5G user identification network model establishment method, device and storage medium

Abstract

The present application provides a method, device and storage medium for establishing a network model for 5G user identification. The method includes: obtaining historical data of users within a predetermined historical time period; constructing a user knowledge graph based on the historical data, the user knowledge graph including relationship triple data between users and user attribute triple data; based on the relationship triple data between users, obtaining the user's network position vector using a random walk algorithm; inputting the user attribute triple data and the network position vector into a graph attention network model for training, and generating a graph attention network model for identifying 5G users. The method of the present application conducts a comprehensive analysis of users through user attribute triple data and relationship triple data between users, and generates a graph attention network model that can identify 5G users. Therefore, the graph attention network model is used to identify whether a user is a 5G user, thereby improving the recognition accuracy.

Description

Technical Field

[0001] The present application relates to the field of communication technology, and in particular to a method, device and storage medium for establishing a network model for 5G user identification. Background Art

[0002] The fifth generation mobile communication technology (5G) has the characteristics of high speed and low latency. Accelerating the growth of 5G users through informatization and intelligence can not only improve the sustainable development of cities, but also promote the rapid construction of smart cities. In order to promote the transformation of 4G users to 5G users, it is necessary to identify potential 5G users.

[0003] The existing technology identifies whether a user is a potential 5G user based on the traffic usage of the user, but the identification accuracy is low. Summary of the invention

[0004] The present application provides a network model establishment method, device and storage medium for 5G user identification, which are used to solve the problem that the prior art identifies whether a user is a potential 5G user based on the traffic used by the user, but the identification accuracy is low.

[0005] In a first aspect, the present application provides a method for establishing a network model for 5G user identification, comprising:

[0006] Obtain the user's historical data within a predetermined historical time period;

[0007] Build a user knowledge graph based on historical data. The user knowledge graph contains triple data of relationships between users and triple data of user attributes.

[0008] Based on the triplet data of the relationship between users, the random walk algorithm is used to obtain the user's network position vector;

[0009] The user attribute triplet data and the network position vector are input into the graph attention network model for training to generate a graph attention network model for identifying 5G users.

[0010] In a possible implementation, a user knowledge graph is constructed based on historical data, including:

[0011] Generate a wide table of feature sets based on historical data;

[0012] Knowledge extraction is performed on the wide table of feature sets to obtain triple data, and a user knowledge graph is constructed based on the triple data.

[0013] In a possible implementation, the historical data includes: user basic information data, user business information data and user interaction behavior data; generating a feature set wide table based on the historical data includes:

[0014] Calculate the data missing rate of user characteristics based on the user basic information data, user business information data and user interaction behavior data;

[0015] The historical data is screened according to the missing rate of user features and the number of user features. The user features include numerical user features and character user features.

[0016] Perform feature superposition processing on numerical user features and perform quantization coding processing on character user features;

[0017] Generate a feature set wide table based on historical data after feature superposition processing and quantization coding processing.

[0018] In a possible implementation, the user attribute triplet data and the network position vector are input into the graph attention network model for training to generate a graph attention network model for identifying 5G users, including:

[0019] Determine neighbor users of a user based on relationship triplet data between users;

[0020] Construct user attribute features based on user attribute triple data;

[0021] Concatenate user attribute features with network position vectors to generate initial semantic representation;

[0022] Input the initial semantic representation into the fully connected layer of the graph attention network model and output an embedded semantic representation, which includes the embedded semantic representation of the user and the embedded semantic representation of neighboring users;

[0023] Input the user's embedded semantic representation and the neighbor user's embedded semantic representation into the graph attention layer of the graph attention network model, and output the user's semantic representation and the neighbor user's semantic representation based on the graph attention;

[0024] The semantic representation of the user based on graph attention and the semantic representation of the neighboring users are input into the output layer of the graph attention network model, and the probability that the user is a 5G user is output.

[0025] In a possible implementation, the semantic representation of the user based on graph attention and the semantic representation of the neighboring user are input into the output layer of the graph attention network model, and the probability of the user being a 5G user is output, including:

[0026] Integrate the user's semantic representation output by each graph attention layer in the graph attention network model to generate the user's final semantic representation;

[0027] The final semantic representation of the user is converted into the probability that the user is a 5G user through the activation function.

[0028] In a possible implementation, the historical data also includes user label data; after the final semantic representation of the user is converted into the probability that the user is a 5G user through an activation function, it also includes:

[0029] The loss function is used to calculate the loss between the probability that a user is a 5G user and the user label data, and the graph attention network model is optimized based on the loss.

[0030] In a possible implementation, the user attribute triplet data and the network position vector are input into the graph attention network model for training. Before the graph attention network model for identifying 5G users is generated, the following steps are also included:

[0031] Divide the triple data and user label data into training set, validation set and test set;

[0032] Use the validation set to optimize the parameters of the graph attention network model;

[0033] The test set is used to test and evaluate the graph attention network model.

[0034] In a second aspect, the present application provides a network model establishment device for 5G user identification, including:

[0035] An acquisition module is used to acquire the historical data of the user within a predetermined historical time period;

[0036] A construction module is used to construct a user knowledge graph based on historical data. The user knowledge graph includes triple data of relationships between users and triple data of user attributes.

[0037] The acquisition module is further used to acquire the network position vector of the user by using a random walk algorithm based on the relationship triplet data between users;

[0038] The training module is used to input the user attribute triplet data and the network position vector into the graph attention network model for training, so as to generate a graph attention network model for identifying 5G users.

[0039] In a third aspect, the present application provides a network model establishment device for 5G user identification, including: a processor and a memory, wherein code is stored in the memory, and the processor runs the code stored in the memory to execute a network model establishment method for 5G user identification as described in any one of the first aspects.

