A Clustering Analysis Method for 6G Service Scenarios

Abstract

The present invention discloses a 6G business scenario clustering analysis method, including: collecting 6G business sample data to determine multiple KPIs; obtaining an incomplete multi-view data set; determining an initial value and multiple test values ​​of the number of clusters; determining multiple regularization parameters and anchor point numbers, and the anchor point numbers are determined according to different multiples of the number of clusters, and obtaining multiple initial clustering parameter combinations formed by the test values ​​of the regularization parameters, the number of anchor points, and the number of clusters; clustering the incomplete multi-view data set using an incomplete multi-view clustering method under each combination, and determining the optimal clustering parameter combination for the result by rough screening and fine screening, and retrieving the corresponding clustering result; performing t-sne visualization to obtain a clustering effect diagram, and combining the original data to compare and analyze the clustering results to obtain a scene KPI feature radar diagram and scene knowledge to cooperate with on-demand services. The present invention can efficiently perform clustering scenario analysis on massive and default data, and the results can be used as scenario knowledge to cooperate with on-demand services to cope with dynamically changing environments.

Description

Technical Field

[0001] The present invention belongs to the field of communications, and specifically relates to a 6G service scenario clustering analysis method. Background Art

[0002] The fifth generation of mobile communications (5G) proposes three major scenarios: enhanced Mobile Broadband (eMBB), ultra-reliable and low-latency communications (uRLLC), and massive Machine Type Communications (mMTC). These three scenarios cover almost all application scenarios under 5G, and the formulation of the three scenarios is of guiding significance for the subsequent development of some 5G services.

[0003] The classification of the three major 5G scenarios is mainly derived by industry experts who refine and classify services based on their business characteristics. Most of the refinement and clustering process is manually implemented by experts. The sixth-generation mobile communication (6G) network has diverse and complex scenarios. Compared with 5G, the services in its scenarios show the characteristics of higher dimensions and larger scope. For the complex business conditions of 6G in all domains and scenarios, it is extremely inefficient and costly for experts to formulate 6G scenario business categories.

[0004] At present, in the field of communications, only a simple clustering analysis can be performed on a small number of services of more than 20. For example, a small number of 6G services can be clustered using the hierarchical clustering method, but there are some defects: First, the clustering of these small-scale business scenarios is very limited, and it is impossible to handle a large amount of default data, which is not applicable to the real massive 6G services. Secondly, the hierarchical clustering method cannot combine all the characteristics of each service for clustering, and the clustering efficiency is low. Thirdly, the hierarchical clustering method has high clustering complexity and low practicality. In addition, since 6G technology is under research, the services related to 6G technology are also unknown. Not only the clustering results of currently known services are needed, but also the clustering of unknown new services.

[0005] Therefore, it is necessary to find a new clustering analysis method that can refine and cluster massive 6G services with default data according to the different characteristics of diversified services in the scenario, and perform detailed analysis on them to meet the scenario clustering analysis needs of known and unknown services. Summary of the invention

[0006] In order to solve the above problems existing in the prior art, the present invention provides a 6G service scenario clustering analysis method. The technical problem to be solved by the present invention is achieved by the following technical solutions:

[0007] Collect a number of 6G service sample data, and determine a plurality of KPIs describing the service perspective; perform data processing on the 6G service sample data to obtain an incomplete multi-view data set; wherein the incomplete multi-view data set includes view data of a plurality of KPIs, and the view data of each KPI includes corresponding data of all 6G service sample data under the KPI, and default data is used to represent the case where no corresponding data exists;

[0008] Determine an initial value of the number of clusters, and expand the range based on the initial value of the number of clusters to determine multiple test values ​​of the number of clusters;

[0009] Determine a first number of regularization parameters and a second number of anchor points, and each number of anchor points is determined according to different multiples of the number of clusters, to obtain a plurality of initial clustering parameter combinations formed by test value combinations of the regularization parameters, the number of anchor points, and the number of clusters;

[0010] Under each initial clustering parameter combination, the incomplete multi-view clustering method is used to cluster the incomplete multi-view data set, and all clustering results are roughly screened and accurately screened in turn, and the regularization parameter, number of anchor points, and number of clusters with the best clustering effect are determined as the optimal clustering parameter combination, and the target clustering result corresponding to the optimal clustering parameter combination is retrieved; wherein each cluster label in each clustering result corresponds to a 6G scenario;

[0011] The target clustering results are visualized by t-sne to obtain a clustering effect diagram, and the clustering results are compared and analyzed with the original data to obtain a scene KPI feature radar diagram and scene knowledge to cooperate with on-demand services.

[0012] In one embodiment of the present invention, the multiple KPIs include: rate, latency, reliability, mobility, connection density, coverage, positioning accuracy and availability.

[0013] In one embodiment of the present invention, the step of processing the 6G service sample data to obtain an incomplete multi-view data set includes:

[0014] After vectorizing the 6G service sample data, the data is layered according to multiple KPIs;

[0015] Generate an index matrix with data values ​​according to the data under each KPI, which includes the serial number of the 6G service sample data with corresponding data under the KPI;

[0016] According to the index matrix under the KPI and the vectorized 6G service sample data, a data vectorization matrix is ​​obtained as the view data of the KPI, and the view data of all KPIs constitute an incomplete multi-view data set.

[0017] In one embodiment of the present invention, the determining of the initial value of the number of clusters, and the range expansion based on the initial value of the number of clusters to determine multiple test values ​​of the number of clusters include:

[0018] The initial value of the number of clusters is determined by the brainstorming method. Where n is the total number of samples of 6G service sample data;

[0019] The initial value k0 is used as a reference to expand to both sides, and k0 values ​​are determined as multiple test values ​​of the number of clusters.

[0020] In one embodiment of the present invention, the determining of a first number of regularization parameters and a second number of anchor points, wherein each anchor point number is determined according to a different multiple of the number of clusters, and a plurality of initial clustering parameter combinations formed by test value combinations of regularization parameters, anchor point numbers, and cluster numbers are obtained, includes:

[0021] Determining a first number of regularization parameters μ includes: 0.001, 0.1, 1, 10, 100 and 1000;

[0022] Determining the second number of anchor points m includes: k, 2k, 3k, 4k, 5k, 6k and 7k; wherein k represents the number of clusters;

[0023] A combination of any regularization parameter μ, the number of anchor points m, and the number of clusters k taking any value from multiple test values ​​is taken as an initial clustering parameter combination, and multiple initial clustering parameter combinations are obtained based on all combinations of the regularization parameter μ, the number of anchor points m, and the number of clusters k.

