User tag optimization method and device, equipment and storage medium

Abstract

The application discloses a user label optimization method, device and equipment and a computer readable storage medium. The user label optimization method comprises the following steps: obtaining industry interest information corresponding to each terminal user every interval preset time; constructing an interest topology data set according to the industry interest information, wherein the interest topology data set comprises a user node, an industry node and a label node, and the interest topology data set is represented by an adjacency matrix and a feature matrix; inputting the adjacency matrix and the feature matrix into a graph auto-encoder to obtain a vector representation corresponding to each user node; performing common clustering on each user node according to the vector representation to generate a plurality of common clusters; determining a user node located at a clustering center in each common cluster; and optimizing user label information of other user nodes located in the same common cluster according to user label information of the user node located at the clustering center. The user label optimization method provided by the application can improve the accuracy of user labels.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of communication, in particular to a user label optimization method and device, equipment and computer readable storage medium. BACKGROUND

[0002] The existing 5G message delivery is mainly realized according to the corresponding user label of the terminal user, and the users under the related label are targeted for delivery. The user label is generally defined by the user himself or obtained by the operator through big data. Therefore, the accuracy of the user label is directly related to the accuracy of the message delivery. However, the existing user label accuracy is not high, resulting in low accuracy of industry message delivery.

[0003] Therefore, there is an urgent need for a new user label optimization method, device, equipment and computer readable storage medium to solve the above technical problems. SUMMARY

[0004] The main purpose of the present application is to provide a user label optimization method, device, equipment and computer readable storage medium, which aims to solve the technical problem of low accuracy of the existing user label.

[0005] To achieve the above purpose, the present application provides a user label optimization method, comprising the following steps:

[0006] Obtain the industry interest information corresponding to each terminal user every interval preset time, and construct an interest topology data set according to the industry interest information, wherein the interest topology data set comprises user nodes, industry nodes and label nodes, and the interest topology data set is characterized by an adjacency matrix and a feature matrix;

[0007] Input the adjacency matrix and the feature matrix into a graph autoencoder to obtain a vector representation corresponding to each user node;

[0008] Commonly cluster each user node according to the vector representation to generate a plurality of common clusters;

[0009] Determine the user node located at the cluster center in each common cluster, and optimize the user label information of other user nodes located in the same common cluster according to the user label information of the user node located at the cluster center.

[0010] Optionally, the step of obtaining the industry interest information corresponding to each terminal user every interval preset time and constructing an interest topology data set according to the industry interest information comprises:

[0011] Obtain the user label information and industry interest information corresponding to each terminal user every interval preset time, and the static attribute information of the user label, terminal user and industry interest;

[0012] generate a user node corresponding to the end user, an industry node corresponding to the industry interest information, and a label node corresponding to the user label according to the user label information, the industry interest information, and the attribute information;

[0013] calculate a connection relationship between the user node, the industry node, and the label node according to the user label information, the industry interest information, and the attribute information, and generate an adjacency matrix according to the connection relationship;

[0014] generate a feature description corresponding to each of the user node, the industry node, and the label node according to the user label information, the industry interest information, and the attribute information, and generate a feature matrix according to the feature description;

[0015] generate an interest topology dataset with the end user as a center node according to the adjacency matrix and the feature matrix.

[0016] Optionally, the step of calculating a connection relationship between the user node, the industry node, and the label node according to the user label information, the industry interest information, and the attribute information, and generating an adjacency matrix according to the connection relationship comprises:

[0017] calculating a connection relationship between the user node, the industry node, and the label node according to the user label information, the industry interest information, and the attribute information;

[0018] performing one-hot encoding processing on the connection relationship, and generating an adjacency matrix according to the connection relationship that has been subjected to one-hot encoding processing.

[0019] Optionally, the step of generating a feature description corresponding to each of the user node, the industry node, and the label node according to the user label information, the industry interest information, and the attribute information, and generating a feature matrix according to the feature description comprises:

[0020] generating a feature description corresponding to each of the user node, the industry node, and the label node according to the user label information, the industry interest information, and the attribute information;

[0021] setting the longest text length in the feature description as a serialization length, and performing padding processing on each of the feature descriptions to generate the feature description of the serialization length;

[0022] generating a feature matrix according to the feature description of the serialization length.

