Graph convolutional neural network-based state grid user classification method, system and device and storage medium

Abstract

The invention relates to the technical field of big data processing, in particular to an online state grid user classification method, system and device based on a graph convolutional neural network and a storage medium. The invention discloses an online state grid user classification method based on a graph convolutional neural network. The method comprises the following steps: extracting user explicit and implicit features; the method specifically comprises the following steps: firstly, constructing a user graph model by taking online state grid user data as input, extracting feature information in graph data by using a stack network formed by a convolutional neural network, and propagating vector representations of aggregated users and neighbor users thereof by using a cyclic convolution operation; classifying users; according to the method, the graph convolutional neural network model is used for constructing the portrait of the state grid user on the Internet, dominant and implicit features of the user are deeply mined, and an accurate user feature vector is formed. Meanwhile, user vector subdivision is achieved according to probability representation output by the full connection layer, and potential user groups are effectively recognized.

Description

Technical Field

[0001] The present invention relates to the field of big data processing technology, and in particular to an online state grid user classification method, system, device and storage medium based on a graph convolutional neural network. Background Art

[0002] The RFM model for online state-owned grid marketing services is an important tool and means to measure customer value and customer's ability to create benefits. Figure 1 As shown in the figure, R represents the time since the user's last payment. Generally speaking, users who paid recently are more likely to pay again. F represents the amount of payment made by the customer over a period of time. Users who pay frequently may be loyal users or have higher payment needs. M represents the amount of payment made by the customer over a period of time. High-value customers tend to have higher payment amounts. Through the RFM model, users can be divided into different groups, such as high-value users, new users, and lost users, which helps companies formulate marketing strategies and services based on different user groups.

[0003] The way that the online State Grid built an RFM customer classification model around the core business of bill payment has a relatively simple portrait of users, without deeply exploring the implicit characteristics such as basic information and behavioral data of users. In addition, users are not segmented, resulting in the inability to accurately reach the target group, locate the drainage population, and find the best drainage time when carrying out operational activities. Manual screening of user classification not only wastes time, but also limits the flexibility and timeliness of strategy formulation.

[0004] In addition, traditional graph convolutional neural networks assign the same and fixed weight to each edge, and do not consider the different effects of different neighboring users when updating user vectors. Moreover, simply introducing graph convolutional neural networks cannot guarantee that the extracted feature vectors are necessarily beneficial to the classification task.

[0005] In the current technical context, although some customer operation strategies have improved the RFM user classification model by introducing clustering algorithms, they are still based on the segmentation method of the RFM model. They only conduct group operations on users through three types of payment data, and do not integrate other user attributes and behavior data. The model characterization is relatively one-sided and lacks mining of user invisible features. Although some solutions use traditional graph convolutional neural networks to extract features, they manually define the weights of the edges in the graph, do not reflect the strength of the association between users, and separate feature extraction from classification tasks, which fails to effectively extract more discriminative user features.

[0006] The shortcomings of the existing technology are mainly manifested in the following aspects:

[0007] (I) Lack of mining of users’ implicit characteristics: The current online State Grid’s selection of operating target groups is based on explicit characteristics such as payment times and payment amounts, and there is no mining and aggregation of users’ implicit characteristics based on scientific algorithms. The model characterization method is superficial and single, and it is unable to explore the deep-level associations such as the purpose and preference of user behavior.

[0008] (II) User grouping is too simple and not detailed enough: The RFM model only classifies users through three types of payment data and conducts grouping operations, without considering the different stages of users in their entire life cycle (such as new users, old users, active users, etc.). In a rapidly changing market environment, customer preferences and needs may change at any time. The existing classification model cannot reflect changes in customer behavior in a timely manner, and has not achieved a deep understanding of customer groups and effective operations.

[0009] (III) Low automation level of customer operation: At present, the online State Grid manually screens and classifies users based on their basic attributes and business attributes, with many manual intervention links and low automation and intelligence levels. This not only affects the execution efficiency of customer operations and wastes human resources, but also reduces customer experience and leads to customer loss.

