A knowledge graph-based online state grid potential loss user identification method

Abstract

The present application relates to the technical field of artificial intelligence, in particular to a method for identifying potential loss users of online state grid based on a knowledge graph. The method comprises the following processes: constructing a user knowledge graph; based on the user knowledge graph, constructing a relational graph convolutional neural network to deeply mine and analyze graph structure data, learning low-dimensional vector representation of entities and relationships in the knowledge graph, and capturing complex relationships between entities and deep features of graph structure data; referring to a multi-task learning network, based on the extracted user features, constructing two classifiers, designing an independent output layer for each classification task, one for identifying whether a user has a loss risk, and the other for predicting the level of the user's loss risk. The present application solves the problem of how to improve the accuracy of online state grid user loss risk identification in a big data environment; effectively identifies the power user group with potential loss risk, and realizes the risk level subdivision of users, and then formulates differentiated recall strategies.

Description

Technical Field

[0001] This invention relates to the field of artificial intelligence technology, specifically to a method for identifying potential lost users of the State Grid online based on knowledge graphs. Background Technology

[0002] State Grid Online App is a large-scale digital energy service platform in China. However, with the rapid development of the digital economy, various mobile applications providing "electricity bill payment" and "life services" have emerged in the market. State Grid Online is facing the pressure of users being diverted through multiple channels. Therefore, it urgently needs to accurately identify potential lost users from its large user base and take effective user recapture strategies.

[0003] Traditional user feature extraction models mostly focus on shallow linear features, neglecting in-depth mining of implicit user characteristics. User data is often one-dimensional, failing to capture the complexity and diversity of behavior, lacking comprehensive user profiles, and unable to accurately identify electricity users at risk of churn in real time. Furthermore, existing user churn risk warning models lack segmentation of potential user risk levels and fail to develop targeted recall strategies for different user risk levels, resulting in poor recall effectiveness.

[0004] Explanation of relevant terms:

[0005] Knowledge graph: A knowledge graph is a structured knowledge base used to describe entities and the relationships between them in the form of a graph. It helps machines better understand and interpret complex data and information, thereby providing more intelligent services and applications.

[0006] Graph Convolutional Neural Networks (GCNs) are deep learning models specifically designed for processing graph-structured data. In graph data, nodes are interconnected by edges, forming complex network structures. These networks perform feature extraction and pattern recognition on graph-structured data to analyze and mine its potential.

[0007] Multi-task learning (MTL): Multi-task learning can train multiple related tasks simultaneously. By sharing representations, it optimizes the objectives of multiple tasks, allowing the model to borrow information from other tasks when processing one task, thereby improving the model's performance on each task and enhancing its generalization ability.

[0008] The ConvE model is a deep learning model for knowledge graph representation learning, originally proposed by Dettmers et al. in 2018. It learns the semantic associations between entities and relations by mapping entities and relations to a low-dimensional vector space and utilizing a convolutional neural network (CNN). Below is a detailed introduction to the ConvE model:

[0009] Entity and Relation Representation: In the ConvE model, each entity and relation is represented as a fixed-length vector. These vectors are randomly initialized and learned through gradient descent during training.

[0010] Input representation: For a given triple, namely: head entity, relation, tail entity, the model concatenates the vectors of the head entity and relation together to form a large input vector. This input vector is flattened into a two-dimensional matrix and fed into the convolutional layer of the model.

[0011] Convolution operations: The model uses one-dimensional convolution operations to extract features from the input matrix. Convolution operations are performed on a two-dimensional tiled input matrix, similar to convolution operations in image processing. These convolution operations are able to capture semantic associations between entities and relationships.

[0012] Feature mapping: The features extracted by the convolutional layers are mapped into the final vector representation space. This mapping process involves fully connected layers and non-linear activation functions to obtain the final entity and relation vector representations.

[0013] Loss function: During training, ConvE models typically use an appropriate loss function to measure the difference between the model output and the true triples.

[0014] Training Process: The ConvE model learns vector representations of entities and relationships by training on knowledge graph data. The training process typically employs optimization algorithms such as stochastic gradient descent to minimize the loss function. Summary of the Invention

[0015] This invention fully integrates multi-source heterogeneous user data to construct a power user knowledge graph, and proposes a power user knowledge graph representation learning method based on a relational graph convolutional neural network. It extracts user feature vectors and uses two classifiers to predict potential user churn risk and segment churn risk levels, respectively. The specific technical solution is as follows:

[0016] A knowledge graph-based method for identifying potential churned users of the State Grid online platform includes the following process:

[0017] S100: Construct a user knowledge graph; combine user attribute and behavior data from the State Grid APP and external multi-source data to sort out data information related to user churn risk; construct a power user knowledge graph by identifying key entities in multi-dimensional data and extracting the relationships between entities.

