CN117196033A - Wireless communication network knowledge graph representation learning method based on heterogeneous graph neural network - Google Patents
Wireless communication network knowledge graph representation learning method based on heterogeneous graph neural network Download PDFInfo
- Publication number
- CN117196033A CN117196033A CN202311257111.6A CN202311257111A CN117196033A CN 117196033 A CN117196033 A CN 117196033A CN 202311257111 A CN202311257111 A CN 202311257111A CN 117196033 A CN117196033 A CN 117196033A
- Authority
- CN
- China
- Prior art keywords
- wireless communication
- communication network
- graph
- heterogeneous
- neural network
- Prior art date
- Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
- Pending
Links
- 238000004891 communication Methods 0.000 title claims abstract description 225
- 238000000034 method Methods 0.000 title claims abstract description 47
- 238000013528 artificial neural network Methods 0.000 title claims abstract description 29
- 238000003062 neural network model Methods 0.000 claims abstract description 51
- 238000005070 sampling Methods 0.000 claims abstract description 10
- 230000000694 effects Effects 0.000 claims abstract description 7
- 238000005516 engineering process Methods 0.000 claims abstract description 7
- 230000000295 complement effect Effects 0.000 claims abstract 2
- 239000011159 matrix material Substances 0.000 claims description 49
- 238000012549 training Methods 0.000 claims description 31
- 239000013598 vector Substances 0.000 claims description 30
- 238000012360 testing method Methods 0.000 claims description 28
- 230000006870 function Effects 0.000 claims description 18
- 238000010586 diagram Methods 0.000 claims description 13
- 238000011156 evaluation Methods 0.000 claims description 6
- 230000011218 segmentation Effects 0.000 claims description 6
- 230000002159 abnormal effect Effects 0.000 claims description 3
- 238000012937 correction Methods 0.000 claims description 3
- 238000010606 normalization Methods 0.000 claims description 3
- 238000007781 pre-processing Methods 0.000 claims description 3
- 238000006243 chemical reaction Methods 0.000 claims 1
- 239000013604 expression vector Substances 0.000 claims 1
- 238000011160 research Methods 0.000 abstract description 6
- 238000013473 artificial intelligence Methods 0.000 abstract description 4
- 238000013507 mapping Methods 0.000 description 6
- 238000010276 construction Methods 0.000 description 3
- 238000005065 mining Methods 0.000 description 3
- 238000005457 optimization Methods 0.000 description 3
- 238000012952 Resampling Methods 0.000 description 2
- 230000009286 beneficial effect Effects 0.000 description 2
- 238000011161 development Methods 0.000 description 2
- 230000018109 developmental process Effects 0.000 description 2
- 238000013499 data model Methods 0.000 description 1
- 238000002474 experimental method Methods 0.000 description 1
- 238000012986 modification Methods 0.000 description 1
- 230000004048 modification Effects 0.000 description 1
- 238000000547 structure data Methods 0.000 description 1
- 238000011144 upstream manufacturing Methods 0.000 description 1
- 230000000007 visual effect Effects 0.000 description 1
Classifications
-
- Y—GENERAL TAGGING OF NEW TECHNOLOGICAL DEVELOPMENTS; GENERAL TAGGING OF CROSS-SECTIONAL TECHNOLOGIES SPANNING OVER SEVERAL SECTIONS OF THE IPC; TECHNICAL SUBJECTS COVERED BY FORMER USPC CROSS-REFERENCE ART COLLECTIONS [XRACs] AND DIGESTS
- Y02—TECHNOLOGIES OR APPLICATIONS FOR MITIGATION OR ADAPTATION AGAINST CLIMATE CHANGE
- Y02D—CLIMATE CHANGE MITIGATION TECHNOLOGIES IN INFORMATION AND COMMUNICATION TECHNOLOGIES [ICT], I.E. INFORMATION AND COMMUNICATION TECHNOLOGIES AIMING AT THE REDUCTION OF THEIR OWN ENERGY USE
- Y02D30/00—Reducing energy consumption in communication networks
- Y02D30/70—Reducing energy consumption in communication networks in wireless communication networks
Landscapes
- Mobile Radio Communication Systems (AREA)
Abstract
The invention 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 auxiliary wireless communication. According to the invention, the wireless communication network knowledge graph is converted into the heterogeneous graph containing the multi-source heterogeneous nodes, a wireless communication network knowledge graph representation learning model based on the heterogeneous graph neural network is constructed, the constructed heterogeneous graph neural network model is trained, the effect of the model is verified by adopting a downstream task of link prediction, the problem of unbalanced positive and negative sides in the link prediction is solved by adopting a negative sampling technology, the constructed heterogeneous graph neural network model can effectively learn and mine the wireless communication network knowledge graph, optimize and complement the wireless communication network knowledge graph, understand the association characteristics of each communication index in the wireless communication network, and have important research significance for realizing the internal intelligence of the wireless communication network.
Description
Technical Field
The invention 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 auxiliary wireless communication.