[0023] Optionally, the step of inputting the interest topology dataset into a graph autoencoder to obtain a vector representation corresponding to each end user comprises:

[0024] inputting the interest topology dataset into a graph autoencoder to output a reconstructed adjacency matrix, and performing iterative training according to the error between the reconstructed adjacency matrix and the real adjacency matrix until the training is completed.

[0025] According to the trained graph auto-encoder, vector representations corresponding to each user node are obtained.

[0026] Optionally, the step of determining user nodes at cluster centers in each of the co-class clusters and optimizing user label information of other user nodes in the same co-class cluster according to the user label information of the user nodes at the cluster centers comprises:

[0027] determining user nodes at cluster centers in each of the co-class clusters according to user label information of the user nodes at the cluster centers;

[0028] associating the user label information of the user nodes at the cluster centers with other user nodes in the same co-class cluster.

[0029] Optionally, the step of determining user nodes at cluster centers in each of the co-class clusters and optimizing user label information of other user nodes in the same co-class cluster according to the user label information of the user nodes at the cluster centers comprises:

[0030] obtaining the number of user labels corresponding to each of the user nodes in each of the co-class clusters, and determining a recommended label set corresponding to each of the co-class clusters according to the number of user labels;

[0031] sending the recommended label set to each terminal user corresponding to the co-class cluster;

[0032] receiving a selected label fed back by a user according to the recommended label set, and storing the selected label associated with the terminal user.

[0033] In addition, to achieve the above object, the application further provides a user label optimization device, which comprises:

[0034] a data set generation module configured to obtain industry interest information corresponding to each terminal user every interval of a preset time, and construct an interest topology data set according to the industry interest information, wherein the interest topology data set comprises user nodes, industry nodes and label nodes, and the interest topology data set is characterized by an adjacency matrix and a feature matrix;

[0035] a graph auto-encoding module configured to input the adjacency matrix and the feature matrix into a graph auto-encoder, and obtain vector representations corresponding to each user node;

[0036] a clustering module configured to perform co-class clustering on each of the user nodes according to the vector representations, and generate a plurality of co-class clusters;

[0037] An optimization module is configured to determine a user node at a cluster center in each of the common clusters, and optimize user label information of other user nodes in the same common cluster according to user label information of the user node at the cluster center.

[0038] In addition, to achieve the above object, the present application also provides a user label optimization device, which comprises a memory, a processor, and a user label optimization program stored in the memory and executable on the processor, and the user label optimization program implements the steps of the user label optimization method when executed by the processor.

[0039] In addition, to achieve the above object, the present application also provides a computer readable storage medium, which stores a user label optimization program, and the user label optimization program implements the steps of the user label optimization method when executed by a processor.

[0040] In the present application, the spatial feature characteristics of the interest topology dataset can be effectively extracted by using the graph autoencoder, so as to extract the interest topology representation corresponding to the user node, and different weights are assigned to the influence of the neighbor nodes in the interest topology dataset by combining the attention mechanism, and then the user nodes in which the industry interest information is aggregated are commonly clustered. Therefore, the user label information of the user node at the cluster center can be used to optimize the user label information of other user nodes in the same common cluster, so that the user label setting is more accurate. The label accuracy of the 5G message terminal user can be further improved, and the 5G industry message delivery is more accurate. BRIEF DESCRIPTION OF DRAWINGS

[0041] Figure 1 is a terminal\device structure schematic diagram of a hardware running environment involved in the embodiment scheme of the present application;

[0042] Figure 2 is a flowchart of the first embodiment of the user label optimization method of the present application;

[0043] Figure 3 is a detailed flowchart of step S50 in the user label optimization method of the present application;

[0044] Figure 4 is a detailed flowchart of step S10 in the user label optimization method of the present application;

[0045] Figure 5 is an architecture schematic diagram of the interest topology dataset in the user label optimization method of the present application;

[0046] Figure 6 is a device module schematic diagram of the user label optimization device of the present application.

[0047] The objectives, functional characteristics and advantages of the present application will be further illustrated in conjunction with embodiments, with reference to the accompanying drawings. DETAILED DESCRIPTION

[0048] It should be understood that the specific embodiments described herein merely exemplify the present application and do not limit the present application.

