Method, apparatus, and server for determining risk

Abstract

This specification provides a method, apparatus, and server for risk determination. Based on this method, before specific implementation, by introducing and utilizing subspace consistency constraints and dynamic pairwise constraints, sample data with and without a preset label can be used simultaneously to train a preset processing model with good performance and high accuracy; during specific implementation, after obtaining multiple feature data of a target object, first, according to a preset mapping rule, by mapping the multiple feature data into multiple preset feature subspaces, multiple feature groups with lower dimensions of the target object can be obtained; then, the preset processing model is called to process the above multiple feature groups to obtain corresponding processing results; furthermore, based on the processing results, it can be determined efficiently and accurately whether the target object has a preset risk. Thus, it can be better applied to data prediction scenarios with higher and more complex feature dimensions, and accurately predict whether the target object has a preset risk.

Description

Technical Field

[0001] This specification belongs to the technical field of artificial intelligence, and particularly relates to a method, device, and server for determining risks. Background Art

[0002] In some relatively complex data prediction scenarios (for example, the scenario of predicting the risk of overdue loans), the types of characteristic data of the data objects whose risks are to be predicted collected are often relatively numerous (for example, it may include more than 100 different types of characteristic data), and the characteristic dimensions are relatively high.

[0003] In view of the above-mentioned data prediction scenarios with relatively high characteristic dimensions and complexity, it is often difficult to train a risk prediction model with good effects and small errors based on existing methods. As a result, it is difficult to apply to the above-mentioned data prediction scenarios with relatively high characteristic dimensions and complexity based on existing methods, and it is difficult to accurately determine whether there is a preset risk for the data objects in the above scenarios.

[0004] In response to the above problems, no effective solution has been proposed yet. Summary of the Invention

[0005] This specification provides a method, device, and server for determining risks. By using a preset processing model pre-trained based on subspace consistency constraints and dynamic pairwise constraints, it can be better applied to data prediction scenarios with relatively high characteristic dimensions and complexity, and accurately predict whether a target object has a preset risk.

[0006] This specification provides a method for determining risks, including:

[0007] Obtain multiple characteristic data of a target object;

[0008] According to a preset mapping rule, map the multiple characteristic data into multiple preset characteristic subspaces to obtain multiple characteristic groups of the target object; wherein, each characteristic group corresponds to a preset characteristic subspace;

[0009] Call a preset processing model to process the multiple characteristic groups to obtain corresponding processing results; wherein, the preset processing model at least includes multiple preset classifiers, and each preset classifier corresponds to a preset characteristic subspace; the preset processing model is pre-trained based on subspace consistency constraints and dynamic pairwise constraints;

[0010] Determine whether the target object has a preset risk according to the processing results.

[0011] In one embodiment, the target object includes a trading account; correspondingly, the preset risk includes the risk of overdue transaction data.

[0012] In one embodiment, the feature data includes: identity type feature data of a trading account, historical trading behavior type feature data of the trading account, current trading behavior type feature data of the trading account, and associated behavior type feature data of the trading account.

[0013] In one embodiment, the preset processing model further includes a discrimination structure; wherein, the discrimination structure is connected to the multiple preset classifiers; the discrimination structure is configured to generate the processing result according to the classification results output by the multiple preset classifiers.

[0014] In one embodiment, the preset processing model is established in the following manner:

[0015] Obtain a plurality of sample data; wherein, the sample data corresponds to a sample object, and the sample data includes a plurality of feature data of the corresponding sample object; the sample object includes a first type of sample object with a preset label and a second type of sample object without a preset label;

[0016] According to a preset mapping rule, map the multiple feature data of the first type of sample object into multiple preset feature subspaces to obtain first type of training data; map the multiple feature data of the second type of sample object into multiple preset feature subspaces to obtain second type of training data;

[0017] Use the first type of training data to train a plurality of initial classifiers to obtain corresponding multiple intermediate classifiers; wherein, each of the initial classifiers corresponds to a preset feature subspace;

[0018] Use the multiple intermediate classifiers, the first type of training data, and the second type of training data to construct an objective function for the multiple intermediate classifiers; wherein, the objective function includes a subspace consistency constraint formula and a dynamic pairwise constraint formula;

[0019] According to the objective function, determine a plurality of preset classifiers that meet the requirements to construct a preset processing model.

[0020] In one embodiment, using the multiple intermediate classifiers, the first type of training data, and the second type of training data to construct an objective function for the multiple intermediate classifiers; wherein, the objective function includes a subspace consistency constraint formula and a dynamic pairwise constraint formula, including:

[0021] Call the multiple intermediate classifiers to process the second type of training data to obtain dynamic labels for the second type of sample objects;

[0022] Construct a dynamic pairwise constraint formula according to the preset labels of the first type of sample objects and the dynamic labels of the second type of sample objects;

[0023] Call multiple intermediate classifiers to process the first-class training data, and obtain multiple classification results of the first-class sample objects;

[0024] Construct a subspace consistency constraint formula according to the multiple classification results of the first-class sample objects;

[0025] Construct the objective function according to the dynamic pairwise constraint formula and the subspace consistency constraint formula.

[0026] In one embodiment, constructing a dynamic pairwise constraint formula according to the preset labels of the first-class sample objects and the dynamic labels of the second-class sample objects includes:

[0027] Establish a connection set and a non-connection set according to the preset labels of the first-class sample objects and the dynamic labels of the second-class sample objects; wherein, the connection set includes multiple connection groups, and each connection group includes two sample objects with the same label; the non-connection set includes multiple non-connection groups, and each non-connection group includes two sample objects with different labels;

[0028] Construct a dynamic pairwise matrix according to the connection set and the non-connection set;

[0029] Construct the dynamic pairwise constraint formula according to the dynamic pairwise matrix and the multiple intermediate classifiers.

[0030] In one embodiment, constructing a subspace consistency constraint formula according to the multiple classification results of the first-class sample objects includes:

[0031] Divide multiple comparison groups according to the multiple classification results of the first-class sample objects; wherein, each comparison group includes two classification results of the same first-class sample object;

[0032] Construct the subspace consistency constraint formula according to the multiple comparison groups.

[0033] In one embodiment, according to the objective function, determining multiple preset classifiers that meet the requirements to construct a preset processing model includes:

[0034] Use the gradient descent algorithm to solve the optimal solution of the objective function to determine multiple preset classifiers that meet the requirements;

[0035] Connect multiple preset classifiers to the discriminant structure to obtain a preset processing model.

[0036] In one embodiment, after obtaining multiple sample data, the method further includes:

[0037] Check whether the number of missing feature data in the detection sample data is less than a preset missing quantity threshold;

[0038] In the case where it is determined that the number of missing feature data in the sample data is less than the preset missing quantity threshold, perform zero-filling processing on the missing feature data.

[0039] This specification also provides a method for determining risk, including:

[0040] Obtain multiple feature data of the target object;

[0041] According to the preset mapping rules, map the multiple feature data into multiple preset feature subspaces to obtain multiple feature groups of the target object; wherein, each feature group corresponds to a preset feature subspace;

[0042] Call a preset processing model to process the multiple feature groups to obtain corresponding processing results; wherein, the preset processing model at least includes multiple preset classifiers, and each preset classifier corresponds to a preset feature subspace; the preset processing model is pre-trained based on subspace consistency constraints or dynamic pairwise constraints;

[0043] According to the processing results, determine whether the target object has a preset risk.

