Wireless service prediction method and device based on knowledge graph and graph neural network

Abstract

The present invention provides a method and device for predicting wireless services based on a knowledge graph and a graph neural network, wherein the method includes: extracting network environment attributes and mobile terminal attributes related to wireless communication services at different times based on a knowledge graph of endogenous factors of wireless communication that records the status and correlation of various wireless communication indicators in a wireless network, and a business data warehouse that records application layer business data involved in a mobile terminal in a wireless network; constructing a space-time attribute graph that is suitable for a space-time graph neural network model and is composed of wireless communication graphs at different times based on the network environment attributes and mobile terminal attributes; constructing a space-time graph neural network model for wireless service prediction based on the space-time attribute graph and training it. The trained space-time graph neural network model can use the correlation between all factors that affect wireless service prediction to make efficient and accurate predictions of wireless communication services.

Description

Technical Field

[0001] The present invention relates to the field of communication technology, and in particular to a method and device for predicting wireless services based on a knowledge graph and a graph neural network. Background Art

[0002] In recent years, many researchers have used graph neural networks to optimize or predict related services in wireless communication networks, such as power allocation tasks, link scheduling tasks, network traffic prediction tasks, etc. The rise of graph neural networks is due to their excellent spatial dependency mining capabilities and permutation and other variability. Wireless communication networks have natural graph topology information, and the communication devices in the network are disordered. At present, the application of graph neural networks in wireless communication networks is limited to regression tasks and does not involve wireless communication network service prediction (classification tasks). Therefore, the application of graph neural networks in wireless communication network service prediction is still blank and has great research significance. In addition, efficient and accurate wireless service prediction will also help achieve the intelligent application goals of vertical industries such as smart parks and smart industries, and accelerate the research and development of key communication technologies for smart wireless communications.

[0003] Considering the complexity of the attributes of mobile terminals and network environment in wireless communication networks, simply determining graph structure data based on expert experience knowledge may not fully consider all factors and their correlations that may affect wireless service predictions, which will not be conducive to the efficient and accurate prediction of wireless communication network service needs. As a knowledge base for interpreting and reasoning data, knowledge graphs are conducive to using deep learning technology to mine the mutual influence and deep relationships between endogenous factors in the data, and help solve problems such as intelligent search, intelligent decision-making, and intelligent scheduling. In addition, high dynamics is one of the significant characteristics of wireless communication networks. When wireless communication networks are processed into graph structure data, the structure of the graph and the feature information of nodes or edges in the graph may change dynamically over time. Therefore, in addition to considering the spatial dependency relationship of wireless communication networks at a specific time point, it is also necessary to consider the temporal dependency relationship of wireless communication networks in time series. Summary of the invention

[0004] Purpose of the invention: The purpose of the present invention is to provide a method and device for predicting wireless services based on knowledge graphs and graph neural networks, while utilizing knowledge graphs, graph neural networks and related time series models to achieve efficient and accurate prediction of wireless communication network services.

[0005] Technical solution: The present invention provides a wireless service prediction method based on knowledge graph and graph neural network, comprising:

[0006] Based on the knowledge graph of wireless communication endogenous factors and the business data warehouse, extract network environment attributes and mobile terminal attributes related to wireless communication services at different times;

[0007] Based on the network environment attributes and mobile terminal attributes, a space-time attribute graph suitable for a space-time graph neural network model is constructed; the space-time attribute graph is organized and constructed by wireless communication graphs at different times, the nodes in the wireless communication graph are mobile terminal objects, and the attributes of each node include mobile terminal attributes and network environment attributes possessed by the corresponding mobile terminal object;

[0008] A space-time graph neural network model for wireless service prediction is constructed based on the space-time attribute graph and model training is performed, and wireless service prediction is performed based on the trained model.