[0049] As shown in Figure 1 , the terminal structure diagram of the hardware running environment involved in the embodiment scheme of the present application. Figure 1

[0050] The terminal of the embodiment of the present application is a user tag optimization device.

[0051] As shown in Figure 1 , the terminal can include a processor 1001, such as a CPU, a network interface 1004, a user interface 1003, a memory 1005, and a communication bus 1002. The communication bus 1002 is used to realize the connection and communication between these components. The user interface 1003 can include a display screen (Display), an input unit such as a keyboard (Keyboard), and the optional user interface 1003 can also include a standard wired interface, a wireless interface. The network interface 1004 can optionally include a standard wired interface, a wireless interface (such as a WI-FI interface). The memory 1005 can be a high-speed RAM memory, or a stable memory (non-volatile memory), such as a magnetic disk memory. The memory 1005 can also be an independent storage device from the aforementioned processor 1001.

[0052] Those skilled in the art can understand that Figure 1 the terminal structure shown in the foregoing is not a limitation on the terminal, and can include more or fewer components than the diagram, or combine certain components, or different component arrangements.

[0053] As shown in Figure 1 , the memory 1005 as a computer storage medium can include an operating system, a network communication module, a user interface module, and a user tag optimization program.

[0054] In the terminal shown in Figure 1 , the network interface 1004 is mainly used to connect to a background server and communicate data with the background server; the user interface 1003 is mainly used to connect to a client (user end) and communicate data with the client; and the processor 1001 can be used to call the user tag optimization program stored in the memory 1005 and perform the following operations:

[0055] ​Obtain industry interest information corresponding to each terminal user every preset time interval, and construct an interest topology dataset according to the industry interest information, wherein the interest topology dataset comprises user nodes, industry nodes and label nodes, and the interest topology dataset is characterized by an adjacency matrix and a feature matrix;

[0056] Input the adjacency matrix and the feature matrix into a graph autoencoder to obtain a vector representation corresponding to each user node;

[0057] Commonly cluster each user node according to the vector representation to generate a plurality of common clusters;

[0058] Determine a user node located at a cluster center in each common cluster, and optimize user label information of other user nodes in the same common cluster according to user label information of the user node located at the cluster center.

[0059] Reference Figure 2 The present application provides a user label optimization method, in a first embodiment of the user label optimization method, the user label optimization method comprises the following steps:

[0060] Step S10, obtain industry interest information corresponding to each terminal user every preset time interval, and construct an interest topology dataset according to the industry interest information, wherein the interest topology dataset comprises user nodes, industry nodes and label nodes, and the interest topology dataset is characterized by an adjacency matrix and a feature matrix;

[0061] In the present application, the industry interest information corresponding to each terminal user is obtained through a 5G message open platform. The following specifically describes the user label optimization method provided by the present application by taking 5G message delivery as an example. It can be understood by those skilled in the art that the user label optimization method provided by the present application can also be used in other targeted information pushing or other user portrait label optimization processes.

[0062] The 5G message service is based on the original short message entry of the terminal, and provides sending and receiving of media content such as text, pictures, audio, video, location and contacts for users. The existing 5G message provides enhanced personal and application message service for industry customers, and introduces a new message interaction mode - Chatbot chat robot. Everyone can intuitively and conveniently enjoy various 5G application services such as payment and recharge, ticket ordering, hotel reservation, logistics inquiry, catering reservation and take-out ordering in the message window. The Chatbot is a service provided by an industry customer to a terminal user in the form of a dialogue. The service is usually based on artificial intelligence software, simulates human intelligent dialogue, and provides specific service functions to users. The 5G message open platform is used to assist industry customers to realize communication in multiple scenarios on demand, and the industry customers can quickly complete the deployment of the message application through the platform.

[0063] The industry interest information corresponding to each terminal user can be user tag information associated with the terminal user, industry interest information followed by the terminal user, and industry interest information refused to follow, and the like. As shown in the following table, in the embodiment, three types of nodes are constructed, namely user nodes, industry nodes, and tag nodes, wherein the user nodes correspond to the terminal users, the industry nodes correspond to the industry interest information, and the user tags correspond to the tag nodes. Figure 3

[0064] The interest topology dataset is represented by an adjacency matrix and a feature matrix, that is, the connection relationship of each node in the interest topology dataset is represented in the form of a matrix through the adjacency matrix, and the feature description of each node in the interest topology dataset is performed through the feature matrix.

