Paired constrained semi-supervised potential high-quality telephone subscriber identification method
By constructing rule equivalent similarity and pairwise constraint matrix, designing a comparative learning loss function, using graph neural network to identify potential high-quality users in mobile phone networks, solving the problem of insufficient identification in the existing technology, achieving high accuracy and enhanced supervision user identification, and supporting operators to optimize user strategies.
Patent Information
- Application Number
- CN202510486977.7
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2025-04-17
- Publication Date
- 2025-08-05
AI Technical Summary
The prior art is difficult to effectively identify potential high-quality users in mobile phone networks, resulting in user churn and loss of marketing opportunities.
Build a rule equivalent similarity matrix and pairwise constraint matrix between users, design a comparative learning loss function, embed a mobile call network structure using graph neural network, and train a graph neural network model with reconstruction loss and cluster loss to achieve semi-supervised recognition of potential high-quality users.
Without leaking user information, it improves the identification accuracy of potential high-quality users, reduces dependence on tags, enhances supervision of information, discovers and monitors the abnormal behavior of potential high-quality users, and helps operators formulate preferential policies to reduce churn.
Smart Images
Figure CN120429784A_ABST
Abstract
Description
Technical Field
[0001] The present invention belongs to the fields of complex networks and deep learning technologies, and in particular to a pairwise constrained semi-supervised potential high-quality telephone user identification method. Background Art
[0002] Currently, major operators primarily identify high-quality users based on basic information such as the plans they purchase and the frequency or duration of their mobile phone calls. However, this simplistic approach can miss potential high-quality mobile phone users (hereinafter referred to as users). Calling behavior between users can be represented as a mobile call network. These users may not have a significant advantage in terms of basic information such as plans, call frequency, or duration. However, from the perspective of the overall mobile call network structure, their departure or number change can significantly alter the network structure and even affect connectivity (of course, this impact is short-lived, but it is highly likely to cause losses), which in turn can impact other users in the network and lead to significant user churn. Furthermore, these users may also be key users for connecting to other networks, making targeted marketing targeting them helpful in attracting more users. Therefore, potential high-quality users play a crucial role in the call network, improving user stickiness and increasing user volume. Identifying them has significant practical significance and application value.
[0003] Existing methods for identifying high-quality users typically directly identify them based on explicit features such as the packages purchased or the amount spent. However, among users whose features are less pronounced, there are potentially high-quality users. These users play a vital role in the normal operation and future development of the entire mobile communication network. However, there is currently a lack of research on identifying potential high-quality users in mobile communication networks, and no effective identification methods have emerged. This paper addresses this issue by proposing a semi-supervised identification method based on pairwise constrained contrastive learning to effectively identify potential high-quality users without leaking user feature information. Summary of the Invention
[0004] The present invention aims to solve the above problems in the prior art. It proposes a pairwise constrained semi-supervised method for identifying potentially high-quality telephone users. The technical solution of the present invention is as follows:
[0005] A pairwise constrained semi-supervised method for potentially high-quality telephone user identification comprises the following steps:
[0006] Construct a rule equivalence similarity matrix between users and generate a pairwise constraint matrix;
[0007] Designing a contrastive learning loss function based on the pairwise constraint matrix to optimize the vector representation learning of users;
[0008] Using the graph neural network to embed the network structure of the mobile call network and the features constructed based on rule equivalence similarity, we can get the vector representation Z of each user i. i : Use clustering algorithm on user embedding vector Z to divide users into two categories: high-quality users and ordinary users;
[0009] A total loss function is constructed by combining the reconstruction loss, the contrastive learning loss, and the clustering loss, and the graph neural network model is trained by minimizing the total loss function.
