User label labeling method, credit scoring method and related devices

Abstract

The embodiment of the invention discloses a user label labeling method, a credit scoring method and related devices, and relates to the field of model training. In the description, the problem of how to label the second labels for the second users in the target domain data is solved through the target model. The target model is obtained by migrating an initial model obtained based on source domain data training to target domain data for learning, the type of the target model is a graph neural network model, a source domain graph is constructed through the source domain data, a target domain graph is constructed through the target domain data, and the graph neural network model is trained until the target model is obtained; and the loss function of the target model in the training process is constrained by the smoothness among a plurality of target domain nodes on the target domain graph and is constrained by the feature distribution difference between the source domain graph and the target domain graph.

Description

Technical Field

[0001] This specification relates to the field of model training, and in particular to a user labeling method, a credit scoring method, and related devices. Background Art

[0002] Credit scoring is designed to measure a customer's probability of credit default using machine learning methods. Based on the assessed creditworthiness, financial institutions such as banks and online lending companies can decide whether to approve or reject credit loan applications.

[0003] When a customer applies for a credit loan, the customer's application may be approved or rejected. If the application is approved, it will become an approval sample and the customer will get a loan. After a period of time, if the customer repays on time, it will become a non-default sample; if the customer does not repay on time, it will become a default sample. On the contrary, if the application is not approved, it will become a rejection sample and the customer will not be able to get a loan. Since the rejected sample did not get a loan, we cannot observe whether it will default or not. Therefore, credit scoring models are usually built based on approval samples, because we do not have real data of rejection samples labeled with default / non-default labels. The rejection / approval strategy is usually based on machine learning models or expert rules of customer characteristics, so the approval sample and the rejection sample have different feature distributions. This exposes us to the problem of non-random selection bias in the sample data.

[0004] When serving online, the credit scoring model needs to infer the credit of a loan applicant based on the feature distribution of approved samples and rejected samples. Training the credit scoring model with such biased data will lead to biased parameters of the credit scoring model, that is, the relationship between the predicted input features and the probability of default is incorrect, resulting in significant economic losses. Therefore, in order to reliably score credit, in addition to modeling the approved samples, we also need to consider the rejected samples and infer their true credit.

[0005] Domain adaptation is used to solve the problem of knowledge transfer between data distributions in different domains. Through domain adaptation technology, when the feature distributions of the source domain (such as approved samples) and the target domain (such as rejected samples) are different, the knowledge of the source domain can be transferred to the target domain, eliminating the distribution difference between the two domains, thereby building a more robust credit scoring model and improving the accuracy of the credit scoring model for customers.

[0006] However, existing user labeling methods have problems such as insufficient transfer ability, inability to accurately eliminate the distribution differences between the source and target domains, and overfitting the source domain. Summary of the invention

[0007] The embodiments of this specification provide a user tagging method, a credit scoring method and related devices, which can solve the above problems. The technical solution is as follows:

[0008] In a first aspect, an embodiment of the present specification provides a method for marking a user tag, the method comprising:

[0009] According to the multiple first users included in the source domain data and the first labels annotated by the first users, a source domain graph including multiple source domain nodes and the source domain nodes annotated with node labels is constructed, wherein the multiple first users correspond to the multiple source domain nodes respectively;

[0010] Training the graph neural network model to be trained according to the source domain graph to obtain an initial model;

[0011] According to the multiple second users included in the target domain data, construct a target domain graph including multiple target domain nodes, wherein the multiple second users correspond to the multiple target domain nodes respectively;

[0012] The initial model is transferred to the target domain graph for training until a target model is obtained, and a loss function of the target model during the training process is constrained by the smoothness between multiple target domain nodes on the target domain graph and by the feature distribution difference between the source domain graph and the target domain graph, and the target model is used to respectively label second labels for multiple second users included in the target domain data.

[0013] In a second aspect, the embodiments of this specification provide a credit scoring method, the method comprising:

[0014] Instructing the credit scoring model to be trained to output a predicted credit score of the second user for the second user according to the plurality of second users included in the target domain data and the second tags annotated by the second users, wherein the second tags annotated by the second user are obtained by the user tagging method described in the first aspect;

[0015] Acquiring an actual credit score of the second user, where the actual credit score of the second user is determined in the process of providing credit services to the second user based on the predicted credit score of the second user;

[0016] The credit score model to be trained is trained according to the predicted credit scores and the actual credit scores respectively corresponding to the plurality of second users until the credit score model is obtained, and the credit score model is used to output the credit score of the third user for a third user.

[0017] In a third aspect, an embodiment of the present specification provides a user tagging device, the device comprising:

[0018] A source domain construction module, configured to construct a source domain graph including a plurality of source domain nodes and the source domain nodes being labeled with node labels according to a plurality of first users included in the source domain data and a first label labeled by the first user;

[0019] An initial training module, used to train the graph neural network model to be trained according to the source domain graph to obtain an initial model;

[0020] A target construction module, configured to construct a target domain graph including a plurality of target domain nodes according to a plurality of second users included in the target domain data;

[0021] A migration training module is used to migrate the initial model to the target domain graph for training until a target model is obtained, and the loss function of the target model in the training process is constrained by the smoothness between multiple target domain nodes on the target domain graph, and by the feature distribution difference between the source domain graph and the target domain graph. The target model is used to respectively label second labels for multiple second users included in the target domain data.

[0022] In a fourth aspect, an embodiment of the present specification provides a credit scoring device, the device comprising:

[0023] a label prediction module, configured to instruct the credit score model to be trained to output a predicted credit score of the second user for the second user according to the plurality of second users included in the target domain data and the second labels annotated by the second users, wherein the second labels annotated by the second users are obtained by the user label annotating method described in the first aspect;

[0024] a label acquisition module, configured to acquire an actual credit score of the second user, where the actual credit score of the second user is determined in a process of providing credit services to the second user based on a predicted credit score of the second user;

[0025] A model training module is used to train the credit score model to be trained according to the predicted credit scores and actual credit scores respectively corresponding to the plurality of second users until the credit score model is obtained, and the credit score model is used to output the credit score of the third user for a third user.

[0026] In a fifth aspect, an embodiment of the present specification provides a computer storage medium, wherein the computer storage medium stores a plurality of instructions, wherein the instructions are suitable for being loaded by a processor and executing the above-mentioned method steps.

[0027] In a sixth aspect, an embodiment of the present specification provides a computer program product, wherein the computer program product stores a plurality of instructions, wherein the instructions are suitable for being loaded by a processor and executing the above-mentioned method steps.

[0028] In a seventh aspect, an embodiment of the present specification provides an electronic device, which may include: a processor and a memory; wherein the memory stores a computer program, and the computer program is suitable for being loaded by the processor and executing the above-mentioned method steps.

[0029] The beneficial effects brought by the technical solutions provided by some embodiments of this specification include at least:

[0030] In this specification, the problem of how to respectively label multiple second users in the target domain data with second labels is solved by a target model. The target model is obtained by migrating an initial model trained based on source domain data to target domain data for learning, and the type of the target model is a graph neural network model, and the target model is obtained by training the graph neural network model by constructing a source domain graph using source domain data and a target domain graph using target domain data.

[0031] Specifically, according to the multiple first users included in the source domain data and the first labels annotated by the first users, a source domain graph including multiple source domain nodes and the source domain nodes annotated with node labels is constructed, and the graph neural network model to be trained is further trained according to the source domain graph to obtain an initial model; according to the multiple second users included in the target domain data, a target domain graph including multiple target domain nodes is constructed; the initial model is migrated to the target domain graph for training until the target model is obtained, and the loss function of the target model in the training process is constrained by the smoothness between the multiple target domain nodes on the target domain graph, and by the difference in feature distribution between the source domain graph and the target domain graph.

[0032] In other words, in this specification, a user labeling model based on a graph neural network is first constructed as an initial model so that the target model can consider the relationship between multiple second users to label the second users. Secondly, in the process of transfer learning, the target model learns the feature representation of the source domain graph constructed by the source domain data and the feature representation of the target domain graph constructed by the target domain data, capturing the relationship characteristics between multiple first users and multiple second users. And the loss function used in the target model over-training process is constrained by the smoothness between multiple target domain nodes on the target domain graph and the difference in feature distribution between the source domain graph and the target domain graph, thereby enhancing the robustness and generalization ability of the target model, thereby improving the accuracy and reliability of the target model in labeling the second labels for multiple second users included in the target domain data. BRIEF DESCRIPTION OF THE DRAWINGS

[0033] In order to more clearly illustrate the embodiments of this specification or the technical solutions in the prior art, the drawings required for use in the embodiments or the description of the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of this specification. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying any creative work.