[0065] In step S20, the adjacency matrix and the feature matrix are input into a graph autoencoder to obtain a vector representation corresponding to each user node.

[0066] The graph autoencoder is a calculation model based on a graph attention network. By introducing an attention mechanism in the propagation layer, different weights are assigned to different adjacent nodes in the aggregation process of the center node features, and the attention of the adjacent nodes is different, so that some slice nodes with relatively large correlation are focused on, and some slice nodes with relatively small correlation are ignored. In the graph autoencoder, an encoder and a decoder are set. The encoder extracts the interest topology space features of the user node corresponding to a time period by using a graph convolutional neural network, projects the node and the connection relationship into a low-dimensional vector space, and the decoder reconstructs the interest topology dataset by calculating the inner product of the node pairs. Then, the vector representation of the user node is selected from the interest topology latent space vector representation of the user node output by the trained decoder.

[0067] In step S30, the vector representation is used to cluster each user node according to the commonality, and a plurality of common clusters are generated.

[0068] Specifically, each user node can be assumed to be a clustering center in turn, and the probability that each node belongs to a certain common category is calculated, so as to obtain a plurality of common clusters. A common cluster includes a plurality of user nodes, one of which is a clustering center.

[0069] In step S40, the user node located at the clustering center in each common cluster is determined, and the user tag information corresponding to the user node located at the clustering center is used to optimize the user tag information corresponding to the other user nodes in the same common cluster.

[0070] In an embodiment, step S40 includes:

[0071] ​Step a, determining a user node at a cluster center in each of the co-class clusters, and obtaining user label information corresponding to the user node at the cluster center;

[0072] Step b, associating the user label information corresponding to the user node at the cluster center with other user nodes in the same co-class cluster.

[0073] In the present application, the spatial feature characteristics of the interest topology dataset can be effectively extracted by using the graph autoencoder, thereby extracting the interest topology representation corresponding to the user node, and different weights are assigned to the differences in the influence of the neighbor nodes in the interest topology dataset in combination with the attention mechanism, and then the user nodes in which the industry interest information is aggregated are clustered. Therefore, the user label information corresponding to the user nodes in the same co-class cluster can be optimized by the user label information corresponding to the user nodes at the cluster center, so that the user label setting is more accurate. The label accuracy of the 5G message terminal user can be further improved, and the 5G industry message delivery is more accurate.

[0074] Please refer to Figure 4 In the second embodiment of the user label optimization method, step S10 includes:

[0075] Step S11, obtaining user label information and industry interest information corresponding to each terminal user, and static attribute information of the user label, terminal user and industry interest at a predetermined interval;

[0076] The static attribute information of the user label can specifically include the user label name, the label type to which it belongs, and whether the label source is selected by the terminal user or obtained by the operator through big data; The static attribute information of the terminal user can specifically include the opening location of the terminal user, the online time, the current registered location, the monthly average revenue (ARPU), etc.; The static attribute information of the industry interest (chatbot) includes the industry to which the industry user belongs, the user group it faces, etc. The relationship between different nodes can include the relationship between the user label and the terminal user, and the relationship between the terminal user and the industry interest, such as the relationship that the terminal user follows or refuses one or more industry interests.

[0077] Step S12, generating a user node corresponding to the terminal user, an industry node corresponding to the industry interest information, and a label node corresponding to the user label according to the user label information, the industry interest information and the attribute information;

[0078] Step S13, calculating the connection relationship between the user node, the industry node and the label node, and generating an adjacency matrix according to the connection relationship;

[0079] Specifically, step S13 includes:

[0080] Step c, according to the user label information, industry interest information and attribute information, the connection relationship between the user node, industry node and label node is calculated;

[0081] Step d, the connection relationship is processed by one-hot encoding, and the adjacency matrix is generated according to the connection relationship processed by one-hot encoding.

[0082] Specifically, e ij represents the connection relationship between nodes V i and V j , eij=(vi,r,vj), where r is the type of edge, that is, the edge between nodes V i and V j . In this embodiment, the connection relationship includes the affiliation relationship, the attention relationship, the rejection relationship and no connection. Four kinds of connection relationships are one-hot encoded to form an adjacency matrix with the shape of N*N, N is the number of nodes.

