User information classification method and device

Abstract

The invention discloses a user information classification method and device, and relates to the technical field of super deep learning. The specific embodiment of the method comprises the following steps: receiving an information classification request; wherein the information classification request comprises attribute data of a plurality of users; constructing an initial feature matrix and an initial adjacency matrix by utilizing the attribute data of the plurality of users, and preprocessing the initial feature matrix and the initial adjacency matrix to obtain an improved feature matrix and an improved adjacency matrix; inputting the improved feature matrix and the improved adjacent matrix into a node feature optimization model, and determining an optimized feature matrix according to the output of the node feature optimization model; and inputting the optimized feature matrix and the improved adjacency matrix into a user information classification model, and determining a target group of each user. According to the embodiment, the complex network can be accurately divided, the target group division result suitable for the user is obtained, the authenticity and accuracy of global division are improved, and user personalized recommendation in practical application is further optimized.

Description

Technical Field

[0001] The present invention relates to the field of super deep learning technology, and in particular, to a method and apparatus for classifying user information. Background Art

[0002] Complex networks are used to vividly represent individuals in society and the relationships between individuals. Individuals are usually represented by nodes, and individual relationships are represented by edges between nodes. By partitioning the network into communities according to the degree of connection tightness between nodes, the community structure of the complex network can be obtained.

[0003] In the existing process of community structure partitioning, community detection algorithms are usually used to detect complex networks in order to obtain appropriate community partitions from complex networks. Community detection algorithms include traditional community detection algorithms and deep learning community detection algorithms using graph neural networks.

[0004] In the process of implementing the present invention, the inventors found that there are at least the following problems in the prior art:

[0005] Traditional community detection algorithms use heuristic or guided optimization objectives to iterate continuously, but they cannot effectively evaluate the partitioning results of community structures; deep learning community detection algorithms use modularity, which measures the structural degree of community partitioning results, as a measurement index, introduce objective functions for optimization, and use the backpropagation method to iterate continuously, but they will merge communities smaller than a certain scale into other large communities, resulting in lower authenticity and accuracy of community partitioning results. Summary of the Invention

[0006] In view of this, embodiments of the present invention provide a method and apparatus for classifying user information, which can accurately partition a complex network, obtain appropriate target group partitioning results for users, improve the authenticity and accuracy of global partitioning, and further optimize user personalized recommendations in practical applications.

[0007] To achieve the above object, according to one aspect of the embodiments of the present invention, there is provided a method for classifying user information, including:

[0008] Receiving one or more information classification requests; wherein, the information classification request includes attribute data of multiple users;

[0009] Constructing an initial feature matrix and an initial adjacency matrix using the attribute data of the multiple users, and preprocessing the initial feature matrix and the initial adjacency matrix to obtain an improved feature matrix and an improved adjacency matrix;

[0010] Input the improved feature matrix and the improved adjacency matrix into a node feature optimization model. The node feature optimization model uses a graph convolutional neural network and takes the mutual information between the local information and the global information of each node being less than a preset loss threshold as the optimization objective, and determines an optimized feature matrix according to the output of the node feature optimization model;

[0011] Input the optimized feature matrix and the improved adjacency matrix into a user information classification model to determine the target group of each user; wherein, the user information classification model takes the sum of the modularity and a regularization term for correcting the modularity as the objective function.

[0012] Optionally, the determining the optimized feature matrix includes:

[0013] Set an initial current optimization weight matrix, perform convolution on the improved adjacency matrix, the improved feature matrix, and the current optimization weight matrix to obtain a current optimization convolution result;

[0014] Use a non-linear activation function to transform the current optimization convolution result to obtain a current optimization pooling result, and determine the loss function of the point feature optimization model according to the mutual information between the current optimization convolution result and the current optimization pooling result;

[0015] Iteratively calculate the current loss value of the loss function. When the current loss value is less than or equal to the preset loss threshold, take the current optimization weight matrix as the target optimization weight matrix, calculate the current optimization convolution result of the target optimization weight matrix, and obtain the optimized feature matrix.

[0016] Optionally, the determining the loss function of the point feature optimization model according to the mutual information between the current optimization convolution result and the current optimization pooling result includes:

[0017] Take the current optimization convolution result as a positive sample, reassign the positive sample to generate a negative sample corresponding to the positive sample;

[0018] Calculate the first mutual information between the positive sample and the current optimization pooling result and the second mutual information between the negative sample and the current optimization pooling result respectively, and superimpose the first mutual information and the second mutual information as the loss function of the node feature optimization model.

[0019] Optionally, determining the objective function of the user information classification model includes:

[0020] Set an initial current partitioning weight matrix, perform convolution on the improved adjacency matrix, the optimized feature matrix, and the current partitioning weight matrix to obtain a current partitioning convolution result;

[0021] Calculate the first norm of the improved adjacency matrix, and use the difference obtained by subtracting the ratio of the matrix product of the degree matrix of the improved adjacency matrix and the transpose matrix of the degree matrix of the improved adjacency matrix from the improved adjacency matrix as the transition matrix;

[0022] Perform a trace operation on the result of multiplying the current partition convolution result, the transition matrix, and the transpose matrix of the current partition convolution result, and take the negative of the result of dividing the trace result by the ratio of the first norm to obtain the modularity.

[0023] Optionally, it further includes:

[0024] Take the reciprocal of the column dimension of the current partition weight matrix as a constant term, and correct each element of the current partition convolution result to obtain the regularization term of the modularity.

[0025] Introduce a harmonic parameter, and form the objective function of the user information classification model from the modularity, the product of the harmonic parameter and the regularization term.

[0026] Optionally, the preprocessing of the initial feature matrix and the initial adjacency matrix includes:

[0027] Determine the filter matrix according to the initial adjacency matrix and the first filtering parameter;

[0028] Stack multiple layers of the filter matrix and perform filtering on the initial feature matrix to obtain the improved feature matrix; where the number of layers of the filter matrix is the second filtering parameter.

[0029] Optionally, the preprocessing of the initial adjacency matrix includes:

[0030] Calculate the influence coefficients between each node according to the preset diffusion coefficient and the order relationship between nodes in the initial adjacency matrix, and form an influence matrix;

[0031] Perform a Laplace transform on the initial adjacency matrix to obtain the transformed adjacency matrix of the initial adjacency matrix;

[0032] Calculate the product of the influence matrix and the transformed adjacency matrix, and determine the improved adjacency matrix according to the Laplace transform result of the product of the influence matrix and the transformed adjacency matrix.