[0040] In a fourth aspect, the present application provides a computer-readable storage medium, which stores computer execution instructions. When the computer execution instructions are executed by a processor, they are used to implement a network model establishment method for 5G user identification as described in any one of the first aspects.

[0041] The present application provides a method, device and storage medium for establishing a network model for 5G user identification. The method includes: obtaining historical data of users within a predetermined historical time period, and constructing a user knowledge graph based on the historical data. The user knowledge graph may include relationship triple data between users and user attribute triple data. According to the relationship triple data between users, a random walk algorithm is used to obtain the network position vector of the user. The user attribute triple data and the network position vector are input into a graph attention network model for training to generate a graph attention network model that can identify 5G users. The method provided in the present application comprehensively analyzes the user through user attribute triple data that describes the user's own attributes and relationship triple data between users that describes the relationship between other users and the user, and generates a graph attention network model that can identify 5G users. Therefore, the graph attention network model is used to identify whether the user is a 5G user, thereby improving the recognition accuracy. BRIEF DESCRIPTION OF THE DRAWINGS

[0042] The accompanying drawings, which are incorporated in and constitute a part of this specification, illustrate embodiments consistent with the present application and, together with the description, serve to explain the principles of the present application.

[0043] Figure 1 A schematic diagram of establishing a network model for 5G user identification provided in an embodiment of the present application;

[0044] Figure 2 A flow chart of a method for establishing a network model for 5G user identification provided in an embodiment of the present application;

[0045] Figure 3 A flow chart of a method for constructing a user knowledge graph provided in an embodiment of the present application;

[0046] Figure 4 A schematic diagram of a user knowledge graph provided in an embodiment of the present application;

[0047] Figure 5 A flow chart of a method for generating a graph attention network model provided in an embodiment of the present application;

[0048] Figure 6 A schematic diagram of a network model establishment device for 5G user identification provided in an embodiment of the present application Figure 1 ;

[0049] Figure 7A schematic diagram of a network model establishment device for 5G user identification provided in an embodiment of the present application Figure 2 .

[0050] The above drawings have shown clear embodiments of the present application, which will be described in more detail later. These drawings and text descriptions are not intended to limit the scope of the present application in any way, but to illustrate the concept of the present application to those skilled in the art by referring to specific embodiments. DETAILED DESCRIPTION

[0051] Exemplary embodiments will be described in detail herein, examples of which are shown in the accompanying drawings. When the following description refers to the drawings, unless otherwise indicated, the same numbers in different drawings represent the same or similar elements. The implementations described in the following exemplary embodiments do not represent all implementations consistent with the present application. Instead, they are merely examples of devices and methods consistent with some aspects of the present application as detailed in the appended claims.

[0052] As social transformation accelerates, in order to more accurately control user needs and improve the marketing success rate of enterprises, it is imperative to build smart cities. The high-speed transmission capacity of 5G networks can support the construction of smart cities. Accelerating the growth of 5G users through information and intelligent means can not only improve the sustainable development of cities, but also promote the rapid construction of smart cities. However, how to tap into potential 5G users and promote the transformation of 4G users to 5G users is the key research and urgent problem to be solved in realizing 5G smart cities. At present, the traditional method is to identify whether a user is a potential 5G user based on the traffic used by the user, and the accuracy rate is low.

[0053] In order to improve the accuracy of identifying 5G users, this application obtains historical data of users within a predetermined historical time period. The historical data includes relevant data of multiple aspects of users, such as basic information data of users, business information data of users, and interactive behavior data of users. A user knowledge graph is constructed based on historical data. The user knowledge graph includes relationship triple data between users and user attribute triple data. Based on the relationship triple data between users, the network position vector of the user is obtained using a random walk algorithm. The user attribute triple data and the network position vector are input into the graph attention network model for training to generate a graph attention network model that can identify 5G users. The method provided in this application comprehensively analyzes the user by using the user attribute triple data that describes the user's own attributes and the relationship triple data between users that describes the relationship between other users and the user. Therefore, the graph attention network model is used to identify whether the user is a 5G user, thereby improving the recognition accuracy.

[0054] Figure 1A schematic diagram of establishing a network model for 5G user identification provided in an embodiment of the present application. Figure 1 As shown, the acquired user basic information data, user business information data and user interaction behavior data are processed. According to the processed data, the entities, relationships and attributes of the user knowledge graph architecture are defined, and knowledge extraction is performed to obtain triple data. The user knowledge graph can be constructed based on the triple data. The user knowledge graph may include triple data of the relationship between users, i.e., user-user-relationship, and user attribute triple data, i.e., user-attribute-attribute value. The initial semantic representation of the user is generated based on the triple data of the relationship between users and the triple data of user attributes. The initial semantic representation of the user may include the initial semantic representation of user A and the initial semantic representation of user A's neighbor users. The dimension of the initial semantic representation of the user is expanded through the fully connected layer to generate the embedded semantic representation of the user. Similarly, the embedded semantic representation of the user includes the embedded semantic representation of user A and the embedded semantic representation of neighbor users. The embedded semantic representation of user A and the embedded semantic representation of neighbor users are input into the graph attention layer for training, and the output layer can output the probability that user A is a 5G user.

[0055] The technical solution of the present application and how the technical solution of the present application solves the above-mentioned technical problems are described in detail below with specific embodiments. The following specific embodiments can be combined with each other, and the same or similar concepts or processes may not be repeated in some embodiments. The embodiments of the present application will be described below in conjunction with the accompanying drawings.

[0056] Figure 2 A flowchart of a method for establishing a network model for 5G user identification provided in an embodiment of the present application. The method of this embodiment can be executed by a device for establishing a network model for 5G user identification, and can be implemented by hardware, software, or a combination of hardware and software. Figure 2 As shown, the method provided by the present application may include:

[0057] S201: Obtain historical data of the user within a predetermined historical time period.