[0083] After step d, step S14 is executed: according to the user label information, industry interest information and attribute information, the feature description corresponding to each user node, industry node and label node is generated, and the feature matrix is generated according to the feature description;

[0084] Specifically, step S14 includes:

[0085] Step e, according to the user label information, industry interest information and attribute information, the feature description corresponding to each user node, industry node and label node is generated;

[0086] Step f, set the longest text length in the feature description as the serialization length, and fill each feature description to generate the feature description with the serialization length;

[0087] Step g, generate a feature matrix according to the feature description with the serialization length.

[0088] The length of the longest text in the data set is the serialization length F, and the length of each data is filled to F, and the dictionary size is K. An N*F feature matrix is generated.

[0089] Step S15, generate an interest topology data set with the end user as the center node according to the adjacency matrix and the feature matrix.

[0090] Through steps S11 to S15, the data is processed, and an interest topology data set containing user labels, industry interests and end user relationships is obtained.

[0091] Please refer to Figure 5 , in the third embodiment of the user label optimization method, step S40 includes:

[0092] In step S41, the number of user tag information corresponding to each user node in each of the co-class clusters is obtained, and a recommended tag set corresponding to each of the co-class clusters is determined according to the number of user tags.

[0093] In step S42, the recommended tag set is sent to each terminal user corresponding to the co-class cluster.

[0094] In step S43, the selected tag fed back by the user according to the recommended tag set is received, and the selected tag associated with the terminal user is stored.

[0095] Specifically, the user tags with the largest number or a preset number of larger numbers can be counted to form a recommended tag set, and the recommended tag set is sent to the terminal user, so that the user can select according to the recommended tag set and customize the tags of interest based on the recommendation, thereby further improving the accuracy of the user tags associated with the terminal user.

[0096] In the fourth embodiment of the user tag optimization method, step S30 includes:

[0097] In step h, the interest topology data set is input into the graph auto-encoder, the reconstructed adjacency matrix is output, and iterative training is performed according to the error between the reconstructed adjacency matrix and the real adjacency matrix until the training is completed.

[0098] In step i, the vector representation corresponding to each user node is obtained according to the trained graph auto-encoder.

[0099] In this embodiment, the graph auto-encoder (GAE) is divided into an encoder and a decoder. The encoder is composed of a graph convolutional neural network (GCN), and the decoder is responsible for performing an inner product operation to reconstruct the adjacency matrix.

[0100] Specifically, the encoder is used to output the latent space vector representation Z of the nodes and edges according to the input adjacency matrix A and feature matrix X. The first layer is an input layer for inputting the adjacency matrix A and the feature matrix X; the second layer is a graph attention layer, where the number of convolution kernels is 128 and the activation function is a relu function. The graph attention layer function can be represented as:

[0101]

[0102] where z i l is the vector representation of the ith node in the lth layer, and z i l+1 is the vector representation of the ith node in the l+1th layer that aggregates the features of the neighbor nodes, and Ni α represents the set of neighboring nodes corresponding to node i. ij Let σ represent the attention weight between node pairs (i, j), σ be the activation function, and W be the weight matrix for node feature transformation.

[0103] The third layer is a dropout layer, with a dropout probability of 0.2 in this embodiment. During training, the input neurons will be randomly disconnected with a certain probability of 20% each time the parameters are updated to prevent overfitting. The fourth layer is a graph attention layer with 64 convolutional kernels and a lambda activation function. The fifth layer is a dropout layer with a dropout probability of 0.2.

[0104] The decoder is used to output a reconstructed adjacency matrix based on the latent space vector representation Z of the user node interest topology at a certain time period. Specifically, the sixth layer is an inner product layer, which multiplies the latent space vector representation Z by its transpose and inputs it to the activation function. The activation function is set to the sifmoid function, which can be expressed as follows:

[0105]

[0106] Then calculate the reconstructed adjacency matrix and the actual adjacency matrix A ij The training objective is to minimize the error between the given values. That is, the loss function is:

[0107] Lr is the loss function of the graph autoencoder.