[0033] Optionally, the calculating the influence coefficients between each node according to the preset diffusion coefficient and the order relationship between nodes in the initial adjacency matrix includes:

[0034] Determine the neighbor order between the nodes according to the connection paths between different nodes in the initial adjacency matrix;

[0035] Calculate the power result of the difference between the diffusion constant and the diffusion coefficient, and multiply the power result by the diffusion coefficient as the influence coefficient between the nodes; wherein, the power of the power result is equal to the neighbor order between the nodes.

[0036] According to another aspect of the embodiments of the present invention, there is provided a classification device for user information, including:

[0037] A receiving module, configured to receive one or more information classification requests; wherein, the information classification request includes attribute data of multiple users;

[0038] A preprocessing module, configured to construct an initial feature matrix and an initial adjacency matrix by using the attribute data of multiple users, and preprocess the initial feature matrix and the initial adjacency matrix to obtain an improved feature matrix and an improved adjacency matrix;

[0039] A feature optimization module, configured to input the improved feature matrix and the improved adjacency matrix into a node feature optimization model, the node feature optimization model uses a graph convolutional neural network, and takes the mutual information between the local information and the global information of each node being less than a preset loss threshold as the optimization target, and determine an optimized feature matrix according to the output of the node feature optimization model;

[0040] A partitioning module, configured to input the optimized feature matrix and the improved adjacency matrix into a user information classification model to determine the target group of each user; wherein, the user information classification model takes the sum of the modularity and a regularization term for correcting the modularity as the objective function.

[0041] According to another aspect of the embodiments of the present invention, there is provided an electronic device for classifying user information, including:

[0042] One or more processors;

[0043] A storage device, configured to store one or more programs,

[0044] When the one or more programs are executed by the one or more processors, the one or more processors implement the user information classification method provided by the present invention.

[0045] According to still another aspect of the embodiments of the present invention, there is provided a computer-readable medium, on which a computer program is stored, and when the program is executed by a processor, the user information classification method provided by the present invention is implemented.

[0046] One embodiment of the above invention has the following advantages or beneficial effects: By preprocessing the initial adjacency matrix and the initial feature matrix, and using the preprocessed improved adjacency matrix and improved feature matrix as the input of the graph convolutional neural network in the first stage, an optimized feature matrix is obtained; in the second stage, the modularity of a single target is corrected, and the objective function of the user information classification model is constructed by using both modularity and regularization terms. The objective function is solved according to the optimized feature matrix and the improved adjacency matrix, and the technical means of the target group to which the user belongs is obtained according to the output of the graph convolutional neural network in the second stage. Therefore, the technical problems of being unable to effectively evaluate the division result of the community structure and having low authenticity and accuracy of the division result are overcome. Furthermore, the technical effect of being able to accurately divide a complex network, obtain a suitable target group division result for users, improve the authenticity and accuracy of the global division, and further optimize the user personalized recommendation in practical applications is achieved.

[0047] The further effects of the above non-conventional optional methods will be described below in combination with specific embodiments. BRIEF DESCRIPTION OF THE DRAWINGS

[0048] The drawings are used to better understand the present invention and do not constitute an improper limitation to the present invention. Among them:

[0049] Figure 1 is a schematic diagram of the main process of the classification method of user information according to an embodiment of the present invention;

[0050] Figure 2 is a schematic diagram of the main process of the method for determining the improved feature matrix according to an embodiment of the present invention;

[0051] FIG. 3(a) is a schematic diagram of the main process of the method for determining the improved adjacency matrix according to an embodiment of the present invention;

[0052] FIG. 3(b) is a schematic diagram of the graph structure according to an embodiment of the present invention;

[0053] Figure 4 is a schematic diagram of the main process of the method for determining the optimized feature matrix according to an embodiment of the present invention;

[0054] Figure 5 is a schematic diagram of the main process of the method for determining the target group of users according to an embodiment of the present invention;

[0055] Figure 6 is a schematic diagram of the main modules of the classification device of user information according to an embodiment of the present invention;

[0056] Figure 7 shows an exemplary system architecture diagram suitable for application to an embodiment of the present invention;

[0057] Figure 8 It is a schematic structural diagram of a computer system suitable for a terminal device or a server for implementing the embodiments of the present invention. Detailed implementation manners

[0058] The following describes exemplary embodiments of the present invention with reference to the accompanying drawings. Various details of the embodiments of the present invention are included to facilitate understanding, and they should be considered merely exemplary. Therefore, those of ordinary skill in the art should recognize that various changes and modifications can be made to the embodiments described herein without departing from the scope and spirit of the present invention. Similarly, for clarity and conciseness, descriptions of well-known functions and structures are omitted in the following description.

[0059] It should be noted that in the technical solution of the present invention, the processing of collection, use, storage, sharing, and transfer of user personal information complies with the provisions of relevant laws and regulations, and users need to be informed and obtain their consent or authorization. When applicable, technical processing such as de-identification and / or anonymization and / or encryption is performed on user personal information. (For example: after collecting user contact information, address, etc., we will perform de-identification processing on the data through technical means.)

[0060] GCN: Graph Convolution Networks, also known as graph convolutional neural networks, apply convolutional operations in Euclidean space to graph structures (Graph structures), extract the features of graph structure data, use function mapping, and obtain new node representations by giving the topological structure information of the network and the current node information, and repeat this process to obtain appropriate classification results.

[0061] In traditional community detection algorithms, heuristic or guided optimization objectives are used and iterated continuously until the target iteration number is reached or the desired effect is achieved. However, the community division results cannot be judged.

[0062] Deep learning community detection algorithms use modularity as the optimization objective, use the backpropagation method, aggregate information to obtain node representations, and obtain division results. Modularity can measure the structural degree of community division results. The stronger the structure, the clearer the community division results. However, on the one hand, modularity cannot accurately define communities smaller than a certain scale and usually merges them into large communities, resulting in a resolution limit problem; on the other hand, the data source of algorithms using modularity as the objective is unlabeled data. Even if the objective is to maximize modularity, the division results still have a large difference from the true community structure, with low accuracy, the division results are not optimal, and they do not match the true community structure and do not conform to the true community structure.

[0063] Figure 1 It is a schematic diagram of the main process of the user information classification method according to the embodiments of the present invention, asFigure 1 As shown in Figure 1 , the classification method for user information of the present invention includes the following steps:

[0064] Step S101, receiving one or more information classification requests; wherein, the information classification request includes attribute data of multiple users.