[0010] (IV) Traditional graph convolutional neural networks set the same and fixed weights for each edge, which cannot reflect the importance of different neighboring users to the current user. Moreover, they only use labels as supervision information and rely on a large number of training samples. The trained model is not ideal and it is difficult to effectively extract the discriminative features of user data, resulting in low classification accuracy.

[0011] Explanation of relevant terms:

[0012] User portrait: Analyze and describe multi-dimensional data such as users’ personal information, interests, hobbies, behavioral habits, etc. to form a user feature model, and formulate different user operation strategies based on the characteristics of the customer group.

[0013] Graph Convolutional Network (GCN): Graph Convolutional Neural Network is a deep learning model specifically designed to process graph structured data. In graph data, nodes are connected to each other through edges to form a complex network structure. The network realizes the analysis and mining of graph data by extracting features and recognizing patterns from graph structured data. Summary of the invention

[0014] The present invention proposes an online State Grid user classification model based on graph convolutional neural network. The graph convolutional network model is used to deeply explore the explicit and implicit characteristics of users, and generate a comprehensive user feature description that integrates user attribute characteristics, behavioral characteristics, etc., to solve the problem of single user model characterization. The technical solution is as follows:

[0015] First, the online state grid user classification method based on graph convolutional neural network includes the following process:

[0016] S100: Extracting explicit and implicit features of the user; specifically including:

[0017] S110: First, the online State Grid user data is used as input to construct a user graph model G(U,E).

[0018] Among them, U={u i |i=1,2,...,N} represents the set of all users, and each user u i It consists of basic information of users, electricity usage behavior, payment behavior and active information;

[0019] E={e ij |i,j=1,2,..,N} is a set of edges, representing the interaction relationship between users, that is, e ij Represents user u i With user u j If users have similar behaviors, then e ij =1, otherwise e ij =0;

[0020] S120: Use a stacked network composed of convolutional neural networks to extract feature information from graph data, and use circular convolution operations to propagate vector representations of aggregated users and their neighboring users; update the user feature vector in the graph through L rounds of iterations, and gradually integrate global and local information; the feature vector integrates features from multiple dimensions including user basic information, electricity consumption, and activity, and fully explores the hidden features of online State Grid users;

[0021] S300: User classification; specifically including:

[0022] The extracted user feature vector The data is input into a fully connected neural network for feature combination and high-level abstract feature mining. The output is converted into a probability representation to obtain the probability that the data belongs to each cluster. The category with the highest probability value is the final classification result of the model. The output of each user vector representation after the classification module is:

[0023]

[0024] Among them, y represents the user category, W and b are the parameters of the fully connected layer; the user feature vector is divided into 9 clusters, including new users, older users, old users, active users with high electricity consumption, inactive users with high electricity consumption, active users with low electricity consumption, inactive users with low electricity consumption, young users, and middle-aged and elderly users.

[0025] Preferably, between S100 and S300, the following steps are also included:

[0026] S200: Integrate the instance-prototype contrast learning framework to further train the model's feature extraction capabilities, including: introducing the idea of ​​clustering, by representing user vectors Perform k-means clustering to obtain the cluster to which each user belongs. The centroid of the cluster to which the user feature vector belongs is the prototype C of the user instance. i ; Each user feature Its prototype C i Constitute a positive pair, and other prototypes constitute a negative pair; the objective function is as follows:

[0027]

[0028] Among them, s(@) represents the similarity measurement function, which is used to measure the similarity between positive and negative example pairs, and K is the number of negative example pairs;

[0029] The loss function is used to train the model, and the main features that are beneficial to the classification task are further extracted from the input data. The user vector representation is obtained by aggregating features such as age and registration time in the basic information dimension, average number of login days and average number of clicks per month in the active dimension, and average monthly electricity consumption in the electricity consumption dimension.

