An Air Interface Parameter Recommendation Method Based on User-Air Interface Knowledge Graph

Abstract

This invention provides an air interface parameter recommendation method based on a user-air interface knowledge graph, comprising: Step S1, defining air interface entities and user entities, constructing an air interface knowledge graph and a user knowledge graph, and obtaining user feature vectors and air interface feature vectors; Step S2, using a self-attention mechanism to extract features from the user feature vectors and air interface feature vectors, enhancing the extracted features using high-order cross-feature extraction, and reducing the dimensionality of the enhanced features using convolution; Step S3, designing a channel type identification module and a user behavior prediction module, and inferring the user's current environment and new needs in the physical world based on the channel identification classification results and the predicted behavior; Step S4, conducting a predicted score of the air interface by the target user under the current environment and new needs, and providing an air interface recommendation list based on the predicted score. This invention can achieve real-time performance, accuracy, rationality, and stability in air interface parameter recommendation.

Description

Technical Field

[0001] This invention relates to an air interface parameter recommendation method based on a user-air interface knowledge graph. Background Technology

[0002] In 6G wireless network communication scenarios characterized by high dynamics, wide coverage, and intelligence, traditional adaptive transmission mechanisms such as adaptive modulation and coding, and adaptive retransmission rely on complex and cumbersome channel estimation and feedback processes. They require adjustments to the air interface configuration based on user-reported channel quality indicators and signal-to-noise ratio (SNR) parameters. However, this approach has drawbacks: the scheme itself introduces a certain delay in air interface strategy adjustment, leading to mismatches in system air interface transmission strategies and a decrease in communication performance. To perceive and monitor changes in the channel environment and user behavior in real time, periodic transmission of channel measurement signals and transceiver synchronization signals is necessary, further increasing system signaling overhead and reducing system spectral efficiency. Existing air interface configuration recommendation schemes rely on brute-force search algorithms, traversing pre-defined environment-air interface configuration tables based on experience. The available environmental parameters and user requirements are limited, and the search complexity increases linearly with the number of communication air interface configuration schemes.

[0003] Recommendation algorithms can analyze and mine information such as users' historical behavior to recommend personalized needs and interests to users, achieving accurate recommendations tailored to user requirements while reducing the computational complexity of the algorithm. Their aim is to recommend one or a series of unobserved options to users. A recommendation system is a multi-stage collaborative system, and the accuracy of each stage's computation directly affects the quality of the recommendation result. Leng Yajun et al. focused on the neighbor selection stage, making neighbor selection more reasonable and improving recommendation quality.

[0004] Collaborative recommendation algorithms have no special requirements for the content of the recommended objects and support multi-user collaborative recommendation, making them suitable for 6G air interface parameter recommendation. However, existing collaborative recommendation technologies still have problems: the potential preferences of target users and their neighbors are similar, which can easily lead to multiple users matching the same air interface parameters, thus exacerbating communication interference and congestion; the k-means weighted algorithm used in recommendation does not consider the personalized differences in users' basic ratings and fails to mine potential user air interface factors, affecting the accuracy of personalized recommendations; the determination of the number of neighbors k is too subjective, thus affecting the accuracy of the recommendation technology; classic recommendation algorithms take low-dimensional mapping vectors of user and item IDs as input, and only learn the interaction between users and items to determine whether to make a recommendation, ignoring the information of users and items themselves. However, autonomous recommendation of air interface parameters needs to be tailored to the given channel propagation and interference environment conditions, and recommend applicable air interface parameters based on specific requirements such as data rate, bit error rate, and signal-to-noise ratio. Therefore, it is necessary to represent the feature information of users and items in the form of a knowledge graph and introduce it into the recommendation model, so that the model can fully learn and enhance the features of users and items, thereby increasing the interpretability and reliability of the recommendations. Furthermore, we need to propose a new recommendation method based on user-air interface knowledge graph to achieve real-time and accurate recommendation of air interface parameters, thereby improving the adaptability and intelligence level of 6G air interface to large dynamic, wide-domain, and heterogeneous network environments. Summary of the Invention

[0005] This invention proposes an air interface parameter recommendation method based on a user-air interface knowledge graph, which can achieve real-time performance, accuracy, rationality and stability of air interface parameter recommendation.

