimension of the feature vector. For the key performance indicator uplink throughput, there are 82 nodes, which are divided into 10 categories by the present application. The key performance indicator uplink throughput and the factor index node in the wireless communication network constitute a heterogeneous node in the heterogeneous graph, and are defined as is a node type mapping function, and for each node v i ∈V(i=1,...,N) has a different category, and the nodes are divided into different categories according to the category of the node, i.e. Similarly, in order to construct a heterogeneous graph suitable for the wireless communication network knowledge graph, the semantic relationship between the nodes in the wireless communication network is fully considered, and the present application defines a type of node to another type of node as a heterogeneous edge required by the model constructed by the present application, is an edge type mapping function, where

[0060] Step 2.2: Constructing the candidate adjacency matrix tensor of the heterogeneous graph HetG A piece of the candidate adjacency matrix tensor represents the adjacency matrix of the heterogeneous edge type .

[0061] Step 2.3: Constructing the 1x1 non-negative selection weight of the candidate adjacency matrix tensor required for the lth layer of a single channel,

[0062] Step 2.4: The candidate adjacency matrix tensor and a softmax(W convThe 1×1 non-negative selected weight convolution in the l-th layer graph structure P of a single channel (l) Heterogeneous edge types of the l-th layer weight As selected, setting l=4 in this invention can be expressed as:

[0063]

[0064] Step 2.5: Automatically learn the meta-paths of the constructed heterogeneous graph neural network model. For a class of meta-paths p, the adjacency matrix of meta-path p is... The adjacency matrix of the metapath p is obtained by multiplying the matrices of the selected first (l-1) layer graph structure and the l-th layer graph structure. It is composed of softmax(W conv The weighted sum of the candidate adjacency matrices obtained by 1×1 non-negative weighted convolution in ():

[0065]

[0066] Step 2.6: To automatically learn multiple types of metapaths simultaneously, multi-channel sampling is used. The output channel of the 1×1 filter is changed from a single channel to C. In this invention, C=4 is set to select the adjacency matrix tensor of multiple types of metapaths. The adjacency matrix of these C metapaths Each output channel is processed using a graph convolutional network to obtain node representations for different meta-paths. These node representations are then concatenated to obtain the representation vectors of key performance indicator nodes and factor indicator nodes in the knowledge graph of the wireless communication network.

[0067]

[0068] Here, || represents the concatenation operation. express The adjacency matrix of the c-th channel, D c express The degree matrix is ​​W, which is a channel-shared trainable weight matrix.

[0069] Furthermore, step 3, which involves learning the representation of the knowledge graph of the wireless communication network and training the model based on the constructed heterogeneous graph neural network model, specifically includes:

[0070] Step 3.1: Collect real-time data on communication performance indicators in the wireless communication network of the B5G / 6G experimental network, and preprocess the data using interpolation and normalization algorithms to obtain the node feature matrix X of key performance indicator nodes and factor indicator nodes in the knowledge graph of the wireless communication network.

[0071] Step 3.2: design a heterogeneous graph neural network model to embed the key performance indicator nodes and factor indicator nodes in the wireless communication network knowledge graph into a continuous vector space, to obtain the representation vectors of the key performance indicator nodes and factor indicator nodes in the wireless communication network knowledge graph.

[0072] Step 3.3: train the constructed heterogeneous graph neural network model, and optimize the model weights by minimizing the cross-entropy through back propagation and gradient descent.

[0073] Step 3.4: test and verify the constructed heterogeneous graph neural network model, and test the trained model on the test set in the wireless communication network communication performance indicator dataset, and verify the performance of the constructed heterogeneous graph neural network model by using the downstream task of link prediction.

[0074] Step 3.5: obtain the effective knowledge representation vectors of the wireless communication network communication performance indicators through the trained model.

[0075] Specifically, the data of the communication performance indicators in the wireless communication network in step 3 can be divided into a training data set D train and a test data set D test .

[0076] Specifically, the downstream task of link prediction in step 3 includes the following specific steps:

[0077] Since all communication performance indicators in the actual wireless communication network do not have the same category characteristics, and each communication performance indicator has different category information, it is not suitable to use a homogeneous graph neural network model to represent learning of the wireless communication network knowledge graph. The present application considers the different category information of the nodes in the wireless communication network, and constructs a heterogeneous graph neural network model to embed the key performance indicator nodes and factor indicator nodes in the wireless communication network knowledge graph into a continuous vector space. For the link prediction task, the present application uses the constructed heterogeneous graph neural network model to represent the nodes in the training set, to obtain the representation vectors of the nodes, and then uses these representation vectors to perform supervised learning on the positive and negative samples in the training set (re-sample negative samples at each training round), to predict whether there is a connection between two nodes, and to infer or predict the unobserved connection relationship in the graph. The negative sampling technique is used to balance the imbalance of positive and negative edges in link prediction. The F1 score result graph and the AUC score result graph of the representation learning method of the wireless communication network key performance indicator uplink throughput knowledge graph on the link prediction task are shown in Figure 5 、 Figure 6 and Figure 5 Figure 6It can be seen that the wireless communication network knowledge graph representation learning method based on the heterogeneous graph neural network is much better than the classic KG2E representation learning method.