[0044] This specification also provides a device for determining risk, including:

[0045] An acquisition module, configured to acquire multiple feature data of the target object;

[0046] A mapping module, configured to map the multiple feature data into multiple preset feature subspaces according to the preset mapping rules to obtain multiple feature groups of the target object; wherein, each feature group corresponds to a preset feature subspace;

[0047] A calling module, configured to call a preset processing model to process the multiple feature groups to obtain corresponding processing results; wherein, the preset processing model at least includes multiple preset classifiers, and each preset classifier corresponds to a preset feature subspace; the preset processing model is pre-trained based on subspace consistency constraints and dynamic pairwise constraints;

[0048] A determination module, configured to determine whether the target object has a preset risk according to the processing results.

[0049] This specification also provides a server, including a processor and a memory for storing processor-executable instructions, and the processor implements the relevant steps of the method for determining risk when executing the instructions.

[0050] This specification also provides a computer-readable storage medium, on which computer instructions are stored, and when the instructions are executed, the related steps of the method for determining the risk are implemented.

[0051] A method, apparatus, and server for determining a risk provided by this specification. Based on this method, before specific implementation, for a data prediction scenario with relatively high and complex feature dimensions, according to a preset mapping rule, multi-dimensional features can be divided into multiple preset feature subspaces to reduce the dimension of the feature data of the sample object, so as to reduce the complexity and the amount of data processing involved in subsequent model training; further, in combination with the feature data after the above-mentioned dimension reduction, by introducing and using subspace consistency constraints and dynamic pairwise constraints to construct an objective function that takes into account the interaction between different feature subspaces and the interaction between different sample objects to participate in specific model training, so that sample data with and without a preset label can be fully utilized at the same time, and a preset processing model with better effect and higher accuracy can be trained; when specifically implementing, after obtaining multiple feature data of the target object, according to a preset mapping rule, multiple feature data with higher dimensions can be mapped into multiple preset feature subspaces to obtain multiple corresponding feature groups with lower dimensions of the target object; then call the preset processing model to process the above-mentioned multiple feature groups to obtain the corresponding processing result more quickly; furthermore, according to the processing result, it is possible to efficiently and accurately determine whether the target object has a preset risk. Thus, it can be better applied to a data prediction scenario with relatively high and complex feature dimensions and accurately predict whether the target object has a preset risk. Description of the Drawings

[0052] To more clearly illustrate the embodiments of this specification, the drawings required for the embodiments will be briefly introduced below. The drawings in the following description are only some embodiments recorded in this specification. For those of ordinary skill in the art, without creative efforts, other drawings can also be obtained based on these drawings.

[0053] Figure 1 It is a schematic diagram of an embodiment of the structural composition of a system applying the method for determining a risk provided by the embodiments of this specification;

[0054] Figure 2 It is a schematic diagram of an embodiment applying the method for determining a risk provided by the embodiments of this specification in a scenario example;

[0055] Figure 3 It is a flowchart of the method for determining a risk provided by an embodiment of this specification;

[0056] Figure 4It is a schematic diagram of the structural composition of a server provided by an embodiment of this specification;

[0057] Figure 5 It is a schematic diagram of the structural composition of a risk determination device provided by an embodiment of this specification;

[0058] Figure 6 It is a schematic diagram of an embodiment of applying the risk determination method provided by the embodiment of this specification in a scenario example;

[0059] Figure 7 It is a schematic diagram of an embodiment of applying the risk determination method provided by the embodiment of this specification in a scenario example. Detailed implementation manners

[0060] In order to enable those skilled in the art to better understand the technical solutions in this specification, the technical solutions in the embodiments of this specification will be clearly and completely described below with reference to the accompanying drawings in the embodiments of this specification. Obviously, the described embodiments are only a part of the embodiments of this specification, rather than all the embodiments. Based on the embodiments in this specification, all other embodiments obtained by those of ordinary skill in the art without creative efforts shall fall within the protection scope of this specification.

[0061] The embodiment of this specification provides a risk determination method, which can be specifically applied to a system including a server and a terminal device. Specifically, reference can be made to Figure 1 As shown, the server and the terminal device can be connected by wire or wirelessly to perform specific data interaction.

[0062] In this embodiment, the server can specifically include a background server applied to one side of the network platform, which can implement functions such as data transmission and data processing. Specifically, the server can be, for example, an electronic device with data operation, storage functions, and network interaction functions. Alternatively, the server can also be a software program running in the electronic device to provide support for data processing, storage, and network interaction. In this embodiment, the number of servers included in the server is not specifically limited. The server can specifically be one server, or several servers, or a server cluster formed by several servers.

[0063] In this embodiment, the terminal device may specifically include a front-end electronic device disposed on the user side and capable of implementing functions such as data collection and data transmission. Specifically, the client may be, for example, a desktop computer, a tablet computer, a laptop computer, a smart phone, a self-service machine, a counter terminal, etc. Alternatively, the terminal device may also be a software application that can run on the above-mentioned electronic devices. For example, it may be a relevant APP running on a smart phone, etc.

[0064] In specific implementation, the user can use the self-service machine disposed in the bank business hall as the terminal device to handle relevant services involving transaction data (such as housing provident fund loans, etc.).

[0065] The user can log in to their transaction account (such as the target object) according to the instructions of the guidance interface displayed on the self-service machine and initiate a service handling request for transaction data.

[0066] Correspondingly, the self-service machine can respond to the above operation of the user, generate and display an information input interface to the user. Through the above information input interface, the self-service machine can receive first user data input by the user, such as the user's name, the user's address, the user's phone number, etc.

[0067] Then, the self-service machine can send the service handling request carrying the first user data to the server.

[0068] Correspondingly, the server receives the above service handling request and determines the transaction account requesting to handle relevant services involving transaction data and the first user data by parsing the service handling request.

[0069] Furthermore, the server can respond to the service handling request, query its own user database and the user database of the cooperation party according to the transaction account, and obtain second user data such as historical transaction records, credit service records, integrity data, etc. of the transaction account.

[0070] Further, the server can extract various relevant feature data from a large amount of first user data and second user data, including: identity type feature data of the transaction account (such as the name, work unit, education level, etc. of the user of the transaction account), historical transaction behavior type feature data of the transaction account (such as the purchase turnover in the past year, the total income data in the past year, the total expenditure data in the past, etc.), current transaction behavior type feature data of the transaction account (such as the current credit type service data of the transaction account, the current housing provident fund contribution data of the transaction account, the current insurance type service data participated in by the transaction account, etc.), associated behavior type feature data of the transaction account (such as the credit rating of the transaction account in other institutions, the transaction overdue record of the transaction account in other institutions, etc.).

[0071] As a result, the server can obtain a variety of feature data with a high dimension for the trading account.

[0072] Furthermore, the server can map the multiple feature data of the trading account into multiple preset feature subspaces according to a preset mapping rule, obtaining multiple feature groups. Among them, each of the multiple feature groups corresponds to a preset feature subspace.

[0073] In addition, each feature group can specifically contain one or more feature data of the trading account. Moreover, each of the multiple feature data of the trading account is at least divided into one feature group.

[0074] Specifically, for example, the first feature group corresponding to the 1st preset feature subspace contains three types of feature data: the user name of the trading account, the user age, and the current housing provident fund deposit data of the trading account. The second feature group corresponding to the 2nd preset feature subspace contains five types of feature data: the user's work unit, the user's education level, the user age, the credit rating of the trading account in other institutions, and the trading overdue record of the trading account in other institutions...