[0018] S200: Deeply mine user features; based on user knowledge graph, construct relational graph convolutional neural network to deeply mine and analyze graph structure data, learn low-dimensional vector representations of entities and relationships in knowledge graph, and capture the deep features of complex relationships between entities and graph structure data;

[0019] S300: Subdivide user risk levels; draw on multi-task learning networks, construct two classifiers based on the extracted user features, and design independent output layers for each classification task, one for identifying whether a user is at risk of churn, and the other for predicting the level of user churn risk.

[0020] The data information in S100 includes five dimensions: user basic information, industry information, electricity consumption information, interaction behavior, and billing and payment data.

[0021] The basic user information includes: user ID, gender, age, occupation, education level, industry category, whether the user is a member, whether the user has activated their subscription, and whether the user is a temporary user.

[0022] The industry information includes: account number, distribution area, power supply station, access method, and industry type; the industry type includes: residential, enterprises and institutions, shops, charging piles, and distributed photovoltaic.

[0023] Electricity consumption information includes: meter number, meter type, electricity consumption category, average monthly electricity consumption, peak electricity consumption, electricity consumption time distribution, and meter fault alarms;

[0024] Interactive behaviors include: number of times users log in within the app, activity frequency, bill inquiry, energy efficiency analysis, activity participation, message click rate, and push response status;

[0025] Billing and payment include: amount due, discount amount, actual amount paid, payment channel, payment timeliness rate, historical arrears records, number of reminders, and reminder response status; payment channels include: online payment through the State Grid, bank direct debit, and third-party payment.

[0026] Furthermore, the entity types in S100 include:

[0027] "User": refers to a registered user of the State Grid APP on the Internet. Attributes include: user ID, gender, age, occupation, education level, registration time, whether the user is a member, and other user-related information and the user's behavior data on the State Grid APP on the Internet.

[0028] "Industry": This refers to the industries owned by the user. Attributes include: account number, substation area, industry type and housing type, length of residence, housing attributes, industry scale, installed capacity, industry category, access voltage level, shop type, business area, business hours, charging pile type, number of installed units, access method, photovoltaic installed capacity, power generation, and whether an energy storage system is configured, etc., which are relevant attribute information for different industry types. Industry types include: residential, enterprises and institutions, shops, charging piles, and distributed photovoltaic.

[0029] "Electricity Meter": This refers to the electricity meter used by the user. Its attributes include: meter number, meter type, electricity consumption category, meter fault alarm, electricity consumption, electricity consumption time, peak electricity consumption, and other related electricity consumption data.

[0030] "Bill": This refers to the electricity bill generated by the user. Attributes include: amount due, discount amount, actual amount paid, payment channel, payment deadline, actual payment date, number of overdue days, number of overdue payments, maximum amount owed, number of reminders, reminder response time, and other payment-related data.

[0031] The types of relationships include: users "own" property, users "use" electricity meters, and users "generate" bills.

[0032] Furthermore, in S200, the graphs under the three relationships of "owning", "using", and "generating" are each regarded as an isomorphic subgraph. The GCN is used to aggregate the features of the subgraphs respectively, and the extracted multi-relationship feature vectors are superimposed to transform the multi-relationship graph into a problem of feature aggregation of multiple isomorphic subgraphs. For each subgraph, the vector representations of entities and relations are first flattened into a two-dimensional matrix and then processed by a convolutional layer for feature extraction. Then, the extracted features are mapped to the final vector representation space through a fully connected layer.

[0033] Furthermore, for the constructed multi-relationship power user knowledge graph, the node set V represents the entities of "user", "industry", "electricity meter" and "bill", and the edge set R represents the relationship between different entities of "owner", "use" and "generate".