[0078] Specifically, the step 3 adopts a negative sampling technique to balance the imbalance of positive and negative edges in link prediction, and specifically includes:

[0079] The wireless communication network knowledge graph is converted into a heterogeneous graph, which is sparse, and link prediction is performed in a supervised manner, that is, it is regarded as a binary classification task, and all existing edges in the network are regarded as positive samples, and non-existing edges are regarded as negative samples; segmentation is performed on edge_index to ensure that there is no target leakage on the node embedding when predicting the test data, two new attributes edge_label and edge_label_index are added to each segmented data, which are the edge label and edge index corresponding to each segmentation, edge_label_index is used to calculate the error, and edge_label is used for model evaluation, the same number of negative edges as positive edges are added to the test set, which are added to the edge_label and edge_label_index attributes, but not to the edge_index, the present application uses 30% of the edges as positive samples of the test set, and randomly breaks 70% of the edges, and the remaining edges are used as training samples; the node pairs are constructed according to the edges.

[0080] Specifically, the step 3 of training the constructed heterogeneous graph neural network model specifically includes:

[0081] After importing the heterogeneous graph data including node information and edge information, the graph data is segmented into a training set and a test set, the heterogeneous graph neural network model constructed by the present application obtains the representation vector of the node, and defines a proper loss function to measure the difference between the prediction result of the heterogeneous graph neural network model constructed by the present application and the true label; the parameters of the model are continuously updated by the back propagation algorithm to reduce the value of the loss function; the loss function is a binary cross-entropy loss function, is the value of the loss function, η is a correction factor (η>1), y is the true label, is the prediction probability value of the model:

[0082]

[0083] Further, in the step 4, after being fully trained and optimized, the trained heterogeneous graph neural network model is used to realize the representation learning of the wireless communication network knowledge graph based on the constructed heterogeneous graph neural network model according to the specific application requirement, and the representation vectors of the key performance indicator nodes and the factor indicator nodes in the wireless communication network knowledge graph are obtained, and the cosine similarity between two nodes is calculated By re-modifying the adjacency matrix of the graph data, new edges in the graph can be predicted, the correlation characteristics of various communication indicators in the wireless communication network can be effectively understood, the wireless communication network knowledge graph is optimized and completed, and it is of great significance to realize the endogenous intelligence of the wireless communication network.

[0084] The above describes the preferred embodiments of the present application in detail. It should be understood that those skilled in the art can make many modifications and changes without creative labor according to the concept of the present application. Therefore, any technical solution obtained by logical analysis, reasoning or limited experiment on the basis of the prior art according to the concept of the present application should be within the protection scope determined by the claims.

Claims

1. A wireless communication network knowledge graph representation learning method based on a heterogeneous graph neural network, characterized in that, The method comprises the following steps: A wireless communication network knowledge graph is constructed, and the wireless communication network knowledge graph is converted into a heterogeneous graph; in the heterogeneous graph, a set of entities representing communication performance indicators in the wireless communication network is mapped to each node in the heterogeneous graph, and a set of relationships representing relationships between the communication performance indicators in the wireless communication network is mapped to edges in the heterogeneous graph; Heterogeneous graph data composed of node information and edge information of the heterogeneous graph is divided into a training set and a test set; the node information comprises a node set, a node type and a node feature matrix; the edge information comprises an edge set and an edge feature matrix; A heterogeneous graph neural network model is constructed according to the heterogeneous graph data, comprising: constructing a candidate adjacency matrix tensor, constructing weights of the candidate adjacency matrix tensor required by each layer of a single channel, constructing a single-channel graph structure, obtaining an adjacency matrix of a type of meta path, simultaneously and automatically learning a plurality of types of meta path adjacency matrices by multi-channel sampling, and learning a representation vector of a node according to information of the plurality of types of meta path adjacency matrices; The constructed heterogeneous graph neural network model is trained by using the heterogeneous graph data; and a downstream task of link prediction is used to verify the model effect, so as to construct an effective knowledge representation of related communication performance indicators in the wireless communication network; The communication performance indicators of each different category attribute in the wireless communication network comprise: physical layer key performance indicators, medium access control layer key performance indicators, wireless link control layer key performance indicators, upper and lower limit indicators of each layer key performance indicator, factor indicators of a modulation and coding and resource block control feedback process, factor indicators of each factor of initial non-retransmission, factor indicators affecting uplink and downlink bit error rates, factor indicators affecting uplink and downlink resource block numbers, factor indicators affecting each layer key performance indicator, and other user level evaluation indicators in the wireless communication network, general non-adjustable data parameter indicators and adjustable data parameter indicators; The constructed heterogeneous graph neural network model is trained by using the heterogeneous graph data, and specifically comprises: Real-time real acquisition obtains the data of the communication performance index in the wireless communication network in the B5G / 6G test network, and uses interpolation algorithm and normalization algorithm for pretreatment to obtain the initial node feature matrix of the key performance indicator node and the factor indicator node in the wireless communication network knowledge graph ; The heterogeneous graph neural network model is designed to embed key performance indicator nodes and factor indicator nodes in the wireless communication network knowledge graph into a continuous vector space, to obtain representation vectors of the key performance indicator nodes and the factor indicator nodes in the wireless communication network knowledge graph; The constructed heterogeneous graph neural network model is trained, and the model weight is optimized by minimizing cross-entropy through back propagation and gradient descent.