[0075] Through the above method, mapping multiple feature data into multiple preset feature subspaces according to the preset mapping rule can convert the multiple high-dimensional feature data into multiple low-dimensional feature groups, realizing dimensionality reduction of the high-dimensional feature data and facilitating subsequent data processing.

[0076] Then, the server can use the above multiple feature groups as model inputs and input them into a preset processing model for specific processing to determine whether there is a preset risk for the trading account (for example, the overdue repayment risk of the housing provident fund loan).

[0077] Among them, the above preset processing model can be specifically understood as a model that is pre-trained through semi-supervised learning based on subspace consistency constraints and dynamic pairwise constraints using the dimensionality-reduced sample data, and can predict whether the target object has a preset risk according to the input multiple feature groups.

[0078] It can be referred to Figure 2 As shown, the above preset processing model at least includes multiple preset classifiers. Among them, the multiple preset classifiers correspond to a preset feature subspace respectively and are used to process the corresponding feature group. For example, the 1st preset classifier corresponds to the 1st preset feature subspace and is used to access and process the first feature group. In contrast, the 2nd preset classifier corresponds to the 2nd preset feature subspace and is used to access and process the second feature group.

[0079] In addition, the above-mentioned preset processing model further includes a discrimination structure. Among them, the discrimination structure is connected to a plurality of preset classifiers, and is used to access and generate and output a final processing result according to the classification results output by the plurality of preset classifiers.

[0080] After the preset processing model receives multiple feature groups of the input model, in specific implementation, each of the multiple preset classifiers can process the responsible feature groups respectively and output corresponding multiple classification results. After the discrimination structure obtains the multiple classification results output by the multiple preset classifiers, it can generate a final processing result and output a model through means such as weighted summation.

[0081] The server can determine whether there is a preset risk for the trading account according to the processing result.

[0082] Specifically, if the server determines according to the processing result that there is a preset risk for the trading account, it can stop handling relevant services for the user and generate a first type of prompt message of "This user has an overdue risk and cannot handle relevant services".

[0083] The server can send the above-mentioned first type of prompt message to the self-service machine. Correspondingly, the self-service machine can display the above-mentioned first type of prompt message to the user to prompt the user that the service handling fails.

[0084] If the server determines according to the processing result that there is no preset risk for the trading account, it can continue to handle relevant services for the user, and after the handling is completed, generate a second type of prompt message of "Service handling is successful".

[0085] The server can send the above-mentioned second type of prompt message to the self-service machine. Correspondingly, the self-service machine can display the above-mentioned second type of prompt message to the user to prompt the user that the service handling is successful.

[0086] Through the above method, the server can utilize a preset processing model with good effect obtained by pre-introducing and training based on subspace consistency constraints and dynamic pairwise constraints, which is applicable to data prediction scenarios with higher and more complex feature dimensions, accurately predict whether there is a preset risk for the user requesting to handle relevant services, and then can accurately determine whether to handle relevant services for the user according to the risk situation of the user, reducing the risk borne by the business side.

[0087] Refer to Figure 3 As shown in the content, the embodiment of the present specification provides a method for determining risk. Among them, this method is specifically applied to the server side. In specific implementation, this method may include the following content:

[0088] S301: Obtain multiple feature data of the target object;

[0089] S302: Map the multiple pieces of feature data into multiple preset feature subspaces according to a preset mapping rule, to obtain multiple feature groups of the target object; wherein, each of the feature groups corresponds to a preset feature subspace.

[0090] S303: Invoke a preset processing model to process the multiple feature groups, to obtain corresponding processing results; wherein, the preset processing model at least includes multiple preset classifiers, and each of the preset classifiers corresponds to a preset feature subspace; the preset processing model is pre-trained based on subspace consistency constraints and dynamic pairwise constraints.

[0091] S304: Determine whether the target object has a preset risk according to the processing results.

[0092] Through the above embodiments, a preset processing model with good effects pre-trained based on subspace consistency constraints and dynamic pairwise constraints can be used, which is well applicable to data prediction scenarios with relatively high and complex feature dimensions, accurately predicts whether the target object has a preset risk, and reduces the prediction error.

[0093] In some embodiments, the above target object can be specifically understood as a data object for which it is to be predicted whether there is a preset risk. Corresponding to different application scenarios and different processing requirements, the above target object can be different types of data objects.

[0094] In some embodiments, the above target object can include a trading account; correspondingly, the preset risk can include the risk of overdue trading data.

[0095] Specifically, the above trading account can be an account used by a user who wants to apply for relevant services involving trading data (such as credit loans, housing provident fund loans, small and micro enterprise loans, etc.).

[0096] Through the above embodiments, the method for determining risks provided in this specification can be applied to predict whether there is a risk of overdue trading data in a trading account applying for relevant services involving trading data, so as to provide a reference basis with relatively high value for the business handling party, so that the business handling party can more reasonably and accurately judge whether to handle relevant services for the trading account according to the predicted overdue risk situation of the trading account, reducing the risks borne by the business handling party.

[0097] Of course, it should be noted that the above-listed target objects and preset risks are only illustrative. For different application scenarios and processing requirements, the above target objects may also include other types of data objects, and correspondingly, the above preset risks may also include other types of risks. This specification does not make any limitations in this regard.

[0098] In some embodiments, the above-mentioned multiple feature data can be specifically understood as data related to the target object that can characterize certain attribute characteristics of the target object of concern. Usually, based on the above-mentioned multiple feature data, it can be predicted whether the target object has a preset risk.

[0099] In some embodiments, for data prediction scenarios with a relatively high and complex feature dimension, in order to more accurately predict whether the target object has a preset risk, the number of types of feature data of the target object is often relatively large and the feature dimension is relatively high.

[0100] In some embodiments, the feature data may specifically include: identity-related feature data of the trading account (for example, the name, work unit, education level, etc. of the user of the trading account), historical trading behavior-related feature data of the trading account (for example, the purchase transaction records of the trading account in the past year, the total income data in the past year, the total expenditure data in the past year, etc.), current trading behavior-related feature data of the trading account (for example, the current credit-related business data of the trading account, the current housing provident fund contribution data of the trading account, the current insurance-related business data participated in by the trading account, etc.). In addition, the above feature data may also include associated behavior-related feature data of the trading account (for example, the credit rating of the trading account in other institutions, the trading overdue record of the trading account in other institutions, etc.). Of course, it should be noted that the above-listed feature data is only an illustrative description.

[0101] Through the above embodiments, the method for determining risks provided in this specification can be applied to data prediction scenarios with a relatively high and complex feature dimension, making full and comprehensive use of the high-dimensional feature data of the target object, so as to more accurately predict whether the target object has a preset risk.

[0102] In some embodiments, the above-mentioned acquisition of multiple feature data of the target object may specifically include: receiving the first user data actively input by the target object and extracting relevant feature data from the first user data; and / or, receiving and based on the identity identifier provided by the target object (for example, the user name or the trading account name, etc.), retrieving the databases held by oneself and the databases held by the cooperation parties, mining the data associated with the identity identifier of the target object as the second user data, and extracting relevant feature data from the second user data.