[0010] Furthermore, the construction of the rule equivalence similarity matrix between users and the generation of the pairwise constraint matrix specifically include:
[0011] S1. Let S n×n is the rule equivalence similarity matrix between users, where n is the number of users, S ij It represents the rule equivalence similarity between users i and j, and measures the similarity between users i and j in terms of structural characteristics. The calculation formula of S is: S = (I-αA) -1 , where A is the adjacency matrix of the mobile phone network, I is the identity matrix, α∈(0,1 / λ max ) is an adjustable parameter, λ max is the largest eigenvalue of A;
[0012] S2. Let L i Indicates the label of user i, the label value is high-quality user or ordinary user. For the displayed high-quality user or ordinary user, L i The value of is known; for other users, L i The value of is unknown;
[0013] S3. Let matrix C n×n is a pairwise constraint matrix, where C ij Represents the constraint relationship between users i and j. If C ij =1, it means that users i and j should belong to the same category of users, that is, they are both high-quality users or ordinary users; if C ij = -1, it means that users i and j should belong to different categories of users, that is, one is a high-quality user and the other is an ordinary user; if C ij =0, it means there is no constraint between users i and j.
[0014] Furthermore, for any pair of users i and j, C is determined by the following steps: ij Value:
[0015] When the labels of users i and j are both known, then:
[0016] If Li = = L j , that is, if they are both high-quality users or ordinary users, then C ij =1;
[0017] If L i ≠L j , one is a high-quality user and the other is an ordinary user, then C ij =-1;
[0018] When at least one of the tags of user i and j is unknown, then:
[0019] If S ij ∈max3({S ij |j=1,2,…,N}) and S ij ∈max3({S ij |i=1,2,…,N}), max3() is used to select the three maximum values from the specified sequence. This condition means that the rule equivalence similarity value between users i and j is one of the three largest rule equivalence similarity values between them and all nodes, then C ij =1; if S ij ∈min3({S ij |j=1,2,…,N}) and S ij ∈min3({S ij |i=1,2,…,N}), min3() is used to select the three minimum values from the specified sequence. This condition means that the rule equivalence similarity value between users i and j is one of the three smallest rule equivalence similarity values between them and all nodes, then C ij =-1;
[0020] If users i and j do not meet the above conditions, then C ij =0.
[0021] Furthermore, the design of the comparative learning loss function specifically includes:
[0022] For any user i, let Z i Representing its vector representation, the contrastive learning loss function based on pairwise constraints is designed as:
[0023]
[0024] Furthermore, the network structure of the mobile call network and the features constructed based on rule equivalence similarity are embedded in the graph neural network to obtain the vector representation Z of each user i. i , specifically including:
[0025] User feature construction: Perform eigendecomposition on the rule equivalence similarity matrix S, and take the eigenvectors corresponding to the largest k eigenvalues and the degree vector to form the user feature matrix X:
[0026] X=[e1,e2,...,e k ,d]
[0027] where e i is the eigenvector corresponding to the i-th largest eigenvalue of S, d is an n-dimensional column vector, each element of which corresponds to the degree of a user in the mobile calling network;
[0028] User vector embedding: Using X and A as input, a three-layer graph convolutional neural network is used to obtain the embedding vectors Z of all users. i A row of Z is the embedding vector of user i. The calculation formula of the l-th layer graph convolutional neural network is as follows:
[0029]
[0030] Among them H (l) is the output result of the lth layer, W (l) is the weight matrix of layer l, is the adjacency matrix with self-loops added, yes The corresponding degree diagonal matrix, Φ is the activation function; H 0 =X,Z=H 3 .