[0034] Figure 1 This is a scenario diagram of a user tagging method provided by an embodiment of this specification;

[0035] Figure 2 It is a flowchart of a user tagging method provided by an embodiment of this specification;

[0036] Figure 3 This is a schematic diagram of a scenario in which a user applies for a credit service provided in an embodiment of this specification;

[0037] Figure 4 It is a flowchart of a user tagging method provided by an embodiment of this specification;

[0038] Figure 5 is a schematic diagram of the structure of a target domain graph provided in an embodiment of this specification;

[0039] Figure 6 It is a flowchart of a user tagging method provided by an embodiment of this specification;

[0040] Figure 7 It is a flowchart of a user tagging method provided by an embodiment of this specification;

[0041] Figure 8 It is a flowchart of a credit scoring method provided in an embodiment of this specification;

[0042] Fig. 9 It is a structural schematic diagram of a user label marking device provided in an embodiment of this specification;

[0043] Fig.10 It is a structural diagram of a credit scoring method provided in an embodiment of this specification;

[0044] Fig.11 It is a structural schematic diagram of an electronic device provided in an embodiment of this specification. DETAILED DESCRIPTION

[0045] The following will be combined with the drawings in the embodiments of this specification to clearly and completely describe the technical solutions in the embodiments of this specification. Obviously, the described embodiments are only part of the embodiments of this specification, not all of the embodiments. Based on the embodiments in this specification, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of this specification.

[0046] In the description of this specification, it should be understood that the terms "first", "second", etc. are only used for descriptive purposes and cannot be understood as indicating or implying relative importance. In the description of this specification, it should be noted that, unless otherwise clearly specified and limited, "including" and "having" and any of their variations are intended to cover non-exclusive inclusions. For example, a process, method, system, product or device that includes a series of steps or units is not limited to the listed steps or units, but optionally also includes steps or units that are not listed, or optionally also includes other steps or units inherent to these processes, methods, products or devices. For those of ordinary skill in the art, the specific meanings of the above terms in this specification can be understood in specific circumstances. In addition, in the description of this specification, unless otherwise specified, "multiple" refers to two or more. "And / or" describes the association relationship of associated objects, indicating that there can be three relationships, for example, A and / or B, which can represent: A exists alone, A and B exist at the same time, and B exists alone. The character " / " generally indicates that the associated objects before and after are an "or" relationship.

[0047] The present specification is described in detail below with reference to specific embodiments.

[0048] It should be noted that the information (including but not limited to user device information, user personal information, etc.), data (including but not limited to data used for analysis, stored data, displayed data, etc.) and signals involved in the embodiments of this specification are all authorized by the user or fully authorized by all parties, and the collection, use and processing of relevant data must comply with the relevant laws, regulations and standards of relevant countries and regions. For example, the features, information and data involved in this specification are all obtained with full authorization.

[0049] like Figure 1 As shown, Figure 1 It is a schematic diagram of the architecture of a user tagging method provided in an embodiment of this specification. Figure 1 The system at least includes a server 102 for collecting source domain data and / or target domain data, a server 101 for executing a user tagging method, and a plurality of electronic devices for uploading and sending source domain data and / or target domain data. The plurality of electronic devices at least includes an electronic device 1031, an electronic device 1032, and an electronic device 1033. It is understandable that Figure 1 The number of servers and electronic devices shown in the figure is for illustration only, and this specification does not impose any limitation on this.

[0050] The above-mentioned servers include but are not limited to physical or virtual processors, mobile stations (MS), mobile terminals, mobile phones, handsets and portable equipment, Bluetooth headsets, smart watches and other types. The electronic devices can communicate with one or more core networks via the Radio Access Network (RAN).

[0051] For example, the server is a plurality of physical servers, and the plurality of physical servers are independent in hardware. Or, the server is a plurality of virtual servers, and the plurality of virtual servers are deployed in the same hardware resource pool, and the deployment methods of the virtual servers include but are not limited to: VMware, Virtual Box and Virtual PC.

[0052] It is understandable that the server and other servers also have other service capabilities and functions to complete the tasks in the following embodiments. For example, the server also provides portal services, resource management services, and CI / CD services.

[0053] In an embodiment of the present specification, the server 102 is used to send the source domain data and / or the target domain data to the server 101. The server 101 is used to execute the user labeling method to label multiple users in the target domain data with a second label. In another embodiment, the server 102 is also used to receive the target domain data labeled with the second label sent by the server 101, and to execute the credit scoring method to perform a credit scoring on the third user to obtain the credit score of the third user.

[0054] In the embodiments of the present specification, electronic devices such as electronic device 1021, electronic device 1022 and electronic device 1023 may also be equipped with display devices, and the display devices may be various devices capable of realizing display functions, for example, the display devices may be cathode ray tube display (Cathode ray tube display, referred to as CR), light-emitting diode display (Light-emitting diode display, referred to as LED), electronic ink screen, liquid crystal display (Liquid crystal display, referred to as LCD), plasma display panel (Plasma display panel, referred to as PDP), etc. For example, a user may use the display device on electronic device 1021 to send a credit service application to server 102, or view the information sent by server 102.

[0055] Multiple electronic devices and multiple servers can communicate through communication links established by communication protocols, such as the gRPC protocol. gRPC is a high-performance, general-purpose open source remote procedure call (RPC) framework, which is mainly designed for mobile application development and based on the HTTP / 2 protocol standard, based on the protocol buffer (PB) serialization protocol development, and supports many development languages. In addition, the communication link can also be a wireless communication link or a wired communication link, for example, a wired communication link includes an optical fiber, a twisted pair or a coaxial cable, and a wireless communication link includes a Bluetooth communication link, a wireless fidelity (WIreless-FIdelity, Wi-Fi) communication link or a microwave communication link.

[0056] In one embodiment, Figure 2 The figure is a flow chart of a user labeling method provided in an embodiment of the present specification, which can be implemented by a computer program and can be run on a user labeling device based on the von Neumann system. The computer program can be integrated into an application or run as an independent tool application.

[0057] Specifically, the user tagging method includes:

[0058] S102: construct a source domain graph including multiple source domain nodes and source domain nodes annotated with node labels according to multiple first users included in the source domain data and first labels annotated by the first users.

[0059] Multiple first users correspond to multiple source domain nodes respectively. The source domain nodes in the source domain graph represent various entities in the source domain data. In this embodiment, users in the source domain data are represented by a source domain node in the source domain graph. The node label of the source domain node is obtained by the first label of the first user corresponding to the source domain node. The first label of the first user represents the characteristic information of the user, and the node label represents the category or attribute of the source domain node. For example, the characteristic information of the user includes the user's family relationship, assets owned, type of credit product applied for, loan amount, etc.

[0060] The source domain graph can be viewed as a directed or undirected graph, where nodes represent entities in the source domain and edges represent relationships between nodes. For example, the edge between node A and node B may represent the following types of relationships: similarity relationship, dependency relationship, interaction relationship, and conversion relationship. For another example, node A and node B represent user A and user B respectively. Since there is a loan relationship between user A and user B, it is represented that there is an edge relationship between node A and node B on the source domain graph.

[0061] In this embodiment, the method for constructing a source domain graph based on source domain data can be based on graph embedding models (such as DeepWalk, Node2Vec, LINE, etc.), graph convolutional neural networks (Graph Convolutional Networks, GCN), graph autoencoders (Graph Autoencoder), graph attention networks (Graph Attention Networks, GAT), graph generation adversarial networks, etc., and can also be other methods for constructing graphs.

[0062] In one embodiment, the source domain data is data related to a first user applying for a credit service and being approved, and a first label marked by the first user is default or non-default; the target domain data is data related to a second user applying for a credit service and being rejected, and a second label marked by the second user is default or non-default.

[0063] like Figure 3 As shown, Figure 3 201 is a schematic diagram of a scenario in which a user applies for a loan service provided by an embodiment of this specification. When user 201 applies for a loan application, the user's application may be approved or rejected. If the application is approved, it will become an approved sample Approved Samples, as part of the source domain data 202, and the user will obtain the loan service. After a period of time, if user 201 repays the loan on time, it will become a non-default sample, and the first label corresponding to the user 201 is non-default. If user 201 fails to repay the loan on time, it will become a default sample, and the first label corresponding to the user 201 is default.