[0034] Each relation type r has its corresponding adjacency matrix. The adjacency matrix represents the relationships between entities in the subgraph corresponding to the relationship type, with each node... There is an eigenvector , representing the attributes of an entity, and the feature matrix H is composed of the feature vectors of all nodes;

[0035] For each node By iteratively aggregating the feature vectors of all neighboring nodes under different relationships, as well as the node's own attributes, the node update formula is as follows:

[0036] ,

[0037] Where node j is a neighboring node of node i. Let h represent the set of neighboring nodes of node i under relation r. j It is the feature vector of node j. It is the degree of node i under relation r, that is, the number of nodes and edges in the isomorphic subgraph. It is the weight matrix of the corresponding relation type r in the l-th layer. It is the self-loop weight matrix in the l-th layer, used to preserve the node's own characteristics;

[0038] Using the above formula, the feature vector of the user node is continuously updated, iteratively aggregating the attribute features of all neighboring nodes and itself in the three relationships, and obtaining a vector representation that fully integrates the multi-dimensional data information of the power user.

[0039] Furthermore, in S300,

[0040] For the user vector representation mined by the relation graph convolutional neural network, it is first passed through a shared fully connected layer for feature mapping, and then the low-dimensional feature vector is input into two classifiers respectively.

[0041] The binary classifier outputs a real value, and the sigmoid function maps this value to a range of 0 to 1, representing the probability of user churn.

[0042] A multi-classifier can output multiple values, which, after being processed by the Softmax function, represent the probability that a user belongs to different churn risk levels.

[0043] The advantages of this invention over the prior art are as follows:

[0044] (i) This invention utilizes existing Internet resources to obtain basic user information and behavioral data such as electricity consumption and payment within the power system, as well as multi-source user data outside the system. It designs knowledge graph entity types and relationship types, establishes relevant entity and relationship attributes, and integrates multi-source heterogeneous data to construct a power user knowledge graph. This solves the problem of how to improve the accuracy of identifying the risk of State Grid user loss on the Internet in a big data environment.

[0045] (II) For multi-source heterogeneous power user knowledge graphs, this invention constructs a multi-relation graph convolutional neural network for knowledge graph representation learning, analyzes the features of nodes and edges in different relationships, and infers the interaction behavior and related properties between different entities. This effectively captures the associations between different types of entities, generates a more comprehensive user entity vector representation, and achieves in-depth mining and accurate representation of user features, thereby providing strong support for predicting potential churn risks and formulating personalized recall strategies.

[0046] (III) This invention draws on the multi-task learning mechanism to construct two classifiers and adopts a joint optimization framework. By jointly training the potential churn user prediction task and the risk level segmentation task, user vector segmentation is achieved, effectively identifying the power user group with potential churn risk and realizing the risk level segmentation of users, thereby formulating differentiated recall strategies. Attached Figure Description

[0047] Figure 1This is a schematic diagram of the process for identifying potential lost users of the State Grid online based on knowledge graphs, according to the present invention.

[0048] Figure 2 This is a schematic diagram illustrating the relationships between entity nodes in the power user knowledge graph in an embodiment of the present invention. Detailed Implementation

[0049] A knowledge graph-based method for identifying potential churned users of the State Grid online platform includes the following process:

[0050] S100: Construct a user knowledge graph; combine user attribute and behavior data from the State Grid APP and external multi-source data to sort out data information related to user churn risk; including data from five dimensions such as basic user information, industry information, electricity consumption information, interaction behavior, billing and payment, fully integrate different types of user characteristics, model user-related auxiliary information, and enrich user data; by identifying key entities in multi-dimensional data and extracting the relationships between entities, construct a power user knowledge graph; use graph structure to represent user-related knowledge, comprehensively depict the multi-dimensional characteristics of users and their interrelationships, and provide data support and decision-making basis for accurately identifying potential churned users.

[0051] The basic user information includes: user ID, gender, age, occupation, education level, industry category, whether the user is a member, whether the user has activated their subscription, whether the user is a temporary user, etc., which are used to characterize the user's basic attributes and correspond to the user entity.

[0052] The industry information includes: user number, distribution area, power supply station, access method, industry type and related attribute fields, which are used to reflect all industry categories and application scenarios of the user and correspond to industry entities; the industry types include: residential, enterprises and institutions, shops, charging piles, distributed photovoltaic and other different types of industries.

[0053] Electricity consumption information includes features such as meter number, meter type, electricity consumption category, average monthly electricity consumption, peak electricity consumption, electricity consumption time distribution, and meter fault alarms. These features are used to reflect the user's energy consumption behavior and equipment operation status, and correspond to the physical meter.