2. The method of claim 1, wherein the method is based on a heterogeneous graph neural network. The wireless communication network knowledge graph is converted into a heterogeneous graph, and specifically comprises: Based on the expert knowledge in the field of wireless communication, the wireless communication network is constructed as a wireless communication network knowledge graph , represents the i-th head node, represents the j-th relationship, represents the k-th tail node, set is an entity set representing the communication performance indicators of various different categories of attributes in the wireless communication network, set is a relationship set representing the relationships between the communication performance indicators of various different categories of attributes in the wireless communication network; the wireless communication network knowledge graph is converted into a heterogeneous graph , set is a node set of the communication performance indicators of various different categories of attributes in the wireless communication network, set is a set of edges between the nodes in the set , is a node type, is an edge type.

3. The method of claim 2, wherein the method further comprises: The heterogeneous graph neural network model is constructed according to the heterogeneous graph data, and specifically comprises: Constructing isomorphic graphs of candidate adjacency matrix tensors of a patch in a candidate adjacency matrix tensor representing the first type isomorphic edges; , representing the number of nodes, ; Constructing a single channel from a plurality of channels a candidate adjacency matrix tensor required for constructing a layer non-negative selection weights, ; The candidate adjacency matrix tensor and In Non-negative selected weight convolution, the first channel of a single channel Layer diagram structure By the Heterogeneous edge types of layers weight The selected result is represented as: ； wherein, , denotes an adjacency matrix of the isomorphic edge type . Before Layer diagram structure and the first Matrix multiplication of layered graph structures yields a class of meta-paths. adjacency matrix , is represented as: ； Multi-channel sampling is used to... The filter's output channel is set from a single channel. This allows for the selection of adjacency matrix tensors for various types of metapaths. Regarding this Adjacency matrix of individual paths Each output channel is processed using a graph convolutional network to obtain node representations for different meta-paths. These node representations are then concatenated to obtain the representation vectors of key performance indicator nodes and factor indicator nodes in the knowledge graph of the wireless communication network. ； wherein, denotes a concatenation operation, denotes the adjacency matrix of the channel, denotes the degree matrix of is the initial node feature matrix, is the dimension of the feature vector, is a trainable weight matrix shared across channels.

4. The method of claim 1, wherein the method further comprises: The test model effect is verified by using a downstream task of link prediction, and the specific steps comprise: The constructed heterogeneous graph neural network model is used to represent nodes in the training set, to obtain representation vectors of the nodes, and then the representation vectors are used for supervised learning of positive and negative samples in the training set, to predict whether there is a connection between two nodes, to infer or predict unobserved connection relationships in the graph, wherein a negative sampling technology is used to balance the imbalance of positive and negative edges in link prediction; negative samples are resampled in each round of training.

5. The method of claim 4, wherein the method further comprises: The negative sampling technique is used to balance the imbalance of positive and negative edges in link prediction, specifically including: All edges existing in the heterogeneous graph are regarded as positive samples, and non-existing edges are regarded as negative samples; segmentation is performed on edge_index to ensure that there is no target leakage on node embedding when predicting test data, two new attributes edge_label and edge_label_index are added to each segmented data, which are edge labels and edge indexes corresponding to each segmentation, edge_label_index is used to calculate error, and edge_label will be used for model evaluation, the same number of negative edges as positive edges are added to the test set, which are added to the edge_label and edge_label_index attributes, but not to the edge_index; node pairs are constructed according to edges.

6. The method of claim 1, wherein the method is implemented by a wireless communication network knowledge graph representation learning method based on a heterogeneous graph neural network. When training the model, the parameters of the model are constantly updated through the back propagation algorithm to reduce the value of the loss function; The loss function is a binary cross-entropy loss function, is a value of the loss function, is a correction factor, is a true label, is a predicted probability value of the model: 。 7. The method of claim 1, wherein the method further comprises: The trained heterogeneous graph neural network model is used to predict new edges in the heterogeneous graph according to specific application requirements, and to infer edges with unclear connection relationships in the heterogeneous graph, so as to optimize and complete the wireless communication network knowledge graph.

Images