[0065] In an embodiment of the present invention, the attribute data of a user includes the user's age, gender, contact information, address, occupation, login record, follow record, search record, browsing record, order record, transaction method, transaction account, transaction time period, recipient, sharing record, storage record, etc.

[0066] Step S102, using the attribute data of multiple users to construct an initial feature matrix and an initial adjacency matrix, and preprocessing the initial feature matrix and the initial adjacency matrix to obtain an improved feature matrix and an improved adjacency matrix.

[0067] In an embodiment of the present invention, the attribute data of a user presents the characteristics of a graph structure. Each node in the graph structure represents a user, and the edges between different nodes represent the relationships between users. The graph structure can be expressed by a feature matrix and an adjacency matrix. The feature matrix represents the feature vectors of each node; the adjacency matrix represents the relationships between nodes.

[0068] In an embodiment of the present invention, the attribute data of all users included in the information classification request is encoded to construct an initial feature matrix X and an initial adjacency matrix A of the information classification request. Among them, when constructing the initial feature matrix, the age, gender, contact information, address, occupation, transaction method, etc. of N users are encoded for C items of attributes to obtain an initial feature matrix X with a dimension of N×C. Each element X of the initial feature matrix X nc That is, it represents the encoded value of the c-th attribute of the n-th node, or rather, the encoded value of the c-th attribute of the n-th user; wherein, n∈1,2,…,N, c∈1,2,…,C. When encoding the attribute data, different age ranges correspond to different encoded values. For example, 10 - 20 is encoded as 01, 21 - 30 is encoded as 02, 31 - 40 is encoded as 03, 41 - 50 is encoded as 04, 51 - 60 is encoded as 05, 61 - 70 is encoded as 06, etc.; different genders correspond to different encoded values. For example, male is encoded as 0, female is encoded as 1; the address is encoded according to province - city - district. For example, the codes of provinces include 01, 02, …, 31, the codes of cities include 01, 02, …, 15, etc., and the codes of districts include 01, 02, …, 10, etc. The corresponding encoded values can be 251008, 130510, 050403, etc.; different transaction methods correspond to different encoded values. For example, bank card payment is encoded as 1, online banking payment is encoded as 2, mobile banking payment is encoded as 3, etc.

[0069] When constructing the initial adjacency matrix, the attribute data of N users is compared and analyzed to determine the relationship weights between each user, obtaining the initial adjacency matrix A with dimensions N×N. Each element A of the initial adjacency matrix A ij is the relationship weight value between the i-th node and the j-th node, or rather, the relationship weight value between the i-th user and the j-th user; where i ∈ {1, 2, …, N} and j ∈ {1, 2, …, N}. When analyzing the attribute data of users, the initial relationship weight value between users can be comprehensively determined based on the follow-up relationship between users (such as new follow-up, unfollow, mutual follow-up), gift-giving relationship (such as comparison of recipients of different order records), transaction relationship (such as paying on behalf), etc., to obtain the initial adjacency matrix A.

[0070] In the embodiments of the present invention, as Figure 2 shown, the method for determining the improved feature matrix of the present invention includes the following steps:

[0071] Step S201: Determine the filter matrix according to the initial adjacency matrix and the first filtering parameter.

[0072] In the embodiments of the present invention, calculate the product of the difference between the degree matrix D of the initial adjacency matrix A A and the initial adjacency matrix A and the first filtering parameter m, and subtract the product of the difference between the degree matrix D A and the initial adjacency matrix A and the first filtering parameter m from the identity matrix I. The resulting difference is the filter matrix G, where the dimension of G is N×N; the identity matrix I is an N×N-dimensional matrix, and the elements on the diagonal of the identity matrix I are 1, and the other elements are 0; the value of the first filtering parameter m can be selectively set according to the actual application scenario. For example, m = 1.

[0073] Step S202: Stack multiple layers of the filter matrix to perform filtering processing on the initial feature matrix to obtain the improved feature matrix; where the number of layers of the filter matrix is the second filtering parameter.

[0074] In the embodiments of the present invention, the initial feature matrix X can be filtered to obtain the improved feature matrix X new , specifically, the product of the initial feature matrix and the stack of t layers of the filter G t (that is, the product of t filter matrices G) can be used as the feature matrix X new .

[0075] where X new has dimensions N×C; the value of the second filtering parameter t can be selectively set according to the actual application scenario.

[0076] In the embodiment of the present invention, by using the method for determining the improved feature matrix of the present invention, the filter matrix can be determined by using the filtering parameters and the initial adjacency matrix, and the initial feature matrix is filtered by stacking multiple filter matrices to obtain the improved feature matrix, so as to realize the smoothing process of the node features with strong correlation, further reflect the overall node features of the graph structure, and accurately divide the groups to which the users belong.

[0077] In the embodiment of the present invention, as shown in Fig. 3(a), the method for determining the improved adjacency matrix of the present invention includes the following steps:

[0078] Step S301, calculate the influence coefficients between each of the nodes according to the preset diffusion coefficient and the order relationship between the nodes in the initial adjacency matrix, and form an influence matrix.

[0079] Step S3011, determine the neighbor order between the nodes according to the connection paths between different nodes in the initial adjacency matrix.

[0080] In the embodiment of the present invention, analyze the connection relationship of the edges between the nodes in the initial adjacency matrix A, count the connection paths Z between different nodes, and respectively determine the order relationship between the nodes under each connection path z where Z represents the total number of paths between the nodes, z represents the path identifier of each path between the nodes, and z ∈ 1, 2, …, Z; That is, it represents the neighbor order of the jth node relative to the ith node under the zth path. In the case that there is only one connection path between the ith node and the jth node, calculate the number of nodes passed by the ith node to reach the jth node as the neighbor order of the jth node relative to the ith node For example, as shown in Fig. 3(b), there is only one connection path between node ① and node ③, and the number of nodes passed by node ① to reach node ③ is 2, then the neighbor order of node ③ relative to node ① is 2 orders.

[0081] Further, in the case that there are multiple connection paths between the ith node and the jth node, calculate the number of nodes passed by the ith node to reach the jth node under each connection path respectively as the multiple orders of the jth node relative to the ith node For example, as shown in Fig. 3(b), there are 5 connection paths between node ① and node ⑥, namely ①-②-③-⑤-⑥, ①-④-⑤-⑥, ①-④-⑦-⑤-⑥, ①-⑦-⑤-⑥, ①-⑦-④-⑤-⑥, and the neighbor orders of node ⑥ relative to node ① under the 5 connection paths are 4 orders, 3 orders, 4 orders, 3 orders, and 4 orders respectively.