[0054] Interactive behaviors include: the number of times a user logs in within the app, activity frequency, bill inquiry, energy efficiency analysis, activity participation, message click rate, push response, etc., which are used to describe the user's behavioral characteristics within the application and correspond to the user entity.

[0055] Bills and payments include: amount due, discount amount, amount actually paid, payment channel, payment timeliness rate, historical overdue payment records, number of reminders, and reminder response status; payment channels include: online billing, bank direct debit, third-party payment, etc., which are used to reflect users' habits in payment and billing processing and correspond to the bill entity.

[0056] The entity types in S100 include:

[0057] "User": refers to a registered user of the State Grid APP on the Internet. Attributes include: user ID, gender, age, occupation, education level, registration time, whether the user is a member, and other user-related information and the user's behavior data on the State Grid APP on the Internet.

[0058] "Industry": This refers to the industries owned by the user. Attributes include: account number, substation area, industry type and housing type, length of residence, housing attributes, industry scale, installed capacity, industry category, access voltage level, shop type, business area, business hours, charging pile type, number of installed units, access method, photovoltaic installed capacity, power generation, and whether an energy storage system is configured, etc., which are relevant attribute information for different industry types. Industry types include: residential, enterprises and institutions, shops, charging piles, and distributed photovoltaic.

[0059] "Electricity Meter": This refers to the electricity meter used by the user. Its attributes include: meter number, meter type, electricity consumption category, meter fault alarm, electricity consumption, electricity consumption time, peak electricity consumption, and other related electricity consumption data.

[0060] "Bill": This refers to the electricity bill generated by the user. Attributes include: amount due, discount amount, actual amount paid, payment channel, payment deadline, actual payment date, number of overdue days, number of overdue payments, maximum amount owed, number of reminders, reminder response time, and other payment-related data.

[0061] The types of relationships include: users "own" property, users "use" electricity meters, and users "generate" bills.

[0062] S200: Deeply mine user characteristics; based on the user knowledge graph, construct a relational graph convolutional neural network to deeply mine and analyze graph structure data, learn low-dimensional vector representations of entities and relationships in the knowledge graph, capture complex relationships between entities and deep-level features of graph structure data; through convolutional aggregation operations, information is propagated between entities, aggregating feature information of nodes and neighboring nodes to update the hidden layer state of nodes. Through continuous iteration, the representation learning of the knowledge graph is completed, user features are extracted, and the new vector representation not only includes the explicit and implicit features of the user itself, but also considers its relationship with neighboring nodes, which can effectively capture complex relationships between entities and similarities between users, thereby more accurately assessing the risk of power user churn.

[0063] In step S200, the graphs under the three relationships of "owning", "using", and "generating" are each regarded as an isomorphic subgraph. The GCN is used to aggregate the features of the subgraphs respectively, and the extracted multi-relationship feature vectors are superimposed to transform the multi-relationship graph into a problem of feature aggregation of multiple isomorphic subgraphs. For each subgraph, the vector representations of entities and relations are first flattened into a two-dimensional matrix, and then feature extraction is performed through a convolutional layer. Then, the extracted features are mapped to the final vector representation space through a fully connected layer.

[0064] For the constructed multi-relationship power user knowledge graph, the node set V represents the entities of "user", "industry", "electricity meter" and "bill", and the edge set R represents the relationship between different entities of "owner", "use" and "generate".

[0065] Each relation type r has its corresponding adjacency matrix. The adjacency matrix represents the relationships between entities in the subgraph corresponding to the relationship type, with each node... There is an eigenvector , representing the attributes of an entity, and the feature matrix H is composed of the feature vectors of all nodes;

[0066] For each node By iteratively aggregating the feature vectors of all neighboring nodes under different relationships, as well as the node's own attributes, the node update formula is as follows:

[0067] ,

[0068] Where node j is a neighboring node of node i. Let h represent the set of neighboring nodes of node i under relation r. j It is the feature vector of node j. It is the degree of node i under relation r, that is, the number of nodes and edges in the isomorphic subgraph. It is the weight matrix of the corresponding relation type r in the l-th layer. It is the self-loop weight matrix in the l-th layer, used to preserve the node's own characteristics;

[0069] Using the above formula, the feature vector of the user node is continuously updated, iteratively aggregating the attribute features of all neighboring nodes and itself in the three relationships, and obtaining a vector representation that fully integrates the multi-dimensional data information of the power user.