[0006] This invention is achieved through the following technical solution:

[0007] An air interface parameter recommendation method based on a user-air interface knowledge graph includes the following steps:

[0008] Step S1: The knowledge representation module defines air interface entities and user entities, constructs air interface knowledge graphs and user knowledge graphs, and obtains the user feature vector E mapped from the semantic, numerical, and relational information of user entities through the Word2vec model. U The air interface feature vector E, which maps the semantic, numerical, and relational information of air interface entities, is obtained through the TransD model and the Word2ve model. A ;

[0009] Step S2: Use the self-attention mechanism to process the user feature vector E obtained in step S1. U and air interface feature vector E A Feature extraction is performed, and the extracted features are enhanced using higher-order cross-feature extraction. Convolution is then used to reduce the dimensionality of the enhanced features to obtain the user features. and air interface features This is used to calculate the matching score R(m×n) of user historical information against air interface features; where R(m×n) represents the scoring matrix of n air interface features based on the historical information of m users.

[0010] Step S3: Based on digital twin technology, design a channel type identification module for identifying and classifying channels in a certain scenario and outputting the results, and a user behavior prediction module for predicting user behavior in future moments. Based on the channel identification and classification results and the predicted behavior, infer the user's current environment and new needs in the physical world.

[0011] Step S4: Based on the matching score R(m×n) from Step S2 and the user's current environment and new needs deduced in Step S3, conduct a predictive score for the air interface of the target user under the current environment and new needs, and sort the predicted scores to provide an air interface recommendation list.

[0012] Furthermore, in step S1, the air interface entity is composed of a triplet consisting of a head entity, a relation set, and a tail entity. An air interface determined by different parameters and technical methods is regarded as the air interface entity. The air interface is defined as the head entity, the various parameters of the air interface are defined as the relation set, and the specific values ​​of the parameters are defined as the tail entity. The technical methods and parameters used by a certain air interface are extracted to form a triplet of air interface resource configuration, technical method, and technical parameters to construct an air interface knowledge graph.

[0013] A user entity is composed of a triplet consisting of a head entity, a set of relations, and a tail entity. A user who uses a certain parameter and communication technology in a certain environment is considered a user entity. The overall environment and performance requirements are defined as the head entity of the user, the parameter types of the environment and performance requirements are defined as the set of relations, and the specific values ​​of the parameters are defined as the tail entity. Based on the link simulation results, the environmental conditions of different air interface simulations are summarized to form a triplet of user, environment and performance requirement type, and corresponding parameter values ​​of environment and performance requirement type, in order to construct a user knowledge graph.

[0014] The Neo4j graph database and tabular format are used to jointly store the air interface parameter knowledge graph, and Excel spreadsheets are used to store air interface and user information.

[0015] Furthermore, in step S1, the user feature vector E U Dimensions Where, N U For the number of user features, The dimension of a single feature mapping is 2, where 2 indicates dual channels; the air interface feature vector E A Dimensions Where, N A The characteristic number of the air passage, The number of bits for a single feature map is 3, where 3 indicates three channels.

[0016] Furthermore, step S2 includes the following steps:

[0017] Step S21, for dual-channel N U The user feature vector of each feature will be E U and three-channel N A The air interface feature vector E of each feature A For each feature in each channel, self-attention weighting is performed by assigning query vectors, key vectors, and value vectors to each feature. The similarity of each feature is then calculated to obtain the weighted feature matrix. These feature matrices are then concatenated to obtain the feature E′ extracted by the multi-head attention mechanism. U and E′ A ;

[0018] Step S22: For the feature E′ extracted in step S21 U and E′ A Features E′ are automatically constructed through multiple cross layers. U and E′ A The finite high-order cross features are used to obtain the enhanced features;

[0019] Step S23: Use two convolutional layers, a non-linear layer, and the corresponding pooling layer to reduce the dimensionality of the enhanced features from step S22 to obtain the user features. and air interface features

[0020] Step S24: Based on the historical information of m users, score n air interface features, that is, use a multilayer perceptron to fit the matching function between user features and air interface features, thus obtaining the matching score matrix R(m×n), where the input of the multilayer perceptron is... . concatenated vectors.

[0021] Furthermore, step S3 includes the following steps:

[0022] Step S31: In a digital twin network that is highly symbiotic with the physical world, construct a dataset of received signals with pilots under different electromagnetic propagation environments, and obtain a channel training dataset using least squares estimation; extract channel features under different electromagnetic propagation conditions from the channel training data as a channel feature parameter dataset, and establish a channel type identification model using machine learning algorithms and the channel feature parameter dataset; input the channel data from the physical world, and the channel type identification model identifies and classifies the channel in this scenario and outputs the results; among which, the channel features include kurtosis, skewness, root mean square delay spread, and Rice K-factor features;

[0023] Step S32: Collect long-term user behavior data in the digital twin network, train it using the Long Short-Term Memory algorithm, and establish an LSTM training model of user behavior; input the user behavior data of the physical world at the current moment, and the LSTM training model predicts the user behavior at future moments to update the user's status and make decisions in real time.