Background
With the rapid development of digitalization and large-scale internet of things, wireless communication networks are focusing more on the needs of specific application scenarios and the performance and architecture of the overall communication network. Current wireless communication networks are not effectively empirical and continue to improve, and wireless communication networks need to have intelligent capabilities to address this problem. With the rapid development of artificial intelligence, the level of intelligence and network quality of wireless communication networks are gradually increasing. The artificial intelligence enables the 6G wireless communication network, and the network configuration is optimized and the network performance is improved by learning the characteristics of wireless network infrastructure and sensor equipment in big data. Knowledge maps can store knowledge in a graphical structure, and can infer and discover new knowledge. The knowledge graph is utilized to describe the wireless communication network, so that the wireless communication network can be effectively interpreted and inferred, wireless communication data are converted into available wireless communication knowledge, and an endogenous intelligent data model is built, so that the method is a technical basis for realizing the endogenous intelligence of the 6G network. At present, related researches on building a wireless communication network knowledge graph to visualize the association characteristics of wireless communication data are rarely performed, so that research on a representation learning method of the wireless communication network knowledge graph has high research significance. The graph neural network is an emerging research hotspot, can process graph structure data, and can adapt to tasks in different fields. Taking into account the multi-source heterogeneous nature of wireless communication network data, learning a wireless communication network knowledge-graph representation with a isomorphic neural network would sacrifice much semantic information, which is not appropriate. The invention constructs a wireless communication network knowledge graph representation learning method based on the heterogeneous graph neural network, which is used for excavating the association characteristic of each communication index in the wireless communication network knowledge graph, optimizing and complementing the wireless communication network knowledge graph, providing help for the subsequent network optimization task and having important significance for realizing the 6G network endogenetic intelligence.
Disclosure of Invention
The invention aims to: aiming at mining and learning of wireless communication network knowledge maps, the invention constructs a wireless communication network knowledge map representation learning method based on a heterogeneous map neural network, adopts the wireless communication network knowledge map to depict visual wireless communication data, considers the multi-source heterogeneous characteristics of the wireless communication network data, builds a heterogeneous map neural network model to carry out heterogeneous representation learning on the wireless communication network knowledge map based on the wireless communication network knowledge map, maps the association relation between key performance index nodes and factor index nodes in the wireless communication network knowledge map to a continuous vector space, learns the node representation vector of each node, calculates the association relation between each node, and uses a downstream task of link prediction to verify a test model for optimizing and complementing the wireless communication network knowledge map, thereby providing help for the subsequent network optimization task and being beneficial to realizing the endogenous intelligence of the 6G network.
In order to achieve the above purpose, the invention adopts the following technical scheme:
a wireless communication network knowledge graph representation learning method based on a heterogeneous graph neural network comprises the following steps:
step 1: the wireless communication network knowledge graph constructed according to expert knowledge comprises the association relation between the communication performance indexes of different category attributes in the wireless communication network and the relation between the data of the communication performance indexes of different category attributes in the wireless communication network; the data in the wireless communication network has multi-source heterogeneous characteristics, and the knowledge graph of the wireless communication network is converted into a different graph; in the heterogram, an entity set representing the communication performance index in the wireless communication network is mapped to each node in the heterogram, and a relation set representing the relation between the communication performance indexes in the wireless communication network is mapped to an edge in the heterogram.
Step 2: constructing a heterogeneous graph neural network model applicable to the wireless communication network knowledge graph, and mapping entities and relations 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 element paths, and element paths do not need to be divided in advance according to priori knowledge, so that the representation vectors of key performance index nodes and factor index nodes in the wireless communication network knowledge graph are obtained and are used for mining and learning the wireless communication network knowledge graph.
Step 3: realizing the representation learning of the wireless communication network knowledge graph based on the constructed heterogeneous graph neural network model and carrying out model training; the training model uses the downstream tasks of link prediction to verify test model effects for constructing an efficient knowledge representation of relevant communication performance indicators in a wireless communication network.
Step 4: based on a trained heterogeneous graph neural network model suitable for the wireless communication network knowledge graph, the method is used for optimizing and complementing the wireless communication network knowledge graph; the trained heterograph neural network model can effectively understand the correlation characteristics of related communication performance indexes in the wireless communication network and has important significance for realizing the endogenous intelligence of the wireless communication network.
Specifically, the communication performance indexes of each different category attribute in the wireless communication network in the step 1 include:
according to the specific key performance indexes, the constructed wireless communication network knowledge graph is characterized in that the entity set comprises physical layer key performance indexes, media access control layer key performance indexes, wireless link control layer key performance indexes, upper and lower limit indexes of each layer key performance indexes, factor indexes of modulation coding and resource block control feedback processes, initial non-retransmission factor indexes, factor indexes affecting uplink and downlink error rates, factor indexes affecting the number of uplink and downlink resource blocks, factor indexes affecting other key performance indexes of each layer, user level evaluation indexes in other wireless communication networks, general non-adjustment data parameter indexes, adjustable data parameter indexes and the like.
Specifically, in the step 1, the wireless communication network knowledge graph is converted into an abnormal graph, which specifically includes:
based on expert knowledge in the wireless communication field, constructing a wireless communication network as a wireless communication network knowledge graphu i Represents the i-th head node, r j Represents the j-th relation, v k Represents the kth end node, set epsilon is the set of entities representing the communication performance index of each different category attribute in the wireless communication network, set +.>Is a set of relationships representing relationships between communication performance indicators of various category attributes in a wireless communication network; converting the wireless communication network knowledge graph kappa into an isomerism graph +.>Set V is a set of communication performance index nodes of various different category attributes in the wireless communication network, set E is a set of edges between nodes in set V, +.>For node type +.>Is of the edge type.