[0030] In S300, the classification performance of the online State Grid user classification model is further improved by jointly training the target classification task and the auxiliary task based on instance-prototype contrastive learning and updating the graph convolutional neural network parameters.

[0031]

[0032] Among them, L all is the objective function of the joint training framework, L cls is the classification task loss function, λ is the coefficient controlling the loss weight, C is the number of categories, and Y i.j , Y′ i,j Represent the sample labels and the corresponding predicted values ​​respectively.

[0033] Preferably, in S120, the following process is also included:

[0034] In order to realize the automatic learning and updating of weights between users, an attention mechanism is added in the feature extraction stage so that the relevant parameters are updated along with the network model; according to the similarity of behaviors between different users, different neighbor vectors u j Assign different weights θ, and the node update formula is as follows:

[0035]

[0036] in, is the feature vector of the user node at the l+1 layer, σ(·) is the nonlinear activation function, θ represents the weight parameter of the information transmission aggregation between user nodes, W (l) is the neural network weight parameter of the lth layer.

[0037] In the second aspect, the online State Grid user classification system based on graph convolutional neural network running the above method includes the following modules: feature extraction module, user classification module;

[0038] The feature extraction module is used to extract the explicit and implicit features of users. First, the online State Grid user data is used as input to construct a user graph model G(U,E), where U = {u i |i=1,2,...,N} represents the set of all users, and each user u i It consists of basic information of users, electricity consumption behavior, payment behavior and active information; E = {e ij |i,j=1,2,..,N} is a set of edges, representing the interaction relationship between users, that is, e ij Represents user u i With user u j If users have similar behaviors, then e ij =1, otherwise e ij =0;

[0039] Secondly, a stacked network consisting of convolutional neural networks is used to extract feature information from graph data, and the vector representation of aggregated users and their neighboring users is propagated using circular convolution operations. Through L rounds of iterations, the user feature vector in the graph is updated, and global and local information is gradually integrated. The feature vector integrates features from multiple dimensions, including user basic information, electricity consumption, and activity, to fully explore the hidden features of online State Grid users.

[0040] The user classification module extracts the user feature vector The data is input into a fully connected neural network for feature combination and high-level abstract feature mining. The output is converted into a probability representation to obtain the probability that the data belongs to each cluster. The category with the highest probability value is the final classification result of the model. The output of each user vector representation after the classification module is:

[0041]

[0042] Where y represents the user category, W and b are the parameters of the fully connected layer; the category of each user is identified through the user classification module, and the user feature vector is divided into 9 clusters: new users, older users, old users, active users with high electricity consumption, inactive users with high electricity consumption, active users with low electricity consumption, inactive users with low electricity consumption, young users, and middle-aged and elderly users.

[0043] Preferably, the user classification system further includes a contrastive learning auxiliary module; the contrastive learning auxiliary module is used to further train the feature extraction capability of the model; the idea of ​​clustering is introduced, by representing the user vector Perform k-means clustering to obtain the cluster to which each user belongs. The centroid of the cluster to which the user feature vector belongs is the prototype C of the user instance. i ; Each user feature Its prototype C i Constitute a positive pair, and other prototypes constitute a negative pair; the objective function is as follows:

[0044]

[0045] Where s(·) represents a similarity measurement function, which is used to measure the similarity between positive and negative pairs, and K is the number of negative pairs. The model is trained by using this loss function to further extract the main features in the input data that are beneficial to the classification task, and obtain a user vector representation that aggregates features such as age and registration time in the basic information dimension, average monthly login days and average monthly clicks in the active dimension, and average monthly electricity consumption in the electricity consumption dimension.

[0046] The user classification module jointly trains the target classification task and the auxiliary task based on instance-prototype contrast learning, updates the graph convolutional neural network parameters, and further improves the classification performance of the online State Grid user classification model.