[0009] The wireless communication endogenous factor knowledge graph and business data warehouse in the present invention are constructed using wireless network protocol knowledge and domain expert knowledge. The wireless endogenous factor knowledge graph details the status of each wireless communication indicator in the wireless network and its correlation, while the business data warehouse records the application layer business data involved in the mobile terminal in the wireless network. By comprehensively using the wireless communication network data warehouse, a space-time attribute graph that can more comprehensively reflect the state of the wireless network can be constructed, thereby achieving the goal of efficiently and accurately predicting wireless communication service needs. The construction of the wireless communication endogenous factor knowledge graph and business data warehouse can refer to the specific implementation steps of the inventor's existing patent "Method, system, device and medium for constructing a wireless communication network data warehouse" (patent number: ZL202110634448.9), which will not be repeated in the present invention.

[0010] Preferably, the extracting of network environment attributes and mobile terminal attributes related to the wireless communication service at different times includes:

[0011] Determine relevant mobile terminal objects according to the wireless service prediction task;

[0012] Based on the knowledge graph of wireless communication endogenous factors, the endogenous element nodes of the mobile terminal object that affect the wireless service prediction at different times are determined, namely, the channel state information, wireless link control throughput, physical throughput, media access control throughput and other network environment indicators of the communication network environment where the mobile terminal object is located, and these are used as network environment attributes;

[0013] Based on the business data warehouse, the mobile terminal attributes such as behavior habits, actions, and locations of mobile terminal objects that affect wireless business prediction at different times are determined.

[0014] Preferably, constructing a space-time attribute graph includes:

[0015] Constructing a wireless communication graph at each moment. At a certain moment, the nodes in the wireless communication graph are the determined mobile terminal objects, and the attributes of each node include mobile terminal attributes and network environment attributes possessed by the corresponding mobile terminal object;

[0016] Construct a space-time attribute graph. By organizing wireless communication graphs at different times, a space-time attribute graph over a time series can be constructed;

[0017] Determine the connection relationship between nodes in the space-time attribute graph based on the knowledge graph of endogenous factors of wireless communication, and generate the adjacency matrix on the time series i∈{0,…,n} and attribute feature matrix i∈{0,…,n}, where N is the number of mobile terminal objects, d is the dimension of the attribute feature vector, and n is the time period for data collection.

[0018] Preferably, if there is a connection between two nodes in the space-time attribute graph and there is at least one pair of related endogenous element nodes in the wireless communication endogenous factor knowledge graph, then it is determined that the two nodes in the space-time attribute graph have a connection relationship, otherwise, the two nodes in the space-time attribute graph have no connection relationship.

[0019] Preferably, the weight of the edge between two nodes in the adjacency matrix of the space-time attribute graph can be determined according to the mobile terminal attributes and network environment attributes possessed by the two nodes.

[0020] Preferably, the space-time graph neural network model for wireless service prediction based on the space-time attribute graph comprises:

[0021] Preprocessing of the space-time attribute graph: selecting a standardization criterion (such as linear function normalization, zero mean normalization, etc.) to standardize the attribute features of nodes in the space-time attribute graph;

[0022] According to the specific wireless service prediction task, a corresponding space-time graph neural network model is designed; the space-time graph neural network model includes two modules, namely the spatial dependency mining module and the temporal dependency mining module. Among them, the spatial dependency mining module can be implemented by a graph convolutional neural network, and the temporal dependency mining module can be implemented by a recursive neural network or a temporal convolutional network. The graph convolutional neural network mines the spatial dependencies between nodes at different times from the dimension of space, while the recursive neural network or the temporal convolutional network mines the temporal dependencies between different nodes at all times from the dimension of time. The input of the space-time graph neural network is and The output is in q is the number of business prediction categories, and T is the business prediction time slice size.

[0023] The loss function of the spatio-temporal graph neural network model is determined according to the specific wireless service prediction task, and training is performed in a supervised manner.

[0024] Based on the same inventive concept, the present invention also provides a wireless service prediction device based on a knowledge graph and a graph neural network, comprising:

[0025] The attribute acquisition module is used to extract network environment attributes and mobile terminal attributes related to wireless communication services at different times based on the knowledge graph of endogenous factors of wireless communication and the business data warehouse.

[0026] The space-time attribute graph construction module is used to construct a space-time attribute graph suitable for a space-time graph neural network based on the network environment attributes and mobile terminal attributes.