[0058] The historical data is data related to the user in the historical time period, and may include: user basic information data, user business information data, user interaction behavior data, user tag data, etc. In one implementation scenario, the historical time period may be three months.

[0059] Among them, the user's basic information data is the user's identity information in multiple dimensions, such as age, gender, name, mobile phone number and other characteristic information.

[0060] User business information data includes user consumption data, usage data of different applications, package data, etc. User business information data is dynamic data and may change continuously according to user preferences and click volume.

[0061] User interaction behavior data refers to data related to the interaction between users and other users, and may include the frequency of calls between users and other users on a monthly, ten-day, weekly, and daily basis.

[0062] User tag data refers to data indicating whether the user has a 5G package.

[0063] S202: Construct a user knowledge graph based on historical data, where the user knowledge graph includes triple data of relationships between users and triple data of user attributes.

[0064] In one implementation scenario, a feature set wide table can be generated based on historical data, knowledge extraction is performed on the feature set wide table to obtain triple data, and a user knowledge graph is constructed based on the triple data.

[0065] Since the historical data contains user features in multiple dimensions, the features in multiple dimensions can be sorted and stored in a table. In this case, the table can be called a feature set wide table.

[0066] Knowledge extraction is the process of extracting knowledge contained in information sources through processes such as identification, understanding, screening, and induction, and storing it to form a knowledge meta-repository. Knowledge extraction can include entity extraction, relationship extraction, and attribute extraction.

[0067] A triplet is a set of (x, y, z). For example, the relationship triplet data between users can be (user A, user B, the correlation index between user A and user B), which can indicate the correlation between users. The user attribute triplet data can be (user A, user_age, age), where user_age is a field indicating the user's age.

[0068] S203: Based on the relationship triplet data between users, a random walk algorithm is used to obtain the network position vector of the user.

[0069] The random walk algorithm is a network-based algorithm that starts from a node, randomly selects an adjacent point of the node, and then starts from the adjacent point to the next adjacent point, repeating this step. Therefore, a path starting from each node can be obtained through the random walk algorithm, that is, the structural information of the node. It can be understood that in order to obtain the location information of each user, it is necessary to use the random walk algorithm to obtain the network location vector of each user.

[0070] In one implementation scenario, a DeepWalk algorithm may be used to obtain a user's network location vector.

[0071] S204: Input the user attribute triplet data and the network position vector into the graph attention network model for training, and generate a graph attention network model for identifying 5G users.

[0072] The graph attention network model introduces the attention mechanism into the graph neural network, where the attention mechanism allows a neural network to focus only on the information required for task learning and select specific inputs. Graph neural network refers to the general term for algorithms that use neural networks to learn graph structured data, extract and discover features and patterns in graph structured data, and meet the needs of graph learning tasks such as clustering, classification, prediction, segmentation, and generation.

[0073] Introducing the attention mechanism into graph neural networks can enable the neural network to focus on task-related nodes, which is conducive to improving the effectiveness and accuracy of model training.

[0074] In one implementation scenario, before inputting the user attribute triple data and the network position vector into the graph attention network model for training, the triple data and the user label data can also be divided into a training set, a validation set, and a test set; the validation set is used to optimize the parameters of the graph attention network model; and the test set is used to test and evaluate the graph attention network model. When dividing the triple data and the user label data, if the triple data and the user label data are 3 months of data, the data of the first two months can be used as the training set, 50% of the data of the last month can be used as the validation set, and 50% can be used as the test set.

[0075] The embodiment of the present application provides a method for establishing a network model for 5G user identification, obtains historical data of users within a predetermined historical time period, and constructs a user knowledge graph based on the historical data. The user knowledge graph may include relationship triple data between users and user attribute triple data. Based on the relationship triple data between users, a random walk algorithm is used to obtain the user's network position vector. The user attribute triple data and the network position vector are input into the graph attention network model for training to generate a graph attention network model for identifying 5G users. The method provided in the present application not only takes into account the user attribute triple data describing the user's own attributes, but also uses the relationship triple data between users that describe the relationship between other users and the user, that is, taking into account the influence of other users on the user, it is possible to conduct a comprehensive analysis of the user. Therefore, the graph attention network model obtained according to the user attribute triple data and the network position vector is used to identify whether the user is a 5G user, thereby improving the recognition accuracy.

[0076] Based on the above embodiment, the process of constructing a user knowledge graph based on historical data is introduced in detail. The historical data may include: user basic information data, user business information data and user interaction behavior data.

[0077] Figure 3 A flow chart of a method for constructing a user knowledge graph provided in an embodiment of the present application, such as Figure 3 As shown, the method provided by this application is as follows:

[0078] S301: Calculate the data missing rate of user characteristics based on the user basic information data, the user business information data and the user interaction behavior data.

[0079] It should be noted that the user basic information data, user business information data and user interaction behavior data stored in the database are stored in the format of fields. Tables 1 to 3 respectively introduce the fields contained in the user basic information data, user business information data and user interaction behavior data, as well as the meanings represented by the fields, i.e., user characteristics:

[0080] Table 1: User basic information data table

[0081] Field Name type Indicator Explanation msisdn string User's mobile phone number user_sex string gender user_age int64 User age type_vip int64 VIP Level

[0082] Table 2: User business information data table

[0083]

[0084]

[0085] Table 3: User u0 and any other i-th user interaction behavior data table

[0086] <![CDATA[Current user u0]]> type Indicator Explanation <![CDATA[Msisdn_u i ]]> string Use each user's mobile phone number instead of user ID Com_mouth int64 Number of contact months between two users by month Com_Tendays int64 The number of contacts between two users per decade Com_week int64 Number of weeks of contact between two users by week Per_day int64 Number of days the two users talk Com_time int64 Total call duration between two users

[0087] In addition to user basic information data, user business information data and user interaction behavior data, historical data also includes: user tag data, as shown in Table 4:

[0088] Table 4: User tag data table

[0089] Field Name type Indicator Explanation msisdn string User's mobile phone number month int64 month if_5g_dinner string Is it a 5g package?