[0047]

[0048] Among them, L all is the objective function of the joint training framework, L cls is the classification task loss function, λ is the coefficient controlling the loss weight, C is the number of categories, and Y i.j , Y′ i,j Represent the sample labels and the corresponding predicted values ​​respectively.

[0049] In a third aspect, an electronic device includes: a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor implements the above-mentioned method for classifying online State Grid users using a graph convolutional neural network when executing the computer program.

[0050] In a fourth aspect, a non-transitory computer-readable storage medium stores a computer program, which, when executed by a processor, implements the above-mentioned method for classifying online State Grid users using a graph convolutional neural network.

[0051] In order to solve the problem of insufficient extraction of online State Grid user features, the present invention introduces a graph convolutional neural network to explore user behavior features, achieve deep mining and accurate representation of user features, and enrich user portraits. At the same time, since it is difficult for graph neural networks to directly process original discrete user data, it is necessary to convert the original user data into graph data. In this process, the weight of the edge is usually defined as a fixed value, but node aggregation based on fixed weights is difficult to meet the complex online State Grid user data feature extraction requirements. To this end, an attention mechanism is introduced to automatically update the weight of the edge in the graph with the network.

[0052] Since the online State Grid user data has a high dimension and contains a lot of noise, it is difficult to extract discriminative feature representations from the data by relying solely on convolutional neural networks. Therefore, in order to enhance the discriminative ability of user feature representations to improve classification accuracy, an auxiliary task-improved graph convolution user classification model based on contrastive learning is proposed in the embodiment. Through instance-prototype contrastive learning, the model learns feature representations that are more conducive to the classification of online State Grid users, and then correctly divides user vectors into corresponding clusters, and formulates personalized marketing strategies for different user groups. The comparative information between users and different user groups is used as different perspectives to achieve information complementarity from different perspectives and explore the differences between users and user clusters. By considering the specific distribution of sample features within the cluster where the user is located, the ability of the model to extract discriminative features is enhanced.

[0053] The embodiment of the present invention constructs a multi-task joint optimization framework, updates the graph convolutional neural network parameters by jointly training the target classification task and the auxiliary task based on instance-prototype comparative learning, enhances the discriminability of the learned features, and thus further improves the classification performance of the online State Grid user classification model.

[0054] In summary, the present invention uses a graph convolutional neural network model to construct an online State Grid user portrait, deeply mines the explicit and implicit characteristics of users, effectively integrates user basic information and behavioral data, gradually improves the richness of user characteristics, and forms an accurate user feature vector. At the same time, user vector segmentation is achieved based on the probability representation of the output of the fully connected layer, and potential user groups are effectively identified. This model can not only handle the uncertainty in the online State Grid user data, but also capture the diversity of users between different categories. By analyzing the classification results, more personalized activity recommendations are provided for each user group, thereby improving user satisfaction and the conversion rate of the platform. BRIEF DESCRIPTION OF THE DRAWINGS

[0055] Figure 1 It is a schematic diagram of the existing user classification model - RFM model.

[0056] Figure 2 It is a schematic diagram of the method of the present invention. DETAILED DESCRIPTION

[0057] The present invention utilizes the relationship capture and feature aggregation characteristics of graph convolutional neural networks, comprehensively integrates user behavior characteristics, deeply mines implicit characteristics, and comprehensively characterizes the online State Grid user portrait, thereby realizing user segmentation and formulating personalized operation strategies for different clusters. The online State Grid user classification system based on graph convolutional neural networks specifically includes the following modules: feature extraction module, comparative learning auxiliary module, and user classification module.

[0058] The feature extraction module combines the RFM model with the behavior map model input items to sort out the required data information, namely the basic information dimensions of online State Grid users, including age, gender, region, registration time, and account number binding; payment behavior dimensions, including payment frequency, latest payment time, and payment amount; electricity consumption behavior dimensions, including electricity type, industry type, average monthly electricity consumption, and average monthly electricity bill; active information dimensions, including average monthly login days, average monthly clicks, and number of activities participated in; fully integrate different types of user data. Through data processing and format conversion, user feature vectors and adjacency matrices are formed, and a user graph model G(U,E) is established.