[0103] In some embodiments, considering that in many data prediction scenarios, especially those with relatively high and complex feature dimensions, the types of feature data of the target object obtained and used are often numerous, and the feature dimensions are also relatively high. In this case, if the above-mentioned high-dimensional feature data is directly processed, the processing process will inevitably be relatively complex and cumbersome, involving a relatively large amount of data processing, a relatively high processing cost, and it is also more likely to have errors. Therefore, in this embodiment, instead of directly using the multiple feature data of the target object obtained, the multiple feature data will first be mapped into multiple preset feature subspaces according to a preset mapping rule to obtain multiple feature groups regarding this target object. Among them, each feature group corresponds to a preset feature subspace respectively. And each feature group may specifically include one or more feature data with a relatively small number of types and a relatively low feature dimension.

[0104] In this way, multiple feature groups with relatively low feature dimensions can be used to replace the multiple feature data with relatively high original feature dimensions for subsequent data processing, realizing dimensionality reduction of the high-dimensional feature data so as to be able to process the feature data more efficiently and accurately.

[0105] In some embodiments, the above-mentioned preset mapping rule may specifically be a rule set including the mapping relationship between each feature data and the preset feature subspace. Specifically, based on this preset mapping rule, each feature data in the multiple feature data can be mapped into at least one preset feature subspace; at the same time, it can also be ensured that each feature group corresponding to each preset feature subspace contains at least one feature data. It should be noted that based on the preset mapping rule, the same feature data can be simultaneously mapped into multiple different preset feature subspaces.

[0106] In some implementations, the above-mentioned preset processing model can specifically be understood as a model that is pre-trained through semi-supervised learning using the dimensionality-reduced sample data based on subspace consistency constraints and dynamic pairwise constraints, and can predict whether the target object has a preset risk according to the input multiple feature groups.

[0107] It should be supplemented and explained that conventional models for predicting risks often only consider the mutual relationship between the feature data and labels of a single sample object, ignore the interaction between different sample objects, and even more do not consider the interaction between different feature subspaces. However, the preset processing model provided in this specification is pre-trained by introducing and based on subspace consistency constraints involving different feature subspaces and dynamic pairwise constraints involving different sample objects, and has better coverage and higher model accuracy. The specific training method of the above-mentioned preset processing model will be described separately later.

[0108] In some embodiments, specifically, refer to Figure 2 As shown, the above-mentioned preset processing model at least includes a plurality of preset classifiers. Among them, the preset classifiers respectively correspond to a preset feature subspace. Correspondingly, the preset classifiers respectively correspond to a feature group.

[0109] Specifically, multiple feature groups input into the preset processing model will be split and input into the corresponding preset classifiers for processing, and a classification result based on the corresponding feature group will be output from the preset classifiers.

[0110] In some embodiments, further, the above-mentioned preset processing model may further include a discriminant structure. Among them, the discriminant structure is connected to the plurality of preset classifiers; the discriminant structure is used to generate the processing result according to the classification results output by the plurality of preset classifiers.

[0111] Specifically, multiple classification results output from the plurality of preset classifiers will be input into the discriminant structure, and then the discriminant structure will comprehensively analyze the multiple classification results based on the built-in discriminant function, calculate and output the final processing result as the model output of the preset processing model.

[0112] Through the above embodiments, the preset processing model can parallel process multiple feature groups corresponding to multiple preset feature subspaces to obtain multiple classification results; furthermore, the final processing result can be obtained according to the above multiple classification results.

[0113] In some embodiments, the above processing result may specifically be a prediction label, which is used to characterize whether the target object has a preset risk.

[0114] In some embodiments, the above processing result may specifically also be a prediction probability used to characterize that the target object has a preset risk.

[0115] In some embodiments, according to the processing result, determining whether the target object has a preset risk may specifically include: according to the processing result, comparing the prediction probability with a preset probability threshold; in the case of determining that the prediction probability is greater than or equal to the preset probability threshold, determining that the target object has a preset risk; in the case of determining that the prediction probability is less than the preset probability threshold, determining that the target object does not have a preset risk.

[0116] In some embodiments, in the case of determining that the target object has a preset risk, a corresponding risk mark may be set for the target object, so that subsequent business data processing can be performed on the target object according to the risk mark of the target object.

[0117] In some embodiments, before specific implementation, the above-mentioned preset processing model can be constructed and trained in the following manner:

[0118] S1: Obtain a plurality of sample data; wherein, the sample data corresponds to a sample object, and the sample data includes a plurality of feature data of the corresponding sample object; the sample object includes a first type of sample object carrying a preset label and a second type of sample object not carrying a preset label;

[0119] S2: According to a preset mapping rule, map the multiple feature data of the first type of sample object into a plurality of preset feature subspaces to obtain first type of training data; map the multiple feature data of the second type of sample object into a plurality of preset feature subspaces to obtain second type of training data;

[0120] S3: Use the first type of training data to train a plurality of initial classifiers to obtain corresponding intermediate classifiers; wherein, the initial classifiers respectively correspond to a preset feature subspace;

[0121] S4: Use the plurality of intermediate classifiers, the first type of training data, and the second type of training data to construct an objective function for the plurality of intermediate classifiers; wherein, the objective function includes a subspace consistency constraint formula and a dynamic pairwise constraint formula;

[0122] S5: According to the objective function, determine a plurality of preset classifiers that meet the requirements to construct a preset processing model.

[0123] Through the above embodiments, for a data prediction scenario with a relatively high feature dimension and complexity, a preset processing model that meets the requirements and has a good effect can be established and trained.

[0124] In some embodiments, for some data prediction scenarios with a relatively high feature dimension and complexity, it is often impossible to directly obtain all the sample data carrying preset labels. In most cases, only part of the large amount of sample data obtained carries a preset label, or the corresponding preset label can be easily determined; while the remaining other sample data does not carry a preset label, or it is not easy to determine the corresponding preset label.

[0125] In this embodiment, the sample object carrying a preset label can be recorded as the first type of sample object. The sample object not carrying a preset label can be recorded as the second type of sample object.

[0126] In some embodiments, during specific implementation, after obtaining a plurality of sample data, it is also possible to label the sample objects for which the preset label can be easily determined according to a preset labeling rule, so as to determine and label a plurality of first type of sample objects from the second type of sample objects.

[0127] Specifically, taking the scenario of overdue risk prediction of transaction data as an example, the repayment records of the historical business of the sample object can be obtained and used to label the sample object. For example, if it is determined according to the repayment records of the historical business of the sample object that the sample object has had an overdue record in the historical business, a preset label with a value of "1" (corresponding to the existence of a preset risk) can be set for the sample object, and the sample object can be recorded as a first-class sample object. If it is determined according to the repayment records of the historical business of the sample object that the sample object has never had an overdue record in all historical businesses and there are still other outstanding businesses for the current sample object, a preset label with a value of "0" (corresponding to the non-existence of a preset risk) can be set for the sample object, and the sample object can be recorded as a first-class sample object. If it is determined according to the repayment records of the historical business of the sample object that the sample object has never had an overdue record in all historical businesses and there are still other outstanding businesses for the current sample object, it is not possible to simply set a preset label for the sample object. In this case, the sample object can be recorded as a second-class sample object.

[0128] In some embodiments, during specific training, a plurality of initial classifiers corresponding to the preset feature subspaces can be constructed first; then the feature data in the first training data can be separately input into the corresponding initial classifiers in a differentiated manner to train the plurality of initial classifiers respectively, so as to obtain a plurality of intermediate classifiers trained by using the feature data of the first-class sample objects carrying the preset labels. Among them, each intermediate classifier corresponds to a preset feature subspace.