[0024] Step S33: Based on the channel identification classification results of the channel type identification model and the behavioral requirements predicted by user behavior prediction, deduce the user's current environment and new requirements in the physical world.

[0025] Furthermore, step S4 includes the following steps:

[0026] Step S41: Based on the matching score R(m×n), calculate the Euclidean distance matrix D(m×m) for the user, and calculate the Euclidean similarity matrix OS(m×m) based on the Euclidean distance matrix; where the row vector of the i-th row of the Euclidean matrix D(m×m) is the user u i Distance vectors to other users;

[0027] Step S42: Calculate the mean of the sum of distances between all users based on the Euclidean distance matrix, and perform heuristic clustering of users based on the mean of the sum of distances to obtain multiple cluster sets;

[0028] Step S43, Target User u a The cluster set G it belongs to g Given the target user's neighbor set, the target user's predicted rating for the currently unrated target airway l is: in, For target user u a For the currently unrated target interface l, the basic score value is Isim(l,j'), which represents the target user u. a The evaluated category similarity between air interface j' and the target air interface l, where A represents the target user u. a Number of air ports evaluated, R a (j') represents the target user u a For the score of empty j', 1≤j'≤A, In order to connect with target user u a Neighboring users with similar preferences j For the scoring items of target air interface l, Sim(a,j) represents the target user u. a and neighboring user u j The similarity between them, where K is the target user u a The number of users in adjacent user groups, R j (l) represents adjacent user u j The score of the air interface l, Indicates adjacent user uj The mean score for the empty speech;

[0029] Step S44: Based on the predicted score P a,i Sort by size for target user u a Provides a list of recommended air interface providers.

[0030] Furthermore, in step S23, the number of convolution kernels in the two convolutional layers is 8, and the size is 3×3. The nonlinear layer uses the LeakyReLU function, and the pooling kernel size of the pooling layer is 1×4.

[0031] The present invention has the following beneficial effects:

[0032] 1. This invention employs a self-attention mechanism to weight user features, making more important features have a greater impact on the output. Simultaneously, it enhances the extracted features using high-order cross-feature extraction, and uses convolution to reduce the dimensionality of these enhanced features, obtaining the most representative representation vector after dimensionality reduction. This strengthens the features of users and air interface vectors in the training matching score. A heuristic clustering algorithm is used, which, compared to the traditional K-means method, can more objectively identify neighbors with similar preferences to the target user. A negative feedback factor for such neighbors is introduced to further reduce the probability of multiple users matching the same air interface. Furthermore, considering the potential relationships between air interface resource allocations, a weighting factor of air interface category similarity is introduced into the predicted score, making the neighbor-weighted scoring more accurate. The training matching score also improves the rationality of the predicted matching score for air interface parameters and the stability of the recommendation.

[0033] 2. The training data for both the channel type identification module and the user behavior prediction module of this invention originate from a digital twin network. Since the digital twin network is highly synchronized and symbiotic with the physical world, it significantly reduces model training costs. Furthermore, by iteratively optimizing and testing the two trained models within the digital twin network, the user's environment and behavioral needs are effectively and dynamically updated. The derived new user feature vectors further improve the real-time performance and accuracy of the air interface parameter recommendation. Attached Figure Description

[0034] The present invention will now be described in further detail with reference to the accompanying drawings.

[0035] Figure 1 This is a flowchart of the present invention.

[0036] Figure 2 This is a framework diagram of the present invention.

[0037] Figure 3 This is a schematic diagram of the multi-head self-attention mechanism model of the present invention.

[0038] Figure 4A schematic diagram of the collaborative filtering recommendation algorithm based on user density clustering invented by Benfen. Detailed Implementation

[0039] like Figure 1 and Figure 2 As shown, the air interface parameter recommendation method based on the user-air interface knowledge graph includes the following steps:

[0040] Step S1: The knowledge representation module defines air interface entities and user entities, constructs air interface knowledge graphs and user knowledge graphs, and obtains the user feature vector E mapped from the semantic, numerical, and relational information of user entities through the Word2vec model. U The air interface feature vector E, which maps the semantic, numerical, and relational information of air interface entities, is obtained through the TransD model and the Word2ve model. A ;

[0041] Specifically, considering the parameter configuration and performance of the air interface, different parameter configuration methods can determine the performance of an air interface, which can be regarded as an "air interface" entity. In a certain environment, using this communication technology for data transmission and achieving the desired data rate, bit error rate, and other performance characteristics, this environment and performance requirements can be regarded as a "user" entity. The air interface achieving a certain performance under such parameter configuration establishes an interactive relationship between the air interface and the user entity, forming a connection between nodes.