[0024] In one embodiment of the present invention, clustering the incomplete multi-view data set using an incomplete multi-view clustering method under each initial clustering parameter combination includes:

[0025] Under each initial clustering parameter combination, m anchor points are learned from the view data of each KPI through the anchor point selection strategy of direct alternating sampling;

[0026] For each KPI corresponding to m anchor points, use the bipartite graph method to construct an m*n-order matrix as the anchor graph under the KPI view;

[0027] The anchor graphs under each KPI are fused using different preset weights to construct an m*n-order matrix as the consensus anchor graph;

[0028] The consensus anchor graph is used to perform spectral clustering based on the consensus anchor graph to obtain cluster labels of multiple clusters, wherein each cluster represents a category of a 6G scenario.

[0029] In one embodiment of the present invention, the method of performing rough screening and precise screening on all clustering results in sequence to determine the regularization parameter, the number of anchor points, and the number of clusters with the best clustering effect as the optimal clustering parameter combination includes:

[0030] The clustering results obtained from the test values ​​of all cluster numbers under the same regularization parameter μ and anchor point number m are used to generate an elbow coefficient graph using the elbow method; among all the elbow coefficient graphs obtained, the clustering parameter combinations that meet the shape of the elbow curve are retained to complete the rough screening;

[0031] The silhouette coefficient method is used to generate a silhouette coefficient diagram for the clustering parameter combinations that are roughly screened and retained. The clustering parameter combinations are accurately screened by selecting the maximum value in the silhouette coefficient diagram, and the regularization parameter, number of anchor points and number of clusters with the best clustering effect are obtained as the optimal clustering parameter combination.

[0032] In one embodiment of the present invention, in the elbow coefficient graph, the calculation formula of the elbow coefficient is:

[0033]

[0034] Among them, D k represents the elbow coefficient under a set of regularization parameters μ, anchor point number m and cluster number k; x represents the sample, which is the 6G service sample data after vectorization processing; c i It represents the cluster center of the i-th cluster in the clustering result; dist represents the distance function.

[0035] In one embodiment of the present invention, for any sample x i , the process of calculating the silhouette coefficient using the silhouette coefficient method includes:

[0036] Calculate sample x i The average distance a(i) to other samples in the same cluster is taken as the sample x i The intra-cluster dissimilarity of

[0037] Calculate sample x i The average distance b(ij) to all samples in each other cluster C(j) is taken as the distance between sample x i Dissimilarity with cluster C(j);

[0038] According to the sample x i The dissimilarity with all other clusters is calculated for sample x i The inter-cluster dissimilarity is b(i)=min{b i1 ,b i2 ,…,b ik};

[0039] According to the sample x iThe intra-cluster dissimilarity a(i) and inter-cluster dissimilarity b(i) of the sample x are calculated. i The silhouette coefficient is:

[0040] In one embodiment of the present invention, obtaining scene knowledge to cooperate with on-demand services includes:

[0041] Analyze the scene KPI feature radar chart to obtain the KPI features of each scene as corresponding scene knowledge;

[0042] The acquired scenario knowledge is connected to the pre-established 6G knowledge graph library to obtain the correspondence between the business and the relevant on-demand service method in the 6G knowledge graph library, so as to achieve the purpose of any 6G business in a dynamic environment obtaining the corresponding on-demand service based on the correspondence.

[0043] The embodiment of the present invention pioneered a scheme in the field of communications to cluster massive services using incomplete multi-view clustering to obtain 6G scenarios and analyze them. In this scheme, first, a number of 6G service sample data are collected, and data stratification and other processing are performed according to KPI to obtain an incomplete multi-view data set, and complete the data preparation for clustering; since the default data is considered under each KPI, a large amount of 6G service sample data with data default can be processed; secondly, multiple initial clustering parameter combinations formed by test value combinations of regularization parameters, anchor points, and cluster numbers are obtained as parameter preparation for parameter adjustment; next, for each initial clustering parameter combination, the incomplete multi-view data set is clustered using an incomplete multi-view clustering method, and all clustering results are roughly screened and accurately screened in turn, and the regularization parameter, anchor point number, and cluster number with the best clustering effect are determined as the optimal clustering parameter combination, the parameter adjustment process is completed, and the optimal clustering result is determined. Finally, the optimal clustering result is analyzed, including obtaining a visual clustering result diagram and a scene KPI feature radar diagram.

[0044] Compared with the prior art, the embodiments of the present invention mainly have the following beneficial effects:

[0045] 1. Ability to perform clustering scenario analysis on massive 6G service data with high efficiency and short time;

[0046] 2. Ability to handle clustering with a large amount of default data;

[0047] 3. The final result of cluster analysis can be used as a kind of scenario knowledge to cooperate with on-demand services to cope with dynamically changing environments, and can meet the scenario clustering analysis needs of known and unknown businesses. BRIEF DESCRIPTION OF THE DRAWINGS

[0048] Figure 1A schematic diagram of a process of a 6G service scenario clustering analysis method provided by an embodiment of the present invention;

[0049] Figure 2 This is a schematic diagram of the 6G scenario clustering idea of ​​an embodiment of the invention;

[0050] Figure 3 A schematic diagram of a flow chart of an incomplete multi-view clustering method used in an embodiment of the invention;

[0051] Figure 4 A plurality of elbow coefficient graphs obtained for Example 1 of an embodiment of the invention;

[0052] Figure 5 A plurality of contour coefficient graphs obtained for Example 1 of the embodiment of the invention;

[0053] Figure 6 This is a visualization clustering result diagram obtained in Example 1 of an embodiment of the invention;

[0054] Figure 7 This is a radar chart of the scene KPI characteristics obtained in Example 1 of the embodiment of the invention;

[0055] Figure 8 A schematic diagram of 6G service clustering in the 6G service scenario clustering analysis method provided in an embodiment of the present invention;

[0056] Fig. 9 This is a schematic diagram of the overall process of 6G scenario clustering according to an embodiment of the present invention. DETAILED DESCRIPTION

[0057] The following will be combined with the drawings in the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are only 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.

[0058] The embodiments of the present invention aim to find a method for refining and clustering the rich and complex 6G services, so as to inherit the three major application scenarios of 5G and at the same time divide the scenarios in more detail according to the 6G service conditions, and after dividing the scenarios, analyze the characteristics and practical applications of each type of scenario based on the KPI-related knowledge of each type of scenario.

[0059] like Figure 1 As shown, a 6G service scenario clustering analysis method provided by an embodiment of the present invention may include:

[0060] S1, collect a number of 6G service sample data, and determine a plurality of KPIs describing the service perspective; perform data processing on the 6G service sample data to obtain an incomplete multi-view data set;

[0061] Among them, the incomplete multi-view data set includes view data of multiple KPIs, and the view data of each KPI includes the corresponding data of all 6G service sample data under the KPI, and the default data is used to represent the situation where there is no corresponding data.