[0064] On the contrary, if user 201 applies for credit services but is rejected, it will become a rejected sample RejectedSamples, and user 201 will not be able to obtain a loan. Since the rejected sample does not receive the application service, we cannot observe whether it will default or not.

[0065] Therefore, in this embodiment, it is necessary to annotate a second label for multiple second users in the target domain data related to the second user applying for credit services and being rejected based on the source domain data including the first user applying for credit services and being approved, and the second label is default or non-default.

[0066] S104. Train the graph neural network model to be trained according to the source domain graph to obtain an initial model.

[0067] The graph neural network model to be trained is trained according to the source domain graph to obtain an initial model, and the initial model is used to predict the first label for other users who meet the preset conditions corresponding to the source domain data. For example, the source domain data is data related to the first user applying for credit services and being approved, and the first label is default or non-default. The initial model obtains the feature information related to the credit behavior of the user in the source domain graph, as well as the edge features between the source domain node of the user and the source domain nodes of other users, and marks the first label of default or non-default for the user.

[0068] The graph neural network model to be trained proposed in this embodiment can be any desired graph neural network model, for example, a graph convolutional network GCN, a graph attention network GAT, etc. This specification embodiment does not impose any limitation on this.

[0069] Taking the graph neural network model GNN as an example, the process of training the graph neural network model to be trained includes: obtaining multiple embedding vectors embedding based on the feature information conversion corresponding to multiple source domain nodes on the source domain graph; in each layer of the graph neural network model GNN, message forward propagation Message Passing will be performed, that is, each node collects information from its neighboring nodes, and the node updates its own feature representation. This process can be repeated between multiple layers so that each node can obtain information from more distant neighbors; after several rounds of message passing, the graph neural network model GNN needs to convert the learned node representation into the final output to obtain the category probability distribution; and for the user label prediction task, a loss function is constructed, and the difference between the model prediction result and the actual label is calculated as the loss value. When the loss value meets the loss threshold, back propagation and optimization are performed in the graph neural network model GNN, the weight of the graph neural network model GNN is adjusted according to the gradient of the loss relative to the model parameters, and the parameters of the graph neural network model GNN are updated using an optimizer (such as Adam, SGD, etc.); after multiple iterative training, the initial model is obtained.

[0070] It is understandable that the above training process simplifies many details, such as the selection of hyperparameters, the application of regularization techniques such as dropout and early stopping, and different training methods for different types of graph neural network models. The above are only examples.

[0071] In one embodiment, the multiple source domain nodes are source domain nodes of different types, and the edges connecting the multiple source domain nodes include multiple types; the graph neural network model to be trained is trained according to the heterogeneous features included in the source domain graph to obtain an initial model, and the heterogeneous features included in the source domain graph are obtained through the multiple source domain nodes and the edges connecting the multiple source domain nodes.

[0072] In other words, in this embodiment, since the multiple source domain nodes are different types of source domain nodes constructed according to different entities in the source domain data, for example, the entities included in the source domain data include users, institutions, firms, banks, etc., the types of the source domain nodes include multiple types. Since the types of the multiple source domain nodes mapped by the multiple types of entities are different, the edges connected between the multiple source domain nodes include multiple types, for example, the type of the edge constructed between two source domain nodes corresponding to two users respectively represents a friend relationship, and the type of the edge constructed between two source domain nodes corresponding to two firms respectively represents a friendship relationship.

[0073] Therefore, directly applying the standard homogeneous graph neural network Homogeneous GNNs may not be able to effectively capture all information. In this embodiment, heterogeneous graph neural network Heterogeneous GNNs are used to process the features of different types of nodes and edges.

[0074] Specifically, the heterogeneous graph neural network model to be trained is trained according to the heterogeneous features included in the source domain graph, including capturing the edge relationships between different types of source domain nodes through Meta-path, and introducing the attention mechanism to assign different weights to different types of edge relationships, so that the heterogeneous graph neural network model can automatically select important relationship types for learning according to task requirements, and multi-view learning can also be introduced to enable the heterogeneous graph neural network model to be observed from multiple angles (or views), each view corresponding to a specific type of node or edge.

[0075] In this embodiment, the heterogeneous graph neural network model is trained to obtain an initial model to perform the task of labeling users, which can improve the generalization ability of the initial model, effectively handle the complex structure of the source domain graph and the target domain graph, and adapt to a variety of application scenarios. In particular, in terms of cross-domain migration, since the initial model needs to be migrated from the source domain graph to the target domain graph for transfer learning training to obtain the target model, the initial model based on the heterogeneous graph neural network is easier to adjust to the new environment, reducing the need to redesign the model.

[0076] S106. Construct a target domain graph including multiple target domain nodes according to the multiple second users included in the target domain data.

[0077] The plurality of second users correspond to the plurality of target domain nodes respectively. A target domain graph including the plurality of target domain nodes is constructed according to the target domain data including the plurality of second users. The construction method is as follows: according to the plurality of first users included in the source domain data and the first label annotated by the first user, a source domain graph including the plurality of source domain nodes and the source domain nodes annotated with the node label is constructed in S102. The target domain nodes are also annotated with the node label, which represents the user features of the second user and is used to construct the edge between the target domain nodes, thereby instructing the target model to learn.

[0078] S108. Migrate the initial model to the target domain graph for training until a target model is obtained, and during the training process, a loss function of the target model is constrained by the smoothness between multiple target domain nodes on the target domain graph and by the feature distribution difference between the source domain graph and the target domain graph. The target model is used to respectively label second labels for multiple second users included in the target domain data.

[0079] The target model is used to label the second tags for the multiple second users included in the target domain data. Figure 3 As shown, when user 201 applies for a credit service target model and is rejected, the target model labels a second label for a second user as target domain data 203, and the second label labeled by the second user is default or non-default.

[0080] Unsupervised Graph Domain Adaptation (UGDA) aims to transfer knowledge from a source domain graph with a first label to a target domain graph without a label to solve the distribution transfer problem between graph domains. Its core principle is to learn domain-invariant representations to ensure that node representations are not affected by environmental changes during knowledge transfer.

[0081] Most existing research focuses on aligning the data of source domain graphs and target domain graphs in the representation space through graph neural networks (GNNs). However, graph neural networks (GNNs) are highly dependent on the stability of local structural features. Even a small change in the structure of the target domain graph will cause a significant change in the node representation of a target domain node on the target domain graph, making it difficult for other target domain nodes to comply with the invariance principle.

[0082] In order to solve this problem, in the embodiment of this specification, when the initial model is migrated to the target domain graph for training until the target model is obtained, the loss function of the target model during the training process is constrained by the smoothness between multiple target domain nodes on the target domain graph, and by the feature distribution difference between the source domain graph and the target domain graph, as shown in the following formula:

[0083] ε Q (f)≤ε P (f)+2∈+2MTV(P,Q)+K;

[0084] Among them, Q represents the sample distribution of the target domain map, P represents the sample distribution of the source domain map, ε represents smoothness, f represents the target model, MTV(P,Q) represents the feature distribution difference between the source domain map and the target domain map, and K represents the intermediate variable.

[0085] In the embodiments of the present specification, the loss function of the target model is constrained by the smoothness between multiple target domain nodes on the target domain graph. Smoothness generally refers to the continuity or smoothness of data distribution. Specifically, smoothness reflects the relationship between multiple nodes. If the nodes are smooth in some feature space, then similar nodes usually have similar labels or outputs. In other words, when the data distribution is relatively smooth, the prediction results of the target model will not change drastically in the feature space.

[0086] In this embodiment, multiple target domain nodes on the target domain graph have smoothness, which means that target domain nodes adjacent to a target domain node on the target domain graph have similar node labels, which can be understood as target domain data that a second user similar to a second user has a similar second label. Therefore, the target model can use this smoothness for better learning, and the smoothness constraint encourages the target model to generate consistent predictions on adjacent nodes, thereby improving the generalization ability of the target model.

[0087] And in transfer learning, there is usually a difference between the data distribution of the target domain data and the data distribution of the source domain data, which may lead to unstable training or overfitting of the target model. Therefore, by introducing a smoothness constraint, it aims to limit the prediction of the target model to be consistent between similar inputs. This constraint can help the model adapt to the distribution of the target domain, thereby better transferring the source domain knowledge to the target domain.

[0088] Smoothness constraints can be implemented in the following ways: regularization, such as L2 regularization, smoothing function, etc., to encourage smooth changes in model parameters; constructing a specific loss function, such as a distance-based loss function, etc. The embodiments of this specification may also include other smoothness constraint methods.