[0129] In some embodiments, when constructing an objective function for the plurality of intermediate classifiers by using the plurality of intermediate classifiers, the first-class training data, and the second-class training data, the specific implementation may include the following contents:

[0130] S1: Invoke a plurality of intermediate classifiers to process the second-class training data to obtain dynamic labels for the second-class sample objects;

[0131] S2: Construct a dynamic pairwise constraint formula according to the preset labels of the first-class sample objects and the dynamic labels of the second-class sample objects;

[0132] S3: Invoke a plurality of intermediate classifiers to process the first-class training data to obtain multiple classification results of the first-class sample objects;

[0133] S4: Construct a subspace consistency constraint formula according to the multiple classification results of the first-class sample objects;

[0134] S5: Construct the objective function according to the dynamic pairwise constraint formula and the subspace consistency constraint formula.

[0135] Through the above embodiments, based on the intermediate classifier trained with the feature data of the first type of sample objects, the feature data of the second type of sample objects can be further introduced and utilized, and at the same time, the subspace consistency constraint and the dynamic pairwise constraint are introduced and utilized to construct a qualified objective function, so that a preset processing model with better effects can be trained based on this objective function subsequently.

[0136] In some embodiments, during specific implementation, multiple intermediate classifiers can be called to process the corresponding feature groups of the second type of sample objects in the second type of training data respectively to obtain corresponding classification results; and then, through methods such as voting, the classification results output by multiple intermediate classifiers are integrated to obtain the dynamic label for the second type of sample objects.

[0137] Among them, the above dynamic label can be specifically understood as a pseudo-label determined by integrating the classification results output by multiple intermediate classifiers, which is different from the preset label.

[0138] In some embodiments, during specific implementation, the classification results output by multiple intermediate classifiers can be integrated according to the following formula to determine the corresponding dynamic label:

[0139]

[0140] Among them, K is the number of preset feature subspaces (i.e., the number of intermediate classifiers), and f k (·) is the intermediate classifier corresponding to the preset feature subspace numbered k, and x u is the feature data of the second type of sample object numbered u, and l u is the dynamic label of the second type of sample object numbered u.

[0141] In some embodiments, when constructing the dynamic pairwise constraint formula according to the preset label of the first type of sample objects and the dynamic label of the second type of sample objects, during specific implementation, it may include: establishing a connection set and a non-connection set according to the preset label of the first type of sample objects and the dynamic label of the second type of sample objects; wherein, the connection set contains multiple connection groups, and each connection group contains two sample objects with the same label; the non-connection set contains multiple non-connection groups, and each non-connection group contains two sample objects with different labels; constructing a dynamic pairwise matrix according to the connection set and the non-connection set; and constructing the dynamic pairwise constraint formula according to the dynamic pairwise matrix and the multiple intermediate classifiers.

[0142] Through the above embodiments, a dynamic pairwise matrix can be constructed by determining and based on the above connection set and non-connection set to introduce the mutual relationship between different sample objects, and then a corresponding dynamic pairwise constraint equation can be established.

[0143] In some embodiments, during specific implementation, the connection set can be established according to the following equation: ML = {(x i , x j ) | l i = l j}. Wherein, x i represents the feature data of the sample object numbered i, x j represents the feature data of the sample object numbered j. The above sample objects can be the first type of sample objects or the second type of sample objects, and i and j can be the same or different. l i represents the label of the sample object numbered i, l j represents the label of the sample object numbered j. The above labels can be preset labels or dynamic labels. Similarly, the non-connection set can be established according to the following equation: CL = {(x i , x j ) | l i ≠ l j}.

[0144] Then, based on the above connection set and non-connection set, a dynamic pairwise matrix S including the mutual interactions between all sample objects pairwise is constructed according to the following equation:

[0145]

[0146] Wherein, s i,j is the element in the i-th row and j-th column of the dynamic pairwise matrix S, and is used to represent the mutual interaction between the sample object numbered i and the sample object numbered j.

[0147] In some embodiments, during specific implementation, the dynamic pairwise constraint equation can be constructed according to the following equation based on the dynamic pairwise matrix and the multiple intermediate classifiers:

[0148]

[0149] Wherein, R1 represents the dynamic pairwise constraint, S represents the dynamic pairwise matrix, f k (·) is the intermediate classifier corresponding to the preset feature subspace numbered k, and X, X' represent the feature data of any two sample objects (which can be the same sample object or different sample objects).

[0150] In some embodiments, when specifically implementing the construction of the subspace consistency constraint equation based on the multiple classification results of the first type of sample objects, it may include: dividing multiple comparison groups according to the multiple classification results of the first type of sample objects; wherein, each comparison group contains two classification results of the same first type of sample object; constructing the subspace consistency constraint equation according to the multiple comparison groups.

[0151] Through the above embodiments, by comparing and utilizing the classification results for the same first type of sample object obtained by intermediate classifiers based on corresponding different feature subspaces, the interaction of different feature subspaces based on the same sample object can be introduced, and thus the corresponding subspace consistency constraint equation can be established.

[0152] In some embodiments, when specifically implementing, the subspace consistency constraint equation can be constructed according to the following equation based on the multiple comparison groups:

[0153]

[0154] Among them, R2 represents the subspace consistency constraint, and f p (·), f q (·) represent two intermediate classifiers corresponding to different feature subspaces, and X L represents the feature data of the first type of sample object carrying a preset label.

[0155] In some embodiments, when specifically implementing, the objective function can be constructed according to the following equation based on the dynamic pairwise constraint equation and the subspace consistency constraint equation:

[0156] L = R emp +α·R1 + β·R2

[0157] Among them, L represents the function value (or loss value) of the objective function, R emp represents the empirical loss, and α, β represent hyperparameters for adjusting relevant weights. The above empirical loss can be specifically determined according to the preset label of the first type of sample object.

[0158] The above objective function can be further expanded into the following form:

[0159]

[0160] Among them, the above Y represents the preset label of the first type of sample object X L .

[0161] In some embodiments, according to the above objective function, multiple preset classifiers that meet the requirements are determined to construct a preset processing model. Specifically, in implementation, it may include: solving the optimal solution (e.g., the minimum loss value) of the objective function by using the gradient descent algorithm to determine multiple preset classifiers that meet the requirements; and then connecting the multiple preset classifiers to the discriminant structure respectively to obtain the preset processing model.

[0162] Through the above embodiments, the model training process can be converted into a process of solving the optimal solution of the objective function, and by finding and determining the optimal solution of the objective function, the corresponding multiple preset classifiers can be determined, so that a preset processing model that meets the requirements and has good effects can be constructed.

[0163] In some embodiments, the above discriminant structure may specifically be built-in with the following discriminant function:

[0164]

[0165] Among them, F(x) represents the processing result output by the discriminant structure, x∈ω1 indicates that the data object has a preset risk, and x∈ω2 indicates that the data object does not have a preset risk.

[0166] In some embodiments, during the specific training process, the contribution degree of multiple preset classifiers to the processing result can also be calculated and, corresponding weight parameters are set for each preset classifier in the discriminant function to make the processing result output by the discriminant function relatively more accurate.

[0167] In some embodiments, after obtaining multiple sample data, when the method is specifically implemented, the following content may further be included: detecting whether the missing quantity of the feature data of the sample data is less than a preset missing quantity threshold; and in the case of determining that the missing quantity of the feature data of the sample data is less than the preset missing quantity threshold, performing zero-padding processing on the missing feature data.