[0062] Specifically, this step may first collect a number of 6G service sample data, and then determine multiple KPIs as indicators based on the characteristics of the 6G services, which are used to describe different perspectives of the services or to describe different characteristics of the services.

[0063] In an optional implementation, multiple KPIs include: rate, latency, reliability, mobility, connection density, coverage, positioning accuracy and availability.

[0064] It can be understood that each 6G service sample data contains corresponding data of at least one of the KPIs.

[0065] Generally speaking, when the set of 6G service sample data is complete, the more KPIs are clustered, the better the characteristics of each service can be reflected, the better the clustering effect will be, and the closer the final clustering result will be to the real scenario.

[0066] In an optional implementation manner, data processing is performed on the 6G service sample data to obtain an incomplete multi-view data set, including steps A1 to A3:

[0067] A1, after vectorizing the 6G service sample data, stratify it according to multiple KPIs;

[0068] Since many 6G services are still under exploration, there will be a lot of data missing, and a large amount of service data is textual description data, so the embodiment of the present invention extracts a large amount of collected 6G service sample data according to KPI and performs corresponding vectorization processing, wherein the dimension of data vectorization under each KPI is different. By adopting vectorization processing, all types of data can be clustered together, which can increase the richness and completeness of the data.

[0069] In addition, the vectorized 6G service sample data can be layered according to different KPIs, and each layer is the corresponding data of all vectorized 6G service sample data under this KPI.

[0070] A2, generating an index matrix with data values ​​according to the data under each KPI, which includes the serial numbers of the 6G service sample data with corresponding data under the KPI;

[0071] Since each 6G service sample data may not contain corresponding data under all KPIs, but may have data missing under individual KPIs, for each KPI, the numbers of the 6G service sample data with corresponding data can be determined, and an index matrix is ​​obtained from these numbers.

[0072] A3, according to the index matrix under the KPI and the 6G service sample data after vectorization processing, a data vectorization matrix is ​​obtained as the view data of the KPI, and the view data of all KPIs constitute an incomplete multi-view data set.

[0073] When the specific algorithm is executed, this step first generates an indicator matrix with data defaults for each KPI according to the index matrix under the KPI.

[0074] Among them, the data in the indicator matrix is ​​arranged according to the numbering order of the 6G business sample data. When the 6G business sample data has corresponding data under the KPI, it is 1, and when the 6G business sample data has no corresponding data under the KPI, it is 0. Therefore, the indicator matrix can also be called a (0, 1) state matrix.

[0075] Then, for each KPI, the indicator matrix under the KPI and the matrix formed by the vectorized 6G service sample data under the KPI are multiplied to obtain a data vectorization matrix as the view data of the KPI. The dimension of the view data of the KPI is the dimension of the data vectorization under the KPI * the total number of samples of the 6G service sample data.

[0076] To facilitate understanding of the solution, an example 1 is used for illustration. Assume that the total number of samples of 6G service sample data collected by S1 is 489, 8 KPIs are determined, and the number of samples for each KPI view is shown in Table 1 below:

[0077] Table 1

[0078]

[0079] For each KPI, the sum of the number of samples with data and the number of samples with default data is the total number of samples, 489. The vectorization dimensions of each KPI are different. As can be seen from Table 1, the default rate of data for some KPIs exceeds 50% or more.

[0080] S2, determining an initial value of the number of clusters, and expanding the range based on the initial value of the number of clusters to determine multiple test values ​​of the number of clusters;

[0081] In an optional implementation manner, the step may include:

[0082] S21, using the brainstorming method to determine the initial value of the number of clusters

[0083] Wherein, n is the total number of samples of 6G service sample data, and its value is a natural number greater than 0.

[0084] The embodiment of the present invention can quickly obtain the initial value k0 of the number of clusters by using the brainstorming method, which means that the number of clusters obtained by clustering, that is, the number of clusters, is approximately k0, that is, near k0.

[0085] For example, in the above example 1, n is 489, and k0 is approximately calculated to be 15.

[0086] S22, expanding to both sides based on the initial value k0, and determining k0 values ​​as multiple test values ​​of the number of clusters.

[0087] In an optional implementation, k0 is used as a reference and expanded to both sides with a step size of 1 to obtain a total of k0 values ​​as multiple test values ​​of the number of clusters. The number of k0 expanded to both sides may be different. The embodiment of the present invention determines multiple test values ​​of the number of clusters in order to subsequently observe the impact of changes in the number of clusters k on the adjustment of clustering parameters, thereby achieving a good parameter adjustment effect.

[0088] For example, in Example 1 above, 15 values ​​are selected around k0 = 15, and the test value of the cluster number k obtained can be k = [6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19, 20].

[0089] S3, determining a first number of regularization parameters and a second number of anchor points, wherein each number of anchor points is determined according to different multiples of the number of clusters, to obtain a plurality of initial clustering parameter combinations formed by test value combinations of the regularization parameters, the number of anchor points, and the number of clusters;

[0090] It can be understood that the regularization parameter and the number of anchor points are necessary parameters in the incomplete multi-view clustering method. In an embodiment of the present invention, a plurality of specific values ​​can be preset accordingly so as to obtain the optimal parameters through subsequent parameter adjustment, wherein the first number and the second number are both natural numbers greater than 0 and can be set as needed.

[0091] In an optional implementation manner, the step may include:

[0092] S31, determining a first number of regularization parameters μ including: 0.001, 0.1, 1, 10, 100 and 1000;

[0093] For the concept of regularization parameter, please refer to the related art for understanding, which will not be explained here. The first number is not limited to 6 as shown above, that is, the regularization parameter μ includes but is not limited to the above 6 values.

[0094] For the above example 1, the regularization parameter is determined to be μ=[0.001, 0.1, 1, 10, 100, 1000].

[0095] S32, determining that the second number of anchor points m includes: k, 2k, 3k, 4k, 5k, 6k and 7k;

[0096] Here, k represents the number of clusters, which is a natural number greater than 0. In the embodiment of the present invention, the numerical structure of the second number of anchor points m is constructed using the number of clusters, and the specific numerical value thereof can be obtained after substituting different numbers of clusters k.

[0097] For the concept of anchor point number, please refer to the related art for understanding, which will not be explained here. The second number is not limited to 7 shown above, that is, the anchor point number m includes but is not limited to the above 7 numerical structures.

[0098] For the above example 1, the number of anchor points is determined to be m = [k, 2*k, 3*k, 4*k, 5*k, 6*k, 7*k].

[0099] S33, taking a combination of any regularization parameter μ, the number of anchor points m, and the number of clusters k as any value among multiple test values ​​as an initial clustering parameter combination, and obtaining multiple initial clustering parameter combinations based on all combinations of the regularization parameter μ, the number of anchor points m, and the number of clusters k.