[0089] And in the embodiments of the present specification, the loss function of the target model is also constrained by the difference in feature distribution between the source domain graph and the target domain graph. In the process of transfer learning, the data distribution of the source domain data and the target domain data is often different, which is characterized by the difference in feature space, label distribution and structure between the source domain data and the target domain data. This difference will affect the training process of the initial model on the target domain graph and the calculation of the loss function. This is because the loss function is usually calculated based on the source domain data, and if the data distribution of the target domain is inconsistent with the source domain data distribution, the initial model trained on the source domain may not be able to achieve the same performance on the target domain graph. Therefore, the loss function needs to take into account the difference in feature distribution between the source domain data and the target domain data in order to effectively make the trained target model perform better on the target domain graph.

[0090] Methods for solving the feature distribution differences between the source domain graph and the target domain graph in transfer learning may be: adversarial training, so that the feature distributions of the source domain data and the target domain data are as close as possible; Maximum Mean Discrepancy (MMD), which is a method for measuring the feature distribution differences between the source domain data and the target domain data. By minimizing the MMD value between the source domain data and the target domain data, the feature distributions of the source domain and the target domain can be made more similar; Feature Alignment, aligning the feature spaces of the source domain data and the target domain data; Self-supervised Learning, by constructing self-supervised tasks (e.g., data enhancement, prediction tasks, etc.) to help the target model learn effective features in the unlabeled target domain. The embodiments of this specification may also include other methods for reducing the feature distribution differences between the source domain graph and the target domain graph.

[0091] Based on the above steps and formulas, the initial model is migrated to the target domain graph for migration training, and the parameters of the target model are fine-tuned according to the loss function of the target model until the target model is obtained when the preset conditions are met.

[0092] In this specification, the problem of how to respectively label multiple second users in the target domain data with second labels is solved by a target model. The target model is obtained by migrating an initial model trained based on source domain data to target domain data for learning, and the type of the target model is a graph neural network model, and the target model is obtained by training the graph neural network model by constructing a source domain graph using source domain data and a target domain graph using target domain data.

[0093] In other words, in this specification, a user labeling model based on a graph neural network is first constructed as an initial model so that the target model can consider the relationship between multiple second users to label the second users. Secondly, in the process of transfer learning, the target model learns the feature representation of the source domain graph constructed by the source domain data and the feature representation of the target domain graph constructed by the target domain data, capturing the relationship characteristics between multiple first users and multiple second users. And the loss function used in the target model over-training process is constrained by the smoothness between multiple target domain nodes on the target domain graph and the difference in feature distribution between the source domain graph and the target domain graph, thereby enhancing the robustness and generalization ability of the target model, thereby improving the accuracy and reliability of the target model in labeling the second labels for multiple second users included in the target domain data.

[0094] In one embodiment, Figure 4 The figure is a flow chart of a user labeling method provided in an embodiment of the present specification, which can be implemented by a computer program and can be run on a user labeling device based on the von Neumann system. The computer program can be integrated into an application or run as an independent tool application.

[0095] Specifically, the user tagging method includes:

[0096] S202: Construct a source domain graph including multiple source domain nodes and source domain nodes annotated with node labels according to multiple first users included in the source domain data and first labels annotated by the first users.

[0097] Refer to the above S102, which will not be repeated here.

[0098] S204. Train the graph neural network model to be trained according to the source domain graph to obtain an initial model.

[0099] Refer to the above S104 and will not be repeated here.

[0100] S206: Construct a target domain graph including multiple target domain nodes according to the multiple second users included in the target domain data.

[0101] Refer to the above S106, which will not be repeated here.

[0102] S208, migrating the initial model to the target domain graph for training, and determining at least one target domain node adjacent to the target domain node based on a random walk method during the training process as required for calculating smoothness, until the target model is obtained.

[0103] Random Walk refers to a series of random node jumps in the target domain. Each time starting from the current target domain node, a target domain node adjacent to the target domain node is selected with a certain probability, and then the same process is continued.

[0104] Specifically, if Figure 5 As shown, Figure 5 It is a structural schematic diagram of a target domain graph provided by an embodiment of this specification. The target domain node A on the target domain graph represents the user in the target domain data, and each edge represents the relationship between multiple target domain nodes. Starting from the target domain node A, a target domain node adjacent to the target domain node A is randomly selected with a certain probability, and then the jump continues until a certain stop condition (such as the maximum number of steps, a certain smoothness condition, etc.) is met. The multiple target domain nodes determined based on the random walk are used as the multiple target domain nodes required for calculating the smoothness in the loss function of the target model, so that the edge relationship between the multiple target domain nodes is used to calculate the smoothness in the loss function of the target model when training the target model.

[0105] The Laplace smoothing term is calculated based on the random walk method. See the following formula:

[0106]

[0107] (i, j) represents a pair of neighbor nodes obtained by random walk sampling on the target domain graph; f(x i ) represents the characterization vector of node i; Tr is the trace of the matrix, that is, the sum of all elements on the diagonal of the matrix, d i is the degree of node i, d j is the degree of node j, A ij is a graph adjacency matrix, A ij =1 means (i, j) is a pair of neighbors obtained by random walk sampling, A ij =0 means not a neighbor.

[0108] In this embodiment, directly calculating the smoothness between all target domain nodes on the target domain graph to constrain the loss function of the target model requires a lot of computing resources, especially when the target domain graph is large. Therefore, the target domain nodes required for calculating the smoothness are determined by a local random walk, the global information is approximately calculated, and the dimensionality disaster faced by the full graph calculation is avoided, the number of node pairs that need to be calculated is effectively reduced, the amount of calculation is significantly reduced, and the training efficiency of the target model is improved.

[0109] In this specification, the problem of how to respectively label multiple second users in the target domain data with second labels is solved by a target model. The target model is obtained by migrating an initial model trained based on source domain data to target domain data for learning, and the type of the target model is a graph neural network model, and the target model is obtained by training the graph neural network model by constructing a source domain graph using source domain data and a target domain graph using target domain data.

[0110] In this specification, a user labeling model based on a graph neural network is first constructed as an initial model so that the target model can consider the relationship between multiple second users to label the second users. Secondly, in the process of transfer learning, the target model learns the feature representation of the source domain graph constructed by the source domain data and the feature representation of the target domain graph constructed by the target domain data, capturing the relationship characteristics between multiple first users and multiple second users. And the loss function used in the target model over-training process is constrained by the smoothness between multiple target domain nodes on the target domain graph and the difference in feature distribution between the source domain graph and the target domain graph, thereby enhancing the robustness and generalization ability of the target model, thereby improving the accuracy and reliability of the target model in labeling the second labels for multiple second users included in the target domain data.

[0111] In one embodiment, Figure 6 The figure is a flow chart of a user labeling method provided in an embodiment of the present specification, which can be implemented by a computer program and can be run on a user labeling device based on the von Neumann system. The computer program can be integrated into an application or run as an independent tool application.

[0112] Specifically, the user tagging method includes:

[0113] S302: Construct a source domain graph including multiple source domain nodes and source domain nodes labeled with node labels according to multiple first users included in the source domain data and first labels labeled by the first users.

[0114] Refer to the above S102, which will not be repeated here.

[0115] S304. Convert the source domain graph into a feature representation corresponding to the source domain graph through a preset graph representation learning method, train the graph neural network model to be trained according to the feature representation corresponding to the source domain graph, and obtain an initial model.

[0116] The core idea of ​​graph representation learning methods is to model data through graph structures and learn data representation by using the connection relationships between nodes.

[0117] In unsupervised domain adaptation, this embodiment hopes to map samples of the source domain and the target domain to the same feature space through the same preset graph representation learning method. The purpose of this is to make the feature distribution differences of the source domain and the target domain as close as possible in this feature space, thereby reducing the distribution differences between them.

[0118] In this embodiment, the source domain data and the target domain data are converted into the same representation space through the graph convolutional network model GCN or the graph attention network model GAT.

[0119] This is because the core idea of ​​methods such as graph convolutional network models GCN or graph attention network models GAT is to propagate information through the structure of the graph, and the embedding (representation) of a node is obtained by aggregating the information of its neighboring nodes. This aggregation method is sensitive to local structure, that is, whether it is the source domain graph or the target domain graph, the representation of the node depends on the information of their neighboring nodes and the connection relationship of the graph.