[0108] The reconstruction error of the graph autoencoder can be used to determine whether the graph autoencoder has converged. If not, iterative training continues. If so, the latent space vector representation Z of the message terminal user interest topology in the trained graph autoencoder is obtained. The vector representation of the user node is selected from the latent space vector representation Z and input into the clustering module for self-training. The clustering result of the commonality of terminal users in that time period is output. There are various clustering methods, and those skilled in the art can choose according to their own needs. In this embodiment, the following method is provided:

[0109] Assuming node μ is the cluster center, the probability q of user node i belonging to a certain common cluster μ is... iu Represented as:

[0110] Z i The vector representation of user node i, μ u μ represents the u-th cluster center. k This represents the k-th cluster center.

[0111] To introduce the clustering information to realize the clustering-oriented node representation, each node needs to be forced to be closer to the corresponding cluster center to realize the minimum intra-class distance and the maximum inter-class distance. Therefore, the target distribution p iu As follows:

[0112] Where q iu represents the probability that the user node i belongs to a certain common cluster μ, p iu represents the target distribution.

[0113] Finally, the mutual constraint, i.e., self-training, is realized by calculating the KL divergence between the two target distributions P and the assumed distribution Q, and the target is to minimize the KL divergence between the two distributions:

[0114] Lc represents the loss function of the clustering module.

[0115] Therefore, the final loss function of the model includes the loss functions of the two parts of the graph autoencoder and the clustering module:

[0116] L = L r + γL c .

[0117] In the embodiment, the number of training rounds is set to 2000 (epochs = 2000), and the gradient descent optimization algorithm selects the adam optimizer for improving the learning speed of the traditional gradient descent (optimizer = 'adam'). The neural network can find the optimal weight value that minimizes the objective function through gradient descent, and the neural network can autonomously learn the weight value through training. The training set is used for training, so that the objective function is as small as possible, and the test set is used to evaluate and verify the model after each round of training. After the model converges, the weight of the model is derived.

[0118] The certain common cluster s i to which the user node i belongs can be calculated by the following formula:

[0119]

[0120] In addition, with reference to Figure 6 , the embodiment of the present application also provides a user label optimization device, comprising:

[0121] A data set generation module 10 is configured to acquire industry interest information of each terminal user at each interval of a preset time, and construct an interest topology data set according to the industry interest information, wherein the interest topology data set comprises a user node, an industry node and a label node, and the interest topology data set is characterized by an adjacency matrix and a feature matrix;

[0122] a graph auto-encoding module 20, configured to input the adjacency matrix and the feature matrix into a graph auto-encoder to obtain a vector representation corresponding to each user node;

[0123] a clustering module 30, configured to cluster the user nodes according to the vector representation to generate a plurality of clusters;

[0124] an optimization module 40, configured to determine a user node at a cluster center in each of the clusters, and optimize user label information of other user nodes in the same cluster according to user label information of the user node at the cluster center.

[0125] Optionally, the data set generation module 10 is configured to:

[0126] acquire user label information and industry interest information corresponding to each terminal user, and static attribute information of the user label, the terminal user, and the industry interest, at a preset interval;

[0127] generate a user node corresponding to the terminal user, an industry node corresponding to the industry interest information, and a label node corresponding to the user label according to the user label information, the industry interest information, and the attribute information;

[0128] calculate a connection relationship between the user node, the industry node, and the label node, and generate an adjacency matrix according to the connection relationship;

[0129] generate a feature description corresponding to each of the user node, the industry node, and the label node according to the user label information, the industry interest information, and the attribute information, and generate a feature matrix according to the feature description;

[0130] generate an interest topology data set with the terminal user as a center node according to the adjacency matrix and the feature matrix.

[0131] Optionally, the data set generation module 10 is further configured to:

[0132] calculate a connection relationship between the user node, the industry node, and the label node according to the user label information, the industry interest information, and the attribute information;

[0133] perform one-hot encoding processing on the connection relationship, and generate an adjacency matrix according to the connection relationship subjected to the one-hot encoding processing.

[0134] Optionally, the data set generation module 10 is further configured to:

[0135] generate a feature description corresponding to each of the user node, the industry node, and the label node according to the user label information, the industry interest information, and the attribute information;

[0136] The longest text length in the feature description is set as a serialization length, and each feature description is filled to generate a feature description with the serialization length;

[0137] A feature matrix is generated according to the feature description with the serialization length.