[0031] Furthermore, the combining of the reconstruction loss, the contrastive learning loss, and the clustering loss to construct a total loss function, and training the graph neural network model by minimizing the total loss function, specifically includes:
[0032] S1. Loss function construction: The loss function consists of three parts: reconstruction loss function L rc , contrastive learning loss function L based on pairwise constraints con And the clustering loss function L KL , L rc is the error of reconstructing the adjacency matrix of the mobile call network through the embedding vector, L KL It is the Kullback-Leibler divergence between the user's category distribution Q in the clustering result and the target category distribution P to be approximated, which measures the difference between the two distributions. rc and L KL The calculation formulas are:
[0033] L rc =||A-σ(Z T Z)||2
[0034]
[0035] σ represents the sigmod function, which is as follows: σ(Z T Z) for the matrix Z T Each element of Z performs a sigmod operation. In the Q distribution, the probability q of each user i belonging to each cluster category c is ic The calculation formula is:
[0036]
[0037] c = 1 or 0, indicating high-quality user class and ordinary user class respectively; u c represents the central representation vector of category c, which is obtained by averaging the vector representations of users of category c;
[0038] In the P distribution, the target probability p for each user i to belong to each category c is ic The calculation formula is:
[0039]
[0040] The total loss function L is:
[0041] L=L rc +L con +L KL
[0042] S2. Complete the training of graph convolutional neural network by minimizing the loss function L: First, minimize the reconstruction loss function L rc The graph convolutional neural network is pre-trained, and then the final training of the graph convolutional neural network is completed by minimizing L.
[0043] An electronic device comprises a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein when the processor executes the program, the pairwise constrained semi-supervised potentially high-quality telephone user identification method as described in any one of the preceding claims is implemented.
[0044] A non-transitory computer-readable storage medium stores a computer program thereon, wherein when the computer program is executed by a processor, the method for implementing the pairwise constrained semi-supervised potentially high-quality telephone user identification method described in any one of the preceding claims is implemented.
[0045] The advantages and beneficial effects of the present invention are as follows:
[0046] 1) The present invention can discover potential high-quality users with the help of a small number of displayed users without exposing the private information of mobile phone users (this method only requires user call relationships and a small number of displayed high-quality users, and this information can be desensitized through anonymization).
[0047] 2) This invention enriches pairwise constraints by showing a small amount of rule equivalence similarity between users. Compared with existing unsupervised and semi-supervised technologies, it has stronger supervision information and improves the accuracy of potential high-quality user identification.
[0048] Innovation steps: 1) For the application problem of high-quality mobile phone identification, the entire solution is innovative: I have not found any relevant research on identifying potential high-quality mobile phone users based on graph neural network models; even for the scientific problem behind it (influence node identification), I have not found any semi-supervised models based on graph neural networks or even other basic models. The advantage of this solution: high-quality users can be identified when only the call relationship information between users and a small number of high-quality users are required (or not required). 2) As for the technical details, the innovation lies in the construction of the pairwise constraint matrix (specifically, the determination of C in S3). ij , that is, determining the corresponding element values of the node pairs with missing labels in the pairwise constraint matrix through the rule equivalence similarity between users); this operation utilizes the structural position similarity between users to enhance the supervision information, thereby reducing the dependence on labels and effectively improving the accuracy of potential high-quality user identification. BRIEF DESCRIPTION OF THE DRAWINGS
[0049] Figure 1 Schematic diagram of a semi-supervised potential high-quality mobile phone user identification method based on pairwise constrained contrastive learning according to a preferred embodiment of the present invention;
[0050] Figure 2 This is an example of how a mobile phone network is constructed;
[0051] Figure 3 This is an example diagram of using rule equivalence similarity to evaluate the similarity advantage of users in structural characteristics. DETAILED DESCRIPTION
[0052] The following will describe the technical solutions in the embodiments of the present invention in detail with reference to the accompanying drawings. The described embodiments are only a part of the embodiments of the present invention.
[0053] The technical solution of the present invention to solve the above technical problems is:
[0054] Application scenarios and problems:
[0055] The calls between mobile phone users form a large mobile call network. The mobile call network only contains the IDs (i.e., phone numbers) of mobile phone users and the call relationships between these users (no other information about the users and the content of the calls is required). For the construction of the mobile call network, please refer to Figure 2The purpose of the present invention is to identify potential high-quality users by only displaying high-quality users and mobile call networks.