[0120] Therefore, in cross-domain transfer learning (Domain Adaptation), although the data distribution of the source domain data and the target domain data may be different, the graph structures of the source domain graph and the target domain graph may be similar or can be aligned. Pre-set graph representation learning methods such as the graph convolutional network model GCN or the graph attention network model can use this similar graph structure to learn similar node representations, thereby mapping the source domain nodes on the source domain graph and the target domain nodes on the target domain graph to the same representation space.

[0121] The similarity of connections between multiple nodes on the source domain graph and the target domain graph and the similarity of graph structure enable the target model to learn similar embedding representations in the source domain graph and the target domain graph, thereby reducing the feature distribution difference between the target domain data and the source domain data, and improving the success rate of transfer learning from the initial model to the target domain to obtain the target model.

[0122] The step of converting the source domain graph into the feature representation corresponding to the source domain graph, and further training the graph neural network model to be trained according to the feature representation corresponding to the source domain graph, and obtaining the initial model refers to the above S104, which will not be repeated here.

[0123] S306: Construct a target domain graph including multiple target domain nodes according to the multiple second users included in the target domain data.

[0124] Refer to the above S106, which will not be repeated here.

[0125] S308. Convert the target domain graph into a feature representation corresponding to the target domain graph through a graph representation learning method, and migrate the initial model to the feature representation corresponding to the target domain graph for training until the target model is obtained.

[0126] In S304, a preset graph representation learning method such as a graph convolutional network model GCN or a graph attention network model GAT is used to convert the source domain graph into a feature representation corresponding to the source domain graph. In this step, the target domain graph is converted into a feature representation corresponding to the target domain graph based on the same graph representation learning method. For example, the source domain graph is converted into a feature representation corresponding to the source domain graph, and the target domain graph is converted into a feature representation corresponding to the target domain graph at the same time through a graph convolutional network model.

[0127] Through the preset graph representation learning method, it is possible to share the same learning mechanism through the graph structure, map the source domain graph and the target domain graph to the same representation space, reduce the feature distribution difference between the target domain data and the source domain data, and achieve cross-domain data alignment and transfer learning of the initial model.

[0128] The initial model is transferred to the feature representation corresponding to the target domain graph for training until the target model is obtained, see S108 above, which will not be repeated here.

[0129] In this specification, the problem of how to respectively label multiple second users in the target domain data with second labels is solved by a target model. The target model is obtained by migrating an initial model trained based on source domain data to target domain data for learning, and the type of the target model is a graph neural network model, and the target model is obtained by training the graph neural network model by constructing a source domain graph using source domain data and a target domain graph using target domain data.

[0130] In this specification, a user labeling model based on a graph neural network is first constructed as an initial model so that the target model can consider the relationship between multiple second users to label the second users. Secondly, in the process of transfer learning, the target model learns the feature representation of the source domain graph constructed by the source domain data and the feature representation of the target domain graph constructed by the target domain data, capturing the relationship characteristics between multiple first users and multiple second users. And the loss function used in the target model over-training process is constrained by the smoothness between multiple target domain nodes on the target domain graph and the difference in feature distribution between the source domain graph and the target domain graph, thereby enhancing the robustness and generalization ability of the target model, thereby improving the accuracy and reliability of the target model in labeling the second labels for multiple second users included in the target domain data.

[0131] In one embodiment, Figure 7 The figure is a flow chart of a user labeling method provided in an embodiment of the present specification, which can be implemented by a computer program and can be run on a user labeling device based on the von Neumann system. The computer program can be integrated into an application or run as an independent tool application.

[0132] Specifically, the user tagging method includes:

[0133] S402: Construct a source domain graph including multiple source domain nodes and source domain nodes annotated with node labels according to multiple first users included in the source domain data and first labels annotated by the first users.

[0134] Refer to the above S102, which will not be repeated here.

[0135] S404: Train the graph neural network model to be trained according to the source domain graph to obtain an initial model.

[0136] Refer to the above S104 and will not be repeated here.

[0137] S406: Construct a target domain graph including multiple target domain nodes according to the multiple second users included in the target domain data.

[0138] Refer to the above S106, which will not be repeated here.

[0139] S408, migrating the initial model to the target domain graph for training, and reducing the feature distribution difference between the source domain graph and the target domain graph through a preset domain alignment method during the training process until the target model is obtained.

[0140] Domain alignment is a key step to achieve transfer learning by reducing the difference in feature distribution between the source domain and the target domain. This is crucial for transfer learning because there are usually differences in the feature distribution of source domain data and target domain data. Alignment can reduce these differences, so that the target model can also perform better when processing target domain data.

[0141] In one embodiment, the initial model is transferred to the target domain graph for training, and the feature distribution difference between the source domain graph and the target domain graph is reduced during the training process until the target model is obtained, including:

[0142] The initial model is transferred to the target domain map for training, and during the training process, the feature distribution difference between the source domain map and the target domain map is narrowed by a preset domain alignment method until the target model is obtained.

[0143] The preset domain alignment methods include: maximum mean difference (MMD) alignment method, adversarial training-based alignment method (such as DANN), etc. In graph structured data, the topological structure information of the graph can also be used for alignment, such as subgraph-based alignment, graph embedding alignment, etc.

[0144] Specifically, the maximum mean difference (MMD) is a common statistic used to measure the difference between two distributions. The maximum mean difference (MMD) measures the distribution difference between the source domain and the target domain by calculating the mean difference between them in the feature space. In transfer learning, the maximum mean difference (MMD) is used to minimize the distribution difference between the source domain and the target domain. In transfer learning of graph neural networks, the maximum mean difference (MMD) between the source domain data and the target domain data can be calculated in the graph embedding space to optimize the target model so that the feature distributions of the source domain and the target domain are closer, thereby achieving the purpose of domain alignment.

[0145] Adversarial Training (DANN) is a domain alignment method based on adversarial training. This method introduces a domain classifier and adopts an adversarial training mechanism to make the features of the source domain and the target domain indistinguishable, that is, the domain classifier cannot distinguish between samples in the source domain and the target domain. In this way, the representation learned by the feature learner does not depend on the domain information, thereby achieving alignment between the source domain and the target domain. In the transfer learning of graph neural networks, adversarial training can be used to perform domain alignment of graph embeddings. Specifically, the feature representation in the graph neural network can be aligned through the adversarial training method, so that the graph structure and node representation of the source domain and the target domain become more consistent.

[0146] It can be understood that there are many possible domain alignment methods, and the above are only examples.

[0147] In this embodiment, according to the structure of the source domain graph and the target domain graph and the requirements of transfer learning, a required domain alignment method can be determined from multiple preset domain alignment methods, or multiple domain alignment methods can be determined to be used in combination to perform transfer learning on the initial model.

[0148] For example, when the feature distribution difference between the source domain map and the target domain map is large, the maximum mean difference (MMD) alignment method is used; when the labels on the target domain map are scarce or different from those on the source domain map, adversarial training DANN is used for unsupervised alignment; when the local structure similarity is strong, the subgraph-based domain alignment method is used; when the global structure difference is large, the graph embedding alignment method is used.

[0149] By selecting a suitable domain alignment method to effectively align the source domain data and the target domain data, and migrating the initial model to the target domain graph for training until the target model is obtained, the performance of the target model in the target domain can be improved, and the robustness of the target model can be improved.

[0150] In this specification, the problem of how to respectively label multiple second users in the target domain data with second labels is solved by a target model. The target model is obtained by migrating an initial model trained based on source domain data to target domain data for learning, and the type of the target model is a graph neural network model, and the target model is obtained by training the graph neural network model by constructing a source domain graph using source domain data and a target domain graph using target domain data.

[0151] In this specification, a user labeling model based on a graph neural network is first constructed as an initial model so that the target model can consider the relationship between multiple second users to label the second users. Secondly, in the process of transfer learning, the target model learns the feature representation of the source domain graph constructed by the source domain data and the feature representation of the target domain graph constructed by the target domain data, capturing the relationship characteristics between multiple first users and multiple second users. And the loss function used in the target model over-training process is constrained by the smoothness between multiple target domain nodes on the target domain graph and the difference in feature distribution between the source domain graph and the target domain graph, thereby enhancing the robustness and generalization ability of the target model, thereby improving the accuracy and reliability of the target model in labeling the second labels for multiple second users included in the target domain data.

[0152] In one embodiment, Figure 8 The figure is a flow chart of a credit scoring method provided in an embodiment of the present specification, which can be implemented by a computer program and can be run on a credit scoring device based on the von Neumann system. The computer program can be integrated into an application or run as an independent tool application.