[0138] Optionally, the graph auto-encoding module 20 is configured to:

[0139] The interest topology dataset is input into a graph auto-encoder, and a reconstructed adjacency matrix is output; iterative training is performed according to the error between the reconstructed adjacency matrix and the real adjacency matrix until the training is completed;

[0140] According to the trained graph auto-encoder, a vector representation corresponding to each user node is obtained.

[0141] Optionally, the optimization module 40 is configured to:

[0142] Determine a user node located at a cluster center in each of the co-class clusters, and determine user label information corresponding to the user node located at the cluster center;

[0143] The user label information corresponding to the user node located at the cluster center is associated with other user nodes located in the same co-class cluster.

[0144] Optionally, the optimization module 40 is further configured to:

[0145] Obtain the number of user labels corresponding to all user nodes in each of the co-class clusters, and determine a recommended label set corresponding to each of the co-class clusters according to the number of user labels;

[0146] The recommended label set is sent to each terminal user corresponding to the co-class cluster.

[0147] Receive a selected label fed back by a user according to the recommended label set, and store the selected label associated with the terminal user.

[0148] The processor is configured to execute the user label optimization program to implement the steps of each embodiment of the user label optimization method.

[0149] The application further provides a computer-readable storage medium storing one or more programs, which can be executed by one or more processors to implement the steps of each embodiment of the user label optimization method.

[0150] The computer-readable storage medium of the application has substantially the same implementation as the user label optimization method described above, and thus will not be described again.

[0151] It is to be understood that the terminology "including", "comprising", or any other variation thereof, is intended to cover a non-exclusive inclusion such that process, method, article, or apparatus that comprises a list of elements does not include only those elements but can also include other elements not expressly listed or inherent to such process, method, article, or apparatus. An element proceeded by "comprises a... " does not, without more constraints, exclude the presence of additional identical elements in the process, method, article, or apparatus that comprises the element.

[0152] The above-mentioned embodiment numbers of the present application are only for description, and do not represent the advantages and disadvantages of the embodiments.

[0153] Those skilled in the art can clearly understand the above-mentioned embodiment methods by means of software and the necessary general hardware platform, of course, they can also be realized by hardware, but in many cases, the former is a better embodiment. Based on such understanding, the technical solutions of the present application can be embodied in the form of a software product, which is stored in a storage medium (such as a ROM / RAM, a magnetic disk, or an optical disk) and includes a plurality of instructions for causing a terminal device (which can be a mobile phone, a computer, a server, an air conditioner, or a network device) to execute the methods described in the various embodiments of the present application.

[0154] The above is only the preferred embodiment of the present application, and does not limit the patent scope of the present application. Any equivalent structure or equivalent flow transformation, or direct or indirect application in other related technical fields, which is made by using the content of the specification and drawings of the present application, is also included in the patent protection scope of the present application.

Claims

1. A method of optimizing a user tag, the method comprising: The optimization method of the user label comprises the following steps: Obtain the industry interest information corresponding to each terminal user every preset time, and construct an interest topology data set according to the industry interest information, wherein the interest topology data set comprises user nodes, industry nodes and label nodes, and the interest topology data set is characterized by an adjacency matrix and a feature matrix; Input the adjacency matrix and the feature matrix into a graph autoencoder to obtain a vector representation corresponding to each user node; Common clustering is performed on each user node according to the vector representation to generate a plurality of common clusters; Determine the user nodes located at the cluster centers in each common cluster, and optimize the user label information of other user nodes located in the same common cluster according to the user label information of the user nodes located at the cluster centers; The step of obtaining the industry interest information corresponding to each terminal user every preset time and constructing an interest topology data set according to the industry interest information comprises: Obtain the user label information and the industry interest information corresponding to each terminal user every preset time, and the static attribute information of the user label, the terminal user and the industry interest; Generate the user nodes corresponding to the terminal users, the industry nodes corresponding to the industry interest information and the label nodes corresponding to the user labels according to the user label information, the industry interest information and the static attribute information; Calculate the connection relationship between the user nodes, the industry nodes and the label nodes, and generate an adjacency matrix according to the connection relationship; Generate the feature descriptions corresponding to each of the user nodes, the industry nodes and the label nodes according to the user label information, the industry interest information and the static attribute information; Set the longest text length in the feature description as a serialization length, fill and process each feature description to generate the feature description with the serialization length; Generate a feature matrix according to the feature description with the serialization length; Generate an interest topology data set with the terminal user as the center node according to the adjacency matrix and the feature matrix.