[0168] Through the above embodiments, before training the preset processing model by using the sample data, the above preprocessing can be performed on the sample data first to reduce data errors, obtain sample data with good training effects, and participate in subsequent model training.

[0169] In some embodiments, when it is determined that the number of missing feature data in the sample data is less than a preset missing quantity threshold and the degree of missingness is not serious, the missing feature data can be filled with the value 0 or the character "unknown". When it is determined that the number of missing feature data in the sample data is greater than or equal to the preset missing quantity threshold and the degree of missingness is relatively serious, the sample data can be deleted to avoid data errors introduced by using the sample data in subsequent model training and improve the model training accuracy.

[0170] In some embodiments, after obtaining a plurality of sample data, specifically in implementation, it may further include: performing multivariate derivative variable exploration on the sample data to obtain richer data features; and then, in combination with the richer data features, training a preset processing model.

[0171] Specifically, the features can be evolved based on the feature data included in the sample data. For example, the statistical features of numerical features (including: maximum value, minimum value, mean value, variance, etc.), the deviation value features of numerical features (including: the differences between the original feature and the minimum value, maximum value, and mean value of this column), the cross features between numerical features (including: new columns obtained by performing relevant addition, subtraction, multiplication, and division operations between numerical features), etc. can be grouped and statistically analyzed according to category features, so as to obtain richer data features.

[0172] As can be seen from the above, for the risk determination method provided in the embodiments of this specification, before specific implementation, for data prediction scenarios with relatively high and complex feature dimensions, the multi-dimensional features can be first divided into multiple preset feature subspaces according to the preset mapping rules to reduce the dimension of the feature data of the sample object and reduce the complexity and the amount of data processing involved in subsequent model training; and, in combination with the dimension-reduced feature data, the target function is constructed by introducing and utilizing the subspace consistency constraint and the dynamic pairwise constraint to participate in model training, so that the sample data with and without the preset label can be effectively utilized to train a preset processing model with good effect and high accuracy; specifically in implementation, after obtaining a plurality of feature data of the target object, the plurality of feature data can be first mapped into multiple preset feature subspaces according to the preset mapping rules to obtain multiple corresponding feature groups of the target object; then, the preset processing model is called to process the multiple feature groups of the target object to obtain the corresponding processing results; and then, based on the processing results, it can be efficiently and accurately determined whether the target object has a preset risk. Thus, it can be better applicable to data prediction scenarios with relatively high and complex feature dimensions and accurately predict whether the target object has a preset risk.

[0173] The embodiments of this specification also provide another risk determination method. Specifically in implementation, the method may include the following content:

[0174] S1: Obtain multiple feature data of a target object;

[0175] S2: According to a preset mapping rule, map the multiple feature data into multiple preset feature subspaces to obtain multiple feature groups of the target object; wherein, each of the feature groups corresponds to a preset feature subspace;

[0176] S3: Invoke a preset processing model to process the multiple feature groups to obtain corresponding processing results; wherein, the preset processing model at least includes multiple preset classifiers, and each of the preset classifiers corresponds to a preset feature subspace; the preset processing model is pre-trained based on subspace consistency constraints or dynamic pairwise constraints;

[0177] S4: Determine whether the target object has a preset risk according to the processing results.

[0178] Through the above embodiments, a preset processing model obtained by training based on subspace consistency constraints alone or dynamic pairwise constraints alone can be used to balance the processing cost and processing efficiency and more accurately determine whether the target object has a preset risk.

[0179] In some embodiments, the objective function used to train the above preset processing model may specifically be an objective function that only includes a subspace consistency constraint formula or an objective function that includes a dynamic pairwise constraint formula.

[0180] The embodiments of this specification further provide a server, including a processor and a memory for storing processor-executable instructions. When specifically implemented, the processor may execute the following steps according to the instructions: Obtain multiple feature data of a target object; according to a preset mapping rule, map the multiple feature data into multiple preset feature subspaces to obtain multiple feature groups of the target object; wherein, each of the feature groups corresponds to a preset feature subspace; invoke a preset processing model to process the multiple feature groups to obtain corresponding processing results; wherein, the preset processing model at least includes multiple preset classifiers, and each of the preset classifiers corresponds to a preset feature subspace; the preset processing model is pre-trained based on subspace consistency constraints and dynamic pairwise constraints; determine whether the target object has a preset risk according to the processing results.

[0181] In order to be able to complete the above instructions more accurately, refer to Figure 4As shown in the figure, the embodiments of the present specification also provide another specific server. The server includes a network communication port 401, a processor 402, and a memory 403. The above structures are connected by internal cables so that each structure can perform specific data interactions.

[0182] Among them, the network communication port 401 can be specifically used to obtain multiple feature data of a target object.

[0183] The processor 402 can be specifically used to map the multiple feature data into multiple preset feature subspaces according to a preset mapping rule to obtain multiple feature groups of the target object; among them, each feature group corresponds to a preset feature subspace; call a preset processing model to process the multiple feature groups to obtain corresponding processing results; among them, the preset processing model at least includes multiple preset classifiers, and each preset classifier corresponds to a preset feature subspace; the preset processing model is pre-trained based on subspace consistency constraints and dynamic pairwise constraints; determine whether the target object has a preset risk according to the processing result.

[0184] The memory 403 can be specifically used to store corresponding instruction programs.

[0185] In this embodiment, the network communication port 401 can be bound to different communication protocols, so as to send or receive different data virtual ports. For example, the network communication port can be a port responsible for web data communication, or a port responsible for FTP data communication, or a port responsible for mail data communication. In addition, the network communication port can also be a physical communication interface or communication chip. For example, it can be a wireless mobile network communication chip, such as GSM, CDMA, etc.; it can also be a Wifi chip; it can also be a Bluetooth chip.

[0186] In this embodiment, the processor 402 can be implemented in any suitable manner. For example, the processor can take the form of, for example, a microprocessor or a processor and a computer-readable medium storing computer-readable program code (such as software or firmware) executable by the (micro)processor, logic gates, switches, application specific integrated circuit (ASIC), programmable logic controller, and embedded microcontroller, etc. The present specification does not make any limitations.

[0187] In this embodiment, the memory 403 may include multiple levels. In a digital system, anything that can store binary data can be a memory. In an integrated circuit, a circuit with a storage function that has no physical form is also called a memory, such as RAM, FIFO, etc. In a system, a storage device with a physical form is also called a memory, such as a memory module, a TF card, etc.

[0188] The embodiments of this specification also provide a computer-readable storage medium based on the above risk determination method. The computer-readable storage medium stores computer program instructions, which when executed, implement: obtaining multiple feature data of a target object; according to a preset mapping rule, mapping the multiple feature data into multiple preset feature subspaces to obtain multiple feature groups of the target object; wherein, each feature group corresponds to a preset feature subspace; calling a preset processing model to process the multiple feature groups to obtain corresponding processing results; wherein, the preset processing model at least includes multiple preset classifiers, and each preset classifier corresponds to a preset feature subspace; the preset processing model is pre-trained based on subspace consistency constraints and dynamic pairwise constraints; determining whether the target object has a preset risk according to the processing results.

[0189] In this embodiment, the above storage medium includes but is not limited to Random Access Memory (RAM), Read-Only Memory (ROM), Cache, Hard Disk Drive (HDD), or Memory Card. The memory can be used to store computer program instructions. The network communication unit can be set according to the standards specified by the communication protocol and is used for the interface of network connection communication.