[0056] Basic ideas of the proposal:
[0057] A pairwise constraint matrix is constructed using the structural similarity between displayed high-quality users and users in the mobile call network, and a contrastive learning loss function is designed using the pairwise constraint matrix. Finally, a semi-supervised potential high-quality recognition model is established by combining the contrastive learning loss function with a graph neural network to complete the classification of potential high-quality users from ordinary users.
[0058] Technical solutions such as Figure 1 The specific steps are as follows:
[0059] 1. Pairwise constraints based on partial label and rule equivalence similarity:
[0060] S1. Let S n×n is the rule equivalence similarity matrix between users, where n is the number of users, S ij represents the rule equivalence similarity between users i and j, which measures the similarity between users i and j in terms of structural characteristics (compared to other similarity indicators, rule equivalence similarity has the advantage of capturing structural similarity and does not require any association between users, for example Figure 3 The association here refers to other users who have spoken to each other or have had a common call. The calculation formula of S is: S = (I-αA) -1 , where A is the adjacency matrix of the mobile phone network, I is the identity matrix, α∈(0,1 / λ max ) is an adjustable parameter (λ max is the largest eigenvalue of A), the mobile call network is constructed through the call behavior between users, for example Figure 2 shown.
[0061] S2. Let L i Indicates the label of user i (label value is high-quality user or ordinary user). For high-quality users or ordinary users displayed, L i The value of is known; for other users, L i The value of is unknown. S3. Let the matrix C n×n is a pairwise constraint matrix, where C ij Represents the constraint relationship between users i and j. If C ij =1, it means that users i and j should belong to the same category of users, that is, they are both high-quality users or ordinary users; if C ij = -1, it means that users i and j should belong to different categories of users, that is, one is a high-quality user and the other is an ordinary user; if C ij= 0, it means there is no constraint between users i and j. For any pair of users i and j, determine C by the following steps: ij Value:
[0062] 1) When the labels of users i and j are both known, then:
[0063] a) If L i = = L j (i.e. both are high-quality users or ordinary users), then directly C ij =1.
[0064] b) If L i ≠L j (One is a high-quality user shown, and the other is an ordinary user), then C ij =-1.
[0065] 2) When at least one of the tags of user i and j is unknown, then:
[0066] a) If S ij ∈max3({S ij |j=1,2,…,N}) and S ij ∈max3({S ij |i=1,2,…,N})(The function of max3() is to select the three maximum values from the specified sequence. This condition means that the rule equivalence similarity value between users i and j is one of the three largest rule equivalence similarity values between them and all nodes), then C ij =1.
[0067] b) If S ij ∈min3({S ij |j=1,2,…,N}) and S ij ∈mij3({S ij |i=1,2,…,N})(min3() is used to select the three minimum values from the specified sequence. This condition means that the rule equivalence similarity value between users i and j is one of the three smallest rule equivalence similarity values between them and all nodes). Then C ij =-1.
[0068] c) If users i and j do not meet conditions a) and b), then C ij =0.
[0069] 2. Contrastive learning loss function based on pairwise constraints: For any user i, let Z i Represents its vector representation (Z i The calculation of is as shown in step S1 in step 3), then the contrastive learning loss function based on pairwise constraints is designed as:
[0070]
[0071] 3. Semi-supervised Identification of Potential High-quality Users Based on Pairwise Constraint Contrastive Learning:
[0072] S1. Use rule equivalence similarity and graph neural network to embed the network structure of the mobile call network to obtain the vector representation Z of each user i i :
[0073] 1) User feature construction: Perform eigenvalue decomposition on the rule equivalence similarity matrix S, and take the eigenvectors corresponding to the largest k eigenvalues and the degree vector to form the user feature matrix X:
[0074] X = [e1, e2,..., e k , d]
[0075] where e i is the eigenvector corresponding to the i-th largest eigenvalue of S, and d is an n-dimensional column vector, and each element corresponds to the degree of a user in the mobile call network.