[0153] Specifically, the credit scoring method includes:

[0154] S502: Instruct the credit scoring model to be trained to output a predicted credit score of the second user according to the multiple second users included in the target domain data and the second tags annotated by the second users.

[0155] The second label annotated by the second user is obtained by any user label annotating method in the above embodiments.

[0156] Therefore, when serving online, the credit scoring model needs to infer the credit of a loan applicant based on the feature distribution of approved samples and rejected samples. Training the credit scoring model with such biased data will lead to biased parameters of the credit scoring model, that is, the relationship between the predicted input features and the probability of default is incorrect, resulting in significant economic losses. Therefore, in order to have reliable credit scoring, in addition to modeling the approved samples, we also need to consider the rejected samples and infer their true credit.

[0157] The target domain data is data related to the second user applying for credit services and being rejected, and the second label annotated by the second user is overdue or not overdue. Based on the data of the rejected samples annotated with the second label, the credit score model to be trained is instructed to output the predicted credit score of the second user for the second user.

[0158] For example, user A applied for a credit service and was rejected, and the second label corresponding to the target user A was labeled as overdue by the user label labeling method provided by the embodiment of this specification, and the user characteristic information corresponding to user A also includes occupation, existing debt situation, and the relationship between other users. Here, the second label of "overdue" or "not overdue" can be used as a target variable when training the credit scoring model to be trained, as a label in the supervised learning task, and the ultimate goal is to generate a credit score for the third user output for the third user.

[0159] The credit scoring model to be trained outputs a predicted credit score of user A for user A. The predicted credit score represents the probability that user A will be overdue in the future. The higher the predicted credit score, the lower the possibility of user A being overdue.

[0160] The credit scoring model to be trained may be one or more machine learning models, and commonly used models include logistic regression, decision tree, random forest, gradient boosting machine GBM, support vector machine SVM, or more complex neural network architectures, such as neural networks (GNNs). In this embodiment, the credit scoring model may also be other types of models, and this embodiment does not impose any restrictions on this.

[0161] S504: Obtain the actual credit score of the second user.

[0162] The actual credit score of the second user is determined in the process of providing credit services to the second user based on the predicted credit score of the second user. Specifically, the actual credit score is usually determined by a credit institution or a related financial service provider through a multi-dimensional credit assessment during the credit approval process. These assessment criteria may include multiple factors such as the second user's historical credit record, borrowing behavior, and debt situation.

[0163] S506: Train the credit scoring model to be trained according to the predicted credit scores and actual credit scores respectively corresponding to the plurality of second users, until a credit scoring model is obtained.

[0164] The credit scoring model is used to output the credit score of the third user. When providing online services, the user labeling method provided in the embodiments of this specification is used to label the third user who applies for credit services but may be rejected with an overdue or non-overdue label according to the target model, and the credit score of the third user is further evaluated by the credit scoring model. The credit score represents the possibility of the third user being overdue or non-overdue when providing credit services to the third user in the future.

[0165] The process of training the credit scoring model to be trained, taking the graph neural network model GNN as an example, converts the feature information of the second user to obtain multiple embedding vectors embedding; in each layer of the graph neural network model GNN, forward propagation of messages will be performed, that is, each node collects information from its neighboring nodes, and the node updates its own feature representation. This process can be repeated between multiple layers so that each node can obtain information from more distant neighbors; after several rounds of message passing, the graph neural network model GNN needs to convert the learned node representation into the final output to obtain the category probability distribution; and construct a loss function for the credit scoring task, calculate the difference between the model prediction result and the actual score as the loss value, until the loss value meets the loss threshold, perform back propagation and optimization in the graph neural network model GNN, adjust the weight of the graph neural network model GNN according to the gradient of the loss relative to the model parameters, and use an optimizer (such as Adam, SGD, etc.) to update the parameters of the graph neural network model GNN; after multiple iterative training, the final credit scoring model is obtained.

[0166] It is understandable that the above training process simplifies many details, such as the selection of hyperparameters, the application of regularization techniques such as dropout and early stopping, and different training methods for different types of credit scoring models. The above are only examples.

[0167] In this embodiment, by training the credit model with labeled rejection samples, the behavior patterns and risk characteristics of users applying for credit services can be captured more comprehensively, thereby avoiding the data bias problem that may be caused by training only with approved samples. This is because when only approved samples are used to train the credit scoring model, the credit scoring model may have a preference for certain features in these approved samples and ignore other features that may be equally important, especially those related to rejected applications. Therefore, the credit scoring method provided in this embodiment can improve the accuracy of credit scoring by improving the generalization ability of the credit scoring model, so that the credit scoring model can perform well on new and unseen data.

[0168] The following are device embodiments of this specification, which can be used to implement the method embodiments of this specification. For details not disclosed in the device embodiments of this specification, please refer to the method embodiments of this specification.

[0169] See also Fig. 9 , which shows a schematic diagram of the structure of a user labeling device provided by an exemplary embodiment of the present specification. The user labeling device can be implemented as all or part of the device through software, hardware, or a combination of both. The device includes a source domain construction module 601, an initial training module 602, a target construction module 603, and a migration training module 604.

[0170] A source domain construction module 601 is used to construct a source domain graph including a plurality of source domain nodes and the source domain nodes are labeled with node labels according to a plurality of first users included in the source domain data and a first label labeled by the first user, wherein the plurality of first users correspond to the plurality of source domain nodes respectively;

[0171] An initial training module 602 is used to train the graph neural network model to be trained according to the source domain graph to obtain an initial model;

[0172] A target construction module 603 is used to construct a target domain graph including a plurality of target domain nodes according to a plurality of second users included in the target domain data, wherein the plurality of second users correspond to the plurality of target domain nodes respectively;

[0173] The migration training module 604 is used to migrate the initial model to the target domain graph for training until the target model is obtained. During the training process, the loss function of the target model is constrained by the smoothness between multiple target domain nodes on the target domain graph and the feature distribution difference between the source domain graph and the target domain graph. The target model is used to respectively label second labels for multiple second users included in the target domain data.

[0174] In one embodiment, the migration training module 604 includes:

[0175] The random walk module is used to migrate the initial model to the target domain graph for training, and to determine at least one target domain node adjacent to the target domain node as required for calculating smoothness based on the random walk method during the training process until the target model is obtained.

[0176] In one embodiment, the initial training module 602 includes:

[0177] An initial training unit, used to convert the source domain graph into a feature representation corresponding to the source domain graph by a preset graph representation learning method, and train a graph neural network model to be trained according to the feature representation corresponding to the source domain graph to obtain an initial model;

[0178] The migration training module 604 includes:

[0179] A transfer training unit is used to convert the target domain graph into a feature representation corresponding to the target domain graph through the graph representation learning method, and transfer the initial model to the feature representation corresponding to the target domain graph for training until the target model is obtained.

[0180] In one embodiment, the migration training module 604 includes:

[0181] The difference reduction unit is used to transfer the initial model to the target domain map for training, and to reduce the feature distribution difference between the source domain map and the target domain map during the training process until the target model is obtained.

[0182] In one embodiment, the difference reduction unit includes

[0183] The domain alignment subunit is used to migrate the initial model to the target domain graph for training, and to reduce the feature distribution difference between the source domain graph and the target domain graph through a preset domain alignment method during the training process until the target model is obtained.

[0184] In one embodiment, the multiple source domain nodes are source domain nodes of different types, and the edges connecting the multiple source domain nodes include multiple types;

[0185] The initial training module 602 includes:

[0186] A heterogeneous training unit is used to train the graph neural network model to be trained according to the heterogeneous features included in the source domain graph to obtain an initial model, wherein the heterogeneous features included in the source domain graph are obtained through the edges connecting the multiple source domain nodes and the multiple source domain nodes.

[0187] In one embodiment, the source domain data is data related to the first user applying for credit services and being approved, and the first label marked by the first user is overdue or not overdue. The target domain data is data related to the second user applying for credit services and being rejected, and the second label marked by the second user is overdue or not overdue.

[0188] In this specification, the problem of how to respectively label multiple second users in the target domain data with second labels is solved by a target model. The target model is obtained by migrating an initial model trained based on source domain data to target domain data for learning, and the type of the target model is a graph neural network model, and the target model is obtained by training the graph neural network model by constructing a source domain graph using source domain data and a target domain graph using target domain data.