[0042] An air interface entity consists of a triplet consisting of a head entity, a set of relations, and a tail entity. An air interface determined by different parameters and technical methods is considered as the air interface entity. The air interface is defined as the head entity, the various parameters of the air interface are defined as the set of relations, and the specific values ​​of the parameters are defined as the tail entity. The technical methods and parameters used by a certain air interface are extracted, such as bandwidth and encoding method, to form a triplet of air interface resource configuration, technical method, and technical parameters, so as to construct an air interface knowledge graph.

[0043] A user entity is composed of a triplet consisting of a head entity, a set of relations, and a tail entity. A user using a certain parameter and communication technology in a certain environment is considered a user entity. The overall environment and performance requirements are defined as the user's head entity, the parameter types of the environment and performance requirements are defined as the set of relations, and the specific values ​​of the parameters are defined as the tail entity. Based on the link simulation results, the environmental conditions of different air interface simulations, such as channel type, interference type, and signal-to-noise ratio, are summarized to form a triplet of user, environment and performance requirement type, and corresponding parameter values ​​of the environment and performance requirement type, in order to construct a user knowledge graph. Performance requirements include bit error rate, block error rate, communication rate, throughput, and spectral efficiency communication performance.

[0044] For users, the feature vectors output by the Word2vec model are concatenated with the corresponding numerical feature vectors to form a two-channel embedded representation vector (i.e., the user feature vector) E. U E U Dimensions Where, N U For the number of user features, The dimension of a single feature map is 2, where 2 indicates two channels;

[0045] For the air interface, the features output by the TransD model and the Word2ve model are concatenated into a three-channel representation vector (i.e., the air interface feature vector E). A E A Dimensions Where, N A The characteristic number of the air passage, The number of bits for a single feature map, where 3 indicates three channels;

[0046] The 6G air interface knowledge base contains both semantic data (such as technical methods and environmental information) and numerical information (such as attribute values ​​and demand values). TransD and Word2vec transform both numerical and semantic data into low-dimensional feature vectors through structural mapping, but they do not represent the original features of this data. Consider representing semantic data in dictionary form, such as subcarrier spacing (15kHz, 30kHz, 60kHz, 120kHz, 240kHz, etc.), defining each type of interference as a dictionary form of 0, 1, 2… and normalizing it. For numerical information, normalize the values ​​for each entity type and concatenate them with the semantic data to form a vector E. A The process of using the TransD model and Word2ve model to output the corresponding vector is an existing technology.

[0047] The Neo4j graph database and tabular format are used to jointly store the air interface parameter knowledge graph. Neo4j's advantages lie in its high efficiency in long-range relationship query and its ability to discover hidden relationships between entities. Excel spreadsheets are used to store air interface and user information, as shown in Table 1 (air interface entity) and Table 2 (user entity). The tabular data is clear and easy to understand, and it is convenient to read and generate Word2vec and numerical data, thus making data management more flexible.

[0048] Table 1

[0049]

[0050] Table 2

[0051]

[0052]

[0053] Step S2: Use the self-attention mechanism to process the user feature vector E obtained in step S1. U and air interface feature vector E A Feature extraction is performed, and the extracted features are enhanced using higher-order cross-feature extraction. Convolution is then used to reduce the dimensionality of the enhanced features to obtain the user features. and air interface features This is used to calculate the matching score R(m×n) of user historical information against air interface features; where R(m×n) represents the scoring matrix of n air interface features based on the historical information of m users.

[0054] Specifically, it includes:

[0055] Step S21, for dual-channel N U The user feature vector of each feature will be E U and three-channel N A The air interface feature vector E of each feature A For each feature in each channel, self-attention weighting is performed by assigning query vectors, key vectors, and value vectors to each feature. The similarity of each feature is then calculated to obtain the weighted feature matrix. These feature matrices are then concatenated to obtain the feature E′ extracted by the multi-head attention mechanism. U and E′ A ;

[0056] For example: dual-channel N U The user feature vector of each feature will be E U Represented as E u =[E U1 E U2 ] T , The feature vector of the first channel, For the feature vector of the second channel, for each feature of the feature vectors of the first and second channels respectively k′∈{1,...N U Attention weighting is applied separately to features. Multiply by the weight matrix W respectively K W V W Q And add the corresponding bias b K b V b Q That is, this feature Assigning a Query vector, Key vector, and Value vector: Where X is the feature. K, V, and Q represent the Query vector, Key vector, and Value vector, respectively.