[0076] 2) User vector embedding: Using X and A as inputs, apply a three-layer graph convolutional neural network to obtain the embedding vectors Z of all users (Z i is a row of Z and is the embedding vector of user i). The calculation formula of the l-th (0 < l ≤ 3) layer graph convolutional neural network is as follows:
[0077]
[0078] where H (l) is the output result of the l-th layer, W (l) is the weight matrix of the l-th layer, is the adjacency matrix with self-loops added, is the corresponding degree diagonal matrix, Φ is the activation function; H 0 = X, Z = H 3 . S2. Apply a clustering algorithm to the user embedding vectors Z to divide users into two categories, one is high-quality users and the other is ordinary users.
[0079] Training of the graph convolutional neural network model for obtaining the user embedding vector Z:
[0080] S1. Loss function construction: In order to obtain accurate user embedding vectors, the loss function includes three parts: the reconstruction loss function L rc , the pairwise constraint-based contrastive learning loss function L con and the clustering loss function L KL . Among them, L con is as shown in step 2, L rcis the error of reconstructing the adjacency matrix of the mobile call network through the embedding vector, L KL It is the Kullback-Leibler divergence between the user's category distribution Q in the clustering result and the target category distribution P to be approximated, which measures the difference between the two distributions. rc and L KL The calculation formulas are:
[0081] L r =||A-ρ(Z T Z)||2
[0082]
[0083] In the Q distribution, the probability q that each user i belongs to each cluster category c is ic (c = 1 or 0, indicating high-quality user class and ordinary user class, respectively) is calculated as follows:
[0084]
[0085] In the P distribution, the target probability p for each user i to belong to each category c is ic The calculation formula is:
[0086]
[0087] The total loss function L is:
[0088] L=L rc +L con +L KL
[0089] S2. Complete the training of graph convolutional neural network by minimizing the loss function L: First, minimize the reconstruction loss function L rc The graph convolutional neural network is pre-trained, and then the final training of the graph convolutional neural network is completed by minimizing L.
[0090] The semi-supervised potential high-quality mobile phone user identification method based on pairwise constrained contrastive learning proposed in the present invention is not limited to a certain semi-supervised deep clustering model. The user feature construction and pairwise constrained contrastive learning loss function provided by it can be used in every method that uses a semi-supervised deep clustering model to identify potential high-quality mobile call users; it is not limited to unweighted and undirected homogeneous networks. The user feature construction and pairwise constrained contrastive learning loss function provided by it can also be used in weighted, directed, and heterogeneous networks (for example, the call network can be represented as a directed weighted network).
[0091] The method for identifying potential high-quality mobile phone users provided by the present invention is primarily used to identify potential high-quality users in mobile communication networks. These potential high-quality users may be connected to multiple communication subnets, can quickly contact other users, or frequently communicate with a large number of users. By monitoring these users in real time (monitoring their call behavior, not their call content), it is possible to promptly detect and even predict abnormal call behavior, such as whether their call frequency has decreased, their call patterns have changed, and whether they can cover a large number of users on different networks. This helps operators analyze the causes and formulate policies (for example, customized communication packages can be provided for users at risk of leaving the communication network, offering timely discounts to prevent user churn; customized packages can be provided for users whose call patterns have changed, maximizing discounts for these users without sacrificing operator revenue; and preferential packages can be formulated for these users' call friends to attract more users).
[0092] The potential high-quality mobile phone user identification method provided by the present invention is also applicable to the identification and mining of core nodes of the computing power network.
[0093] The systems, devices, modules or units described in the above embodiments may be implemented by computer chips or entities, or by products with certain functions.