[0070] S300: This system categorizes user risk levels. Drawing inspiration from multi-task learning networks, it constructs two classifiers based on extracted user features. Each classifier has an independent output layer: one to identify whether a user is at risk of churn, and the other to predict the level of churn risk. By jointly training the two classifiers, the model's classification performance is optimized, enabling the categorization of power user risk levels. This accurately identifies potential churned users, and targeted risk warnings and user reactivation strategies are implemented for users with different risk levels, effectively preventing user churn.

[0071] For the user vector representation mined by the relation graph convolutional neural network, it is first passed through a shared fully connected layer for feature mapping, and then the low-dimensional feature vector is input into two classifiers respectively.

[0072] The binary classifier outputs a real value, which is then mapped to a value between 0 and 1 by the sigmoid function, representing the probability of user churn. The binary cross-entropy loss function is used to measure the difference between the output probability and the actual label, thereby training the classifier to accurately identify users at risk of churn.

[0073] A multi-classifier can output multiple values, which, after being activated by the Softmax function, represent the probability that a user belongs to different churn risk levels. Cross-entropy loss is used to measure the difference between the predicted probability distribution and the actual label, training the model to accurately identify the user's churn risk level.

[0074] This invention constructs a multi-source, heterogeneous user knowledge graph by integrating multi-dimensional attributes and behavioral data of electricity users. This comprehensively depicts the behavioral characteristics of electricity users, overcoming the problems of single data sources and incomplete user profiles in traditional methods. This method avoids the limitations of a single data source, enriches the characteristics of electricity customers, and helps the State Grid Corporation of China more accurately and comprehensively identify potential churned users, improving the accuracy of risk prediction models. Through structured knowledge support, the intelligence level of potential churn risk prediction is enhanced, while the fusion of multi-source data further improves the accuracy of user profiles, making risk prediction more forward-looking and reliable.

[0075] Graph convolutional neural networks (GNNs) possess hierarchical learning capabilities, demonstrating significant advantages in deeply mining the explicit and implicit characteristics of electricity customers. Faced with complex electricity user knowledge graphs, relational graph GNNs can effectively handle the heterogeneity and asymmetry within these graphs, enhancing the model's generalization ability on unknown data. By decomposing the electricity user knowledge graph into different isomorphic subgraphs and using convolutional operations to aggregate information from neighboring nodes, semantic relationships between entities and potential user behavior patterns are captured. Furthermore, by aggregating the feature representations of multiple isomorphic subgraphs, entities and relationships are mapped to vector representations in a low-dimensional space. This helps enterprises better understand the electricity consumption characteristics and interaction habits of different customer groups, thereby enabling in-depth analysis of potential churned users and improving the performance of downstream churn risk level segmentation tasks.

[0076] A multi-task learning mechanism is employed to construct two classifiers, one for predicting potential user churn risk and the other for risk level segmentation. By sharing feature representations, the model's adaptability to different tasks is enhanced, improving the accuracy and practicality of user churn risk warning. In the risk level segmentation task, the model outputs the probability that each electricity customer belongs to a specific churn risk level, not only handling data uncertainty but also reflecting the overlap between different risk levels. Through accurate risk level classification, power companies can more rationally allocate operational resources, formulate differentiated activation or recall strategies, and improve operational efficiency and customer satisfaction.

[0077] This document uses specific embodiments to illustrate the principles and implementation methods of this document. The descriptions of the embodiments above are only for the purpose of helping to understand the methods and core ideas of this document. At the same time, for those skilled in the art, there will be changes in the specific implementation methods and application scope based on the ideas of this document. Therefore, the content of this specification should not be construed as a limitation of this document.