[0100] For the test values ​​of the first number of regularization parameters μ, the second number of anchor points m, and the k0 number of clusters determined by S22, when these three parameters are each combined with one of their own value sets, an initial clustering parameter combination can be obtained, and multiple initial clustering parameter combinations can be obtained by traversing all three-parameter combinations.

[0101] S4, under each initial clustering parameter combination, cluster the incomplete multi-view data set using the incomplete multi-view clustering method, and perform rough screening and precise screening on all clustering results in turn, determine the regularization parameter, number of anchor points, and number of clusters with the best clustering effect as the optimal clustering parameter combination, and retrieve the target clustering result corresponding to the optimal clustering parameter combination;

[0102] In order to facilitate understanding of the solution, the clustering idea and the incomplete multi-view clustering method adopted in the embodiment of the present invention are first introduced.

[0103] The embodiment of the present invention can perform clustering on each business perspective at the same time, which is equivalent to clustering each KPI as a dimension, and then performing multi-dimensional clustering of multiple KPIs. After that, the clustering results of all business perspectives are aggregated to obtain the clustering of all businesses. Each cluster of the clustering results can be regarded as a scene, that is, each cluster label in each clustering result corresponds to a 6G scene. The idea of ​​6G scene clustering in the embodiment of the present invention can be found in Figure 2 shown.

[0104] Figure 2 In the figure, 6G is illustrated as business X, business Y, business Z…business N. The different KPI perspectives of each business correspond to the balls of each layer strung together by vertical dotted lines in the gray layer. Different KPIs are represented as KPI 1, KPI 2, KPI 3, KPI 4, etc. One layer represents a KPI, and different KPIs of the same business are represented by balls of different colors in each layer. In other words, the balls in each layer are the same color, and the balls of the same color in a layer represent different businesses with different KPI values ​​under the same KPI perspective. The balls in different layers have different colors. This is the result of stratification based on multiple KPIs. However, it should be noted that Figure 2 The colors are grayed out in the figure. Finally, through such multi-view clustering, we can get the result of clustering the services into clusters, as shown in the figure below: Figure 2 As shown on the right, one region is a cluster, representing a scene, such as scene A, scene B…scene N.

[0105] Since in the process of collecting 6G service sample data, the default rate of some 6G service sample data exceeds half, the clustering effect of ordinary clustering methods is very low when the data default rate is so high. Therefore, the embodiment of the present invention adopts an incomplete multi-view clustering method for clustering according to the characteristics of the clustered data services and the clustering idea of ​​hierarchical multi-dimensional clustering.

[0106] Among them, the incomplete multi-view clustering method can be mainly divided into two parts: 1) Select anchor points, build anchor graphs based on the bipartite graph of each view and the final consensus anchor graph; 2) Obtain cluster labels based on the spectral clustering based on the consensus anchor graph. For the specific process, please refer to Figure 3 understand.

[0107] Specifically, under each combination of initial clustering parameters, the incomplete multi-view clustering method is used to cluster the incomplete multi-view data set, including steps B1 to B4:

[0108] B1, under each initial clustering parameter combination, m anchor points are learned from the view data of each KPI through the anchor point selection strategy of direct alternating sampling;

[0109] The incomplete multi-view clustering method first prepares an original incomplete multi-view dataset, which can be completed after the KPI stratification in the A1 step of S1, that is, the 6G service sample data after vectorization processing under each KPI is the original view data of the KPI view, which may have default data. The original incomplete multi-view dataset includes the original view data of all KPIs. For ease of understanding, it is assumed here that the original incomplete multi-view dataset has the original view data of V KPIs, referred to as V KPI views, corresponding to V KPIs. The original incomplete multi-view dataset is as follows Figure 3 The left box shows a simplified representation of incomplete multi-view data. The original view data of each KPI corresponds to the X in the box. (1) , X (2) …X (V) As shown, and for the convenience of visual understanding and distinction, the original view data of different KPIs are Figure 3 Represented by different shapes.

[0110] See the procedure for step B1. Figure 3 The understanding of the middle box can be simply summarized as: learning a separate anchor point for each view based on the data of each view.

[0111] Among them, the anchor point selection strategy for direct alternating sampling is as follows:

[0112] Alternately sample data points according to their feature representation:

[0113]

[0114]

[0115] Among them, it is understandable that represents the original incomplete multi-view dataset; X (1) ;X (2) ;...;X (V) Indicates the original view data of each KPI; (v) The dimension representing the data features of the original view data of the vth KPI; d T The total dimension of the data features representing the original incomplete multi-view dataset;

[0116] Among them, the process of calculating the sample score in the anchor point selection step is:

[0117] The scores of all samples are calculated according to the following formula (3): s = [s1, s2, ..., s n ] T :

[0118]

[0119] In formula (3), s i represents the score of the i-th sample, where the sample refers to the 6G service sample data after vectorization processing; Represents the data representation of a sample, where i represents the i-th sample and j represents the dimension. It means that it contains data from different views, and the sample dimensions of each view are different. Here, it means that the same sample sums the values ​​of all dimensions of all views. T Represents the total dimension of the data features; It represents the operation of converting the original feature (the original feature is the vectorized data representation, which may contain positive or negative values) into a non-negative feature. The scores s of samples of the same category will be similar, while the scores of samples of different categories will have a large difference. The anchor point selection strategy is to first select the one with the largest score as the initial anchor point. As shown in formula (4):

[0120]

[0121] In order to make the anchor points fully cover the data points of multiple clusters, the embodiment of the present invention selects anchor points one by one and updates the score after selecting an anchor point, with the purpose of alternately obtaining anchor points from different clusters. Before updating the score value, the score of each sample is first normalized by dividing the largest score:

[0122]

[0123] Among them, max(s) represents the maximum value of the scores of all samples. Formula (5) means that for the i-th sample, calculate The update is achieved by normalizing the scores.

[0124] Therefore, in each round, the score of the selected sample is scaled to 1, and samples in clusters similar to the selected sample get larger values. Then, each score is updated in the following way:

[0125] s i ←s i (1-s i )(6)

[0126] Formula (6) means that s i (1-s i ) as the updated s i .

[0127] In the embodiment of the present invention, equations (5) and (6) are executed m times to obtain m anchor points one by one, where m is greater than the actual number of classes or the expected number of clusters c. In addition, since the sample with the highest score is always selected, the score of the selected sample is zero, and no anchor can be selected repeatedly.