[0090] S302: Filtering historical data according to the data missing rate of user features and the number of user features, where the user features include numerical user features and character user features.

[0091] Since there are a lot of missing data on certain features of user basic information data, user business information data and user interaction behavior data, and the correlation with user tag data is weak, in order to improve the accuracy of identifying 5G users, user basic information data, user business information data and user interaction behavior data can be screened.

[0092] For example, user_dinner represents a package type, and its data missing rate is 40%, and if_5g_dinner is judged based on the package type. Therefore, if_5g_dinner covers the information of user_dinner, and the data corresponding to user_dinner can be deleted.

[0093] days_5g indicates the number of days for 5g access to the Internet. The data missing rate of this field is 97.5%. The Pearson correlation coefficient shows that the correlation coefficient between days_5g and if_5g_dinner is less than 0.5, indicating that the correlation between days_5g and if_5g_dinner is small. At this time, the data corresponding to days_5g can be deleted.

[0094] In another implementation scenario, since filtering historical data based on the missing rate of user features will reduce the number of user features, historical data can also be filtered based on the number of user features. For example, Tables 1 to 3 show 41 user features. If the number of missing user features is greater than 7, the user corresponding to the user feature is deleted.

[0095] In one implementation scenario, the original historical data contains 1 million data points, and after screening, 900,000 data points remain. Deleting data with a high missing rate and data with many missing user features can improve the accuracy of the data, which in turn helps to improve the accuracy of identifying 5G users.

[0096] S303: performing feature superposition processing on the numerical user features and performing quantization coding processing on the character user features.

[0097] The user features included in the historical data include numerical user features and character user features. Among them, the numerical user features can be user features with a time field, such as 19 user features including the time field such as im_times and reading_times in Table 2.

[0098] In one implementation scenario, feature superposition processing is performed on numerical user features. First, the all_time feature can be constructed to calculate the sum of the values ​​of the 19 user features with the time field in Table 2. Let the i-th user feature with the time field be represented as time i, then the all_time feature is expressed as:

[0099]

[0100] In the above formula, time i is the i-th user feature with the time field, and all_time is the sum of the values ​​of the user features with the time field.

[0101] Before calculating the all_time feature, since the missing rate of these user feature data with the time field is 27.5%, the median can be used to fill the missing data, and its distribution is approximately skewed.

[0102] Since the data missing rate of fields such as bill, talklen, flow_2g, flow_3g, and flow_4g is 69%, if the missing values ​​are not filled, it will have a great impact on the results of identifying 5G users. Therefore, the median can be used to fill the missing values ​​of the above fields, and its distribution is approximately skewed.

[0103] The user_age field indicates the user's age. If there are missing values, the mean value can be used to fill them. After filling, the data distribution is approximately normal.

[0104] By performing feature superposition processing on numerical user features and filling in missing values, it is helpful to improve the accuracy of the data.

[0105] In one implementation scenario, since the characteristic of character data is that data can be enumerated but cannot be recognized by the machine, it cannot be used as input for machine learning algorithms. Therefore, it is necessary to quantize and encode the character data to construct a new feature representation.

[0106] Character data can include user features corresponding to fields such as fact_name, model_name, user_sex, type_vip, and night_cell_id.

[0107] For example, for the fact_name field that represents the brand type of mobile phones, first build a mobile phone brand index table as follows:

[0108] fact_tabel={fact0:0, fact1:1,...,fact i :i,......,fact n :n}

[0109] Among them, fact_tabel represents the mobile phone brand index table, fact iIndicates the name of the i-th mobile phone brand, i is a fact i The index number is n, where n is the total number of mobile phone brands. By querying the mobile phone brand keywords, the index number is located as the quantitative coding representation of the fact_name field, that is, the mobile phone brands in the fact_name field are converted into numerical data. Therefore, the fact_name field is processed by quantitative coding and converted into the D_fact_name feature, where the quantitative coding representation of the i-th mobile phone brand is D_fact i , as follows:

[0110] D_fact i =fact_table[fact i ]

[0111] In the above formula, D_fact i is the quantized encoding representation of the i-th mobile phone brand, fact_tabel represents the mobile phone brand index table, fact i Represents the name of the i-th mobile phone brand.

[0112] For model_name, which indicates the terminal model name, each mobile phone manufacturer has multiple different mobile phone terminal models. For example, for the same mobile phone brand A, it can include multiple mobile phone terminal models such as A1, A2, and A3. Therefore, we first construct a mobile phone terminal model name index table as shown below:

[0113] model_tabel={model0:0,model1:1,...,model i :i,......,model k :k}

[0114] Among them, model_tabel represents the index table of mobile terminal model names, model i Indicates the name of the i-th mobile terminal model. i can be used as model i There are k mobile terminal models in total. By querying the mobile terminal model keyword, the index number is located as the quantitative coding representation of the model_name field, that is, the mobile terminal model in the model_name field is converted into numerical data. Therefore, the model_name field can be converted into D_model_name features through quantitative coding, where the quantitative coding representation of the i-th mobile terminal model is D_model i , as shown below:

[0115] D_model i =model_table[model i ]

[0116] In the above formula, D_model i is the quantized encoding representation of the i-th mobile terminal model, model_tabel represents the mobile terminal model name index table, model i Indicates the name of the i-th mobile terminal model.

[0117] It can be understood that for the type_vip field representing the user's VIP level, a user level index table is constructed based on the user level, and the index number is located as the quantitative coding representation of the type_vip field in the manner of quantitative coding of the above-mentioned fact_name field, and the user VIP level in the type_vip field is converted into numerical data, where missing values ​​can be filled with the mode.