[0057] Three-channel N A The air interface feature vector E of each feature A The same method was also used for self-attention weighting;

[0058] Apply a formula to each feature Calculate the similarity, normalize it using the softmax function, and then multiply it by V to obtain the weighted score of the feature: Attention(Q,K,V)=Similarity(Q,K)V;

[0059] Multi-head attention mechanism refers to performing the self-attention weighting mechanism multiple times. If the user has two channels, the self-attention weighting mechanism is performed twice; if the air interface has three channels, the self-attention weighting mechanism is performed three times. Figure 3 As shown, the multi-head attention mechanism concatenates the feature matrices to obtain the feature E′ extracted by the multi-head attention mechanism. U and E′ A ;

[0060] Step S22: For the feature E′ extracted in step S21 U and E′ A Features E′ are automatically constructed through multiple cross layers. U and E′ A The finite high-order cross features are used to obtain the enhanced features;

[0061] The cross-feature component automatically constructs a finite number of high-order cross-features for users and air interface parameters through multiple cross-layers. Each layer generates high-order interactions based on existing layers while maintaining interactions with previous layers. For the user features E′ weighted by the attention mechanism... U Air interface parameter feature vector E A Each layer is input into a cross-layer for feature extraction, and the formula for each layer is: Among them, X l′ X l′+1 The outputs of the l′ and l′+1 layers are of dimensions d and W. l ,b l ∈R d These are the weights and bias parameters of the l′th layer. After each feature cross function f(·), the input X of this layer will be... l′ Adding it to the network, analogous to the idea of ​​residual networks, i.e., f(·) = X l′+1 -X l′ Such residual networks can handle gradient degradation and vanishing problems, allowing neural networks to be deeper.

[0062] Step S23: Use two convolutional layers, a non-linear layer, and the corresponding pooling layer to reduce the dimensionality of the enhanced features from step S22 to obtain the user features. and air interface features

[0063] The two convolutional layers have 8 kernels of size 3×3. The non-linear layers use the LeakyReLU function, whose activation expression is: α = 0.2 is the scaling factor, and the pooling kernel size of the pooling layer is 1×4. The essence of pooling is sampling. For the input feature map, the max pooling or average pooling method is selected to compress and reduce the dimensionality, thereby reducing the number of parameters. The first pooling layer adopts the max pooling strategy to remove redundancy and noise from the local features of the feature vector. The second layer adopts the average pooling strategy to extract each feature, thereby contributing to the global features.

[0064] Step S24: Based on the historical information of m users, score n air interface features, that is, use a multilayer perceptron to fit the matching function between user features and air interface features, thus obtaining the matching score matrix R(m×n).

[0065] A multilayer perceptron consists of an input layer, n fully connected layers, and an output layer. It employs three fully connected layers with LeakyReLU activation functions, with units of 100, 50, and 25 respectively. The input is... The concatenated vectors are normalized by the output layer using Softmax to obtain the degree of matching between the user and the waveform. Assume user u... i The scoring term for the k-th opening is r. ik Then the user air interface rating matrix is Let the i-th row of R be defined as user u i The score vector R for all air ports i .

[0066] Step S3: Based on digital twin technology, design a channel type identification module for identifying and classifying channels in a certain scenario and outputting the results, and a user behavior prediction module for predicting user behavior in future moments. Based on the channel identification and classification results and the predicted behavior, infer the user's current environment and new needs in the physical world.

[0067] Specifically:

[0068] Step S31: In a digital twin network that is highly symbiotic with the physical world, construct a dataset of received signals with pilots under different electromagnetic propagation environments, and obtain a channel training dataset using least squares estimation; extract channel features under different electromagnetic propagation conditions from the channel training data as a channel feature parameter dataset. The channel features include kurtosis, skewness, root mean square delay spread, and Rice K-factor features. Since the units of different channel features are different, normalize all types of feature parameter data.

[0069] The collected channel feature parameter dataset is used as the training set for the machine learning algorithm. Each set of channel feature parameters in the training set is labeled with the corresponding electromagnetic transmission environment to establish a channel type identification model. Existing algorithms can be selected for the machine learning algorithm.