Claims

1. A knowledge graph-based online potential loss user identification method of a state grid, characterized in that, The method comprises the following steps: S100: constructing a user knowledge graph; combining user attribute and behavior data of the online State Grid APP and external multi-source data, and sorting out data information related to user loss risk; By identifying key entities in multi-dimensional data and extracting the correlation between entities, a power user knowledge graph is constructed; S200: in-depth mining of user characteristics; based on the user knowledge graph, a relational graph convolutional neural network is constructed to in-depth mine and analyze graph structure data, learn low-dimensional vector representation of entities and relationships in the knowledge graph, capture complex relationships between entities and deep features of graph structure data; The type of the entity includes: "User": representing a registered user of the online State Grid APP, and the attributes include: user ID, gender, age, occupation, education level, registration time, whether the user is a member user related information, and behavior data of the user on the online State Grid APP; "Industry": representing an industry owned by the user, and the attributes include: account number, belonging to the area, industry type and house type, residence time, housing attribute, industry scale, connection capacity, industry category, access voltage level, shop type, business area, business time rule, charging pile type, installed capacity, access mode, photovoltaic installed capacity, power generation, whether to configure a different industry type attribute information of the energy storage system; the industry type includes: residence, enterprise and institution, shop, charging pile, and distributed photovoltaic; "Electricity meter": representing an electricity meter used by the user, and the attributes include: electricity meter number, electricity meter type, electricity consumption category, electricity meter fault alarm, electricity consumption, electricity consumption time, and peak electricity consumption related electricity consumption data; "Bill": representing an electricity bill generated by the user, and the attributes include: amount to be paid, discount amount, actual payment amount, payment channel, payment deadline, actual payment date, overdue days, overdue times, maximum arrears amount, collection times, and collection response time; The correlation type includes: the user "owns" the industry, the user "uses" the electricity meter, and the user "generates" the bill; The graphs under the three relationships of "owning", "using" and "generating" are regarded as an isomorphic subgraph respectively, the extracted multi-relation feature vectors are stacked and processed, and the multi-relation graph is converted into a feature aggregation problem of multiple isomorphic subgraphs; for each subgraph, the vector representation of the entity and the relationship is first tiled into a two-dimensional matrix, and then the feature is extracted through a convolution layer, and then the extracted feature is mapped to the final vector representation space through a fully connected layer; S300: subdividing the user risk level; based on the extracted user features, two classifiers are constructed by referring to a multi-task learning network, and an independent output layer is designed for each classification task, one for identifying whether the user has a loss risk, and the other for predicting the level of the user loss risk; In the step of combining user attribute and behavior data of the online State Grid APP and external multi-source data, and sorting out data information related to user loss risk, the data information includes: User basic information, specifically including: user ID, gender, age, occupation, education level, industry category, whether a member, whether to open my subscription, and whether a temporary energy user; Industry information, specifically including: household number, belonging to the area, belonging to the power supply station, access mode, industry type; the industry type includes: residence, enterprise and institution, shop, charging pile, distributed photovoltaic; Electricity information, specifically including: meter number, meter type, electricity category, monthly average electricity consumption, electricity peak, electricity time distribution, meter fault alarm; Interaction behavior, specifically including: user login times in APP, active frequency, bill query, energy efficiency analysis, activity participation, message click rate, push response; Bill and payment, specifically including: amount payable, discount amount, actual payment amount, payment channel, payment timeliness, historical arrears record, number of reminders, response to reminders; payment channels include: online State Grid, bank deduction, third-party payment; For the constructed multi-relation power user knowledge graph, the node set V represents the "user", "industry", "meter", "bill" entities, and the edge set R represents the relationships between different entities "owns", "uses", "generates"; Each relation type r has its corresponding adjacency matrix The adjacency matrix represents the relations between entities in the subgraph corresponding to the relation type, each node has a feature vector representing the attributes of the entity, and the feature vectors of all nodes form a feature matrix H; For each node , the node updates its feature vector by aggregating all its neighbor nodes' feature vectors under different relations and its own attributes as follows: ， where node j is a neighbor node of node i, denotes the set of neighbor nodes of node i under relation r, h j is the feature vector of node j, is the degree of node i under relation r, i.e., the number of node edges in the isomorphic subgraph, is the weight matrix of the corresponding relation type r in the l-th layer, is the self-loop weight matrix in the l-th layer, used to maintain the node's own features; Using the above formula, the feature vector of the user node is constantly updated, and the attribute characteristics of all neighbor nodes and itself in the three relationships are iteratively aggregated, obtaining a vector representation fully fused with multi-dimensional data information of the power user; For the user vector representation mined by the relation graph convolutional neural network, first pass through the shared fully connected layer for feature mapping, and then input the low-dimensional feature vector into two classifiers respectively; The binary classifier outputs a real value, and the Sigmoid function maps the value to between 0 and 1, representing the probability of user loss. The multi-classifier can output multiple values, and these values are activated by the Softmax function, representing the probability of the user belonging to different loss risk levels.

Images