[0094] Computer-readable media includes permanent and non-permanent, removable and non-removable media that can be implemented by any method or technology to store information. The information can be computer-readable instructions, data structures, program modules or other data. Examples of computer storage media include, but are not limited to, phase change memory (PRAM), static random access memory (SRAM), dynamic random access memory (DRAM), other types of random access memory (RAM), read-only memory (ROM), electrically erasable programmable read-only memory (EEPROM), flash memory or other memory technology, compact disc read-only memory (CD-ROM), digital versatile disc (DVD) or other optical storage, magnetic cassettes, magnetic tape, magnetic disk storage or other magnetic storage devices or any other non-transmission media that can be used to store information that can be accessed by a computing device. As defined herein, computer-readable media does not include transitory computer-readable media (transitory media), such as modulated data signals and carrier waves.
[0095] It should also be noted that the terms "comprises," "includes," or any other variations thereof are intended to encompass non-exclusive inclusion, such that a process, method, commodity, or apparatus that includes a series of elements includes not only those elements but also other elements not explicitly listed, or includes elements inherent to such process, method, commodity, or apparatus. In the absence of further limitations, an element defined by the phrase "comprises a ..." does not exclude the presence of other identical elements in the process, method, commodity, or apparatus that includes the element.
[0096] The above embodiments should be understood as merely illustrating the present invention and not as limiting the scope of protection of the present invention. After reading the contents of the present invention, technicians may make various changes or modifications to the present invention, and these equivalent changes and modifications also fall within the scope defined by the claims of the present invention.
Claims
1. A pairwise constrained semi-supervised method for potentially high-quality telephone user identification, characterized in that: The following steps are involved: Construct a rule equivalence similarity matrix between users and generate a pairwise constraint matrix; Designing a contrastive learning loss function based on the pairwise constraint matrix to optimize the vector representation learning of users; Using the graph neural network to embed the network structure of the mobile call network and the features constructed based on rule equivalence similarity, we can get the vector representation Z of each user i. i : Use clustering algorithm on user embedding vector Z to divide users into two categories: high-quality users and ordinary users; A total loss function is constructed by combining the reconstruction loss, the contrastive learning loss, and the clustering loss, and the graph neural network model is trained by minimizing the total loss function.
2. The pairwise constrained semi-supervised potential high-quality telephone user identification method according to claim 1, characterized in that: The construction of the rule equivalence similarity matrix between users and the generation of the pairwise constraint matrix specifically include: S1. Let S n×n is the rule equivalence similarity matrix between users, where n is the number of users, S ij It represents the rule equivalence similarity between users i and j, and measures the similarity between users i and j in terms of structural characteristics. The calculation formula of S is: S = (I-αA) -1 , where A is the adjacency matrix of the mobile phone network, I is the identity matrix, α∈(0,1 / λ max ) is an adjustable parameter, λ max is the largest eigenvalue of A; S2. Let L i Indicates the label of user i, the label value is high-quality user or ordinary user. For the displayed high-quality user or ordinary user, L i The value of is known; for other users, L i The value of is unknown; S3. Let matrix C n×n is a pairwise constraint matrix, where C ij Represents the constraint relationship between users i and j. If C ij =1, it means that users i and j should belong to the same category of users, that is, they are both high-quality users or ordinary users; if C ij = -1, it means that users i and j should belong to different categories of users, that is, one is a high-quality user and the other is an ordinary user; if C ij =0, it means there is no constraint between users i and j.
3. The pairwise constrained semi-supervised potential high-quality telephone user identification method according to claim 2, characterized in that: For any pair of users i and j, C is determined by the following steps: ij Value: 1) When the labels of users i and j are both known, then: If L i = = L j , that is, if they are both high-quality users or ordinary users, then C ij =1; If L i ≠L j , one is a high-quality user and the other is an ordinary user, then C ij =-1; 2) When at least one of the tags of user i and j is unknown, then: If S ij ∈max3({S ij |j=1,2,…,N}) and S ij ∈max3({S ij |i=1,2,…,N}), max3() is used to select the three maximum values from the specified sequence. This condition means that the rule equivalence similarity value between users i and j is one of the three largest rule equivalence similarity values between them and all nodes, then C ij =1; if S ij ∈min3({S ij |j=1,2,…,N}) and S ij ∈min3({S ij |i=1,2,…,N}), min3() is used to select the three minimum values from the specified sequence. This condition means that the rule equivalence similarity value between users i and j is one of the three smallest rule equivalence similarity values between them and all nodes, then C ij =-1; If users i and j do not meet the above conditions, then C ij =0.