[0082] It should be noted that the node's own neighbor order for it is 0.

[0083] Step S3012, calculate the power result of the difference between the diffusion constant and the diffusion coefficient, and multiply the power result by the diffusion coefficient as the influence coefficient between the nodes; wherein, the power of the power result is equal to the neighbor order between the nodes.

[0084] In the embodiment of the present invention, the diffusion constant t is 1; the value range of the diffusion coefficient α is between 0 and 1. The smaller the diffusion coefficient α, the slower the influence reduction between nodes. Among them, the value of the diffusion coefficient α can be selectively set according to the actual application scenario.

[0085] Furthermore, the influence coefficient of the j-th node relative to the i-th node is shown in the following formula (1):

[0086]

[0087] For example, the diffusion coefficient α can take values such as 0.05, 0.01, 0.1, etc. Assuming α is 0.05, as shown in Fig. 3(b), the influence coefficient of node ③ relative to node ① The influence coefficient of node ⑥ relative to node ①

[0088] Furthermore, the influence coefficients F between each node ij constitute an influence matrix F with a dimension of N×N.

[0089] Step S302, perform a Laplace transform on the initial adjacency matrix to obtain the transformed adjacency matrix of the initial adjacency matrix.

[0090] In the embodiment of the present invention, using the degree matrix D of the initial adjacency matrix A A , perform a Laplace transform on the initial adjacency matrix A to obtain the transformed adjacency matrix L, as shown in the following formula (2):

[0091]

[0092] In the above formula, D A can also be called the diagonal matrix of the initial adjacency matrix A, with the same dimension as the initial adjacency matrix A. The elements on the diagonal of the degree matrix D A are the sum of the row vectors of the initial adjacency matrix A, and the other elements are 0; the dimension of the transformed adjacency matrix L is N×N.

[0093] Step S303: Calculate the product of the influence matrix and the transformed adjacency matrix, and determine the improved adjacency matrix according to the Laplace transform result of the product of the influence matrix and the transformed adjacency matrix.

[0094] In the embodiment of the present invention, calculate the product of the influence matrix F and the corresponding transformed adjacency matrix L wherein, is of dimension N×N.

[0095] Further, perform a Laplace transform on the product of the influence matrix F and the transformed adjacency matrix L to obtain the improved adjacency matrix as shown in the following formula (3):

[0096]

[0097] In the above formula, the improved adjacency matrix is of dimension N×N.

[0098] In the embodiment of the present invention, through the method for determining the improved adjacency matrix of the present invention, it is possible to analyze the connection relationship between each node in the initial adjacency matrix, respectively determine the neighbor order between nodes under different paths, combine the diffusion coefficient to determine the influence coefficient between nodes, and obtain the influence matrix; perform a Laplace transform on the initial adjacency matrix to obtain the transformed adjacency matrix; finally multiply the influence matrix and the transformed adjacency matrix, and perform a Laplace transform to obtain the improved adjacency matrix. Compared with the existing initial adjacency matrix that only includes the local structure information of the first-order neighbors, the improved adjacency matrix of the present application can analyze the influence of all nodes on the current node. Furthermore, when inputting subsequent neural network operations, under the same network structure, it can maximize the extraction of comprehensive global information and accurately divide the groups to which users belong.

[0099] Step S103: Input the improved feature matrix and the improved adjacency matrix into a node feature optimization model. The node feature optimization model uses a graph convolutional neural network and takes the mutual information between the local information and the global information of each node being less than a preset loss threshold as the optimization objective, and determine the optimized feature matrix according to the output of the node feature optimization model.

[0100] In the embodiment of the present invention, the node feature optimization model uses the improved feature matrix and the improved adjacency matrix to determine the local information of each node and the global information of all nodes, takes the mutual information between the local information and the global information as the objective function, roughly predicts the groups to which users belong, and uses the obtained prediction result as the optimized feature matrix, so that the node features of each node have obvious distinguishability, improve the uniqueness of the node features, and thus obtain excellent node representations.

[0101] Furthermore, the node feature optimization model uses a graph convolutional neural network to aggregate the local information of each node, pool each vector of the local information matrix to obtain global information, and iteratively calculate based on the mutual information between the local information and the global information, as well as the mutual information between the negative samples of the local information and the global information, until the target weight matrix of the GCN network is determined, and an optimized feature matrix is obtained.

[0102] In the embodiment of the present invention, as Figure 4 shown, the method for determining the optimized feature matrix of the present invention includes the following steps:

[0103] Step S401, set an initial current optimized weight matrix.

[0104] In the embodiment of the present invention, the initial current optimized weight matrix W op can be set arbitrarily, or the current optimized weight matrix W is set after analyzing the improved adjacency matrix op , so as to improve the iteration speed, shorten the iteration time, and thus save computing resources.

[0105] Furthermore, the dimension of the current optimized weight matrix W op is C×C, so that the dimension of the output optimized feature matrix Q is N×C, ensuring the correct division of the user information classification model.

[0106] Step S402, perform convolution on the improved adjacency matrix, the improved feature matrix, and the current optimized weight matrix to obtain a current optimized convolution result.

[0107] In the embodiment of the present invention, the node feature optimization model uses a graph convolutional neural network, and the input of the graph convolutional neural network is the improved adjacency matrix the improved feature matrix X new and the current optimized weight matrix W op , and the obtained current optimized convolution result E op is shown in the following formula (4):

[0108]

[0109] In the above formula, the dimension of the current optimized convolution result E op is N×C.

[0110] Furthermore, the current optimized convolution result E op represents the aggregation of the local information of all nodes.

[0111] Step S403, use a non-linear activation function to transform the current optimized convolution result to obtain a current optimized pooling result.

[0112] In the embodiment of the present invention, the graph convolutional neural network uses a non-linear activation function to activate the sum of each row vector included in the current optimized convolution result E op to obtain the current optimized pooling result as shown in the following formula (5):

[0113]

[0114] In the above formula, the current optimized pooling result has a dimension of C×1;

[0115] The non-linear activation function σ() can be selectively set as needed. For example, the ReLu activation function (Rectified Linear Unit), the Sigmoid activation function (or the logistic function), etc.

[0116] Furthermore, the current optimized pooling result represents the aggregation of the global information of all nodes.