[0118] For the user_sex field indicating the user's gender, in one implementation scenario, the case where the field value is "male" can be represented by 0, and the case where the field value is "female" can be represented by 1, and the missing values ​​can be filled with the mode.

[0119] For the night_cell_id field indicating the cell where the user resides, in one implementation scenario, the cell IDs with the same first two digits and the same position of the night_cell_id field can be classified into the same cell category according to the characteristics of the data in the night_cell_id field. Finally, the cell categories are summarized to construct a cell category index table, as shown below:

[0120] night_tabel={night0:0, night1:1,..., night i :i,......,night m :m}

[0121] In the above formula, night_tabel represents the cell category index table, night i Indicates the name of the i-th permanent community, i means night i The index number is the total number of cells. By querying the keywords of the resident cells, the positioning index number is used as the quantized coding representation of the night_cell_id field, and the resident cells in the night_cell_id field are converted into numerical data. Therefore, the night_cell_id field is converted into the D_night_cell_id feature through quantized coding, where the quantized coding representation of the i-th resident cell name is D_night i , as shown below:

[0122] D_night i =night_tabel[nighti ]

[0123] In the above formula, D_night i is the quantized code representation of the name of the i-th resident cell, night_table is the cell category index table, night i Indicates the name of the i-th permanent cell.

[0124] S304: Generate a feature set wide table based on the historical data after feature superposition processing and quantization coding processing.

[0125] In the historical data, in addition to the user features that have been processed by feature superposition and quantization coding, there are also user features that do not need feature superposition and quantization coding.

[0126] It should be noted that the original values ​​of user features that have not been processed by feature superposition and quantization coding can be used. For example, the if_5g_dinner field indicates whether it is a 5G package. In the if_5g_dinner field, 1 indicates a 5G package and 0 indicates a non-5G package. There is no need for feature superposition and quantization coding.

[0127] S305: Perform knowledge extraction on the feature set wide table to obtain triple data, and construct a user knowledge graph based on the triple data.

[0128] In one implementation scenario, the entities, attributes and relationships of the user knowledge graph architecture are first defined based on the feature set wide table. Among them, the entity is each user. Attributes are representations of the entity user's own attribute characteristics and behavior characteristics data, such as user gender, age, terminal model, and various traffic consumption information. Relationships refer to the connection between entities, which can be defined by user interaction behavior data. This application connects entity users based on the association index calculated by call records. The association index between users can be calculated using the following formula:

[0129]

[0130] In the above formula, Com(u i ,u j ) for two users u i and u j The correlation index of two users u i and u j The total call duration Com_time is based on the result of normalizing the total call duration of all users to 0-1. a1, a2, a3, and a4 are the weights of each item, which can be set to 40, 32, 16, 4, and M. y , T y , W y , Dy are the number of contacts between two users in the statistical period, namely, the number of contacts per month, the number of contacts per ten days, the number of contacts per week, and the number of days. x 、T x , W x , D x In one implementation scenario, this application uses data from March, April, and May, and selects the first two months of data as the training set for training. Therefore, the statistical period is 2 months, and M x 、T x , W x , D x They are 2, 6, 9, and 61 respectively.

[0131] Knowledge extraction is performed on the feature set wide table to obtain triple data, namely, triple data between users and triple data of user attributes. Taking user u0 as an example, the triple data may include:

[0132] <User u0, User u1, Com(u0, u1)>

[0133] <User u0, User u2, Com(u0, u2)>

[0134] <user u0, user_sex, 1>

[0135] <user u0, user_age, 23>

[0136] <user u0, type_vip, 1>

[0137] <user u0, bill, 1>

[0138] <user u0, talklen, 1>

[0139] <user u0, flow_2g, 0>

[0140] <user u0, flow_3g, 0.57>

[0141] <user u0, flow_4g, 911.34>

[0142] <user u0, fact_name, D_fact1>

[0143] <user u0, model_name, D_model1>

[0144] <user u0, night_cell_id, D_night1>

[0145] <user u0, time, 26952>

[0146] <user u0, month, 202003>

[0147] <user u0, days_5g, NULL>

[0148] <user u0,if_5g_dinner,1>

[0149] Among them, D_fact1 represents a specific mobile phone brand, D_model1 represents a specific mobile phone terminal model, and D_night1 represents a cell with a specific number.

[0150] Figure 4 Schematic diagram of user knowledge graph. Figure 4 The relationship between the four users and some of the attribute characteristics of each user are shown. By constructing a user knowledge graph and sorting out various types of information in the user communication process and the user's own information, the relationship between users can be better explored, and the user's detailed characterization can be achieved, which can improve the accuracy of identifying potential 5G users. For example, if the correlation index of two users is high, it can be judged that the two are closely related. If one user is a 5G user, the probability of the other user becoming a 5G user within a certain period of time will be higher; secondly, many users aged 26 are 5G users, so the probability of other users aged 26 becoming 5G users will also be higher.

[0151] The embodiment of the present application provides a method for constructing a knowledge graph, which calculates the data missing rate of user features based on the user features contained in the user basic information data, user business information data, and user interaction behavior data. The historical data is screened according to the data missing rate of the user features and the number of user features. Since the user features include numerical user features and character user features, the numerical user features are subjected to feature superposition processing, and the character user features are subjected to quantization coding processing. A feature set width table is generated based on the historical data after feature superposition processing and quantization coding processing, and knowledge is extracted from the feature set width table to obtain triple data, and a user knowledge graph is constructed based on the triple data. The method provided in the present application can mine the relationship between users by constructing a user knowledge graph, so that user-related information is more accurate, which is conducive to improving the accuracy of identifying 5G users.

[0152] Based on the above embodiments, a specific embodiment is provided below to describe in detail the process of generating a graph attention network model for identifying 5G users.