[0027] The space-time graph neural network construction module is used to construct a space-time graph neural network model for wireless service prediction based on the space-time attribute graph and perform model training.

[0028] And, a service prediction module is used to perform wireless service prediction based on the trained space-time graph neural network model.

[0029] Based on the same inventive concept, the present invention also provides an electronic device, including a memory, a processor, and a computer program stored in the memory and executable on the processor. When the computer program is loaded into the processor, the wireless service prediction method based on the knowledge graph and graph neural network is implemented.

[0030] Based on the same inventive concept, the present invention also provides a computer-readable storage medium, which stores a computer program. When the computer program is executed by a processor, it implements the wireless service prediction method based on knowledge graph and graph neural network.

[0031] Beneficial effects: Compared with the prior art, the present invention fully considers all endogenous element nodes and their associations in the wireless communication endogenous factor knowledge graph that may affect the wireless service prediction, as well as the application layer service data involved in the mobile terminal, and constructs a space-time attribute graph that can more comprehensively characterize the state of the wireless communication network, thereby achieving the goal of efficiently and accurately predicting the demand for wireless communication services; the present invention describes a wireless service prediction method based on the wireless endogenous factor knowledge graph and graph neural network for the first time, so as to utilize the graph neural network to fully mine the effective information in the topological structure of the wireless communication network. BRIEF DESCRIPTION OF THE DRAWINGS

[0032] In order to more clearly illustrate the technical solutions in the present invention or the prior art, the drawings required for use in the embodiments or the description of the prior art will be briefly introduced below. Obviously, the drawings described below are some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying creative work.

[0033] Figure 1 It is a flowchart of a wireless service prediction method based on a knowledge graph and a graph neural network provided by a specific embodiment of the present invention;

[0034] Figure 2 It is a schematic diagram of specific steps for extracting network environment attributes and mobile terminal attributes related to wireless communication services at different times provided by a specific embodiment of the present invention;

[0035] Figure 3 It is a schematic diagram of specific steps for determining network environment attributes in the wireless communication endogenous factor knowledge graph and mobile terminal attributes in the service data warehouse according to the mobile terminal object predicted by the wireless service at different times provided by a specific embodiment of the present invention;

[0036] Figure 4 It is a flow chart of constructing a space-time attribute graph provided by a specific embodiment of the present invention;

[0037] Figure 5 It is a schematic diagram of specific steps for constructing a space-time attribute graph suitable for a space-time graph neural network model based on the network environment attributes and mobile terminal attributes provided by a specific embodiment of the present invention;

[0038] Figure 6 It is a schematic diagram of specific steps for constructing a space-time graph neural network model for wireless service prediction based on the space-time attribute graph provided by a specific embodiment of the present invention;

[0039] Figure 7 It is a flow chart of constructing a space-time graph neural network provided by a specific embodiment of the present invention;

[0040] Figure 8 It is a schematic diagram of the structure of a wireless service prediction device based on a knowledge graph and a graph neural network provided in a specific embodiment of the present invention. DETAILED DESCRIPTION

[0041] In order to make the purpose, technical solution and advantages of the present invention clearer, the technical solution of the present invention will be clearly and completely described below in conjunction with the drawings of the present invention. Obviously, the described embodiments are part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0042] The specific embodiment of the present invention takes video, voice, email and web page as wireless service prediction tasks.

[0043] Figure 1 is a flow chart of a wireless service prediction method based on a knowledge graph and a graph neural network provided by an embodiment of the present invention, such as Figure 1As shown, an embodiment of the present invention provides a wireless service prediction method based on a knowledge graph and a graph neural network, the method comprising:

[0044] Step 100: Extract network environment attributes and mobile terminal attributes related to wireless communication services at different times based on the constructed wireless communication endogenous factor knowledge graph and service data warehouse.

[0045] The construction of the knowledge graph of the relationship between endogenous factors of wireless communication and the business data warehouse can refer to the specific implementation steps of the patent "Method, system, equipment and medium for constructing a data warehouse based on wireless communication network" already owned by the inventor of this patent, and this patent will not repeat them.