[0117] Step S404: Use the current optimized convolution result as a positive sample, and reassign the positive sample to generate a negative sample corresponding to the positive sample.

[0118] In the embodiment of the present invention, the current optimized convolution result E op is used as the positive sample E op , and the reassignment method of the positive sample E op can be shuffle processing, that is, randomly shuffle each element of the positive sample E op to generate the corresponding negative sample E op′ .

[0119] Furthermore, the reassignment method can be any method as long as the result of the reassignment is a feature matrix different from the positive sample E op .

[0120] Step S405: Calculate the first mutual information between the positive sample and the current optimized pooling result, and the second mutual information between the negative sample and the current optimized pooling result respectively, and superimpose the first mutual information and the second mutual information as the loss function of the node feature optimization model.

[0121] In the embodiment of the present invention, calculate the logarithm of the first mutual information between the positive sample E op and the current optimized pooling result and the logarithm of the second mutual information between the negative sample E op′ and the current optimized pooling result respectively, sum the logarithm of the first mutual information and the logarithm of the difference between 1 and the second mutual information, and divide by the positive sample E opThe ratio to the sum of the number of nodes corresponding to the negative sample E op′ is used as the loss function S, as shown in the following formula (6):

[0122]

[0123] In the above formula, represents the first mutual information between the positive sample E op and the current optimized pooling result that aggregates global information , represents the second mutual information between the negative sample E op′ and the current optimized pooling result ,

[0124] Step S406, calculate the current loss value of the loss function, and determine whether the current loss value is less than or equal to a preset loss threshold. If so, go to step S407; if not, go to step S409.

[0125] In the embodiment of the present invention, according to the above formula (6), calculate the current loss value S of the loss function, and determine whether the current loss value S is less than or equal to the preset loss threshold S t . Among them, the preset loss threshold S t can be 0.05.

[0126] Step S407, use the current optimized weight matrix as the target optimized weight matrix.

[0127] In the embodiment of the present invention, when the current loss value S is less than or equal to the preset loss threshold S t , the corresponding current optimized weight matrix W op is used as the target optimized weight matrix That is

[0128] Step S408, calculate the current optimized convolution result of the target optimized weight matrix to obtain the optimized feature matrix.

[0129] In the embodiment of the present invention, according to the target optimized weight matrix convolve the improved adjacency matrix the improved feature matrix B and the target optimized weight matrix again, and use the obtained optimized convolution result as the optimized feature matrix Q, that is,

[0130] Step S409, update the current optimized weight matrix according to the difference between the current loss value and the preset loss threshold, and go to step S402.

[0131] In an embodiment of the present invention, when the current loss value S is greater than the preset loss threshold S t , according to the difference between the current loss value S and the preset loss threshold S t , the current optimization weight matrix W is updated op , and it goes to step S402 for iterative calculation.

[0132] In an embodiment of the present invention, through the method for determining the optimized feature matrix of the present invention, GCN can be used to perform convolution and pooling operations on the improved feature matrix and the improved adjacency matrix, obtain the local information and global information of each node, determine the loss function by calculating the mutual information between the pooling result and the positive and negative samples, and perform iterative calculation until the target optimization weight matrix is obtained. After convolution processing, the final optimized feature matrix is obtained. Thus, according to the mutual information between the local and global parts of the graph, the optimal node features representing the comprehensive information of the complex network corresponding to the user can be obtained, so that each node can have obvious distinguishability, which further facilitates the accurate division of user groups.

[0133] Step S104: Input the optimized feature matrix and the improved adjacency matrix into the user information classification model to determine the target group of each user; wherein, the user information classification model uses the sum of modularity and a regularization term for correcting the modularity as the objective function.

[0134] In an embodiment of the present invention, the user information classification model uses a graph convolutional neural network, takes the sum of modularity and the regularization term as the objective function, takes the maximum value of the objective function as the solution target, solves the target division weight matrix, and then substitutes it into the graph convolutional neural network to obtain the division result of the final target group to which each user belongs.

[0135] In an embodiment of the present invention, as Figure 5 shown, the method for determining the target group of the user of the present invention includes the following steps:

[0136] Step S501: Set an initial current division weight matrix.

[0137] In an embodiment of the present invention, the initial current division weight matrix W gd can be set arbitrarily, or the current division weight matrix W is set after analyzing the improved adjacency matrix gd to improve the iteration speed, shorten the iteration time, and thus save computing resources.

[0138] Furthermore, the dimension of the current division weight matrix W gd is C×C gd , and the value of C gd can be adjusted as needed during the iteration process.

[0139] Step S502: Convolve the improved adjacency matrix, the optimized feature matrix, and the current partition weight matrix to obtain the current partition convolution result.

[0140] In the embodiment of the present invention, the user information classification model adopts a graph convolutional neural network, and the input of the graph convolutional neural network is the improved adjacency matrix the optimized feature matrix Q and the current partition weight matrix W gd , and the obtained current partition convolution result wherein, the current partition convolution result E gd has a dimension of N×C gd .

[0141] Step S503: Calculate the first norm of the improved adjacency matrix.

[0142] In the embodiment of the present invention, the first norm of the improved adjacency matrix is the sum of the elements of the improved adjacency matrix , that is

[0143] Step S504: Use the difference obtained by subtracting the ratio of the matrix product of the degree matrix of the improved adjacency matrix and the transpose matrix of the degree matrix of the improved adjacency matrix to the first norm from the improved adjacency matrix as the transition matrix.

[0144] In the embodiment of the present invention, the degree matrix of the improved adjacency matrix is According to the matrix product of the degree matrix of the improved adjacency matrix and its transpose matrix multiplied, and the ratio of the matrix product to the first norm of the improved adjacency matrix , subtract the ratio of the matrix product of the transpose matrix of the degree matrix to the first norm from the improved adjacency matrix to obtain the difference as the transition matrix B.

[0145] Step S505: Take the trace of the result of multiplying the current partition convolution result, the transition matrix, and the transpose matrix of the current partition convolution result, and divide the trace result by the ratio of the first norm and take the negative value to obtain the modularity.

[0146] In the embodiment of the present invention, first calculate the current partition convolution result E gd , the transition matrix B, and the transpose matrix of the current partition convolution result E gd (Egd ) T The trace of the multiplication, divide the trace result by the improved adjacency matrix The first norm of After taking the negative process, the modularity M is obtained.