[0153] Figure 5 A flowchart of a method for generating a graph attention network model provided in an embodiment of the present application is as follows: Figure 5 As shown, the method provided by this application is as follows:

[0154] S501: Determine neighbor users of a user based on relationship triplet data between users.

[0155] The relationship triplet data between users contains the association index between users. Due to the existence of takeaway calls, fraud or wrong number dialing, this application obtains the neighbor users of each user based on the association index between users. i , by calculating the association index of the user with all other users, select the top N users with the largest association index as the user's neighbor users N i , this application can choose TOPN=10.

[0156] S502: Constructing user attribute features based on user attribute triplet data.

[0157] The attributes contained in the user attribute triple data: user gender, user age, user VIP level, user bill, user call duration, user 2g / 3g / 4g traffic usage, user mobile phone brand type, terminal model, user resident community, number of times the user browses each APP, number of days the user uses 5G, user tag data, etc. are spliced ​​to construct a 14-dimensional vector to represent the user attribute characteristics of the user. The user attribute characteristics of the i-th user are as follows:

[0158]

[0159] Among them, m q represents the qth user attribute feature, is the user attribute feature of the i-th user, and a represents the user attribute feature.

[0160] It should be noted that if the user is a new user, it is impossible to determine whether the user is a 5G user. At this time, the user tag data can be set to a value other than 0 and 1, where 0 indicates that the user is not a 5G package, and 1 indicates that the user is a 5G package.

[0161] S503: Concatenate the user attribute features with the network position vector to generate an initial semantic representation.

[0162] In one implementation scenario, the dimension of the network location vector of each user is set to 14, where Represents the network position vector of the ith user. The user attribute features of the ith user and the network position vector dimension are concatenated into a 28-dimensional vector as the initial semantic representation of the user.

[0163] S504: Input the initial semantic representation into the fully connected layer of the graph attention network model, and output an embedded semantic representation, where the embedded semantic representation includes the embedded semantic representation of the user and the embedded semantic representation of neighboring users.

[0164] In one implementation scenario, the initial semantic representation of the 28-dimensional vector is expanded to a 100-dimensional vector through a fully connected layer as the embedded semantic representation of the user, which is defined as follows:

[0165]

[0166] Among them, cat[·] represents the concatenation operation, and Dense(·) represents the fully connected layer, which is used to unify the dimension to 100 dimensions. is the embedded semantic representation of the i-th user, is the user attribute feature of the i-th user, a represents the user attribute feature, Represents the network location vector of the i-th user.

[0167] S505: Input the embedded semantic representation of the user and the embedded semantic representation of the neighboring users into the graph attention layer of the graph attention network model, and output the semantic representation of the user and the semantic representation of the neighboring users based on the graph attention.

[0168] In one implementation scenario, the graph attention network model is a deep graph attention network model, that is, the graph attention layer has multiple layers. For the first layer of graph attention layer, its input is the embedded semantic representation of the user and the embedded semantic representation of neighboring users based on different attentions, as shown below:

[0169]

[0170]

[0171]

[0172] in, represents the i-th user u in the first layer i The semantic representation of Com(u i ,u k ) for two users u i and u k The correlation index, and They are user u i and one of its neighbor users u k The embedded semantic representation of , cat[·] is the concatenation operation, a(·) represents the concatenation of user u i and neighbor user u k The embedded semantic representation is mapped to a real value LeakyReLU is the activation function, is the normalized attention result achieved through a softmax function, indicating that in the first layer neighbor user u k For user ui The importance of.

[0173] Except for the first attention layer, the embeddings of other attention layers are as follows:

[0174]

[0175]

[0176]

[0177] in, and They are user u i and its neighbor user u k The embedded semantic representation of the lth layer, cat[·] is the concatenation operation, and a(·) represents the concatenation of the lth layer user u i and its neighbor user u k The embedded semantic representation is mapped to a real value LeakyReLU is the activation function, is the normalized attention result achieved through a softmax function, indicating that the neighbor user u in the lth layer k For user u i The importance of represents user u at layer l-1 i semantic representation of .

[0178] It should be noted that since each user may be a neighbor of a certain user, each layer of the graph attention network model will update the semantic representation of each user. i The semantic representation of the l-th layer user u i The embedded semantic representation of the neighbor user u output by the l-1 layer k The semantic representation of is taken as the input of the lth layer.

[0179] S506: Input the semantic representation of the user based on graph attention and the semantic representation of the neighboring users into the output layer of the graph attention network model, and output the probability that the user is a 5G user.

[0180] In one implementation scenario, the semantic representation of the user output by each graph attention layer in the graph attention network model is integrated to generate the final semantic representation of the user. The final semantic representation of the user is converted into the probability that the user is a 5G user through an activation function.

[0181] The attention network model can be used to obtain the semantic representation of each user at different attention layers. The semantic representation of each layer represents the different importance of the information transmitted by the user at different attention layers. Therefore, the semantic representation of the user output by each layer of the graph attention layer can be integrated, which is defined as follows:

[0182]

[0183] in, Represents user u i The final semantic representation of represents the user u at level l i , cat[·] represents the concatenation operation, and L represents the number of layers of the graph attention network model.

[0184] The activation function can be a softmax function, as shown below:

[0185]

[0186] In the above formula, is the user u output by the graph attention network model i The probability of being a 5G user, Represents user u i The final semantic representation of .

[0187] S507: Calculate the loss between the probability that the user is a 5G user and the user label data using the loss function, and optimize the graph attention network model based on the loss.

[0188] The loss function is as follows:

[0189]

[0190] Among them, Z represents the user set in the training data, is the user u output by the graph attention network model i The probability of being a 5G user, Represents the real user u i User tag data.

[0191] In one implementation scenario, the probability threshold of a user being a 5G user can be set to Th, by Compare with Th to determine whether any user is a 5G user. If It indicates that the i-th user is a 5G user; otherwise, it is not a 5G user.