Further, the heterogeneous graph neural network model constructed in the step 2 specifically includes:
step 2.1: the invention constructs the isomerism mapV is the set of nodes and E is the set of edges. Initial input feature matrix->The node feature matrix is characterized in that N represents the number of nodes, and M is the dimension of the feature vector. Communication performance index nodes with different types of attributes in a wireless communication network form heterogeneous nodes in different patterns and are definedFor each node v, as a node type mapping function i v (i=1,., N) all have different categories, 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 a knowledge graph of a wireless communication network, semantic relationships among nodes in the wireless communication network are fully considered, the method defines that edges between one type of nodes and the other type of nodes are heterogeneous edges which are specially required by the construction model of the method, and the method comprises the steps of>Mapping functions for edge types, wherein->
Step 2.2: constructing candidate adjacency matrix tensors for heterograms HetGOne of the candidate adjacency matrix tensors +.>Representing heterogeneous edge type ++>Is a contiguous matrix of (a) a plurality of (b) a plurality of (c).
Step 2.3: the 1 x 1 non-negative selection weights of the candidate adjacency matrix tensors required to construct the first layer of a single channel,
step 2.4: tensor of candidate adjacency matrixAnd a softmax (W conv ) 1 x 1 non-negative selection weight convolution in a single channel, layer 1 graph structure P (l) By the type of heterogeneous edge of layer I->Weight of +.>The selection can be expressed as:
step 2.5: automatically learning element paths of the constructed heterogeneous graph neural network model, and for one type of element path p, obtaining an adjacent matrix of the element path pFrom the matrix multiplication of the selected previous l-1 layer diagram structure and the first layer diagram structure, the adjacency matrix of the element path p is +.>Is made of softmax (W conv ) A weighted sum of candidate adjacency matrices obtained by 1 x 1 non-negative weight convolution:
step 2.6: to automatically learn multiple types of element paths simultaneously, multiple channel sampling is adopted, and the output channel of a 1×1 filter is set to C from single channel, so that adjacent matrix tensors of multiple types of element paths are selectedAdjacency matrix for these C-element paths +.>The graph convolutional network is applied to each output channel of the wireless communication network, so that node representations under different element paths are obtained, the node representations under different element paths are spliced together, and the representation vectors of the key performance index nodes and the factor index nodes in the knowledge graph of the wireless communication network are obtained:
wherein, the I represents the splicing operation,representation->Adjacency matrix of c-th channel, D c Representation->W is a channel shared trainable weight matrix.
Further, in the step 3, based on the constructed heterogeneous graph neural network model, representation learning of the wireless communication network knowledge graph is realized and model training is performed, which specifically includes:
step 3.1: and acquiring data of communication performance indexes in a wireless communication network in the B5G/6G test network in real time, and preprocessing by using an interpolation algorithm and a normalization algorithm to obtain an initial node characteristic matrix X of key performance index nodes and factor index nodes in a wireless communication network knowledge graph.
Step 3.2: and designing a heterogeneous graph neural network model, and embedding the key performance index nodes and the factor index nodes in the wireless communication network knowledge graph into a continuous vector space to obtain the representation vectors of the key performance index nodes and the factor index nodes in the wireless communication network knowledge graph.
Step 3.3: training the constructed heterogeneous graph neural network model, and optimizing the model weight by back propagation and gradient descent and minimizing cross entropy.
Step 3.4: and testing and verifying the constructed heterogeneous graph neural network model, testing the trained model on a test set in a communication performance index data set in a wireless communication network, and verifying the performance of the constructed heterogeneous graph neural network model by adopting a downstream task of link prediction.
Step 3.5: and obtaining an effective knowledge representation vector of the communication performance index of the wireless communication network through the trained model.
Specifically, the data of the communication performance index in the wireless communication network in the step 3 may be divided into a training data set D train And test dataset D test 。
Specifically, the downstream task of the link prediction in the step 3 includes the following specific steps:
because all communication performance indexes in an actual wireless communication network do not have the same category characteristics, and all communication performance indexes have different category information, the use of a isomorphic neural network model for representing and learning the wireless communication network knowledge graph is unsuitable, and the invention constructs the isomorphic neural network model by taking the different category information of nodes in the wireless communication network into consideration, so that key performance index nodes and factor index nodes in the wireless communication network knowledge graph are embedded into a continuous vector space; for a link prediction task, the nodes in the training set are represented by using the heterogeneous graph neural network model constructed by the method to obtain the representation vectors of the nodes, then the positive and negative samples in the training set (resampling the negative samples in each training round) are subjected to supervised learning by using the representation vectors, whether connection exists between the two nodes is predicted, and the connection relation which is not observed in a prediction graph is deduced or predicted, wherein the unbalance of the positive and negative edges in the link prediction is balanced by adopting a negative sampling technology.
Specifically, the balancing the imbalance of the positive and negative edges in the link prediction by adopting the negative sampling technology in the step 3 specifically includes:
converting the wireless communication network knowledge graph into a heterogeneous graph, wherein the graph is sparse, and link prediction is performed in a supervised mode, namely the graph is regarded as a classification task, edges existing in the network are regarded as positive samples, and edges which are not existing are regarded as negative samples; performing segmentation on edge_index to ensure that no target leakage exists on node embedding when predicting test data, adding two new attributes edge_label and edge_label_index to each segmented data, wherein the edge label and the edge index correspond to each segmentation, the edge_label_index is used for calculating errors, the edge_label is used for model evaluation, the same number of negative edges as positive edges are added to a test set, the negative edges are added to edge_label and edge_label_index attributes, but the negative edges are not added to edge_index, and the invention uses 30% of edges as positive samples of a test set, randomly breaks 70% of edges, and the rest of edges are used as training samples; node pairs constructed from edges.