[0190] In this embodiment, the functions and effects specifically implemented by the program instructions stored in the computer-readable storage medium can be explained by comparison with other embodiments and will not be elaborated here.

[0191] Refer to Figure 5 As shown, at the software level, the embodiments of this specification also provide a risk determination device, which specifically may include the following structural modules:

[0192] An acquisition module 501, which specifically can be used to acquire multiple feature data of a target object;

[0193] A mapping module 502, which specifically can be used to map the multiple feature data into multiple preset feature subspaces according to a preset mapping rule to obtain multiple feature groups of the target object; wherein, each feature group corresponds to a preset feature subspace;

[0194] The calling module 503 can be specifically used to call a preset processing model to process the multiple feature groups to obtain corresponding processing results. Among them, the preset processing model at least includes a plurality of preset classifiers, and each of the preset classifiers corresponds to a preset feature subspace. The preset processing model is pre-trained based on subspace consistency constraints and dynamic pairwise constraints.

[0195] The determination module 504 can be specifically used to determine whether the target object has a preset risk according to the processing result.

[0196] It should be noted that the units, devices, or modules described in the above embodiments can be specifically implemented by computer chips or entities, or by products with certain functions. For the convenience of description, when describing the above devices, various modules are described separately according to their functions. Of course, when implementing this specification, the functions of each module can be implemented in the same or multiple software and / or hardware, or the modules implementing the same function can be realized by a combination of multiple sub-modules or sub-units. The device embodiments described above are merely illustrative. For example, the division of the units is only a logical function division, and there may be other division methods in actual implementation. For example, multiple units or components can be combined or integrated into another system, or some features can be ignored or not executed. Another point is that the displayed or discussed mutual coupling or direct coupling or communication connection can be through some interfaces, and the indirect coupling or communication connection of the device or unit can be in electrical, mechanical or other forms.

[0197] As can be seen from the above, based on the risk determination device provided in the embodiments of this specification, it can be better applied to data prediction scenarios with higher and more complex feature dimensions, and accurately predict whether the target object has a preset risk.

[0198] In a specific scenario example, the risk determination method provided in this specification can be applied. By constructing and using a semi-supervised housing provident fund loan overdue prediction model (a preset processing model) based on subspace consistency and dynamic pairwise constraints, the accuracy of predicting the housing provident fund loan overdue (a preset risk) can be improved. The specific implementation process can refer to the following content.

[0199] When specifically applied, refer to Figure 6 As shown, the following steps can be included: First, obtain feature information related to housing provident fund loan overdue prediction (for example, multiple feature data of the target object) from the data warehouse. Perform data preprocessing and feature engineering processing on the samples. Use the features of the data to be predicted to construct test samples. Input the test samples into a semi-supervised housing provident fund loan overdue prediction model based on subspace consistency and dynamic pairwise constraints to obtain a prediction result.

[0200] Among them, the training process of the above-mentioned semi-supervised housing provident fund loan overdue prediction model based on subspace consistency and dynamic pairwise constraints can be referred to Figure 7 as shown. After data preprocessing and feature engineering, training samples are obtained, including a small number of labeled samples and a large number of unlabeled samples.

[0201] First, map the features of the samples (e.g., multiple sample data) into a total of K random subspaces (e.g., multiple preset feature subspaces) including random subspace 1, random subspace 2... random subspace K; use the features of the labeled samples (e.g., the first type of sample objects carrying preset labels) in the subspaces to train a total of K sub-classifiers (e.g., multiple intermediate classifiers) including classifier 1, classifier 2... classifier K respectively; then use the voting results of the K sub-classifications to assign corresponding labels (e.g., dynamic labels) to the unlabeled samples (e.g., the second type of sample objects not carrying preset labels), and update the labels of the unlabeled samples during the optimization process of the model.

[0202] Second, the dynamic pairwise constraints can be propagated from the labeled data points to the unlabeled data points by using the dynamic labels of the unlabeled samples during the model optimization process, calculate the dynamic pairwise constraints between all samples, and finally realize the constraint information of the entire dataset, and effectively use the dynamic pairwise constraints to iteratively update the classifier, so that the same-class samples are as close as possible in the output space, and the different-class samples are as far away as possible in the output space.

[0203] At the same time, by designing subspace consistency constraints, different subspaces can optimize each other, improve the effect of subspace learning, enhance robustness, and then a semi-supervised housing provident fund loan overdue prediction model based on subspace consistency and dynamic pairwise constraints can be obtained.

[0204] Specifically in implementation, the classifier can be iteratively optimized by minimizing the empirical loss, dynamic pairwise constraints, and subspace consistency constraints.

[0205] In this scenario example, the above-mentioned semi-supervised housing provident fund loan overdue prediction model based on subspace consistency and dynamic pairwise constraints is specifically constructed, which can include the following three parts: data preprocessing, feature engineering, model construction and training. The following will specifically explain each part.

[0206] 1. Regarding "Data Preprocessing"

[0207] 1.1. Data Selection. The data used in this modeling includes the basic identity information of individual users, as well as data information such as the housing provident fund contributions and loans of individuals. The features related to the prediction of overdue housing provident fund loans are divided into three categories. The first category is basic information, such as age, gender, region, etc. The second category is housing provident fund contribution information, such as individual contribution base, personal account balance, monthly individual contribution amount, etc. The third category is loan information, such as loan amount disbursed, loan balance, loan interest rate, etc. The data range can be determined by category, thereby determining the relevant data tables.

[0208] 1.2. Constructing Label Information. For users who have completed the repayment of housing provident fund loans, users who have had overdue repayments are defined as overdue customers, with the label set to 1, representing the first type of sample ω1. Users who have repaid all loans on time are defined as non-overdue users, with the label set to -1, representing the second type of sample ω2. For users who are still in the repayment period, they are defined as unlabeled samples and no label needs to be constructed.

[0209] 2. Regarding "Feature Engineering"

[0210] 2.1. Missing Value Handling. Observe the data columns in the data table. For columns with missing values, they are filled in a certain way. For example, for missing values in numerical features, they are filled with the value '0' in the column. For missing values in non-numerical features, they are filled with "unknown". For columns with particularly severe missing values, the field is directly deleted.

[0211] 2.2. Exploration of Derived Variables for Multiple Variables. Evolve the features, such as grouping and statistically analyzing the statistical information (maximum value, minimum value, mean value, variance, etc.) of numerical features according to categorical features, the deviation value features of numerical features (the differences between the original feature and the minimum value, maximum value, mean value of the column, etc.), and the cross features between numerical features (new columns obtained through relevant addition, subtraction, multiplication, and division operations between numerical features), etc.

[0212] 3. Regarding "Model Construction and Training"

[0213] 3.1. Constructing Label Information for Unlabeled Samples. Map the features of the samples into K random subspaces. Use the labeled samples to train K sub-classifiers respectively in the subspaces. Assign labels to the unlabeled samples using the voting results of the K sub-classifications, and update the labels of the unlabeled samples during the optimization process of the model. Specifically as follows:

[0214]

[0215] Among them, K is the number of subspaces, f k (·) is the base classifier in the k-th subspace, x u is the unlabeled sample, l u is to assign a label to the unlabeled sample x uAssigned labels.