[0192] S508: Use the validation set to optimize the parameters of the graph attention network model, and use the test set to test and evaluate the graph attention network model.

[0193] Using the validation set to optimize the parameters of the graph attention network model can improve the prediction accuracy of the graph attention network model.

[0194] After optimizing the parameters of the graph attention network model using the validation set, the prediction effect of the graph attention network model is further tested using the test set. In one implementation scenario, the test set is input into the optimized graph attention network model, and the graph attention network model is evaluated and verified using evaluation indicators such as recall rate, precision rate, and F1_Score. Finally, the F1_Score value on the test set is 0.804, the recall rate is about 90%, and the precision rate is 70%, indicating that the prediction effect of the graph attention network model is relatively accurate.

[0195] The embodiment of the present application provides a method for generating a graph attention network model, which determines the neighbor users of a user based on the relationship triple data between users, and constructs the user attribute features based on the user attribute triple data. The user attribute features and the network position vector are spliced ​​to generate an initial semantic representation. The initial semantic representation is input into the fully connected layer of the graph attention network model, and the embedded semantic representation of each user is output. The embedded semantic representation of the user and the embedded semantic representation of the neighbor user are input into the graph attention layer of the graph attention network model, and the semantic representation of the user based on the graph attention layer and the semantic representation of the neighbor user are output. The semantic representation of the user based on the graph attention and the semantic representation of the neighbor user are input into the output layer of the graph attention network model, and the probability that the user is a 5G user is output. The loss function is used to calculate the loss between the probability that the user is a 5G user and the user label data, and the attention network model is optimized according to the loss. The parameters of the graph attention network model are further optimized using the validation set, and the graph attention network model is tested and evaluated using the test set. The method provided in the present application determines the importance of neighbor users to the user based on the attention mechanism, determines the probability that the user is a 5G user based on the semantic representation of the user based on the graph attention and the semantic representation of the neighbor user, and improves the accuracy of the prediction of the graph attention network model. At the same time, the loss function and validation set are used to optimize the parameters of the graph attention network model, further improving the prediction accuracy of the graph attention network model.

[0196] Figure 6 A schematic diagram of a network model establishment device for 5G user identification provided in an embodiment of the present application Figure 1 .like Figure 6 As shown, an embodiment of the present application provides a network model building device 600 for 5G user identification, which may include an acquisition module 601, a construction module 602 and a training module 603.

[0197] The acquisition module 601 is used to acquire the historical data of the user within a predetermined historical time period;

[0198] A construction module 602 is used to construct a user knowledge graph based on historical data, wherein the user knowledge graph includes triple data of relationships between users and triple data of user attributes;

[0199] The acquisition module 601 is further used to acquire the network location vector of the user by using a random walk algorithm based on the relationship triplet data between users;

[0200] The training module 603 is used to input the user attribute triplet data and the network position vector into the graph attention network model for training, so as to generate a graph attention network model for identifying 5G users.

[0201] The device of this embodiment can be used to perform the following steps: Figure 2 , Figure 3 and Figure 5 The implementation principle and technical effect of the method embodiment shown are similar and will not be repeated here.

[0202] Figure 7 A schematic diagram of a network model establishment device for 5G user identification provided in an embodiment of the present application Figure 2 .like Figure 7 As shown, an embodiment of the present application provides a network model establishment device 700 for 5G user identification, including a processor 701 and a memory 702, wherein the processor 701 and the memory 702 are connected via a bus 703.

[0203] In the specific implementation process, the code is stored in the memory 702, and the processor 701 runs the code stored in the memory 702 to execute the network model establishment method for 5G user identification in the above method embodiment.

[0204] The specific implementation process of the processor 701 can be found in the above method embodiment, and its implementation principle and technical effect are similar, so this embodiment will not be repeated here.

[0205] In the above Figure 7 In the illustrated embodiment, it should be understood that the processor 701 may be a central processing unit (CPU), or other general-purpose processors, digital signal processors (DSP), application-specific integrated circuits (ASIC), etc. A general-purpose processor may be a microprocessor or any conventional processor. The steps of the method disclosed in the invention may be directly implemented as being executed by a hardware processor, or may be executed by a combination of hardware and software modules in the processor.

[0206] The memory 702 may include a high-speed RAM memory, and may also include a non-volatile storage NVM, such as at least one disk storage.

[0207] The bus 703 may be an Industry Standard Architecture (ISA) bus, a Peripheral Component Interconnect (PCI) bus, or an Extended Industry Standard Architecture (EISA) bus, etc. The bus 703 may be divided into an address bus, a data bus, a control bus, etc. For ease of representation, the bus 703 in the drawings of the present application is not limited to only one bus or one type of bus.

[0208] An embodiment of the present application provides a computer-readable storage medium, in which computer execution instructions are stored. When the computer execution instructions are executed by a processor, they are used to implement the network model establishment method for 5G user identification of the above method embodiment.

[0209] The computer-readable storage medium mentioned above can be implemented by any type of volatile or non-volatile storage device or a combination thereof, such as static random access memory (SRAM), electrically erasable programmable read-only memory (EEPROM), erasable programmable read-only memory (EPROM), programmable read-only memory (PROM), read-only memory (ROM), magnetic memory, flash memory, magnetic disk or optical disk. The readable storage medium can be any available medium that can be accessed by a general or special-purpose computer.

[0210] An exemplary readable storage medium is coupled to a processor so that the processor can read information from the readable storage medium and write information to the readable storage medium. Of course, the readable storage medium can also be a component of the processor. The processor and the readable storage medium can be located in an application specific integrated circuit (Application Specific Integrated Circuits, referred to as: ASIC). Of course, the processor and the readable storage medium can also exist in the device as discrete components.