[0147] Step S506, generate the regularization term of the modularity according to the current partition convolution result and the column dimension of the current partition weight matrix.

[0148] In the embodiment of the present invention, the current partition weight matrix W gd has a column dimension of C gd , take the reciprocal of the column dimension C gd as a constant term, and correct each element of the current partition convolution result E gd to obtain the regularization term R of the modularity M, as shown in the following formula (7): In the above formula, for the i-th node, subtract the sum of the j = 1 to j = C

[0149]

[0150] matrix elements from the constant term, and take the square of the difference as the correction amount for the i-th node; the sum of the correction amounts for the i = 1 to i = N nodes is the regularization term R. gd

[0151] Step S507, introduce a harmonic parameter, and form the objective function of the user information classification model from the product of the modularity, the harmonic parameter and the regularization term.

[0152] In the embodiment of the present invention, sum the product of the modularity M, the harmonic parameter λ and the regularization term R to obtain the objective function S' of the user information classification model.

[0153] Among them, the value of the harmonic parameter λ can be selectively set according to the actual application scenario, or the value of the harmonic parameter λ can be determined by comparative experiments according to the actual application scenario.

[0154] Step S508, calculate the current function value of the objective function, and continuously adjust the current partition weight matrix until the current function value is maximized.

[0155] Step S509, use the current partition weight matrix corresponding to the maximum value of the current function value as the target partition weight matrix.

[0156] In the embodiment of the present invention, during the process of calculating the current function value S' of the objective function, continuously adjust the current partition weight matrix W gd , and use different current partition weight matrices W gd ​Compare with the current function value S' below until the maximum value S' of the current function value S' is obtained max , take S' max The corresponding current partition weight matrix W gd As the target partition weight matrix

[0157] Step S510, calculate the current partition convolution result of the target partition weight matrix to obtain the group partition result of each user's group; wherein, the group partition result includes the target group to which each user belongs.

[0158] In the embodiment of the present invention, according to the target partition weight matrix Convolve the improved adjacency matrix again Optimize the feature matrix Q and the target optimization weight matrix Perform convolution, and take the obtained optimized convolution result As the group partition result U of user information g , that is

[0159] In the embodiment of the present invention, through the classification method of user information of the present invention, it is possible to use GCN to perform convolution processing on the optimized feature matrix and the improved adjacency matrix, determine the modularity according to the first norm of the improved adjacency matrix and the convolution result, correct the modularity, form an objective function from the modularity and the correction term, and determine the target group to which the user belongs according to the convolution result corresponding to the maximum value of the objective function, so that when the present invention is applied to a usage scenario with real labels, it can evaluate the real labels, improve the accuracy of group division. After determining the group to which each user belongs, subsequent personalized recommendations can be given to the user according to the user's target group, improving user satisfaction.

[0160] In an embodiment of the present invention, by receiving one or more information classification requests; wherein, the information classification requests include attribute data of multiple users; constructing an initial feature matrix and an initial adjacency matrix using the attribute data of the multiple users, preprocessing the initial feature matrix and the initial adjacency matrix to obtain an improved feature matrix and an improved adjacency matrix; inputting the improved feature matrix and the improved adjacency matrix into a node feature optimization model, the node feature optimization model using a graph convolutional neural network and taking the mutual information between the local information and the global information of each node being less than a preset loss threshold as the optimization objective, determining an optimized feature matrix according to the output of the node feature optimization model; inputting the optimized feature matrix and the improved adjacency matrix into a user information classification model to determine the target groups of each user; wherein, the user information classification model takes the sum of the modularity and a regularization term for correcting the modularity as the objective function and other steps, can accurately partition a complex network, obtain a target group partition result suitable for users, improve the authenticity and accuracy of the global partition, and further optimize the user personalized recommendation in practical applications.

[0161] Figure 6 is a schematic diagram of the main modules of a classification device for user information according to an embodiment of the present invention, as Figure 6 shown, the classification device 600 for user information of the present invention includes:

[0162] A receiving module 601, configured to receive one or more information classification requests; wherein, the information classification requests include attribute data of multiple users;

[0163] A preprocessing module 602, configured to construct an initial feature matrix and an initial adjacency matrix using the attribute data of the multiple users, preprocess the initial feature matrix and the initial adjacency matrix to obtain an improved feature matrix and an improved adjacency matrix;

[0164] A feature optimization module 603, configured to input the improved feature matrix and the improved adjacency matrix into a node feature optimization model, the node feature optimization model using a graph convolutional neural network and taking the mutual information between the local information and the global information of each node being less than a preset loss threshold as the optimization objective, determining an optimized feature matrix according to the output of the node feature optimization model;

[0165] A partitioning module 604, configured to input the optimized feature matrix and the improved adjacency matrix into a user information classification model to determine the target groups of each user; wherein, the user information classification model takes the sum of the modularity and a regularization term for correcting the modularity as the objective function.

[0166] In the embodiments of the present invention, through modules such as a receiving module, a preprocessing module, a feature optimization module, and a partitioning module, the complex network can be accurately partitioned to obtain a target group partitioning result suitable for users, improving the authenticity and accuracy of the global partitioning, and further optimizing the user personalized recommendation in practical applications.

[0167] Figure 7 The following shows an exemplary system architecture diagram suitable for the classification method of user information or the classification device of user information in the embodiments of the present invention. As Figure 7 shown, the exemplary system architecture of the classification method of user information or the classification device of user information in the embodiments of the present invention includes:

[0168] As Figure 7 shown, the system architecture 700 may include terminal devices 701, 702, 703, a network 704, and a server 705. The network 704 is used to provide a medium for communication links between the terminal devices 701, 702, 703 and the server 705. The network 704 may include various connection types, such as wired, wireless communication links, or fiber optic cables, etc.

[0169] Users can use the terminal devices 701, 702, 703 to interact with the server 705 through the network 704 to receive or send messages, etc. Various communication client applications can be installed on the terminal devices 701, 702, 703, such as group classification applications, shopping applications, web browser applications, search applications, instant messaging tools, email clients, social platform software, etc.

[0170] The terminal devices 701, 702, 703 can be various electronic devices with a display screen and supporting web browsing, including but not limited to smart phones, tablet computers, laptop portable computers, and desktop computers, etc.

[0171] The server 705 can be a server providing various services, such as a background management server that supports the group classification websites browsed by users using the terminal devices 701, 702, 703. The background management server can analyze and process data such as received information classification requests, and feedback the processing results (the target groups of users) to the terminal devices 701, 702, 703.