2. The method of claim 1, wherein, The step of calculating the connection relationship between the user nodes, the industry nodes and the label nodes, and generating an adjacency matrix according to the connection relationship comprises: Calculate the connection relationship between the user nodes, the industry nodes and the label nodes according to the user label information, the industry interest information and the static attribute information; Perform one-hot encoding processing on the connection relationship, and generate an adjacency matrix according to the connection relationship subjected to the one-hot encoding processing.

3. The method of claim 1, wherein, The step of inputting the adjacency matrix and the feature matrix into a graph autoencoder to obtain a vector representation corresponding to each user node comprises: Input the interest topology data set into a graph autoencoder, output a reconstructed adjacency matrix, perform iterative training according to the error between the reconstructed adjacency matrix and the real adjacency matrix until the training is completed, and obtain a vector representation corresponding to each user node according to the graph autoencoder subjected to the training. The step of determining the user nodes located at the cluster centers in each common cluster, and optimizing the user label information of other user nodes located in the same common cluster according to the user label information of the user nodes located at the cluster centers comprises:

4. The method of claim 1, wherein, ​ determine a user node at a cluster center in each of the co-class clusters according to user tag information corresponding to the user node at the cluster center; associate the user tag information corresponding to the user node at the cluster center with other user nodes in the same co-class cluster.

5. The method of claim 1, wherein, The step of determining a user node at a cluster center in each of the co-class clusters according to user tag information corresponding to the user node at the cluster center includes: obtaining the number of user tags corresponding to all user nodes in each of the co-class clusters, and determining a recommended tag set corresponding to each of the co-class clusters according to the number of user tags; sending the recommended tag set to each terminal user corresponding to the co-class cluster; receiving a selected tag fed back by a user according to the recommended tag set, and storing the selected tag associated with the terminal user.

6. An apparatus for optimizing a user tag, the apparatus comprising: The user tag optimization device includes: A dataset generation module is configured to obtain industry interest information corresponding to each terminal user at a preset interval of time, and construct an interest topology dataset according to the industry interest information. The interest topology dataset includes user nodes, industry nodes, and tag nodes. The interest topology dataset is represented by an adjacency matrix and a feature matrix. The dataset generation module is further configured to obtain user tag information and industry interest information corresponding to each terminal user at a preset interval of time, and static attribute information of the user tag, the terminal user, and the industry interest. The dataset generation module is further configured to generate user nodes corresponding to the terminal user, industry nodes corresponding to the industry interest information, and tag nodes corresponding to the user tag according to the user tag information, the industry interest information, and the static attribute information. The dataset generation module is further configured to calculate connection relationships among the user nodes, the industry nodes, and the tag nodes, and generate an adjacency matrix according to the connection relationships. The dataset generation module is further configured to generate feature descriptions corresponding to each of the user nodes, the industry nodes, and the tag nodes according to the user tag information, the industry interest information, and the static attribute information. The dataset generation module is further configured to set the longest text length in the feature descriptions as a serialization length, fill and process each of the feature descriptions to generate the feature descriptions of the serialization length, generate a feature matrix according to the feature descriptions of the serialization length, and generate the interest topology dataset with the terminal user as a center node according to the adjacency matrix and the feature matrix. A graph auto-encoder module is configured to input the adjacency matrix and the feature matrix into a graph auto-encoder to obtain vector representations corresponding to each user node. A clustering module is configured to perform co-class clustering on each of the user nodes according to the vector representations to generate a plurality of co-class clusters. An optimization module is configured to determine a user node at a cluster center in each of the co-class clusters according to user tag information corresponding to the user node at the cluster center, and optimize user tag information corresponding to other user nodes in the same co-class cluster according to the user tag information corresponding to the user node at the cluster center.

7. An apparatus for optimizing a user tag, the apparatus comprising: The user tag optimization device comprises a memory, a processor and a user tag optimization program stored on the memory and executable on the processor, the user tag optimization program, when executed by the processor, implements the steps of the user tag optimization method according to any one of claims 1 to 5.

8. A computer-readable storage medium, characterized in that, The computer readable storage medium stores a user tag optimization program, the user tag optimization program, when executed by the processor, implements the steps of the user tag optimization method according to any one of claims 1 to 5.

Images