[0128] B2, for each KPI corresponding to m anchor points, use the bipartite graph method to construct an m*n-order matrix as the anchor graph under the KPI view;

[0129] Separate anchor chart Z for each view p It is constructed by the following formula (7):

[0130]

[0131] Among them, X p represents the data matrix under the pth KPI view, that is, the original view data of each KPI in the previous text; H p represents the indicator matrix of the pth KPI view; X p H p Indicates the complete data in the p-th KPI view without the default data, that is, the view data of the p-th KPI. The view data of the p-th KPI is dimension d (p) *n is a data matrix, d (p) is the dimension of data vectorization under this KPI, and n is the total number of samples of 6G service sample data; B p Represents the anchor matrix of the p-th view, as above Figure 4 The m anchor points learned in are shown in ; f(.) represents the regularization term; 1 represents the identity matrix.

[0132] See the procedure for step B2. Figure 3 The left column in the right box understands that we get a separate anchor map for each view, which is an m*n matrix.

[0133] B3, fuse the anchor graphs under each KPI using different preset weights to construct an m*n-order matrix as the consensus anchor graph;

[0134] Based on formula (7), the optimization objective of the consensus anchor graph is shown in formula (8):

[0135]

[0136] In formula (8), α is the weight vector of the view, including the weight of each view, where α p is the weight of the p-th view; Z represents the final consensus anchor graph, and μ is a hyperparameter that balances the influence between consensus anchor graph construction and regularization, that is, the regularization parameter.

[0137] It can be seen that there are three variables in formula (8). When all variables are comprehensively considered for optimization, the optimization problem is a non-convex problem. Therefore, the optimization problem is solved by using a three-step alternating optimization iteration method. In each step, two of the three variables are set as constants to solve the remaining variable. This alternating iteration is repeated until convergence. Finally, the optimal target consensus anchor graph Z is obtained, which is an m*n-order matrix, as shown in Figure 3 The “consensus anchor chart” in the middle section of the right box.

[0138] B4, use the consensus anchor graph to perform spectral clustering based on the consensus anchor graph to obtain cluster labels of multiple clusters; each cluster represents a category of a 6G scenario.

[0139] Since the optimization goal of the entire clustering algorithm is to construct a consensus anchor graph, after solving the aforementioned optimization problem, spectral clustering is performed on the consensus anchor graph of this optimization goal to obtain cluster labels.

[0140] In this step, after the consensus anchor graph Z is obtained, spectral clustering based on Z is performed to obtain cluster labels. Specifically, U is first obtained by performing singular value decomposition on Z, and then k-means clustering is performed on U to obtain cluster labels. Among them, those skilled in the art can understand that U is a matrix obtained by performing SVD, i.e., based on singular value decomposition, which is the first step of spectral clustering. For the specific process of spectral clustering, please refer to the relevant technical understanding, and no detailed description is given here.

[0141] The above is an explanation of the incomplete multi-view clustering method. For S4, specifically, under each initial clustering parameter combination, the incomplete multi-view data set is clustered using the incomplete multi-view clustering method. The specific clustering process is the same and will not be described in detail. Then, a clustering result can be obtained for each initial clustering parameter combination.

[0142] In S4, all clustering results will be roughly screened and precisely screened in turn to determine the regularization parameter, number of anchor points, and number of clusters with the best clustering effect as the optimal clustering parameter combination, implement the clustering parameter adjustment process, and find the optimal value combination of the regularization parameter, number of anchor points, and number of clusters. As for the result of the optimal value combination, since it has been obtained before, this step only needs to be called to obtain the target clustering result corresponding to the optimal clustering parameter combination.

[0143] In an optional implementation, all clustering results are subjected to rough screening and precise screening in sequence to determine the regularization parameter, number of anchor points, and number of clusters with the best clustering effect as the optimal clustering parameter combination, including steps C1 to C2:

[0144] C1, the clustering results obtained by the test values ​​of all cluster numbers under the same regularization parameter μ and anchor point number m are used to generate an elbow coefficient graph using the elbow method; among all the elbow coefficient graphs obtained, the clustering parameter combinations that meet the shape of the elbow curve are retained to complete the rough screening;

[0145] For ease of understanding, we will still use Example 1 in the previous article for illustration, where the regularization parameter is μ = [0.001, 0.1, 1, 10, 100, 1000]; the number of anchor points is m = [k, 2*k, 3*k, 4*k, 5*k, 6*k, 7*k]; k = [6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19, 20], and clustering is performed through different combinations of the three parameters.

[0146] For Example 1, since the k value is relatively small, the elbow method can be used for coarse screening first, and some situations that conform to the clustering rules can be screened out from the 42 possible parameter conditions obtained by 6 regularization parameters and 7 anchor point numbers. The elbow method means that when the selected k value is smaller than the actual number of clusters, the elbow coefficient cost value will be greatly reduced for every increase of k by 1, that is, the clustering between each class will be reduced; when the selected k value is larger than the actual number of clusters, the change in cost value will not be so obvious for every increase of k by 1. In this way, the correct k value will appear at the turning point of the elbow coefficient graph, similar to the elbow. Specifically, the elbow method is represented by calculating the square of the clustering of each cluster sample and the cluster center. The calculation formula of the elbow coefficient is:

[0147]

[0148] Among them, D k represents the elbow coefficient under a set of regularization parameters μ, anchor point number m and cluster number k; x represents the sample, which is the 6G service sample data after vectorization processing; c i It represents the cluster center of the i-th cluster in the clustering result; dist represents the distance function.

[0149] For Example 1, we calculated an elbow coefficient graph for each of the 42 combinations of μ and m for 15 k values ​​of k = [6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19, 20]. The results are shown in Figure 4There are 42 elbow coefficient graphs in 6 rows and 7 columns. Among them, parameter1 indicates that μ takes the first value, and anchor1 indicates that m takes the first value; each row from left to right represents the case of regularization parameter μ = [0.001, 0.1, 1, 10, 100, 1000], and each column from top to bottom represents the case of anchor point number m = [k, 2*k, 3*k, 4*k, 5*k, 6*k, 7*k], and in each elbow coefficient graph, k is [6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19, 20], so the curve of each graph is fitted by 15 points. It can be clearly seen that in the first two rows of 14 figures, the elbow coefficient has a trend of decreasing and then remaining unchanged as the k value increases, which is consistent with the shape of the elbow curve, while the following figures all increase as the k value increases, indicating that under the following parameter conditions, as the k value increases, the distance between clusters becomes larger and larger, indicating that the clustering effect is not good. Therefore, the regularization parameter μ = [1, 10, 100, 1000] can be directly discarded, and μ = [0.001, 0.1] is retained, and the parameters are continued to be adjusted. Therefore, it is equivalent to discarding all clustering parameter combinations containing μ = [1, 10, 100, 1000].