4. The pairwise constrained semi-supervised potential high-quality telephone user identification method according to claim 3, characterized in that: The design contrast learning loss function specifically includes: For any user i, let Z i Representing its vector representation, the contrastive learning loss function based on pairwise constraints is designed as:
5. The pairwise constrained semi-supervised potential high-quality telephone user identification method according to claim 4, characterized in that: The network structure of the mobile call network is embedded in the graph neural network and the features constructed based on rule equivalence similarity are used to obtain the vector representation Z of each user i. i , specifically including: 1) User feature construction: Perform eigendecomposition on the rule equivalence similarity matrix S, and take the eigenvectors corresponding to the largest k eigenvalues and the degree vector to form the user feature matrix X: X=[e1,e2,...,e k ,d] where e i is the eigenvector corresponding to the i-th largest eigenvalue of S, d is an n-dimensional column vector, each element of which corresponds to the degree of a user in the mobile calling network; 2) User vector embedding: Using X and A as input, a three-layer graph convolutional neural network is used to obtain the embedding vectors Z of all users, Z i A row of Z is the embedding vector of user i. The calculation formula of the l-th layer graph convolutional neural network is as follows: Among them H (l) is the output result of the lth layer, W (l) is the weight matrix of layer l, is the adjacency matrix with self-loops added, yes The corresponding degree diagonal matrix, Φ is the activation function; H 0 =X,Z=H 3 .
6. The pairwise constrained semi-supervised potential high-quality telephone user identification method according to claim 5, characterized in that: The combining of the reconstruction loss, the contrastive learning loss, and the clustering loss to construct a total loss function, and training the graph neural network model by minimizing the total loss function, specifically includes: S1. Loss function construction: The loss function consists of three parts: reconstruction loss function L rc , contrastive learning loss function L based on pairwise constraints con And the clustering loss function L KL , L rc is the error of reconstructing the adjacency matrix of the mobile call network through the embedding vector, L KL It is the Kullback-Leibler divergence between the user's category distribution Q in the clustering result and the target category distribution P to be approximated, which measures the difference between the two distributions. rc and L KL The calculation formulas are: L rc =||A-σ(Z T Z)||2 σ represents the sigmod function, which is as follows: σ(Z T Z) for the matrix Z T Each element of Z performs a sigmod operation. In the Q distribution, the probability q of each user i belonging to each cluster category c is ic The calculation formula is: c = 1 or 0, indicating high-quality user class and ordinary user class respectively; u c represents the central representation vector of category c, which is obtained by averaging the vector representations of users of category c; In the P distribution, the target probability p for each user i to belong to each category c is ic The calculation formula is: The total loss function L is: L=L rc +L con +L KL S2. Complete the training of graph convolutional neural network by minimizing the loss function L: First, minimize the reconstruction loss function L rc The graph convolutional neural network is pre-trained, and then the final training of the graph convolutional neural network is completed by minimizing L.
7. An electronic device, characterized in that: The method comprises a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein when the processor executes the program, the method for pairwise constrained semi-supervised potential high-quality telephone user identification according to any one of claims 1 to 6 is implemented.
8. A non-transitory computer-readable storage medium having a computer program stored thereon, characterized in that: When the computer program is executed by a processor, the pairwise constrained semi-supervised potential high-quality telephone user identification method according to any one of claims 1 to 6 is implemented.