[0046] In step 100, the network environment attributes and mobile terminal attributes related to the wireless communication service at different times are extracted, such as Figure 2 As shown, the specific steps are as follows:

[0047] Step 110: determining relevant mobile terminal objects according to the wireless service prediction task;

[0048] Specifically, the knowledge graph of wireless communication endogenous factors and the business data warehouse cover information on dimensions such as network, terminal, user, and business processing. Before determining the network environment attributes and mobile terminal attributes related to wireless communication services, it is necessary to determine the relevant mobile terminal objects based on the wireless service prediction tasks, and then determine the relevant network environment attributes and mobile terminal attributes based on the relevant mobile terminal objects.

[0049] Step 120: At different times, determine the network environment attributes in the wireless communication endogenous factor knowledge graph and the mobile terminal attributes in the business data warehouse according to the mobile terminal object predicted by the wireless business.

[0050] Each node in the wireless communication endogenous factor knowledge graph represents an endogenous factor and its attributes related to the performance of the wireless communication network. For example, if a node represents the endogenous factor of signal to interference and noise ratio, its attributes may include the value of the signal to interference and noise ratio, the mobile terminal ID related to it, etc. When determining the relevant mobile terminal object, it is necessary to simultaneously lock the endogenous factor nodes and their attributes related to it to construct a space-time attribute graph.

[0051] In step 120, at different times, the network environment attributes in the wireless communication endogenous factor knowledge graph and the mobile terminal attributes in the business data warehouse are determined according to the mobile terminal object predicted by the wireless business, such as Figure 3 As shown, the specific steps are as follows:

[0052] Step 121: In the spatial dimension, based on the knowledge graph of endogenous factors of wireless communication, determine the network environment attributes such as channel state information, wireless link control throughput, physical throughput, and media access control throughput of the communication network environment in which the mobile terminal object is located that affect the wireless service prediction at different times.

[0053] Specifically, at a specific moment, the relevant network environment information of the communication network ring where the relevant mobile terminal object is located can reflect the state of the relevant mobile terminal object to a certain extent. For example, performance evaluation indicators such as radio link control throughput, physical throughput, and media access control throughput can reflect the type of service executed by the mobile terminal object. According to the wireless communication service type assumed in the specific embodiment of the present invention, when the relevant throughput performance indicator is large, it may be executing a service type that occupies a large amount of network resources, such as a video service. On the contrary, it may be executing voice, email, and web services. Similarly, network environment attributes such as channel state information and wireless resources can also affect the type of service executed by the relevant mobile terminal object to a certain extent. The network environment attributes in the spatial dimension are generally numerical attribute features. If there are relevant non-numerical attribute features, they can be encoded using the relevant non-numerical attribute feature encoding method. Commonly used non-numerical attribute feature encoding methods include serial number encoding, unique hot encoding, and binary encoding. Among them, serial number encoding is suitable for non-numerical attribute features with obvious level differences; unique hot encoding is suitable for non-numerical attribute features with not too many types; binary encoding generally uses relevant hash functions to map non-numerical attribute features into different values, but facing a large number of feature types, a more complex binary encoding method will be required, which will lead to higher computational overhead. In actual application, a suitable non-numerical attribute feature encoding method can be selected according to the specific non-numerical attribute features.

[0054] Step 122: In the time dimension, the mobile terminal attributes such as behavior habits, actions, and locations of the mobile terminal objects that affect the wireless service prediction at different times are determined based on the service data warehouse.

[0055] Specifically, there is a certain correlation between the behavior habits, actions and other mobile terminal attributes at the current moment and the next moment. The behavior habits, actions and other mobile terminal attributes in the time dimension have a crucial impact on wireless service prediction. Considering that the behavior habits, actions and other mobile terminal attributes are non-numerical attributes, which are not conducive to the use of machine learning models, the aforementioned non-numerical attribute encoding method can be used to encode the relevant mobile terminal attributes to make them numerical.