[0070] The channel type identification model and twin network undergo continuous iterative optimization and testing: The channel in the current transmission environment of the twin network is input into the channel identification model. The channel identification module extracts the characteristic parameters of this channel as test data, calculates the Euclidean distance between the test data and each set of channel characteristic parameters in the training set, sorts the Euclidean distances, and selects the K values ​​with the smallest distance to the set of test data. The electromagnetic transmission environment corresponding to these K values, which appears most frequently in the labels, is determined as the electromagnetic transmission environment of the channel in the twin network. The identified transmission environment is then compared with the transmission environment in the twin network. If it is outside the model's allowable error range, it is fed back to the channel identification model. The channel identification model is then retrained by collecting more channel data to further optimize the model. Through continuous iterative optimization and testing of the channel type identification model and twin network, the accuracy of the channel type identification model is continuously improved.

[0071] Input channel data from the physical world, and the channel type identification model identifies and classifies the channels in the scenario and outputs the electromagnetic propagation environment in the physical world.

[0072] Step S32: Collect long-term user behavior data in the digital twin network as a behavior parameter dataset, such as mobile speed, service transmission type, and service quality requirements. Since the units of different channel characteristics are different, normalize the behavior parameter data.

[0073] The LSTM training model of user behavior is established using the Long Short-Term Memory algorithm, specifically as follows:

[0074] Initialize parameters, and further initialize data weights based on the normal distribution characteristics. Set the mean and standard deviation to 0 and 1 respectively, and set all initial bias constants to 0.1.

[0075] When training the LSTM model, gradient descent is used to train the parameters, taking into account the actual situation.

[0076] The LSTM training model and the twin network are continuously iterated, optimized, and tested: The user behavior parameters of the twin network at the previous moment are input into the LSTM training model. The LSTM training model analyzes these parameters, predicts the user behavior parameters of the twin network at the current moment, and then compares the predicted value with the user behavior parameters of the twin network at the current moment. If it is not within the error range allowed by the model, it is fed back to the LSTM training model. The channel identification model is trained again by collecting more user behavior parameters to further improve the accuracy of the LSTM training model.

[0077] Input the user behavior data of the physical world at the current moment, and train the LSTM model to predict the user behavior at future moments in order to update the user's status and make decisions in real time.

[0078] Step S33: Based on the channel identification classification results of the channel type identification model and the behavioral requirements predicted by user behavior prediction, deduce the user's current environment and new requirements in the physical world.

[0079] Step S4: Based on the matching score R(m×n) from step S2 and the user's current environment and new needs deduced in step S3, conduct a predictive score of the target user's air interface under the current environment and new needs, and sort the predicted scores to provide an air interface recommendation list.

[0080] Specifically:

[0081] Step S41: Based on the matching score R(m×n), calculate the Euclidean distance matrix D(m×m) for the user, and calculate the Euclidean similarity matrix OS(m×m) based on the Euclidean distance matrix; where the row vector of the i-th row of the Euclidean matrix D(m×m) is the user u i Distance vectors to other users;

[0082] User u i The score vector for all air ports is R. i =(r i1 ,r i2 ,r i3 ,...r im If user u i and user u j The Euclidean distance between them is This constitutes the Euclidean distance matrix of the user. The row vector L of the i-th row of matrix D (m×m) i (d i,1 ,d i,2 ,d i,3 ,...,d i,m ) for user u iDistance vectors to other users in the system;

[0083] Euclidean similarity is inversely correlated with Euclidean distance, and user u i and user u j The European similarity is This constitutes the Euclidean similarity matrix of the users.

[0084] Step S42: Calculate the mean d of the sum of distances between all users based on the Euclidean distance matrix, and perform heuristic clustering on the users based on the mean d of the sum of distances to obtain multiple cluster sets;

[0085] D(m×m) is an angularly symmetric matrix; we only need to calculate the upper triangular matrix to obtain the result.

[0086] Let the mean d be the distance threshold d T Introducing the point density parameter m i,j For any j∈{1,2,3...m}, d i,j ≥d T Then m i,j =1, otherwise m i,j =0, user u i Point density m i for: From m i Find the optimal cluster center m in (1≤i≤m). k For any integer i ∈ [1, m], m k ≥m i This allows us to find the user u with the highest user point density. k . will u k Let z1 be the first cluster center. Using the k-nearest neighbor algorithm based on Euclidean distance, find the k neighbors most similar to the target user from the dataset to form G1.