Specifically, the training of the model of the constructed heterogeneous graph neural network model in the step 3 specifically includes:
after heterogeneous graph data comprising node information and side information is imported, the heterogeneous graph data is divided into a training set and a testing set, the heterogeneous graph neural network model constructed by the method obtains a representation vector of the node, and a proper loss function is defined to measure the difference between a prediction result and a real label of the heterogeneous graph neural network model constructed by the method; continuously updating 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, eta is the correction factor (eta > 1), y is the true label,/I>Is the predicted probability value of the model:
further, after the training and tuning are performed fully in the step 4, a trained heterographing neural network model is used, a new edge in the graph can be predicted according to specific application requirements, the association characteristics of each communication index in the wireless communication network can be understood effectively, the knowledge graph of the wireless communication network can be optimized and complemented, and the method has important significance for realizing the endogenous intelligence of the wireless communication network.
The beneficial effects are that:
according to the wireless communication network knowledge graph representation learning method based on the heterogeneous graph neural network, the multi-source heterogeneous characteristics of wireless communication data are considered, the heterogeneous graph neural network is not adopted, heterogeneous graph neural network is adopted to carry out heterogeneous representation learning on the wireless communication network knowledge graph, element paths are automatically learned through candidate adjacency matrix tensors, effective representation vectors of key performance index nodes and factor index nodes in the wireless communication network knowledge graph are obtained, and a trained heterogeneous graph neural network model is used for obtaining a good classification effect for a downstream task of link prediction. The invention provides a wireless communication network knowledge graph representation learning method based on the heterogeneous graph neural network for the first time, obtains a good knowledge graph knowledge representation effect, can be used for optimizing and complementing the wireless communication network knowledge graph, can provide help for the subsequent network optimization task, and has important research significance for realizing the endogenous intelligence of the 6G network.
Drawings
Fig. 1 is a diagram showing steps of a learning method based on a knowledge graph of a wireless communication network of a heterogeneous graph neural network in an embodiment of the invention.
Fig. 2 is a simple schematic diagram of an association relationship between a key performance indicator and a factor indicator in a knowledge graph of a wireless communication network according to an embodiment of the present invention.
Fig. 3 is a schematic diagram of an uplink throughput knowledge graph portion of a wireless communication network key performance indicator according to an embodiment of the present invention.
Fig. 4 is a schematic diagram of a heterogeneous neural network suitable for a knowledge graph of a wireless communication network according to an embodiment of the present invention.
Fig. 5 is a F1 score result diagram of a learning method on a link prediction task in an embodiment of the present invention.
Fig. 6 is a graph of AUC score results of a learning method on a link prediction task in an embodiment of the present invention.
Detailed Description
The technical scheme provided by the present invention will be described in detail with reference to the following specific examples, and it should be understood that the following specific examples are only for illustrating the present invention and are not intended to limit the scope of the present invention.
A wireless communication network knowledge graph representation learning method based on a heterogeneous graph neural network is shown in fig. 1, and specifically comprises the following steps:
step 1: the wireless communication network key performance index uplink throughput knowledge graph constructed according to expert knowledge comprises an incidence relation between key performance index uplink throughput and factor indexes in the wireless communication network and a relation between key performance index uplink throughput and factor index data in the wireless communication network; the data in the wireless communication network has multi-source heterogeneous characteristics, and the knowledge graph of the wireless communication network is converted into a different graph; in the heterogram, an entity set representing the uplink throughput of the key performance index and the factor index in the wireless communication network is mapped to each node in the heterogram, and a relation set representing the relation between the uplink throughput of the key performance index and the factor index in the wireless communication network is mapped to an edge in the heterogram; a simple schematic diagram of the association between the key performance indicators and the factor indicators in the wireless communication network knowledge graph is shown in fig. 2, wherein the key performance indicators and the factor indicators all belong to different categories.
Step 2: constructing a heterogeneous graph neural network model applicable to the wireless communication network knowledge graph, and mapping entities and relations 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 element paths, and element paths do not need to be divided in advance according to priori knowledge, so that the representation vectors of key performance index nodes and factor index nodes in the wireless communication network knowledge graph are obtained and are used for mining and learning the wireless communication network knowledge graph.
Step 3: realizing the representation learning of the wireless communication network knowledge graph based on the constructed heterogeneous graph neural network model and carrying out model training; the training model uses the downstream tasks of link prediction to verify test model effects for constructing an efficient knowledge representation of relevant communication performance indicators in a wireless communication network.
Step 4: based on a trained heterogeneous graph neural network model suitable for the wireless communication network knowledge graph, the method is used for optimizing and complementing the wireless communication network knowledge graph; the trained heterograph neural network model can effectively understand the correlation characteristics of related communication performance indexes in the wireless communication network and has important significance for realizing the endogenous intelligence of the wireless communication network.