[0211] An embodiment of the present application provides a computer program product, including a computer program. When the computer program is executed by a processor, the network model establishment method for 5G user identification provided in any embodiment of the above-mentioned embodiment of the present application is implemented.

[0212] Those skilled in the art will readily appreciate other embodiments of the present application after considering the specification and practicing the invention disclosed herein. The present application is intended to cover any modification, use or adaptation of the present application, which follows the general principles of the present application and includes common knowledge or customary techniques in the art that are not disclosed in the present application. The specification and examples are intended to be exemplary only, and the true scope and spirit of the present application are indicated by the following claims.

[0213] It should be understood that the present application is not limited to the precise structures that have been described above and shown in the drawings, and that various modifications and changes may be made without departing from the scope thereof. The scope of the present application is limited only by the appended claims.

Claims

1. A method for establishing a network model for identifying potential 5G users, characterized in that: include: Obtain the user's historical data within a predetermined historical time period; Constructing a user knowledge graph based on the historical data, wherein the user knowledge graph includes triple data of relationships between users and triple data of user attributes; Based on the relationship triplet data between the users, a random walk algorithm is used to obtain the network position vector of the user; Inputting the user attribute triplet data and the network position vector into a graph attention network model for training to generate a graph attention network model for identifying potential 5G users; The step of inputting the user attribute triplet data and the network position vector into a graph attention network model for training to generate a graph attention network model for identifying potential 5G users includes: Determine neighbor users of the user based on the relationship triplet data between the users; Constructing user attribute features based on the user attribute triplet data; splicing the user attribute feature with the network position vector to generate an initial semantic representation; Inputting the initial semantic representation into the fully connected layer of the graph attention network model, and outputting an embedded semantic representation, wherein the embedded semantic representation includes the embedded semantic representation of the user and the embedded semantic representation of the neighboring users; Inputting the embedded semantic representation of the user and the embedded semantic representation of the neighboring user into the graph attention layer of the graph attention network model, and outputting the semantic representation of the user and the semantic representation of the neighboring user based on the graph attention; The semantic representation of the user based on graph attention and the semantic representation of the neighboring users are input into the output layer of the graph attention network model, and the probability that the user is a potential 5G user is output.

2. The method according to claim 1, characterized in that The step of constructing a user knowledge graph based on the historical data includes: generating a feature set wide table according to the historical data; Knowledge extraction is performed on the feature set wide table to obtain triple data, and a user knowledge graph is constructed based on the triple data.

3. The method according to claim 2, characterized in that The historical data includes: user basic information data, user business information data and user interaction behavior data; the feature set wide table generated according to the historical data includes: Calculating the data missing rate of the user characteristics according to the user basic information data, the user business information data and the user interaction behavior data; Filtering the historical data according to the data missing rate of the user features and the number of user features, wherein the user features include numerical user features and character user features; Performing feature superposition processing on the numerical user features, and performing quantization coding processing on the character user features; The feature set wide table is generated according to the historical data after feature superposition processing and quantization coding processing.

4. The method according to claim 1, characterized in that: The step of inputting the semantic representation of the user based on graph attention and the semantic representation of the neighboring user into the output layer of the graph attention network model, and outputting the probability that the user is a potential 5G user, includes: Integrate the semantic representation of the user output by each graph attention layer in the graph attention network model to generate a final semantic representation of the user; The final semantic representation of the user is converted into the probability that the user is a potential 5G user through an activation function.

5. The method according to claim 4, characterized in that The historical data further includes user tag data; after the final semantic representation of the user is converted into the probability that the user is a potential 5G user through the activation function, it also includes: The loss function is used to calculate the loss between the probability that the user is a potential 5G user and the user label data, and the graph attention network model is optimized according to the loss.

6. The method according to claim 5, characterized in that Before inputting the user attribute triplet data and the network position vector into the graph attention network model for training and generating the graph attention network model for identifying potential 5G users, the method further includes: Dividing the triplet data and the user label data into a training set, a validation set and a test set; Optimizing the parameters of the graph attention network model using the validation set; The test set is used to test and evaluate the graph attention network model.

7. A network model building device for identifying potential 5G users, characterized in that: include: An acquisition module is used to acquire the historical data of the user within a predetermined historical time period; A construction module, used to construct a user knowledge graph based on the historical data, wherein the user knowledge graph includes triple data of relationships between users and triple data of user attributes; The acquisition module is further used to acquire the network position vector of the user by using a random walk algorithm based on the relationship triplet data between the users; A training module, used for inputting the user attribute triplet data and the network position vector into a graph attention network model for training, so as to generate a graph attention network model for identifying potential 5G users; The training module is specifically used to determine the neighbor users of the user based on the relationship triplet data between the users; construct user attribute features based on the user attribute triplet data; concatenate the user attribute features with the network position vector to generate an initial semantic representation; input the initial semantic representation to the fully connected layer of the graph attention network model, and output an embedded semantic representation, wherein the embedded semantic representation includes the embedded semantic representation of the user and the embedded semantic representation of the neighbor users; The embedded semantic representation of the user and the embedded semantic representation of the neighbor users are input into the graph attention layer of the graph attention network model, and the semantic representation of the user and the semantic representation of the neighbor users based on graph attention are output; the semantic representation of the user and the semantic representation of the neighbor users based on graph attention are input into the output layer of the graph attention network model, and the probability that the user is a potential 5G user is output.

8. A network model building device for identifying potential 5G users, comprising: A processor and a memory, wherein the memory stores code, and the processor runs the code stored in the memory to execute the network model establishment method for identifying potential 5G users as described in any one of claims 1-6.

9. A computer-readable storage medium, characterized in that: The computer-readable storage medium stores computer execution instructions, which, when executed by a processor, are used to implement the network model establishment method for identifying potential 5G users as described in any one of claims 1-6.

Images