[0087] Then calculate the distance from the remaining users to each cluster center, assuming there are already s cluster centers z1, z2, z3, ..., z s Calculate the i-th user u that is not yet a cluster center. i To the center of class Z j The distance d between (j=1,2,3,...,s) ij and calculate If d J If α*d(z1,z2), then establish the (s+1)th cluster center z. s+1 And z s+1 =x JOtherwise, clustering ends. Where d(z1,z2) is the distance between the first and second cluster centers, and α can be obtained by α = d... T / d max The calculations yielded the following sets: G1, G2, G3, ... G... k ;

[0088] Step S43, Target User u a The cluster set G it belongs to g Given the target user's neighbor set, the target user's predicted rating for the currently unrated target airway l is: in, For target user u a For the currently unrated target interface l, the basic score value is Isim(l,j′), which represents the target user u. a The evaluated category similarity between air interface j′ and target air interface l, where A represents the target user u. a Number of air ports evaluated, R a (j′) represents the target user u a For the score of the empty j′, 1≤j′≤a, In order to connect with target user u a Neighboring users with similar preferences j For the scoring items of target air interface l, Sim(a,j) represents the target user u. a and neighboring user u j The similarity between them, where K is the target user u a The number of users in adjacent user groups, R j (l) represents adjacent user u j The score of the air interface l, Indicates adjacent user u j The mean score for the empty speech;

[0089] Among them, target user u a The category similarity between the evaluated air interface j′ and the target air interface l is expressed by the formula calculate, For the enhanced feature vector of the air port l, The enhanced feature vector of air interface j is obtained using existing technology.

[0090] Step S44: Based on the predicted score P a,i Sort by size for target user u a Provide an air interface recommendation list

[0091] The above description is merely a preferred embodiment of the present invention and should not be construed as limiting the scope of the present invention. All equivalent changes and modifications made in accordance with the scope of the patent application and the contents of the specification of the present invention should still fall within the scope of the patent of the present invention.

Claims

1. A method for recommending air interface parameters based on a user-air interface knowledge graph, characterized in that: Includes the following steps: Step S1: The knowledge representation module defines air interface entities and user entities, constructs air interface knowledge graphs and user knowledge graphs, and obtains the user feature vector E mapped from the semantic, numerical, and relational information of user entities through the Word2vec model. U The air interface feature vector E, which maps the semantic, numerical, and relational information of air interface entities, is obtained through the TransD model and the Word2ve model. A ; Step S2: Use the self-attention mechanism to process the user feature vector E obtained in step S1. U and air interface feature vector E A Feature extraction is performed, and the extracted features are enhanced using higher-order cross-feature extraction. Convolution is then used to reduce the dimensionality of the enhanced features to obtain the user features. and air interface features This is used to calculate the matching score R(m×n) of user historical information against air interface features; where R(m×n) represents the scoring matrix of n air interface features based on the historical information of m users. Step S3: Based on digital twin technology, design a channel type identification module for identifying and classifying channels in a certain scenario and outputting the results, and a user behavior prediction module for predicting user behavior in future moments. Based on the channel identification and classification results and the predicted behavior, infer the user's current environment and new needs in the physical world. Step S4: Based on the matching score R(m×n) from Step S2 and the user's current environment and new needs deduced in Step S3, conduct a predictive score for the air interface of the target user under the current environment and new needs, and sort the predicted scores to provide an air interface recommendation list.

2. The air interface parameter recommendation method based on user-air interface knowledge graph according to claim 1, characterized in that: In step S1, the air interface entity is composed of a triplet consisting of a head entity, a relation set, and a tail entity. An air interface determined by different parameters and technical methods is regarded as the air interface entity. The air interface is defined as the head entity, the various parameters of the air interface are defined as the relation set, and the specific values ​​of the parameters are defined as the tail entity. The technical methods and parameters used by a certain air interface are extracted to form a triplet of air interface resource configuration, technical method, and technical parameters to construct an air interface knowledge graph. A user entity is composed of a triplet consisting of a head entity, a set of relations, and a tail entity. A user who uses a certain parameter and communication technology in a certain environment is considered a user entity. The overall environment and performance requirements are defined as the head entity of the user, the parameter types of the environment and performance requirements are defined as the set of relations, and the specific values ​​of the parameters are defined as the tail entity. Based on the link simulation results, the environmental conditions of different air interface simulations are summarized to form a triplet of user, environment and performance requirement type, and corresponding parameter values ​​of environment and performance requirement type, in order to construct a user knowledge graph. The Neo4j graph database and tabular format are used to jointly store the air interface parameter knowledge graph, and Excel spreadsheets are used to store air interface and user information.