[0056] Preferably, in selecting the relevant mobile terminal attributes and network environment attributes, it is preferred to select the same type of attributes for all relevant mobile terminal objects. For example, if all mobile terminal objects have the signal to interference noise ratio endogenous factor node in the wireless communication endogenous factor knowledge graph, then the signal to interference noise ratio network environment attribute is selected; similarly, if all mobile terminal objects in the business data warehouse have the behavior habit mobile terminal attribute, then the behavior habit is selected.

[0057] Step 200: construct a space-time attribute graph suitable for a space-time graph neural network model based on the network environment attributes and mobile terminal attributes.

[0058] like Figure 4 As shown in the figure, the construction process of the space-time attribute graph is mainly to organize the extracted mobile terminal attributes and network environment attributes from the wireless communication endogenous knowledge graph on the time series into a wireless communication graph, and then convert the mobile terminal attributes and network environment attributes related to the mobile terminal into attribute feature vectors, and finally organize the wireless communication graph on the time series together to construct a space-time attribute graph.

[0059] In step 200, the space-time attribute graph suitable for the space-time graph neural network model is constructed based on the network environment attributes and the mobile terminal attributes, such as Figure 5 As shown, the specific steps are as follows:

[0060] Step 210: construct a wireless communication graph at each moment. At a certain moment, the nodes in the wireless communication graph are the determined mobile terminal objects, and the attributes of each node include the mobile terminal attributes and network environment attributes possessed by the corresponding mobile terminal object;

[0061] Step 220: Constructing a space-time attribute graph. The wireless communication graphs at different times are organized together to construct a space-time attribute graph over a time series.

[0062] Specifically, the nodes in the space-time attribute graph are specific mobile terminal objects for wireless service prediction, and the attribute feature vectors of the nodes in the space-time attribute graph are composed of the extracted network environment attributes and mobile terminal attributes. Considering that the extracted mobile terminal attributes and network environment attributes have been pre-processed by digitization and normalization in step 120, the composition of the attribute feature vector only needs to be cascaded according to the types of the corresponding network environment attributes and mobile terminal attributes.

[0063] Step 230: Determine the connection relationship between nodes in the space-time attribute graph based on the wireless communication endogenous factor knowledge graph, and generate an adjacency matrix on the time series i∈{0,…,n} and attribute feature matrix i∈{0,…,n}, where N is the number of mobile terminal objects, d is the dimension of the attribute feature vector, and n is the time period for data collection.

[0064] Specifically, the triples in the knowledge graph of endogenous factors of wireless communication reflect the connection relationship between different entities. The effective use of the triples can effectively generate the adjacency matrix of the space-time attribute graph. Specifically, the connection relationship between nodes in different space-time attribute graphs is judged according to the endogenous element nodes involved in the node in the space-time attribute graph. If there is a connection between two nodes in the space-time attribute graph in at least one pair of related endogenous element nodes (not necessarily of the same type) in the knowledge graph of endogenous factors of wireless communication, that is, the two endogenous element nodes are reachable, then it is determined that the two nodes in the space-time attribute graph have a connection relationship, otherwise, the two nodes in the space-time attribute graph have no connection relationship. The generation of the attribute feature matrix directly cascades the attribute feature vectors of each node in the space-time attribute graph by row.

[0065] Preferably, the weight of the edge between two nodes in the adjacency matrix of the space-time attribute graph can be determined according to the mobile terminal attributes and network environment attributes possessed by the two nodes. Specifically, the attribute feature vectors of the two nodes can be similarly measured to obtain the weight of the edge between the two nodes. Commonly used similarity measurement methods include Euclidean distance, Manhattan distance, Chebyshev distance, Mahalanobis distance, angle cosine, Pearson correlation coefficient, etc. Different similarity measurement methods have different effects on wireless service prediction. According to the specific wireless service prediction task, a suitable similarity measurement method can be selected.

[0066] Step 300: construct a space-time graph neural network model for wireless service prediction based on the space-time attribute graph and perform model training, and perform wireless service prediction based on the trained model.