[0059] Among them, U={u i |i=1,2,...,N} represents the set of all users, and each user u i It consists of basic information of users, electricity usage behavior, payment behavior and active information;

[0060] E={e ij |i,j=1,2,..,N} is a set of edges, representing the interaction relationship between users, that is, e ij Represents user u i With user u j If users have similar behaviors, then e ij =1, otherwise e ij =0.

[0061] Secondly, a stacked network consisting of convolutional neural networks is used to extract feature information from graph data, and the vector representation of aggregated users and their neighboring users is propagated using circular convolution operations; through L rounds of iterations, the user feature vector in the graph is updated, and global and local information is gradually integrated. The feature vector integrates the features of multiple dimensions including user basic information, electricity consumption, and activeness, fully mining the hidden features of online State Grid users.

[0062] At the same time, in order to realize the automatic learning and updating of weights between users, an attention mechanism is added to the feature extraction module so that the relevant parameters are updated along with the network model. j Assign different weights θ; the node update formula is as follows:

[0063]

[0064] in, is the feature vector of the user node at the l+1 layer, σ(·) is the nonlinear activation function, θ represents the weight parameter of the information transmission aggregation between user nodes, W (l) is the neural network weight parameter of the lth layer.

[0065] The present invention integrates an instance-prototype contrast learning auxiliary module in the model to further train the feature extraction ability of the model. This task introduces the idea of ​​clustering and Perform k-means clustering to obtain the cluster to which each user belongs. The centroid of the cluster to which the user feature vector belongs is the prototype C of the user instance. i , each user feature Its prototype C i Constitute a positive pair, and other prototypes constitute a negative pair; the objective function is as follows:

[0066]

[0067] Among them, s(·) represents a similarity measurement function, which is used to measure the similarity between positive and negative example pairs, and K is the number of negative example pairs. The model is trained by using this loss function to further extract the main features in the input data that are beneficial to the classification task, and obtain a user vector representation that aggregates features such as age and registration time in the basic information dimension, average monthly login days and average monthly clicks in the active dimension, and average monthly electricity consumption in the electricity consumption dimension.

[0068] A joint optimization framework as shown in the following formula is adopted to coordinate the training of feature extraction and classification processes, promote each other, and realize the flexible segmentation of State Grid customers online.

[0069] The present invention constructs a multi-task joint optimization framework, and updates the graph convolutional neural network parameters by jointly training the target classification task and the auxiliary task based on instance-prototype comparative learning, further improving the classification performance of the online State Grid user classification model. It reduces manual participation in the training process and realizes the automation and intelligence of the customer classification process;

[0070]

[0071] Among them, L all is the objective function of the joint training framework, L cls is the classification task loss function, λ is the coefficient controlling the loss weight, C is the number of categories, and Y i.j , Y′ i,j Represent the sample labels and the corresponding predicted values ​​respectively.

[0072] This combination method makes the model more adaptable to different data distributions, with strong scalability and flexibility. It can adapt to different data sets by adjusting dimensions, thereby achieving dynamic user segmentation. When user behavior data changes, the model can be retrained and the results can be adjusted quickly to ensure the real-time and accuracy of user group segmentation.

[0073] The user classification module extracts the user feature vector The data is input into a fully connected neural network for feature combination and high-level abstract feature mining. The output is converted into a probability representation to obtain the probability that the data belongs to each cluster. The category with the highest probability value is the final classification result of the model. The output of each user vector representation after the classification module is:

[0074]

[0075] Among them, y represents the user category, W and b are the fully connected layer parameters.

[0076] The system identifies the category of each user through the classification module, and divides the user feature vector into 9 clusters, including new users, older users, old users, active users with high electricity consumption, inactive users with high electricity consumption, active users with low electricity consumption, inactive users with low electricity consumption, young users, and middle-aged and elderly users. Then, the system analyzes the characteristics, scale, and representative users of the 9 categories, and formulates targeted marketing strategies based on the needs and preferences of different categories.