3. The air interface parameter recommendation method based on user-air interface knowledge graph according to claim 2, characterized in that: In step S1, the user feature vector E U Dimensions Where, N U For the number of user features, The dimension of a single feature mapping is 2, where 2 indicates dual channels; the air interface feature vector E A Dimensions Where, N A The characteristic number of the air passage, The number of bits for a single feature map is 3, where 3 indicates three channels.

4. The air interface parameter recommendation method based on user-air interface knowledge graph according to claim 3, characterized in that: Step S2 includes the following steps: Step S21, for dual-channel N U The user feature vector of each feature will be E U and three-channel N A The air interface feature vector E of each feature A For each feature in each channel, self-attention weighting is performed by assigning query vectors, key vectors, and value vectors to each feature. The similarity of each feature is then calculated to obtain the weighted feature matrix. These feature matrices are then concatenated to obtain the feature E′ extracted by the multi-head attention mechanism. U and E′ A ; Step S22: For the feature E′ extracted in step S21 U and E′ A Features E′ are automatically constructed through multiple cross layers. U and E′ A The finite high-order cross features are used to obtain the enhanced features; Step S23: Use two convolutional layers, a non-linear layer, and the corresponding pooling layer to reduce the dimensionality of the enhanced features from step S22 to obtain the user features. and air interface features Step S24: Based on the historical information of m users, score n air interface features, that is, use a multilayer perceptron to fit the matching function between user features and air interface features, thus obtaining the matching score matrix R(m×n), where the input of the multilayer perceptron is... Serial vectors.

5. A method for recommending air interface parameters based on a user-air interface knowledge graph according to claim 1, 2, or 3, characterized in that: Step S3 includes the following steps: Step S31: In a digital twin network that is highly symbiotic with the physical world, construct a dataset of received signals with pilots under different electromagnetic propagation environments, and obtain a channel training dataset using least squares estimation; extract channel features under different electromagnetic propagation conditions from the channel training data as a channel feature parameter dataset, and establish a channel type identification model using machine learning algorithms and the channel feature parameter dataset; input the channel data from the physical world, and the channel type identification model identifies and classifies the channel in this scenario and outputs the results; among which, the channel features include kurtosis, skewness, root mean square delay spread, and Rice K-factor features; Step S32: Collect long-term user behavior data in the digital twin network, train it using the Long Short-Term Memory algorithm, and establish an LSTM training model of user behavior; input the user behavior data of the physical world at the current moment, and the LSTM training model predicts the user behavior at future moments to update the user's status and make decisions in real time. Step S33: Based on the channel identification classification results of the channel type identification model and the behavioral requirements predicted by user behavior prediction, deduce the user's current environment and new requirements in the physical world.

6. The air interface parameter recommendation method based on user-air interface knowledge graph according to claim 3, characterized in that: Step S4 includes the following steps: Step S41: Based on the matching score R(m×n), calculate the Euclidean distance matrix D(m×m) for the user, and calculate the Euclidean similarity matrix OS(m×m) based on the Euclidean distance matrix; where the row vector of the i-th row of the Euclidean matrix D(m×m) is the user u i Distance vectors to other users; Step S42: Calculate the mean of the sum of distances between all users based on the Euclidean distance matrix, and perform heuristic clustering of users based on the mean of the sum of distances to obtain multiple cluster sets; Step S43, Target User u a The cluster set G it belongs to g Given the target user's neighbor set, the target user's predicted rating for the currently unrated target airway l is: in, For target user u a For the currently unrated target interface l, the basic score value is Isim(l,j'), which represents the target user u. a The evaluated category similarity between air interface j' and the target air interface l, where A represents the target user u. a Number of air ports evaluated, R a (j') represents the target user u a For the score of empty j', 1≤j'≤A, In order to connect with target user u a Neighboring users with similar preferences j For the scoring items of target air interface l, Sim(a,j) represents the target user u. a and neighboring user u j The similarity between them, where K is the target user u a The number of users in adjacent user groups, R j (l) represents adjacent user u j The score for the air interface l, Indicates adjacent user u j The mean score for the empty speech; Step S44: Based on the predicted score P a,l Sort by size for target user u a Provides a list of recommended air interface providers.

7. The air interface parameter recommendation method based on user-air interface knowledge graph according to claim 4, characterized in that: In step S23, the number of convolution kernels in the two convolutional layers is 8, and the size is 3×3. The nonlinear layer uses the LeakyReLU function, and the pooling kernel size of the pooling layer is 1×4.

Images