[0150] C2, the silhouette coefficient method is used to generate a silhouette coefficient diagram for the clustering parameter combinations that are roughly screened and retained. The clustering parameter combinations are accurately screened by selecting the maximum value in the silhouette coefficient diagram, and the regularization parameter, number of anchor points and number of clusters with the best clustering effect are obtained as the optimal clustering parameter combination.

[0151] The silhouette coefficient is a comprehensive coefficient that measures the cohesion and separation of clusters, and the value of this coefficient ranges from [-1, 1]. In the embodiment of the present invention, all samples are processed using the silhouette coefficient method. For simplicity, a sample is used as an example to illustrate the implementation process of the silhouette coefficient method.

[0152] For any sample x i , the process of calculating the silhouette coefficient using the silhouette coefficient method includes:

[0153] ① Calculate sample x i The average distance a(i) to other samples in the same cluster is taken as the sample x i The intra-cluster dissimilarity of

[0154] The calculation method of a(i) is to calculate the sample x i The distance to every other sample in the same cluster, and then average these distances.

[0155] It can be understood that, in fact, an a(i) is calculated for each sample, and the mean a(i) of all samples in cluster C is called the cluster dissimilarity of cluster C.

[0156] ②Calculate sample x i The average distance b(ij) to all samples in each other cluster C(j) is taken as the distance between sample x i Dissimilarity with cluster C(j);

[0157] For another cluster C(j), b(ij) is calculated by calculating sample x i The distance from each sample in cluster C(j) is calculated, and the average of these distances is calculated.

[0158] ③According to sample x i The dissimilarity with all other clusters is calculated for sample x i The inter-cluster dissimilarity is b(i)=min{b i1 ,b i2 ,…,b ik};

[0159] Among them, the larger b(i), the more i The less it belongs to other clusters.

[0160] ④ According to sample x i The intra-cluster dissimilarity a(i) and inter-cluster dissimilarity b(i) of the sample x are calculated. i The silhouette coefficient is:

[0161] According to the above process, the embodiment of the present invention can obtain the silhouette coefficient of all samples and then average them to obtain the silhouette coefficient under the k value clustering. When the silhouette coefficient is closer to 1, the clustering effect is better.

[0162] For Example 1, under the parameters of μ=[0.001, 0.1]m=[k, 2*k, 3*k, 4*k, 5*k, 6*k, 7*k], the silhouette coefficient is as follows Figure 5 shown. Figure 5 The first two lines are the corresponding contour coefficient graphs for μ = 0.001, m is [k, 2*k, 3*k, 4*k, 5*k, 6*k, 7*k], and the last two lines are the corresponding contour coefficient graphs for μ = 0.1, m is [k, 2*k, 3*k, 4*k, 5*k, 6*k, 7*k]. In each contour coefficient graph, parameter1 indicates that μ takes the first value, anchor1 indicates that m takes the first value, and the curve is obtained from the points of k = [6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19, 20].

[0163] from Figure 5It can be seen that some of the graphs have breakpoints. The reason is that under this k value, a sample point is classified into one category, which will cause an uncalculated NaN value (where NaN means that the value cannot be calculated) to appear when calculating this sample point. Therefore, there is no value here, which is a breakpoint.

[0164] The maximum value of the curve of the above-mentioned contour coefficient graph of Example 1 is statistically analyzed, and it is found that the contour coefficient value is closest to 1 when μ=0.001, m=2*k, and k=11, and the contour coefficient value s=0.9968 at this time. Therefore, the best clustering effect is obtained through analysis. Therefore, the regularization parameter, number of anchor points, and number of clusters with the best clustering effect are obtained as the optimal clustering parameter combination, and the corresponding clustering result can be retrieved.

[0165] S5, perform t-sne visualization on the target clustering results to obtain a clustering effect diagram, and compare and analyze the clustering results with the original data to obtain a scene KPI feature radar chart and scene knowledge to cooperate with on-demand services.

[0166] This step is to analyze the scene features and related applications of the obtained clustering results. First, based on the cluster labels obtained from the target clustering results, combined with the original data, a t-SNE visualization of the entire scene clustering can be obtained. For the relevant content of the visualization method t-SNE, please refer to the prior art understanding and will not be explained here. For example, for the above example 1, the visualization clustering result diagram can be as follows Figure 6 As shown, 1 to 11 represent cluster labels, and there are 11 clusters in total as different classes, which were originally indicated by different colors, but here they are grayed out.

[0167] It can be seen Figure 6 In the figure, basically all the samples of all categories are very tightly clustered within the class, while the distance between classes is very far, which shows that the clustering effect is very good overall, but the inter-class clusters of the 4th and 10th classes are relatively close.

[0168] According to data analysis, Category 4 and Category 10 are basically outdoor positioning-related businesses, so many of the data are similar. However, Category 10 focuses more on positioning services in traffic detection. Its reliability and availability data are different from those of Category 4's outdoor positioning services. Therefore, they are divided into two scenarios, but they are very close.

[0169] After obtaining the visualized clustering result graph, we can further sort out the KPI data of the original business according to the clustering labels obtained by clustering, obtain some KPI features of each scenario, and make a scenario KPI feature radar chart based on the KPI features of each scenario, such as Figure 7 As shown, the eight KPI characteristics of each of the 11 scenarios can be clearly compared and presented. Figure 7In the figure, the 11 categories have different colors, and this is the grayscale effect.

[0170] For example 1, from Figure 7 It can be seen that the three KPI characteristics of reliability, mobility and positioning capability of the first type of scenario are very prominent. This type of scenario is mainly related to autonomous driving. The availability data in this type of scenario is all default and no relevant data is counted. The availability capability of this type of scenario is temporarily determined to be 0.

[0171] The third type of scenario has outstanding latency and coverage capabilities, and this type of scenario is mainly related to the periodic deterministic communication of sensors or some panels. The ninth type of scenario has outstanding availability and connection density capabilities, while all other KPI feature capabilities are second, and the overall comprehensive capabilities are the strongest. This type of scenario is mainly 6G-related services such as tactile communication, AR, and synaesthesia. The rate capability of the eleventh type of scenario is relatively outstanding, and its latency capability is also second. This type of scenario is mainly ultra-high-definition medical video or surgery-related services.

[0172] By analyzing the 11 scenarios of the clustering results, it is concluded that the first scenario is mainly about autonomous driving services (availability data is basically missing); the second scenario is mainly about extensive positioning-related services (rate and connection density-related data are basically missing); the third scenario is mainly about periodic deterministic communication-related services of sensors or some panels (connection density data is missing); the fourth scenario is mainly about outdoor 5G positioning service-related services (rate, mobility, connection density, and coverage data are missing); the fifth scenario is mainly about high-speed current differential protection services (positioning data is missing); the sixth scenario is mainly about regular communication service-related services related to railway rail transit (connection density and positioning data are missing); the seventh scenario is mainly about some positioning service-related services in smart factories (rate, connection density, and coverage data are missing); the eighth scenario is mainly about TV video-related services in rail transit (mobility, connection density, and coverage data are missing); the ninth scenario is mainly about 6G services such as tactile communication, AR, and synaesthesia; the tenth scenario is mainly about positioning services related to traffic detection (rate, mobility, connection density, and coverage data are missing); the eleventh scenario is mainly about ultra-high-definition medical video or surgery-related services (connection density is missing).