Specifically, the uplink throughput and factor index of the key performance index in the wireless communication network in step 1 includes:
according to the specific uplink throughput of the key performance index, the constructed wireless communication network knowledge graph is the uplink throughput knowledge graph of the key performance index of the wireless communication network, and the entity set comprises the uplink throughput of the physical layer, the uplink throughput of the media access control layer, the uplink throughput of the wireless link control layer, the upper and lower limit indexes of the uplink throughput of each layer, the factor indexes of the modulation coding and resource block control feedback flow, the factor indexes of the initial non-retransmission factors, the factor indexes affecting the uplink error rate, the factor indexes affecting the uplink resource block number, the factor indexes affecting the other key performance indexes of each layer, the user level evaluation indexes in other wireless communication networks, the general non-adjustable data parameter indexes, the adjustable data parameter indexes and the like.
Specifically, in the step 1, the wireless communication network knowledge graph is converted into an abnormal graph, which specifically includes:
based on expert knowledge in the wireless communication field, constructing a wireless communication network as a wireless communication network knowledge graphu i Represents the i-th head node, r j Represents the j-th relation, v k Representing the kth tail node, the set epsilon is an entity set representing the uplink throughput of key performance indicators and factor indicators in the wireless communication network, and the set +.>A relation set representing the relation between the uplink throughput of the key performance index and the factor index in the wireless communication network; a schematic diagram of a part of the uplink throughput knowledge graph of the key performance indicators of the wireless communication network is shown in fig. 3; converting the wireless communication network knowledge graph kappa into an isomerism graph +.>Set V is a key performance indicator uplink throughput and factor indicator node set in the wireless communication network, set E is a set of edges between nodes in set V, </i >>For node type +.>Is of the edge type.
Further, the heterogeneous graph neural network model constructed in the step 2 and suitable for the wireless communication network knowledge graph, the model graph of which is shown in fig. 4, specifically comprises the following construction steps:
step 2.1: the invention constructs the isomerism mapV is the set of nodes and E is the set of edges. Initial input feature matrix->The node feature matrix is characterized in that N represents the number of nodes, and M is the dimension of the feature vector. For the key performance indicator upstream throughput, which contains 82 nodes, the present invention classifies these 82 nodes into 10 classes. The key performance index uplink throughput and the factor index nodes in the wireless communication network form heterogeneous nodes in a heterogeneous graph, and ++is defined>For each node v, as a node type mapping function i E V (i=1.,. The N) all have different categories, 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 a knowledge graph of a wireless communication network, semantic relationships among nodes in the wireless communication network are fully considered, the method defines that edges between one type of nodes and the other type of nodes are heterogeneous edges which are specially required by the construction model of the method, and the method comprises the steps of>Mapping functions for edge types, wherein
Step 2.2: constructing candidate adjacency matrix tensors for heterograms HetGOne of the candidate adjacency matrix tensors +.>Representing heterogeneous edge type ++>Is a contiguous matrix of (a) a plurality of (b) a plurality of (c).
Step 2.3: candidate moment of adjacency required for constructing first layer of single channelThe 1 x 1 non-negative selection weights of the matrix tensor,
step 2.4: tensor of candidate adjacency matrixAnd a softmax (W conv ) 1 x 1 non-negative selection weight convolution in a single channel, layer 1 graph structure P (l) By the type of heterogeneous edge of layer I->Weight of +.>Optionally, the setting of l=4 in the present invention can be expressed as:
step 2.5: automatically learning element paths of the constructed heterogeneous graph neural network model, and for one type of element path p, obtaining an adjacent matrix of the element path pFrom the matrix multiplication of the selected previous l-1 layer diagram structure and the first layer diagram structure, the adjacency matrix of the element path p is +.>Is made of softmax (W conv ) A weighted sum of candidate adjacency matrices obtained by 1 x 1 non-negative weight convolution:
step 2.6: to automatically learn multiple types of element paths simultaneously, the output channels of a 1×1 filter are sampled from a single channelThe channel is set as C, and C=4 is set in the invention, so that the adjacent matrix tensor of multiple types of element paths is selectedAdjacency matrix for these C-element paths +.>The graph convolutional network is applied to each output channel of the wireless communication network, so that node representations under different element paths are obtained, the node representations under different element paths are spliced together, and the representation vectors of the key performance index nodes and the factor index nodes in the knowledge graph of the wireless communication network are obtained:
wherein, the I represents the splicing operation,representation->Adjacency matrix of c-th channel, D c Representation->W is a channel shared trainable weight matrix.
Further, in the step 3, based on the constructed heterogeneous graph neural network model, representation learning of the wireless communication network knowledge graph is realized and model training is performed, which specifically includes:
step 3.1: and acquiring data of communication performance indexes in a wireless communication network in the B5G/6G test network in real time, and preprocessing by using an interpolation algorithm and a normalization algorithm to obtain a node characteristic matrix X of key performance index nodes and factor index nodes in a wireless communication network knowledge graph.
Step 3.2: and designing a heterogeneous graph neural network model, and embedding the key performance index nodes and the factor index nodes in the wireless communication network knowledge graph into a continuous vector space to obtain the representation vectors of the key performance index nodes and the factor index nodes in the wireless communication network knowledge graph.
Step 3.3: training the constructed heterogeneous graph neural network model, and optimizing the model weight by back propagation and gradient descent and minimizing cross entropy.