[0067] In step 300, the space-time graph neural network model for wireless service prediction is constructed based on the space-time attribute graph and the model training is performed, such as Figure 6 As shown, the specific steps are as follows:

[0068] Step 310: preprocessing the space-time attribute graph, selecting a standardization criterion (such as linear function normalization, zero mean normalization, etc.) to standardize the attribute features of the nodes in the space-time attribute graph.

[0069] Generally speaking, different network environment attributes have large differences in numerical values. If native numerical attributes are used, it will be detrimental to the effective and rapid learning of machine learning models.

[0070] As a preferred method, in order to ensure the effective implementation of the machine learning model, the numerical attribute feature vector can be standardized. The commonly used standardization method is normalization, which generally includes linear function normalization, zero mean normalization, etc. Among them, the form of linear function normalization is fnorm =(ff min ) / (f max -f min ), the zero mean normalization method is f norm =(ff mean ) / σ, σ is the variance.

[0071] Step 320: Design a corresponding space-time graph neural network model according to the specific wireless service prediction task. Specifically, the space-time graph neural network model includes two modules, namely, a spatial dependency mining module and a temporal dependency mining module.

[0072] Preferably, the spatial dependency mining module can be implemented by graph convolutional neural networks (GCNs), and the temporal dependency mining module can be implemented by recurrent neural networks (RNNs) or temporal convolutional networks. The specific design patterns are as follows: Figure 7 As shown in Figure 2. Graph convolutional neural networks mine spatial dependencies between nodes at different times from the spatial dimension, while recursive neural networks or temporal convolutional networks mine temporal dependencies between different nodes at all times from the temporal dimension. The input of the spatial-temporal graph neural network is and The output is in q is the number of business prediction categories, and T is the business prediction time slice size.

[0073] Specifically, the graph convolutional neural network can adopt the graph convolutional neural network based on the spectral domain and the graph convolutional neural network based on the spatial domain. Among them, the execution mode of each graph convolutional layer of the graph convolutional neural network based on the spectral domain can be defined as

[0074]

[0075] in,* G represents the graph convolution operator, W is the learnable graph convolution kernel, and I N represents the identity matrix of size N×N, D is the diagonal matrix associated with the adjacency matrix, where the element in the i-th row and i-th column of matrix D is the vertex v i The number of neighbor vertices. The graph convolution operation of the k-th layer of the spatial domain-based graph convolutional neural network can be defined as

[0076]

[0077]

[0078] Among them, e ij Represents node v i and v j The edge attribute feature vector between represents the node v in the k-1 layer i The attribute feature vector of Represents node v i The set of neighbor nodes of represents the aggregation function in the kth layer, Indicates the node v in the kth layer i The intermediate attribute feature vector obtained by aggregation, represents the combination function in the kth layer.

[0079] The specific embodiment of the present invention designs a space-time graph neural network model based on a variant of a recurrent neural network, GRU (Gate Recurrent Unit), and a graph convolutional neural network based on the spectral domain. Specifically, the execution mode of the space-time graph neural network model at time t is defined as

[0080] r (t) =σ([x (t) ,s (t-1) ]* G Θ r +b r ),

[0081] u (t) =σ([x (t) ,s (t-1) ]* G Θ u +b u ),

[0082] c (t) =tanh([x (t) ,(r (t) ⊙s(t-1))]* G Θ c +b c ),

[0083] s (t) =u (t) ⊙s (t-1) +(1-u (t) )⊙c (t) ,

[0084] y (t+1) =Softmax(W o s (t) ),

[0085] Among them, σ(·) is the sigmoid activation function, and [·,·] represents the concatenation operation. (t) Represents a graph signal at time t (attribute feature matrix X t , s (t) Represents the hidden state of the output at time t. (t) and u(t) They represent the reset gate and update gate at time t, respectively, and c (t) Indicates the current memory content. r , Θ u and θ c W represents the learnable graph filter coefficients. o represents the learning weight matrix of the output layer, and the output layer uses the Softmax(·) activation function to achieve the classification effect of different wireless communication services. r ,b u ,b c represents the bias term in each gate, and ⊙ is the Hadamard product operator. (t+1) represents the output result at time t+1. We call the execution mode GRUGCN.