[0173] The scenario knowledge analysis based on the scenario KPI feature radar chart can be further coordinated with on-demand services.

[0174] Specifically, scene knowledge is obtained to support on-demand services, including:

[0175] 1) Analyze the scene KPI feature radar chart to obtain the KPI features of each scene as the corresponding scene knowledge;

[0176] 2) Establish a connection between the acquired scenario knowledge and the pre-established 6G knowledge graph library to obtain the correspondence between the business and the relevant on-demand service methods in the 6G knowledge graph library, so as to achieve the purpose of obtaining corresponding on-demand services based on the correspondence for any 6G business in a dynamic environment.

[0177] The embodiment of the present invention can study the results of clustering from many aspects through cluster learning and analysis, and different results can be obtained by adjusting cluster features, thresholds, etc. By analyzing the scene KPI feature radar chart of known scenes, the embodiment of the present invention can determine the KPI features of each scene as its scene knowledge. At the same time, a 6G knowledge graph library can be pre-established or obtained, which can also be called a business on-demand service method knowledge base, containing methods for how to provide services for the business corresponding to each scene. By connecting the obtained scene knowledge with the pre-established 6G knowledge graph library, the correspondence between the business and the relevant on-demand service methods in the 6G knowledge graph library can be obtained. Then, for a certain business, it can be determined according to the corresponding relationship which on-demand service method to use for service. At the same time, for new services, the embodiments of the present invention can also obtain some rough on-demand service methods for scenario services based on the clustering results. Specifically, the clustering of the embodiments of the present invention can automatically cluster according to some characteristics of the new services, and use the clustering results to obtain on-demand service methods of other known services in the class where the new services are located. Since they are of the same type as the new services, these known on-demand service methods will be applicable to the new services or can provide a methodological basis and direction for the establishment of new on-demand service methods for the new services. In this way, it can be well adapted to dynamic environments and changing scenarios.

[0178] Of course, the above Example 1 is only presented as an example to facilitate understanding of the embodiment of the present invention. For other 6G service scenarios, the above clustering analysis process is similar.

[0179] For the cluster analysis process of the embodiment of the present invention, please refer to Figure 8 and Fig. 9 understand. Figure 8 A schematic diagram of 6G service clustering in the 6G service scenario clustering analysis method provided in an embodiment of the present invention, Fig. 9 6G scenario clustering overall process diagram of an embodiment of the present invention. The embodiment of the present invention first collects 6G service sample data, performs data processing and completes data preparation. Figure 8 As shown in the "Current All Services" on the left, each circle represents a 6G service sample data. It can be seen that the number of samples in the embodiment of the present invention can be massive. Then the clustering result is obtained by using the incomplete multi-view clustering method, that is, Figure 8After clustering learning, all services are given cluster labels, that is, a large number of services are labeled with cluster labels. Then the clustering results are processed and 6G services are refined and classified according to the cluster labels, such as Figure 8 As shown on the far right, N scenarios can be obtained based on this. Finally, the scenario characteristics of the clustered scenarios are analyzed to obtain the knowledge of the relevant scenarios and establish a connection with the established 6G knowledge graph library, so that on-demand services can be better provided, thereby improving the efficiency of on-demand services. For specific data preparation, incomplete multi-view clustering, and clustering result analysis, please refer to the previous description and Fig. 9 Please understand the contents in the corresponding boxes in the figure, and no detailed explanation will be given here.

[0180] In the solution provided by the embodiment of the present invention, firstly, a number of 6G service sample data are collected, and data stratification and other processing are performed according to KPI to obtain an incomplete multi-view data set, and complete the data preparation for clustering; since the default data situation is considered under each KPI, a large amount of 6G service sample data with data default can be processed; secondly, a plurality of initial clustering parameter combinations formed by test value combinations of regularization parameters, anchor points, and cluster numbers are obtained as parameter preparation for parameter adjustment; next, for each initial clustering parameter combination, the incomplete multi-view data set is clustered using an incomplete multi-view clustering method, and all clustering results are roughly screened and precisely screened in turn, and the regularization parameter, anchor point number, and cluster number with the best clustering effect are determined as the optimal clustering parameter combination, the parameter adjustment process is completed, and the optimal clustering result is determined. Finally, the optimal clustering result is analyzed, including obtaining a visual clustering result diagram and a scene KPI feature radar diagram.

[0181] The embodiment of the present invention pioneered a scheme in the field of communication to cluster massive services using incomplete multi-view clustering to obtain 6G scenarios and analyze them. The 6G business scenario clustering analysis method can obtain a rough scenario classification for the next generation communication network, and the scenario knowledge analyzed by the result can be used as a 6G scenario KPI knowledge base to assist the next generation communication network architecture to improve its intelligence, flexibility and autonomy. At the same time, the classification of the three major scenarios of 5G is mainly obtained by industry experts to refine and classify the services according to the business characteristics. Most of the refinement and clustering processes are manually implemented by experts, and the process is very labor-intensive and resource-intensive, and it is not possible to flexibly adapt to the complex scenarios brought about by the dynamic environment in future communications. The advantage of the embodiment of the present invention is that a high-quality and efficient clustering of large-scale massive services can be achieved through the incomplete multi-view clustering method, and the scene-related knowledge can be obtained by analyzing the clustering results to improve the efficiency of realizing "free and free" on-demand services in the next generation mobile communication network and reduce the waste of resources. And it has certain guiding significance for the construction of the 6G full-scenario knowledge graph.

[0182] In summary, compared with the prior art, the embodiments of the present invention have the following beneficial effects:

[0183] 1. Ability to perform clustering scenario analysis on massive 6G service data with high efficiency and short time;

[0184] 2. Ability to handle clustering with a large amount of default data;

[0185] 3. The final result of cluster analysis can be used as a kind of scenario knowledge to cooperate with on-demand services to cope with dynamically changing environments, and can meet the scenario clustering analysis needs of known and unknown businesses.

[0186] The above description is only a preferred embodiment of the present invention and is not intended to limit the protection scope of the present invention. Any modification, equivalent replacement, improvement, etc. made within the spirit and principle of the present invention are included in the protection scope of the present invention.