[0172] It should be noted that the classification method of user information provided in the embodiments of the present invention is generally executed by the server 705. Correspondingly, the classification device of user information is generally set in the server 705.

[0173] It should be understood that Figure 7 the numbers of terminal devices, networks, and servers in are merely illustrative. According to actual needs, there can be any number of terminal devices, networks, and servers.

[0174] Figure 8 is a schematic structural diagram of a computer system of a terminal device or a server suitable for implementing the embodiments of the present invention, as Figure 8 shown, the computer system 800 of the terminal device or the server according to the embodiments of the present invention includes:

[0175] A central processing unit (CPU) 801, which can perform various appropriate actions and processes according to a program stored in a read-only memory (ROM) 802 or a program loaded from a storage section 808 into a random access memory (RAM) 803. In the RAM 803, various programs and data required for the operation of the system 800 are also stored. The CPU 801, the ROM 802, and the RAM 803 are connected to each other via a bus 804. An input / output (I / O) interface 805 is also connected to the bus 804.

[0176] The following components are connected to the I / O interface 805: an input section 806 including a keyboard, a mouse, etc.; an output section 807 including a cathode ray tube (CRT), a liquid crystal display (LCD), etc. and a speaker, etc.; a storage section 808 including a hard disk, etc.; and a communication section 809 including a network interface card such as a LAN card, a modem, etc. The communication section 809 performs communication processing via a network such as the Internet. A drive 810 is also connected to the I / O interface 805 as required. A removable medium 811, such as a magnetic disk, an optical disk, a magneto-optical disk, a semiconductor memory, etc., is installed on the drive 810 as required, so that a computer program read from it can be installed into the storage section 808 as required.

[0177] Specifically, according to the embodiments disclosed by the present invention, the processes described above with reference to the flowcharts can be implemented as computer software programs. For example, the embodiments disclosed by the present invention include a computer program product, which includes a computer program carried on a computer-readable medium, and the computer program includes program codes for performing the methods shown in the flowcharts. In such an embodiment, the computer program can be downloaded and installed from a network via the communication section 809, and / or installed from the removable medium 811. When the computer program is executed by the central processing unit (CPU) 801, the above functions defined in the system of the present invention are executed.

[0178] It should be noted that the computer-readable medium shown in the present invention can be a computer-readable signal medium, a computer-readable storage medium, or any combination of the above two. A computer-readable storage medium can be, for example, but not limited to, an electrical, magnetic, optical, electromagnetic, infrared, or semiconductor system, apparatus, or device, or any combination of the above. More specific examples of a computer-readable storage medium can include, but are not limited to: an electrical connection with one or more wires, a portable computer disk, a hard disk, a random access memory (RAM), a read-only memory (ROM), an erasable programmable read-only memory (EPROM or flash memory), an optical fiber, a portable compact disk read-only memory (CD-ROM), an optical storage device, a magnetic storage device, or any suitable combination of the above. In the present invention, a computer-readable storage medium can be any tangible medium that contains or stores a program, and this program can be used by or in combination with an instruction execution system, apparatus, or device. In the present invention, a computer-readable signal medium can include a data signal propagated in a baseband or as part of a carrier wave, which carries computer-readable program code. Such a propagated data signal can take various forms, including but not limited to electromagnetic signals, optical signals, or any suitable combination of the above. A computer-readable signal medium can also be any computer-readable medium other than a computer-readable storage medium, and this computer-readable medium can send, propagate, or transmit a program for use by or in combination with an instruction execution system, apparatus, or device. The program code contained on a computer-readable medium can be transmitted using any appropriate medium, including but not limited to: wireless, wire, optical cable, RF, etc., or any suitable combination of the above.

[0179] The flowcharts and block diagrams in the accompanying drawings illustrate the possible architectures, functions, and operations of systems, methods, and computer program products according to various embodiments of the present invention. In this regard, each block in a flowchart or block diagram can represent a module, a program segment, or a part of code, and the above module, program segment, or part of code contains one or more executable instructions for implementing a specified logical function. It should also be noted that in some alternative implementations, the functions marked in the blocks can occur in a different order than that marked in the accompanying drawings. For example, two consecutive blocks shown can actually be executed substantially in parallel, and they can sometimes be executed in the reverse order, depending on the functions involved. It should also be noted that each block in a block diagram or flowchart, as well as the combination of blocks in a block diagram or flowchart, can be implemented by a dedicated hardware-based system for performing the specified functions or operations, or can be implemented by a combination of dedicated hardware and computer instructions.

[0180] The modules involved in the embodiments of the present invention can be implemented in software or in hardware. The described modules can also be provided in a processor. For example, it can be described as: a processor includes a receiving module, a preprocessing module, a feature optimization module, and a partitioning module. Among them, the names of these modules do not constitute a limitation to the module itself in some cases. For example, the partitioning module can also be described as "a module that inputs the optimized feature matrix and the improved adjacency matrix into a user information classification model to determine the target group of each user".

[0181] As another aspect, the present invention also provides a computer-readable medium. The computer-readable medium can be included in the device described in the above embodiments; or it can exist alone without being assembled into the device. The above computer-readable medium carries one or more programs. When the one or more programs are executed by the device, the device includes: receiving one or more information classification requests; wherein, the information classification requests include attribute data of multiple users; constructing an initial feature matrix and an initial adjacency matrix using the attribute data of the multiple users, preprocessing the initial feature matrix and the initial adjacency matrix to obtain an improved feature matrix and an improved adjacency matrix; inputting the improved feature matrix and the improved adjacency matrix into a node feature optimization model, the node feature optimization model uses a graph convolutional neural network and takes the mutual information between the local information and the global information of each node being less than a preset loss threshold as the optimization target, and determining an optimized feature matrix according to the output of the node feature optimization model; inputting the optimized feature matrix and the improved adjacency matrix into a user information classification model to determine the target group of each user; wherein, the user information classification model uses the sum of modularity and a regularization term for correcting the modularity as the objective function.

[0182] According to the technical solution of the embodiments of the present invention, compared with the existing modularity-guided community detection algorithm that cannot process complex networks with true labels, first, the adjacency matrix and the feature matrix are preprocessed at the information level to obtain a smoother improved feature matrix and an improved adjacency matrix, retaining important information features in the complex network; secondly, the feature matrix is further optimized to obtain an optimized feature matrix; finally, the modularity is corrected, and both the modularity and the regularization term are used as the objective function, and the optimized feature matrix is used for solution to obtain the optimal group to which the user belongs, obtaining a globally optimal result, avoiding the disadvantages of poor balance and low accuracy of local optimality, being able to accurately partition the complex network, obtaining a suitable target group partitioning result for the user, improving the authenticity and accuracy of global partitioning, and further optimizing user personalized recommendation in practical applications.