[0086] Step 330: Determine the loss function of the spatio-temporal graph neural network model according to the specific wireless service prediction task, and perform training in a supervised manner.

[0087] Specifically, wireless service prediction belongs to time series classification tasks and supervised learning. Therefore, all samples in the data set of the specific embodiment of the present invention have real labels. The real labels can be manually annotated or obtained from the knowledge graph of wireless communication endogenous factors and the service data warehouse. The loss function of the spatio-temporal graph neural network model uses Mean Per-Class Error (MPCE). Specifically, MPCE is defined as the average error of each class of all samples in the data set, and the mathematical form is:

[0088]

[0089]

[0090] Among them, PCE k represents the error rate of the kth category, and M represents the number of all categories in the data set.

[0091] Figure 8 The wireless service prediction device based on knowledge graph and graph neural network provided by a specific embodiment of the present invention comprises an attribute acquisition module, a space-time attribute graph construction module, a space-time graph neural network construction module and a service prediction module.

[0092] Attribute acquisition module: used to extract network environment attributes and mobile terminal attributes related to wireless communication services at different times based on the constructed wireless communication endogenous factor knowledge graph and business data warehouse.

[0093] Space-time attribute graph construction module: used to construct a space-time attribute graph suitable for a space-time graph neural network model based on the network environment attributes and mobile terminal attributes.

[0094] Space-time graph neural network construction module: used to construct a space-time graph neural network model for wireless service prediction based on the space-time attribute graph and perform model training.

[0095] Service prediction module: used to predict wireless services based on the trained space-time graph neural network model.

[0096] The specific implementation details of each module in the above-mentioned device embodiment refer to the above-mentioned method embodiment, which will not be repeated here. Those skilled in the art can understand that the modules in the embodiment can be adaptively changed and set in one or more devices different from the embodiment. The modules in the embodiment can be combined into one module, or divided into multiple sub-modules.

[0097] Based on the same inventive concept, an embodiment of the present invention provides an electronic device, including a memory, a processor, and a computer program stored in the memory and executable on the processor. When the computer program is loaded into the processor, the wireless service prediction method based on the knowledge graph and graph neural network is implemented.

[0098] Based on the same inventive concept, an embodiment of the present invention provides a computer-readable storage medium, which stores a computer program. When the computer program is executed by a processor, it implements the wireless service prediction method based on knowledge graph and graph neural network.

[0099] It is understood by those skilled in the art that the technical solution of the present invention, in essence or in other words, the part that contributes to the prior art, can be embodied in the form of a software product, which is stored in a storage medium and includes several instructions for enabling an electronic device (which may be a personal computer, a server, or a network device, etc.) to perform all or part of the steps of the method described in the embodiment of the present invention. The storage medium includes: various media that can store computer programs, such as a USB flash drive, a mobile hard disk, a read-only memory ROM, a random access memory RAM, a magnetic disk or an optical disk.

Claims

1. A wireless service prediction method based on knowledge graph and graph neural network, characterized in that: The steps include: Based on the wireless communication endogenous factor knowledge graph that records the status and correlation of each wireless communication indicator in the wireless network, and the business data warehouse that records the application layer business data involved in the mobile terminal in the wireless network, network environment attributes and mobile terminal attributes related to the wireless communication business at different times are extracted; Based on the network environment attributes and mobile terminal attributes, a space-time attribute graph suitable for a space-time graph neural network model is constructed; the space-time attribute graph is organized and constructed by wireless communication graphs at different times, the nodes in the wireless communication graph are mobile terminal objects, and the attributes of each node include mobile terminal attributes and network environment attributes possessed by the corresponding mobile terminal object; A space-time graph neural network model for wireless service prediction is constructed based on the space-time attribute graph and model training is performed, and wireless service prediction is performed based on the trained model.