Claims

1. A 6G service scenario clustering analysis method, characterized in that: include: Collect a number of 6G service sample data and determine multiple KPIs that describe the service perspective; Processing the 6G service sample data to obtain an incomplete multi-view data set; wherein the incomplete multi-view data set includes view data of multiple KPIs, and the view data of each KPI includes corresponding data of all 6G service sample data under the KPI, and default data is used to represent the situation where there is no corresponding data; Determine an initial value of the number of clusters, and expand the range based on the initial value of the number of clusters to determine multiple test values ​​of the number of clusters; Determine a first number of regularization parameters and a second number of anchor points, and each number of anchor points is determined according to different multiples of the number of clusters, to obtain a plurality of initial clustering parameter combinations formed by test value combinations of the regularization parameters, the number of anchor points, and the number of clusters; Under each initial clustering parameter combination, the incomplete multi-view clustering method is used to cluster the incomplete multi-view data set, and all clustering results are roughly screened and accurately screened in turn, and the regularization parameter, number of anchor points, and number of clusters with the best clustering effect are determined as the optimal clustering parameter combination, and the target clustering result corresponding to the optimal clustering parameter combination is retrieved; wherein each cluster label in each clustering result corresponds to a 6G scenario; The target clustering results are visualized by t-sne to obtain a clustering effect diagram, and the clustering results are compared and analyzed with the original data to obtain a scene KPI feature radar diagram and scene knowledge to cooperate with on-demand services.

2. The 6G service scenario clustering analysis method according to claim 1, characterized in that: The multiple KPIs include: rate, latency, reliability, mobility, connection density, coverage, positioning accuracy and availability.

3. The 6G service scenario clustering analysis method according to claim 2, characterized in that: The processing of the 6G service sample data to obtain an incomplete multi-view data set includes: After vectorizing the 6G service sample data, the data is layered according to multiple KPIs; Generate an index matrix with data values ​​according to the data under each KPI, which includes the serial number of the 6G service sample data with corresponding data under the KPI; According to the index matrix under the KPI and the vectorized 6G service sample data, a data vectorization matrix is ​​obtained as the view data of the KPI, and the view data of all KPIs constitute an incomplete multi-view data set.

4. The 6G service scenario clustering analysis method according to claim 3 is characterized in that: The determining of the initial value of the number of clusters, and expanding the range based on the initial value of the number of clusters to determine multiple test values ​​of the number of clusters, includes: The initial value of the number of clusters is determined by the brainstorming method. Where n is the total number of samples of 6G service sample data; The initial value k0 is used as a reference to expand to both sides, and k0 values ​​are determined as multiple test values ​​of the number of clusters.

5. The 6G service scenario clustering analysis method according to claim 4, characterized in that: The determining of a first number of regularization parameters and a second number of anchor points, wherein each anchor point number is determined according to a different multiple of the number of clusters, and a plurality of initial clustering parameter combinations formed by test value combinations of regularization parameters, anchor point numbers, and cluster numbers are obtained, includes: Determining a first number of regularization parameters μ includes: 0.001, 0.1, 1, 10, 100 and 1000; Determining the second number of anchor points m includes: k, 2k, 3k, 4k, 5k, 6k and 7k; wherein k represents the number of clusters; A combination of any regularization parameter μ, the number of anchor points m, and the number of clusters k taking any value from multiple test values ​​is taken as an initial clustering parameter combination, and multiple initial clustering parameter combinations are obtained based on all combinations of the regularization parameter μ, the number of anchor points m, and the number of clusters k.

6. The 6G service scenario clustering analysis method according to claim 5, characterized in that: The step of clustering the incomplete multi-view data set using an incomplete multi-view clustering method under each initial clustering parameter combination includes: Under each initial clustering parameter combination, m anchor points are learned from the view data of each KPI through the anchor point selection strategy of direct alternating sampling; For each KPI corresponding to m anchor points, use the bipartite graph method to construct an m*n-order matrix as the anchor graph under the KPI view; The anchor graphs under each KPI are fused using different preset weights to construct an m*n-order matrix as the consensus anchor graph; The consensus anchor graph is used to perform spectral clustering based on the consensus anchor graph to obtain cluster labels of multiple clusters, wherein each cluster represents a category of a 6G scenario.

7. The 6G service scenario clustering analysis method according to claim 6, characterized in that: The method of performing rough screening and precise screening on all clustering results in turn to determine the regularization parameter, number of anchor points, and number of clusters with the best clustering effect as the optimal clustering parameter combination includes: The clustering results obtained from the test values ​​of all cluster numbers under the same regularization parameter μ and anchor point number m are used to generate an elbow coefficient graph using the elbow method; among all the elbow coefficient graphs obtained, the clustering parameter combinations that meet the shape of the elbow curve are retained to complete the rough screening; The silhouette coefficient method is used to generate a silhouette coefficient diagram for the clustering parameter combinations that are roughly screened and retained. The clustering parameter combinations are accurately screened by selecting the maximum value in the silhouette coefficient diagram, and the regularization parameter, number of anchor points and number of clusters with the best clustering effect are obtained as the optimal clustering parameter combination.

8. The 6G service scenario clustering analysis method according to claim 7, characterized in that: In the elbow coefficient diagram, the calculation formula of the elbow coefficient is: Among them, D k represents the elbow coefficient under a set of regularization parameters μ, anchor point number m and cluster number k; x represents the sample, which is the 6G service sample data after vectorization processing; c i It represents the cluster center of the i-th cluster in the clustering result; dist represents the distance function.

9. The 6G service scenario clustering analysis method according to claim 8, characterized in that: For any sample x i , the process of calculating the silhouette coefficient using the silhouette coefficient method includes: Calculate sample x i The average distance a(i) to other samples in the same cluster is taken as the sample x i The intra-cluster dissimilarity of Calculate sample x i The average distance b(ij) to all samples in each other cluster C(j) is taken as the distance between sample x i Dissimilarity with cluster C(j); According to the sample x i The dissimilarity with all other clusters is calculated for sample x i The inter-cluster dissimilarity is b(i)=min{b i1 ,b i2 ,…,b ik }; According to the sample x i The intra-cluster dissimilarity a(i) and inter-cluster dissimilarity b(i) of the sample x are calculated. i The silhouette coefficient is:

10. The 6G service scenario clustering analysis method according to claim 1, characterized in that: Gain contextual knowledge to support on-demand services, including: Analyze the scene KPI feature radar chart to obtain the KPI features of each scene as corresponding scene knowledge; The acquired scenario knowledge is connected to the pre-established 6G knowledge graph library to obtain the correspondence between the business and the relevant on-demand service method in the 6G knowledge graph library, so as to achieve the purpose of any 6G business in a dynamic environment obtaining the corresponding on-demand service based on the correspondence.

Images