Step 3.4: and testing and verifying the constructed heterogeneous graph neural network model, testing the trained model on a test set in a communication performance index data set in a wireless communication network, and verifying the performance of the constructed heterogeneous graph neural network model by adopting a downstream task of link prediction.
Step 3.5: and obtaining an effective knowledge representation vector of the communication performance index of the wireless communication network through the trained model.
Specifically, the data of the communication performance index in the wireless communication network in the step 3 may be divided into a training data set D train And test dataset D test 。
Specifically, the downstream task of the link prediction in the step 3 includes the following specific steps:
because all communication performance indexes in an actual wireless communication network do not have the same category characteristics, and all communication performance indexes have different category information, the use of a isomorphic neural network model for representing and learning the wireless communication network knowledge graph is unsuitable, and the invention constructs the isomorphic neural network model by taking the different category information of nodes in the wireless communication network into consideration, so that key performance index nodes and factor index nodes in the wireless communication network knowledge graph are embedded into a continuous vector space; for a link prediction task, the nodes in the training set are represented by using the heterogeneous graph neural network model constructed by the method to obtain the representation vectors of the nodes, then the positive and negative samples in the training set (resampling the negative samples in each training round) are subjected to supervised learning by using the representation vectors, whether connection exists between the two nodes is predicted, and the connection relation which is not observed in a prediction graph is deduced or predicted, wherein the unbalance of the positive and negative edges in the link prediction is balanced by adopting a negative sampling technology. The F1 score result graph and the AUC score result graph of the learning method for expressing the key performance index uplink throughput knowledge graph of the wireless communication network on the link prediction task are respectively shown in fig. 5 and 6. As can be seen from fig. 5 and fig. 6, the knowledge graph representation learning method of the wireless communication network based on the heterogeneous graph neural network is far superior to the classical KG2E representation learning method.
Specifically, the balancing the imbalance of the positive and negative edges in the link prediction by adopting the negative sampling technology in the step 3 specifically includes:
converting the wireless communication network knowledge graph into a heterogeneous graph, wherein the graph is sparse, and link prediction is performed in a supervised mode, namely the graph is regarded as a classification task, edges existing in the network are regarded as positive samples, and edges which are not existing are regarded as negative samples; performing segmentation on edge_index to ensure that no target leakage exists on node embedding when predicting test data, adding two new attributes edge_label and edge_label_index to each segmented data, wherein the edge label and the edge index correspond to each segmentation, the edge_label_index is used for calculating errors, the edge_label is used for model evaluation, the same number of negative edges as positive edges are added to a test set, the negative edges are added to edge_label and edge_label_index attributes, but the negative edges are not added to edge_index, and the invention uses 30% of edges as positive samples of a test set, randomly breaks 70% of edges, and the rest of edges are used as training samples; node pairs constructed from edges.
Specifically, the training of the model of the constructed heterogeneous graph neural network model in the step 3 specifically includes:
after heterogeneous graph data comprising node information and side information is imported, the heterogeneous graph data is divided into a training set and a testing set, the heterogeneous graph neural network model constructed by the method obtains a representation vector of the node, and a proper loss function is defined to measure the difference between a prediction result and a real label of the heterogeneous graph neural network model constructed by the method; continuously updating parameters of the model through a back propagation algorithm to reduce the value of the loss function; loss ofThe function is a binary cross entropy loss function,is the value of the loss function, eta is the correction factor (eta > 1), y is the true label,/I>Is the predicted probability value of the model:
further, after the full training and tuning in the step 4, the trained heterograph neural network model is used, according to specific application requirements, the representation learning of the wireless communication network knowledge graph is realized based on the constructed heterograph neural network model, the representation vectors of the key performance index node and the factor index node in the wireless communication network knowledge graph are obtained, and the cosine similarity between the two nodes is calculatedThe adjacency matrix of the graph data is revised again, so that a new edge in the graph can be predicted, the association characteristic of each communication index in the wireless communication network can be effectively understood, the knowledge graph of the wireless communication network is optimized and complemented, and the method has important significance for realizing the endogenous intelligence of the wireless communication network.
The foregoing describes in detail preferred embodiments of the present invention. It should be understood that numerous modifications and variations can be made in accordance with the concepts of the invention by one of ordinary skill in the art without undue burden. Therefore, all technical solutions which can be obtained by logic analysis, reasoning or limited experiments based on the prior art by the person skilled in the art according to the inventive concept shall be within the scope of protection defined by the claims.
Claims (9)
1. A wireless communication network knowledge graph representation learning method based on a heterogeneous graph neural network is characterized by comprising the following steps:
constructing a wireless communication network knowledge graph, and converting the wireless communication network knowledge graph into an abnormal graph; in the heterogeneous graph, an entity set representing the communication performance index in the wireless communication network is mapped to each node in the heterogeneous graph, and a relation set representing the relation between the communication performance indexes in the wireless communication network is mapped to an edge in the heterogeneous graph;
dividing the heterogeneous graph data formed by the node information and the side information of the heterogeneous graph into a training set and a testing set; the node information comprises a node set, a node type and a node characteristic matrix; the side information comprises a side set and a side feature matrix;
constructing a heterogeneous graph neural network model according to the heterogeneous graph data, wherein the constructing comprises the following steps: constructing candidate adjacency matrix tensors, constructing weights of the candidate adjacency matrix tensors required by each single channel layer, constructing a single channel graph structure, obtaining adjacency matrixes of one type of element paths, adopting multi-channel sampling to automatically learn multiple types of element path adjacency matrixes at the same time, and learning expression vectors of nodes according to information of the multiple types of element path adjacency matrixes;
training the constructed heterogeneous graph neural network model by utilizing the heterogeneous graph data; and uses the downstream tasks of the link prediction to verify the model effect for constructing an efficient knowledge representation of the relevant communication performance indicators in the wireless communication network.