[0189] In this specification, a user labeling model based on a graph neural network is first constructed as an initial model so that the target model can consider the relationship between multiple second users to label the second users. Secondly, in the process of transfer learning, the target model learns the feature representation of the source domain graph constructed by the source domain data and the feature representation of the target domain graph constructed by the target domain data, capturing the relationship characteristics between multiple first users and multiple second users. And the loss function used in the target model over-training process is constrained by the smoothness between multiple target domain nodes on the target domain graph and the difference in feature distribution between the source domain graph and the target domain graph, thereby enhancing the robustness and generalization ability of the target model, thereby improving the accuracy and reliability of the target model in labeling the second labels for multiple second users included in the target domain data.

[0190] It should be noted that the user label marking device provided in the above embodiment only uses the division of the above functional modules as an example when executing the user label marking method. In actual applications, the above functions can be assigned to different functional modules as needed, that is, the internal structure of the device is divided into different functional modules to complete all or part of the functions described above. In addition, the user label marking device provided in the above embodiment and the user label marking method embodiment belong to the same concept, and the implementation process thereof is detailed in the method embodiment, which will not be repeated here.

[0191] See also Fig.10 , which shows a schematic diagram of the structure of a credit scoring device provided by an exemplary embodiment of this specification. The credit scoring device can be implemented as all or part of the device through software, hardware or a combination of both. The device includes a label prediction module 701, a label acquisition module 702 and a model training module 703.

[0192] A label prediction module 701 is used to instruct the credit score model to be trained to output a predicted credit score of the second user for the second user according to the plurality of second users included in the target domain data and the second label annotated by the second user, wherein the second label annotated by the second user is obtained by the user label annotating method according to any one of claims 1 to 7;

[0193] The tag acquisition module 702 is used to acquire an actual credit score of the second user, where the actual credit score of the second user is determined in the process of providing credit services to the second user based on the predicted credit score of the second user;

[0194] The model training module 703 is used to train the credit score model to be trained according to the predicted credit scores and actual credit scores corresponding to the plurality of second users, until the credit score model is obtained, and the credit score model is used to output the credit score of the third user for the third user.

[0195] In this embodiment, by training the credit model with labeled rejection samples, the behavior patterns and risk characteristics of users applying for credit services can be captured more comprehensively, thereby avoiding the data bias problem that may be caused by training only with approved samples. This is because when only approved samples are used to train the credit scoring model, the credit scoring model may have a preference for certain features in these approved samples and ignore other features that may be equally important, especially those related to rejected applications. Therefore, the credit scoring method provided in this embodiment can improve the accuracy of credit scoring by improving the generalization ability of the credit scoring model, so that the credit scoring model can perform well on new and unseen data.

[0196] It should be noted that the credit scoring device provided in the above embodiment only uses the division of the above functional modules as an example when executing the credit scoring method. In actual applications, the above functions can be assigned to different functional modules as needed, that is, the internal structure of the device is divided into different functional modules to complete all or part of the functions described above. In addition, the credit scoring device provided in the above embodiment and the credit scoring method embodiment belong to the same concept, and the implementation process thereof is detailed in the method embodiment, which will not be repeated here.

[0197] The serial numbers of the embodiments of this specification are for description only and do not represent the advantages or disadvantages of the embodiments.

[0198] The present specification also provides a computer storage medium, which can store multiple instructions, and the instructions are suitable for being loaded and executed by a processor as described above. Figure 1 - Figure 8 The user tagging method and / or credit scoring method of the embodiment shown in the figure can be specifically implemented by referring to Figure 1 - Figure 8 The specific description of the illustrated embodiment will not be repeated here.

[0199] The present specification also provides a computer program product, which stores at least one instruction, and the at least one instruction is loaded and executed by the processor. Figure 1 - Figure 8 The user tagging method and / or credit scoring method of the embodiment shown in the figure can be specifically implemented by referring to Figure 1 - Figure 8 The specific description of the illustrated embodiment will not be repeated here.

[0200] See also Fig.11 , is a schematic diagram of the structure of an electronic device provided in the embodiment of this specification. Fig.11As shown, the electronic device 800 may include: at least one processor 801 , at least one network interface 804 , a user interface 803 , a memory 805 , and at least one communication bus 802 .

[0201] The communication bus 802 is used to realize the connection and communication between these components.

[0202] The user interface 803 may include a display screen (Display) and a camera (Camera), and the optional user interface 803 may also include a standard wired interface and a wireless interface.

[0203] The network interface 804 may optionally include a standard wired interface or a wireless interface (such as a WI-FI interface).

[0204] Among them, the processor 801 may include one or more processing cores. The processor 801 uses various interfaces and lines to connect various parts within the entire server 800, and executes various functions and processes data of the server 800 by running or executing instructions, programs, code sets or instruction sets stored in the memory 805, and calling data stored in the memory 805. Optionally, the processor 801 can be implemented in at least one hardware form of digital signal processing (Digital Signal Processing, DSP), field programmable gate array (Field-Programmable Gate Array, FPGA), and programmable logic array (Programmable Logic Array, PLA). The processor 801 can integrate one or more combinations of a processor (Central Processing Unit, CPU), an image processor (Graphics Processing Unit, GPU) and a modem. Among them, the CPU mainly processes the operating system, user interface and application programs; the GPU is responsible for rendering and drawing the content to be displayed on the display screen; and the modem is used to process wireless communications. It can be understood that the above-mentioned modem may not be integrated into the processor 801, and it can be implemented separately through a chip.

[0205] Among them, the memory 805 may include a random access memory (Random Access Memory, RAM) and may also include a read-only memory (Read-Only Memory). Optionally, the memory 805 includes a non-transitory computer-readable storage medium. The memory 805 can be used to store instructions, programs, codes, code sets or instruction sets. The memory 805 may include a program storage area and a data storage area, wherein the program storage area may store instructions for implementing an operating system, instructions for at least one function (such as a touch function, a sound playback function, an image playback function, etc.), instructions for implementing the above-mentioned various method embodiments, etc.; the data storage area may store data involved in the above-mentioned various method embodiments, etc. The memory 805 may also be optionally at least one storage device located away from the aforementioned processor 801. As Fig.11 As shown, the memory 805 as a computer storage medium may include an operating system, a network communication module, a user interface module, and a user tagging and / or credit scoring application program.

[0206] exist Fig.11 In the electronic device 800 shown, the user interface 803 is mainly used to provide an input interface for the user and obtain the data input by the user; and the processor 801 can be used to call the user labeling application stored in the memory 805, and specifically perform the following operations:

[0207] According to the multiple first users included in the source domain data and the first labels annotated by the first users, a source domain graph including multiple source domain nodes and the source domain nodes annotated with node labels is constructed, wherein the multiple first users correspond to the multiple source domain nodes respectively;

[0208] Training the graph neural network model to be trained according to the source domain graph to obtain an initial model;

[0209] According to the multiple second users included in the target domain data, construct a target domain graph including multiple target domain nodes, wherein the multiple second users correspond to the multiple target domain nodes respectively;

[0210] The initial model is transferred to the target domain graph for training until a target model is obtained, and a loss function of the target model during the training process is constrained by the smoothness between multiple target domain nodes on the target domain graph and by the feature distribution difference between the source domain graph and the target domain graph, and the target model is used to respectively label second labels for multiple second users included in the target domain data.

[0211] In one embodiment, the processor 801 performs the step of migrating the initial model to the target domain graph for training until a target model is obtained, specifically performing:

[0212] The initial model is transferred to the target domain graph for training, and during the training process, at least one target domain node adjacent to the target domain node is determined based on a random walk method as required for calculating smoothness until a target model is obtained.

[0213] In one embodiment, the processor 801 executes the training of the graph neural network model to be trained according to the source domain graph to obtain an initial model, specifically performing:

[0214] The source domain graph is converted into a feature representation corresponding to the source domain graph by a preset graph representation learning method, and a graph neural network model to be trained is trained according to the feature representation corresponding to the source domain graph to obtain an initial model;

[0215] The processor 801 executes the migration of the initial model to the target domain graph for training until the target model is obtained, specifically performing:

[0216] The target domain graph is converted into a feature representation corresponding to the target domain graph by the graph representation learning method, and the initial model is migrated to the feature representation corresponding to the target domain graph for training until a target model is obtained.

[0217] In one embodiment, the processor 801 performs the step of migrating the initial model to the target domain graph for training until a target model is obtained, specifically performing:

[0218] The initial model is transferred to the target domain graph for training, and the feature distribution difference between the source domain graph and the target domain graph is reduced during the training process until a target model is obtained.