[0077] The trained graph convolutional neural network model can extract features from users' basic information, historical behavior records and other data, output the user's feature vector, build user portraits, and complete user segmentation, revealing the interest preferences and potential needs of different user groups, providing support for precision marketing.

Claims

1. An online state grid user classification method based on graph convolutional neural network is characterized by: The process includes the following: S100: Extracting explicit and implicit features of the user; specifically including: S110: First, the online State Grid user data is used as input to construct a user graph model G(U, E). Among them, U={u i |i=1,2,...,N} represents the set of all users, and each user u i It consists of basic information of users, electricity usage behavior, payment behavior and active information; E={e ij |i, j = 1, 2, .., N} is a set of edges, representing the interaction relationship between users, that is, e ij Represents user u i With user u j If users have similar behaviors, then e ij =1, otherwise e ij =0; S120: Use a stacked network composed of convolutional neural networks to extract feature information from graph data, and use circular convolution operations to propagate vector representations of aggregated users and their neighboring users; update the user feature vector in the graph through L rounds of iterations, and gradually integrate global and local information; the feature vector integrates features from multiple dimensions including user basic information, electricity consumption, and activity, and fully explores the hidden features of online State Grid users; S300: User classification; specifically including: The extracted user feature vector The data is input into a fully connected neural network for feature combination and high-level abstract feature mining. The output is converted into a probability representation to obtain the probability that the data belongs to each cluster. The category with the highest probability value is the final classification result of the model. The output of each user vector representation after the classification module is: Among them, y represents the user category, W and b are the parameters of the fully connected layer; the user feature vector is divided into 9 clusters: new users, older users, old users, active users with high electricity consumption, inactive users with high electricity consumption, active users with low electricity consumption, inactive users with low electricity consumption, young users, and middle-aged and elderly users.

2. According to claim 1, the online state grid user classification method based on graph convolutional neural network is characterized in that: Between S100 and S300 also include: S200: Integrate the instance-prototype contrast learning framework to further train the model's feature extraction capabilities, including: introducing the idea of ​​clustering, by representing user vectors Perform k-means clustering to obtain the cluster to which each user belongs. The centroid of the cluster to which the user feature vector belongs is the prototype C of the user instance. i ; Each user characteristics Its prototype C i Constitute a positive pair, and other prototypes constitute a negative pair; the objective function is as follows: Where s(·) represents the similarity measurement function, which is used to measure the similarity between positive and negative example pairs, and K is the number of negative example pairs; The loss function is used to train the model and further extract the main features in the input data that are beneficial to the classification task. The user vector representation is obtained by aggregating the age and registration time in the basic information dimension, the average number of login days and the average number of clicks per month in the active dimension, and the average monthly electricity consumption in the electricity consumption dimension. In S300, the classification performance of the online State Grid user classification model is further improved by jointly training the target classification task and the auxiliary task based on instance-prototype contrastive learning and updating the graph convolutional neural network parameters. Among them, L all is the objective function of the joint training framework, L cls is the classification task loss function, λ is the coefficient controlling the loss weight, C is the number of categories, and Y i，j , Y′ i，j Represent the sample labels and the corresponding predicted values ​​respectively.

3. The online state grid user classification method based on graph convolutional neural network according to claim 1 or 2 is characterized in that: In S120, the following process is also included: In order to realize the automatic learning and updating of weights between users, an attention mechanism is added in the feature extraction stage so that the relevant parameters are updated along with the network model; according to the similarity of behaviors between different users, different neighbor vectors u j Assign different weights θ, and the node update formula is as follows: in, is the feature vector of the user node at the l+1 layer, σ(·) is the nonlinear activation function, θ represents the weight parameter of the information transmission aggregation between user nodes, W (l) is the neural network weight parameter of the lth layer.