2. The method for learning the knowledge graph representation of the wireless communication network based on the heterogeneous graph neural network according to claim 1, wherein the communication performance indexes of different category attributes in the wireless communication network comprise: physical layer key performance indicators, media access control layer key performance indicators, radio link control layer key performance indicators, upper and lower limit indicators of each layer key performance indicators, factor indicators of modulation coding and resource block control feedback flow, factor indicators of initial non-retransmission, factor indicators affecting uplink and downlink error rates, factor indicators affecting uplink and downlink resource block numbers, factor indicators affecting other layers key performance indicators, user level evaluation indicators in other wireless communication networks, general non-modulation data parameter indicators, and adjustable data parameter indicators.
3. The method for learning the wireless communication network knowledge graph representation based on the heterogeneous graph neural network according to claim 1, wherein the wireless communication network knowledge graph is converted into the heterogeneous graph, and the method specifically comprises the following steps:
based on expert knowledge in the wireless communication field, constructing a wireless communication network as a wireless communication network knowledge graphu i Represents the i-th head node, r j Represents the j-th relation, v k Represents the kth end node, set epsilon is the set of entities representing the communication performance index of each different category attribute in the wireless communication network, set +.>Is a set of relationships representing relationships between communication performance indicators of various category attributes in a wireless communication network; knowledge graph of wireless communication network->Conversion to an isomerism map->Set V is a set of communication performance index nodes of various different category attributes in the wireless communication network, set E is a set of edges between nodes in set V, +.>For node type +.>Is of the edge type.
4. The method for learning the knowledge graph representation of the wireless communication network based on the heterogeneous graph neural network according to claim 3, wherein constructing the heterogeneous graph neural network model according to the sample set specifically comprises the following steps:
constructing candidate adjacency matrix tensors for heterograms HetGOne piece a in candidate adjacency matrix tensor k ∈R N×N An adjacency matrix representing a kth type of heterogeneous edge; k=1..k, N represents the number of nodes, +.>
The 1 x 1 non-negative selection weights of the candidate adjacency matrix tensors required to construct the first layer of a single channel,
tensor of candidate adjacency matrixAnd a softmax (W conv ) 1 x 1 non-negative selection weight convolution in a single channel, layer 1 graph structure P (l) By heterogeneous edge type t of layer I l Weight of +.>The selection results, expressed as:
wherein,representing heterogeneous edge type t l Is a contiguous matrix of (a);
multiplying the matrix of the former l-1 layer diagram structure and the first layer diagram structure to obtain an adjacent matrix G of a type of element path p p Expressed as:
the output channel of the 1X 1 filter is set to C from single channel by adopting multi-channel sampling, so that adjacent matrix tensors of multiple types of element paths are selectedAdjacency matrix for these C-element paths +.>The graph convolutional network is applied to each output channel of the wireless communication network, so that node representations under different element paths are obtained, the node representations under different element paths are spliced together, and the representation vectors of the key performance index nodes and the factor index nodes in the knowledge graph of the wireless communication network are obtained:
wherein, the I represents the splicing operation,representation->Adjacency matrix of c-th channel, D c Representation->Is used for the degree matrix of the (c),is the initial nodeThe feature matrix, M, is the dimension of the feature vector, and W is a channel shared trainable weight matrix.
5. The method for learning the wireless communication network knowledge graph representation based on the heterogeneous graph neural network according to claim 1, wherein training the constructed heterogeneous graph neural network model by utilizing the heterogeneous graph data specifically comprises the following steps:
acquiring data of communication performance indexes in a wireless communication network in a B5G/6G test network in real time, and preprocessing by using an interpolation algorithm and a normalization algorithm to obtain an initial node characteristic matrix X of key performance index nodes and factor index nodes in a wireless communication network knowledge graph;
designing a heterogeneous graph neural network model, and embedding key performance index nodes and factor index nodes in a wireless communication network knowledge graph into a continuous vector space to obtain representation vectors of the key performance index nodes and the factor index nodes in the wireless communication network knowledge graph;
training the constructed heterogeneous graph neural network model, and optimizing the model weight by back propagation and gradient descent and minimizing cross entropy.
6. The method for learning the knowledge graph representation of the wireless communication network based on the heterogeneous graph neural network according to claim 1, wherein the downstream task of the link prediction is used for verifying the effect of the test model, and the specific steps include:
representing nodes in the training set by using the constructed heterogeneous graph neural network model to obtain representation vectors of the nodes, performing supervised learning on positive and negative samples in the training set by using the representation vectors, predicting whether connection exists between the two nodes, and deducing or predicting a connection relationship which is not observed in the graph, wherein unbalance of positive and negative edges in link prediction is balanced by adopting a negative sampling technology; the negative samples were resampled at each round of training.