[0219] In one embodiment, the processor 801 performs the step of migrating the initial model to the target domain graph for training, and reducing the feature distribution difference between the source domain graph and the target domain graph during the training process until the target model is obtained, specifically performing:

[0220] The initial model is transferred to the target domain map for training, and during the training process, the feature distribution difference between the source domain map and the target domain map is reduced by a preset domain alignment method until the target model is obtained.

[0221] In one embodiment, the multiple source domain nodes are source domain nodes of different types, and the edges connecting the multiple source domain nodes include multiple types;

[0222] The processor 801 executes the training of the graph neural network model to be trained according to the source domain graph to obtain an initial model, specifically performing:

[0223] The graph neural network model to be trained is trained according to the heterogeneous features included in the source domain graph to obtain an initial model, wherein the heterogeneous features included in the source domain graph are obtained through the edges connecting the multiple source domain nodes and the multiple source domain nodes.

[0224] In one embodiment, the source domain data is data related to the first user applying for credit services and being approved, and the first label marked by the first user is overdue or not overdue. The target domain data is data related to the second user applying for credit services and being rejected, and the second label marked by the second user is overdue or not overdue.

[0225] In one embodiment, the processor 801 may be used to call the user tagging application stored in the memory 805, and specifically perform the following operations:

[0226] Instructing the credit scoring model to be trained to output a predicted credit score of the second user for the second user according to the plurality of second users included in the target domain data and the second tags annotated by the second users, wherein the second tags annotated by the second user are obtained by the user tagging method described in any of the embodiments of this specification;

[0227] Acquiring an actual credit score of the second user, where the actual credit score of the second user is determined in the process of providing credit services to the second user based on the predicted credit score of the second user;

[0228] The credit score model to be trained is trained according to the predicted credit scores and the actual credit scores respectively corresponding to the plurality of second users until the credit score model is obtained, and the credit score model is used to output the credit score of the third user for a third user.

[0229] Those skilled in the art can understand that all or part of the processes in the above-mentioned embodiments can be implemented by instructing related hardware through a computer program, and the program can be stored in a computer-readable storage medium, and when the program is executed, it can include the processes of the embodiments of the above-mentioned methods. The storage medium can be a disk, an optical disk, a read-only storage memory, or a random access memory, etc.

[0230] The technical features of the above-described embodiments may be arbitrarily combined. To make the description concise, not all possible combinations of the technical features in the above-described embodiments are described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this specification.

[0231] The above disclosure is only the preferred embodiment of this specification, which certainly cannot be used to limit the scope of rights of this specification. Therefore, equivalent changes made according to the claims of this specification are still within the scope covered by this specification.

Claims

1. A user tagging method, the method comprising: According to the multiple first users included in the source domain data and the first labels annotated by the first users, a source domain graph including multiple source domain nodes and the source domain nodes annotated with node labels is constructed, wherein the multiple first users correspond to the multiple source domain nodes respectively; Training the graph neural network model to be trained according to the source domain graph to obtain an initial model; According to the multiple second users included in the target domain data, construct a target domain graph including multiple target domain nodes, wherein the multiple second users correspond to the multiple target domain nodes respectively; The initial model is transferred to the target domain graph for training until a target model is obtained, and a loss function of the target model during the training process is constrained by the smoothness between multiple target domain nodes on the target domain graph and by the feature distribution difference between the source domain graph and the target domain graph, and the target model is used to respectively label second labels for multiple second users included in the target domain data.

2. According to the user labeling method of claim 1, the step of migrating the initial model to the target domain graph for training until the target model is obtained comprises: The initial model is transferred to the target domain graph for training, and during the training process, at least one target domain node adjacent to the target domain node is determined based on a random walk method as required for calculating smoothness until a target model is obtained.

3. According to the user labeling method of claim 1, the step of training the graph neural network model to be trained according to the source domain graph to obtain an initial model comprises: The source domain graph is converted into a feature representation corresponding to the source domain graph by a preset graph representation learning method, and a graph neural network model to be trained is trained according to the feature representation corresponding to the source domain graph to obtain an initial model; The step of migrating the initial model to the target domain graph for training until a target model is obtained includes: The target domain graph is converted into a feature representation corresponding to the target domain graph by the graph representation learning method, and the initial model is migrated to the feature representation corresponding to the target domain graph for training until a target model is obtained.

4. According to the user labeling method of claim 1, the step of migrating the initial model to the target domain graph for training until the target model is obtained comprises: The initial model is transferred to the target domain graph for training, and the feature distribution difference between the source domain graph and the target domain graph is reduced during the training process until a target model is obtained.

5. The user labeling method according to claim 4, wherein the initial model is transferred to the target domain graph for training, and the feature distribution difference between the source domain graph and the target domain graph is reduced during the training process until the target model is obtained, comprising: The initial model is transferred to the target domain map for training, and during the training process, the feature distribution difference between the source domain map and the target domain map is reduced by a preset domain alignment method until the target model is obtained.

6. The user labeling method according to claim 1, wherein the plurality of source domain nodes are source domain nodes of different types, and the edges connecting the plurality of source domain nodes include multiple types; The step of training the graph neural network model to be trained according to the source domain graph to obtain an initial model includes: The graph neural network model to be trained is trained according to the heterogeneous features included in the source domain graph to obtain an initial model, and the heterogeneous features included in the source domain graph are obtained through the edges connecting the multiple source domain nodes and the multiple source domain nodes.

7. According to the user labeling method of claim 1, the source domain data is data related to the first user applying for credit services and being approved, and the first label labeled by the first user is overdue or not overdue; the target domain data is data related to the second user applying for credit services and being rejected, and the second label labeled by the second user is overdue or not overdue.

8. A credit scoring method, the method comprising: Instructing the credit scoring model to be trained to output a predicted credit score of the second user for the second user according to the plurality of second users included in the target domain data and the second tags annotated by the second users, wherein the second tags annotated by the second user are obtained by the user tagging method according to any one of claims 1 to 7; Acquiring an actual credit score of the second user, where the actual credit score of the second user is determined in the process of providing credit services to the second user based on the predicted credit score of the second user; The credit score model to be trained is trained according to the predicted credit scores and the actual credit scores respectively corresponding to the plurality of second users until the credit score model is obtained, and the credit score model is used to output the credit score of the third user for a third user.

9. A user label marking device, the user label marking device comprising: a source domain construction module, configured to construct a source domain graph including a plurality of source domain nodes and the source domain nodes being labeled with node labels according to a plurality of first users included in the source domain data and a first label labeled by the first user, wherein the plurality of first users correspond to the plurality of source domain nodes respectively; An initial training module, used to train the graph neural network model to be trained according to the source domain graph to obtain an initial model; A target construction module, configured to construct a target domain graph including a plurality of target domain nodes according to a plurality of second users included in the target domain data, wherein the plurality of second users correspond to the plurality of target domain nodes respectively; A migration training module is used to migrate the initial model to the target domain graph for training until a target model is obtained, and the loss function of the target model in the training process is constrained by the smoothness between multiple target domain nodes on the target domain graph, and by the feature distribution difference between the source domain graph and the target domain graph. The target model is used to respectively label second labels for multiple second users included in the target domain data.

10. A credit scoring device, comprising: a label prediction module, configured to instruct the credit score model to be trained to output a predicted credit score of the second user for the second user according to the plurality of second users included in the target domain data and the second label annotated by the second user, wherein the second label annotated by the second user is obtained by the user label annotating method according to any one of claims 1 to 7; a label acquisition module, configured to acquire an actual credit score of the second user, where the actual credit score of the second user is determined in a process of providing credit services to the second user based on a predicted credit score of the second user; A model training module is used to train the credit score model to be trained according to the predicted credit scores and actual credit scores respectively corresponding to the plurality of second users until the credit score model is obtained, and the credit score model is used to output the credit score of the third user for a third user.

11. A computer storage medium storing a plurality of instructions, wherein the instructions are suitable for being loaded by a processor and executing the method steps according to any one of claims 1 to 8.

12. A computer program product, wherein the computer program product stores a plurality of instructions, wherein the instructions are suitable for being loaded by a processor and executing the method steps according to any one of claims 1 to 8.

13. An electronic device comprising: A processor and a memory; wherein the memory stores a computer program, and the computer program is suitable for being loaded by the processor and executing the method steps as claimed in any one of claims 1 to 8.

Images