2. The wireless service prediction method based on knowledge graph and graph neural network according to claim 1 is characterized in that: The extracting of network environment attributes and mobile terminal attributes related to the wireless communication service at different times includes: Determine relevant mobile terminal objects according to the wireless service prediction task; Determine the endogenous factor nodes of the mobile terminal object that affect the wireless service prediction at different times based on the knowledge graph of endogenous factors of wireless communication, including the channel state information, wireless link control throughput, physical throughput and media access control throughput of the communication network environment where the mobile terminal object is located, and use them as network environment attributes; The mobile terminal attributes of the mobile terminal objects that affect the wireless service prediction at different times are determined based on the service data warehouse, including behavior habits, actions and locations.

3. The wireless service prediction method based on knowledge graph and graph neural network according to claim 1 is characterized in that: The steps to construct a space-time attribute graph include: Construct wireless communication graphs at each moment; Organize wireless communication graphs at different times to construct a space-time attribute graph over a time series; Determine the connection relationship between nodes in the space-time attribute graph based on the knowledge graph of endogenous factors of wireless communication, and generate the adjacency matrix on the time series and attribute feature matrix Among them, N is the number of mobile terminal objects, d is the dimension of the attribute feature vector, and n is the time period for data collection.

4. The wireless service prediction method based on knowledge graph and graph neural network according to claim 3 is characterized in that: If there is a connection between two nodes in the space-time attribute graph and at least one pair of related endogenous element nodes in the wireless communication endogenous factor knowledge graph, then it is determined that the two nodes in the space-time attribute graph have a connection relationship, otherwise, the two nodes in the space-time attribute graph have no connection relationship.

5. The wireless service prediction method based on knowledge graph and graph neural network according to claim 3 is characterized in that: The weight of the edge between two nodes in the adjacency matrix of the space-time attribute graph is determined according to the mobile terminal attributes and network environment attributes possessed by the two nodes.

6. The wireless service prediction method based on knowledge graph and graph neural network according to claim 1 is characterized in that: The step of constructing a space-time graph neural network model for wireless service prediction based on the space-time attribute graph and performing model training includes: Preprocessing of the space-time attribute graph: selecting a standardization criterion to standardize the attribute features of nodes in the space-time attribute graph; According to the specific wireless service prediction task, a corresponding space-time graph neural network model is designed; the space-time graph neural network model includes a spatial dependency mining module and a temporal dependency mining module; the spatial dependency mining module is implemented by a graph convolutional neural network, and the temporal dependency mining module is implemented by a recursive neural network or a temporal convolutional network; The loss function of the spatio-temporal graph neural network model is determined according to the specific wireless service prediction task, and training is performed in a supervised manner.

7. The wireless service prediction method based on knowledge graph and graph neural network according to claim 1 is characterized in that: The loss function for training the spatio-temporal graph neural network model uses MPCE, which is the average error of each class for all samples in the dataset.

8. A wireless service prediction device based on knowledge graph and graph neural network, characterized in that: include: An attribute acquisition module is used to extract network environment attributes and mobile terminal attributes related to wireless communication services at different times based on a wireless communication endogenous factor knowledge graph that records the status and correlation of various wireless communication indicators in the wireless network, and a business data warehouse that records application layer business data involved in mobile terminals in the wireless network; A space-time attribute graph construction module, used to construct a space-time attribute graph suitable for a space-time graph neural network model based on the network environment attributes and the mobile terminal attributes; the space-time attribute graph is organized and constructed by wireless communication graphs at different times, the nodes in the wireless communication graph are mobile terminal objects, and the attributes of each node include the mobile terminal attributes and network environment attributes possessed by the corresponding mobile terminal object; A space-time graph neural network construction module, used to construct a space-time graph neural network model for wireless service prediction based on the space-time attribute graph and perform model training; And, a service prediction module is used to perform wireless service prediction based on the trained space-time graph neural network model.

9. An electronic device comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that: When the computer program is loaded into the processor, the wireless service prediction method based on the knowledge graph and graph neural network is implemented according to any one of claims 1 to 7.

10. A computer-readable storage medium storing a computer program, characterized in that: When the computer program is executed by a processor, the wireless service prediction method based on knowledge graph and graph neural network is implemented according to any one of claims 1 to 7.

Images