[0183] The above specific embodiments do not constitute a limitation to the protection scope of the present invention. Those skilled in the art should understand that various modifications, combinations, sub - combinations and substitutions can occur depending on design requirements and other factors. Any modifications, equivalent substitutions and improvements made within the spirit and principle of the present invention shall be included within the protection scope of the present invention.

Claims

1. A method for classifying user information, characterized in that, it includes: Receiving one or more information classification requests; wherein, the information classification request includes attribute data of multiple users; Constructing an initial feature matrix and an initial adjacency matrix using the attribute data of multiple users, preprocessing the initial feature matrix and the initial adjacency matrix to obtain an improved feature matrix and an improved adjacency matrix; Inputting the improved feature matrix and the improved adjacency matrix into a node feature optimization model, the node feature optimization model uses a graph convolutional neural network, and takes the mutual information between the local information and the global information of each node being less than a preset loss threshold as the optimization goal, and determines an optimized feature matrix according to the output of the node feature optimization model; Inputting the optimized feature matrix and the improved adjacency matrix into a user information classification model to determine the target group of each user; wherein, the user information classification model takes the sum of modularity and a regularization term for correcting the modularity as the objective function.

2. The method according to claim 1, characterized in that, determining the optimized feature matrix includes: Setting an initial current optimization weight matrix, performing convolution on the improved adjacency matrix, the improved feature matrix and the current optimization weight matrix to obtain a current optimization convolution result; Using a non-linear activation function to transform the current optimization convolution result to obtain a current optimization pooling result, and determining the loss function of the point feature optimization model according to the mutual information between the current optimization convolution result and the current optimization pooling result; Iteratively calculating the current loss value of the loss function, and when the current loss value is less than or equal to the preset loss threshold, taking the current optimization weight matrix as the target optimization weight matrix, calculating the current optimization convolution result of the target optimization weight matrix to obtain the optimized feature matrix.

3. The method according to claim 2, characterized in that, determining the loss function of the point feature optimization model according to the mutual information between the current optimization convolution result and the current optimization pooling result includes: Taking the current optimization convolution result as a positive sample, reassigning the positive sample to generate a negative sample corresponding to the positive sample; Calculating the first mutual information between the positive sample and the current optimization pooling result and the second mutual information between the negative sample and the current optimization pooling result respectively, and superimposing the first mutual information and the second mutual information as the loss function of the node feature optimization model.

4. The method according to claim 1, characterized in that, determining the objective function of the user information classification model includes: Setting an initial current partitioning weight matrix, performing convolution on the improved adjacency matrix, the optimized feature matrix and the current partitioning weight matrix to obtain a current partitioning convolution result; Calculating the first norm of the improved adjacency matrix, using the difference obtained by subtracting the matrix product of the degree matrix of the improved adjacency matrix and the transpose matrix of the degree matrix of the improved adjacency matrix from the improved adjacency matrix and dividing by the first norm as a transition matrix; Perform a trace operation on the result of multiplying the current partition convolution result, the transition matrix, and the transpose matrix of the current partition convolution result, and take the negative value of the ratio of the trace result divided by the first norm to obtain the modularity.

5. The method according to claim 4, wherein, further comprising: Taking the reciprocal of the column dimension of the current partition weight matrix as a constant term, and correcting each element of the current partition convolution result to obtain the regularization term of the modularity; Introduce a harmonic parameter, and form the objective function of the user information classification model from the product of the modularity, the harmonic parameter, and the regularization term.

6. The method according to claim 1, wherein, The preprocessing of the initial feature matrix and the initial adjacency matrix includes: Determine a filter matrix according to the initial adjacency matrix and the first filtering parameter; Stack multiple layers of the filter matrix, and perform filtering processing on the initial feature matrix to obtain the improved feature matrix; wherein, the number of layers of the filter matrix is the second filtering parameter.

7. The method according to claim 1, wherein, The preprocessing of the initial adjacency matrix includes: Calculate the influence coefficients between each node according to the preset diffusion coefficient and the order relationship between the nodes in the initial adjacency matrix, and form an influence matrix; Perform a Laplace transform on the initial adjacency matrix to obtain the transformed adjacency matrix of the initial adjacency matrix; Calculate the product of the influence matrix and the transformed adjacency matrix, and determine the improved adjacency matrix according to the Laplace transform result of the product of the influence matrix and the transformed adjacency matrix.

8. The method according to claim 7, wherein, The calculating the influence coefficients between each node according to the preset diffusion coefficient and the order relationship between the nodes in the initial adjacency matrix includes: Determine the neighbor order between the nodes according to the connection paths between different nodes in the initial adjacency matrix; Calculate the power result of the difference between the diffusion constant and the diffusion coefficient, and multiply the power result by the diffusion coefficient as the influence coefficient between the nodes; wherein, the power of the power result is equal to the neighbor order between the nodes.

9. A classification device for user information, wherein, comprising: A receiving module, configured to receive one or more information classification requests; wherein, the information classification requests include attribute data of multiple users; A preprocessing module, configured to construct an initial feature matrix and an initial adjacency matrix by using the attribute data of multiple users, and perform preprocessing on the initial feature matrix and the initial adjacency matrix to obtain an improved feature matrix and an improved adjacency matrix; A feature optimization module, configured to input the improved feature matrix and the improved adjacency matrix into a node feature optimization model, the node feature optimization model adopts a graph convolutional neural network, and takes the mutual information between the local information and the global information of each node being less than a preset loss threshold as an optimization target, and determine an optimized feature matrix according to the output of the node feature optimization model; A partitioning module, configured to input the optimized feature matrix and the improved adjacency matrix into a user information classification model to determine the target group of each user; wherein the user information classification model uses the sum of modularity and a regularization term for correcting the modularity as an objective function.

10. An electronic device for classifying user information, characterized in that it includes: one or more processors; a storage device for storing one or more programs, when the one or more programs are executed by the one or more processors, the one or more processors implement the method according to any one of claims 1-8.

11. A computer-readable medium, having a computer program stored thereon, characterized in that when the program is executed by a processor, the method according to any one of claims 1-8 is implemented.

Images