7. The method for learning the knowledge graph representation of the wireless communication network based on the heterogeneous graph neural network according to claim 6, wherein the method is characterized in that unbalance of positive and negative sides in link prediction is balanced by adopting a negative sampling technology, and specifically comprises the following steps:
the edges existing in the iso-graph are all regarded as positive samples, and the edges which do not exist are all regarded as negative samples; performing segmentation on edge_index to ensure that there is no target leakage on node embedding when predicting test data, adding two new attributes edge_label and edge_label_index to each segmented data, which are edge labels and edge indexes corresponding to each segmentation, edge_label_index being used to calculate errors, edge_label being used for model evaluation, adding the same number of negative edges as positive edges to the test set, and adding them to edge_label and edge_index attributes, but not to edge_index; node pairs constructed from edges.
8. The method for learning the knowledge graph representation of the wireless communication network based on the heterogeneous graph neural network according to claim 1, wherein the parameters of the model are continuously updated through a back propagation algorithm to reduce the value of a loss function when the model is trained; the loss function is a binary cross entropy loss function,is the value of the loss function, eta is the correction factor, y is the true label,/is>Is the predicted probability value of the model:
9. the wireless communication network knowledge graph representation learning method based on the heterogeneous graph neural network according to claim 1, wherein a trained heterogeneous graph neural network model is used for predicting new edges in the heterogeneous graph according to specific application requirements, and deducing edges with unobvious connection relations in the heterogeneous graph so as to optimize and complement the wireless communication network knowledge graph.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202311257111.6A CN117196033A (en) | 2023-09-27 | 2023-09-27 | Wireless communication network knowledge graph representation learning method based on heterogeneous graph neural network |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202311257111.6A CN117196033A (en) | 2023-09-27 | 2023-09-27 | Wireless communication network knowledge graph representation learning method based on heterogeneous graph neural network |
Publications (1)
| Publication Number | Publication Date |
|---|---|
| CN117196033A true CN117196033A (en) | 2023-12-08 |
Family
ID=89001579
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN202311257111.6A Pending CN117196033A (en) | 2023-09-27 | 2023-09-27 | Wireless communication network knowledge graph representation learning method based on heterogeneous graph neural network |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN117196033A (en) |
Cited By (1)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN117785490A (en) * | 2024-02-27 | 2024-03-29 | 苏州元脑智能科技有限公司 | Training architecture, method, system and server of graph neural network model |
-
2023
- 2023-09-27 CN CN202311257111.6A patent/CN117196033A/en active Pending
Cited By (2)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN117785490A (en) * | 2024-02-27 | 2024-03-29 | 苏州元脑智能科技有限公司 | Training architecture, method, system and server of graph neural network model |
| CN117785490B (en) * | 2024-02-27 | 2024-05-10 | 苏州元脑智能科技有限公司 | Training architecture, method, system and server of graph neural network model |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CN110866190B (en) | Method and device for training neural network model for representing knowledge graph | |
| Jin et al. | Auto-keras: Efficient neural architecture search with network morphism | |
| CN110263227B (en) | Group partner discovery method and system based on graph neural network | |
| CN111753101B (en) | Knowledge graph representation learning method integrating entity description and type | |
| CN112699247A (en) | Knowledge representation learning framework based on multi-class cross entropy contrast completion coding | |
| CN109857457B (en) | Function level embedding representation method in source code learning in hyperbolic space | |
| CN112100403A (en) | Knowledge graph inconsistency reasoning method based on neural network | |
| CN109787821B (en) | Intelligent prediction method for large-scale mobile client traffic consumption | |
| CN117196033A (en) | Wireless communication network knowledge graph representation learning method based on heterogeneous graph neural network | |
| US20220383127A1 (en) | Methods and systems for training a graph neural network using supervised contrastive learning | |
| WO2021035412A1 (en) | Automatic machine learning (automl) system, method and device | |
| CN113297429A (en) | Social network link prediction method based on neural network architecture search | |
| CN115828143A (en) | Node classification method for realizing heterogeneous primitive path aggregation based on graph convolution and self-attention mechanism | |
| CN117009545A (en) | Construction method of continuous multi-mode knowledge graph | |
| CN112036651A (en) | Electricity price prediction method based on quantum immune optimization BP neural network algorithm | |
| CN115982385A (en) | Knowledge graph based relation graph neural network patent quality assessment method | |
| CN115080795A (en) | Multi-charging-station cooperative load prediction method and device | |
| CN112270058B (en) | Optical network multichannel transmission quality prediction method based on echo state network | |
| CN113095480A (en) | Interpretable graph neural network representation method based on knowledge distillation | |
| CN112668633A (en) | Adaptive graph migration learning method based on fine granularity field | |
| IKEDA et al. | Approximation of chaotic dynamics by using smaller number of data based upon the genetic programming and its applications | |
| CN114153996A (en) | Multi-map attention cooperative geoscience knowledge map updating method and device | |
| Gupta et al. | Implementation of genetic algorithm for developing knowledge centric environment in higher education | |
| CN114444517A (en) | Intelligent law judgment method for numerical perception with increased sentencing standard knowledge | |
| CN114722211A (en) | Quality evaluation method and device of network optimization knowledge graph and electronic equipment |
Legal Events
| Date | Code | Title | Description |
|---|---|---|---|
| PB01 | Publication | ||
| PB01 | Publication | ||
| SE01 | Entry into force of request for substantive examination | ||
| SE01 | Entry into force of request for substantive examination |