4. The online state grid user classification method based on graph convolutional neural network according to claim 1 or 2 is characterized in that: The basic user information includes: age, gender, region, registration time, and account number binding; the payment behavior includes: payment frequency, most recent payment time, and payment amount; the electricity usage behavior includes: electricity usage type, industry type, average monthly electricity consumption, and average monthly electricity bill; the active information includes: average monthly login days, average monthly clicks, and number of activities participated in.

5. An online State Grid user classification system based on a graph convolutional neural network that runs the user classification method of claim 1, characterized in that: It includes the following modules: feature extraction module, user classification module; The feature extraction module is used to extract explicit and implicit features of users. First, the online State Grid user data is used as input to construct a user graph model G(U, E), where U = {u i |i=1,2,...,N} represents the set of all users, and each user u i It consists of basic information of users, electricity consumption behavior, payment behavior and active information; E = {e ij |i, j = 1, 2, .., N} is a set of edges, representing the interaction relationship between users, that is, e ij Represents user u i With user u j If users have similar behaviors, then e ij =1, otherwise e ij =0; Secondly, a stacked network consisting of convolutional neural networks is used to extract feature information from graph data, and the vector representation of aggregated users and their neighboring users is propagated using circular convolution operations. Through L rounds of iterations, the user feature vector in the graph is updated, and global and local information is gradually integrated. The feature vector integrates features from multiple dimensions, including user basic information, electricity consumption, and activity, to fully explore the hidden features of online State Grid users. The user classification module extracts the user feature vector The data is input into a fully connected neural network for feature combination and high-level abstract feature mining. The output is converted into a probability representation to obtain the probability that the data belongs to each cluster. The category with the highest probability value is the final classification result of the model. The output of each user vector representation after the classification module is: Where y represents the user category, W and b are the parameters of the fully connected layer; the category of each user is identified through the user classification module, and the user feature vector is divided into 9 clusters: new users, older users, old users, active users with high electricity consumption, inactive users with high electricity consumption, active users with low electricity consumption, inactive users with low electricity consumption, young users, and middle-aged and elderly users.

6. The online state grid user classification system based on graph convolutional neural network according to claim 5 is characterized in that: It also includes a contrastive learning auxiliary module; the contrastive learning auxiliary module is used to further train the feature extraction ability of the model; introduce the idea of ​​clustering, by representing the user vector Perform k-means clustering to obtain the cluster to which each user belongs. The centroid of the cluster to which the user feature vector belongs is the prototype C of the user instance. i ; Each user feature Its prototype C i Constitute a positive pair, and other prototypes constitute a negative pair; the objective function is as follows: Where s(·) represents a similarity measurement function, which is used to measure the similarity between positive and negative pairs, and K is the number of negative pairs. The model is trained by using this loss function to further extract the main features in the input data that are beneficial to the classification task, and obtain a user vector representation that aggregates the age and registration time in the basic information dimension, the average number of login days and clicks per month in the active dimension, and the average monthly electricity consumption in the electricity consumption dimension.

7. The online state grid user classification system based on graph convolutional neural network according to claim 6 is characterized in that: The user classification module jointly trains the target classification task and the auxiliary task based on instance-prototype contrast learning, updates the graph convolutional neural network parameters, and further improves the classification performance of the online State Grid user classification model. Among them, L all is the objective function of the joint training framework, L cls is the classification task loss function, λ is the coefficient controlling the loss weight, C is the number of categories, and Y i，j , Y′ i，j Represent the sample labels and the corresponding predicted values ​​respectively.

8. An electronic device comprising: A memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor implements an online State Grid user classification method using a graph convolutional neural network as described in any one of claims 1 to 3 when executing the computer program.

9. A non-transitory computer-readable storage medium having a computer program stored thereon, characterized in that: When the computer program is executed by a processor, it implements an online state grid user classification method using a graph convolutional neural network as described in any one of claims 1 to 3.

Images