[0216] 3.2. Construct dynamic pairwise constraints. Using the dynamic labels of unlabeled samples during model optimization, calculate the connection set between samples: ML = {(x i , x j ) | l i = l j} and the non-connection set: CL = {(x i , x j ) | l i ≠ l j}, and then calculate the dynamic pairwise matrix S for all samples. The calculation method is as follows:

[0217]

[0218] Then, use the dynamic pairwise matrix to calculate the dynamic pairwise constraints. The calculation method is as follows:

[0219]

[0220] where K is the number of random subspaces, and X, X' are any two training samples (including labeled and unlabeled samples).

[0221] 3.3. Construct subspace consistency constraints. By designing subspace consistency constraints, different subspaces can optimize each other, improve the effect of subspace learning, and enhance robustness. The calculation method of subspace consistency constraints is as follows:

[0222]

[0223] where X L is the set of labeled samples in the training samples.

[0224] 3.4. Objective function design. Iteratively optimize the classifier by minimizing the empirical loss, dynamic pairwise constraints, and subspace consistency constraints. The objective function is as follows: L = R emp + α·R1 + β·R2.

[0225] Furthermore, expand the objective function as follows:

[0226]

[0227] where Y is the label of the labeled sample set, R emp is the empirical loss, and α, β are hyperparameters used to adjust the weights of the above items.

[0228] 3.5. Model Optimization. The gradient descent method is used to solve this optimization problem. The objective function of the model is minimized until the preset number of iterations is reached or the difference between the loss values of the two loss functions is less than the preset threshold. The final classification model is obtained. The specific discriminant function is as follows:

[0229]

[0230] 3.6. Model Testing. For the test sample x, the discriminant function of the classifier is input to obtain the discriminant result of the model.

[0231] In this scenario example, a semi-supervised housing provident fund loan overdue prediction model based on subspace consistency and dynamic pairwise constraints is established through the above method. Its training samples include a small number of labeled samples and a large number of unlabeled samples. First, subspace learning is used instead of splicing the features of all categories together to prevent the "curse of dimensionality" problem from occurring during training due to too many features. Second, by designing subspace consistency constraints, different subspaces can optimize each other, improve the effect of subspace learning, and enhance robustness. Second, during the iterative optimization process of the model, possible dynamic labels are assigned to unlabeled samples according to the voting results of the model in multiple subspaces, further making full use of the spatial structure information contained in unlabeled samples. Finally, a dynamic pairwise constraint is designed to make the same-class samples as close as possible in the output space and the different-class samples as far away as possible in the output space, and the dynamic labels of unlabeled samples in the model optimization process are used to spread the pairwise constraints from labeled data points to unlabeled data points, ultimately realizing the constraint information of the entire dataset and improving the generalization effect of the model.

[0232] Through the above scenario example, the above model obtained through training has better performance in terms of precision, recall, and comprehensive evaluation value in the classification of housing provident fund loan overdue prediction compared with the traditional semi-supervised learning algorithm model, and can more accurately predict housing provident fund loan overdue users. Applying this model to financial institutions such as banks, using data information such as personal basic identity information, personal housing provident fund contributions and loans, etc., an accurate risk control model can be established to scientifically evaluate the safety of credit assets and the debt repayment ability of credit entities, and maximize the prevention of housing provident fund loan overdue risks.

[0233] Although this specification provides method operation steps as described in the embodiments or flowcharts, more or fewer operation steps may be included based on conventional or non-creative means. The order of steps listed in the embodiments is only one way among many orders of step execution and does not represent the only execution order. When the actual device or client product is executed, it can be executed in the order of the method shown in the embodiments or the drawings or in parallel (for example, in a parallel processor or multi-threaded processing environment, or even in a distributed data processing environment). The terms "comprising", "including" or any other variant thereof are intended to cover non-exclusive inclusion, such that a process, method, product or device comprising a series of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such process, method, product or device. Without further limitation, there is no exclusion of additional identical or equivalent elements in the process, method, product or device comprising the said elements. The terms such as first, second, etc. are used to denote names and do not denote any particular order.

[0234] As is also known to those skilled in the art, in addition to implementing the controller in the form of pure computer-readable program code, it is entirely possible to logically program the method steps so that the controller can be implemented in the form of logic gates, switches, application-specific integrated circuits, programmable logic controllers, embedded microcontrollers, etc. to achieve the same functions. Therefore, such a controller can be considered a hardware component, and the devices included therein for implementing various functions can also be regarded as the structures within the hardware component. Or even, the devices for implementing various functions can be regarded as either software modules for implementing the method or structures within the hardware component.

[0235] This specification can be described in the general context of computer-executable instructions executed by a computer, such as program modules. Generally, program modules include routines, programs, objects, components, data structures, classes, etc. that perform specific tasks or implement specific abstract data types. This specification can also be practiced in a distributed computing environment where tasks are performed by remote processing devices connected through a communication network. In a distributed computing environment, program modules can be located in local and remote computer-readable storage media including storage devices.

[0236] From the descriptions of the above embodiments, those skilled in the art can clearly understand that this specification can be implemented by means of software plus a necessary general hardware platform. Based on such an understanding, the technical solution of this specification can essentially be embodied in the form of a software product, which can be stored in a storage medium, such as ROM / RAM, magnetic disk, optical disk, etc., including several instructions for causing a computer device (which can be a personal computer, mobile terminal, server, or network device, etc.) to execute the methods described in various embodiments or some parts of the embodiments of this specification.

[0237] The various embodiments in this specification are described in a progressive manner. For the same or similar parts between the various embodiments, reference can be made to each other. The key point of each embodiment is to illustrate the differences from other embodiments. This specification can be used in many general or specific computer system environments or configurations. For example: personal computers, server computers, handheld or portable devices, tablet devices, multi-processor systems, microprocessor-based systems, set-top boxes, programmable electronic devices, network PCs, minicomputers, mainframe computers, distributed computing environments including any of the above systems or devices, and so on.

[0238] Although this specification is depicted through embodiments, those of ordinary skill in the art know that this specification has many variations and changes without departing from the spirit of this specification. It is hoped that the appended claims will cover these variations and changes without departing from the spirit of this specification.

Claims

1. A method for determining a risk, characterized in that, Including: Obtain multiple feature data of a target object; wherein, the target object includes a trading account; the feature data includes: identity type feature data of the trading account, historical trading behavior type feature data of the trading account, and current trading behavior type feature data of the trading account; According to a preset mapping rule, map the multiple feature data into multiple preset feature subspaces to obtain multiple feature groups of the target object; wherein, each feature group corresponds to a preset feature subspace; Call a preset processing model to process the multiple feature groups to obtain corresponding processing results; wherein, the preset processing model at least includes multiple preset classifiers, and each preset classifier corresponds to a preset feature subspace; the preset processing model is pre-trained based on subspace consistency constraints and dynamic pairwise constraints; According to the processing results, determine whether the target object has a preset risk; the preset risk includes the risk of overdue transaction data; Wherein, the preset processing model is trained based on an objective function; the objective function is constructed by using subspace consistency constraints and dynamic pairwise constraints; the objective function is expressed as the following formula: where, L represents the function value of the objective function, Y represents the preset label of the first type of sample object X L , S represents the dynamic pairwise matrix, f k (·) is the intermediate classifier corresponding to the preset feature subspace numbered k, X and X' represent the feature data of any two sample objects, f p (·), f q (·) represent two intermediate classifiers corresponding to different feature subspaces, X L represents the feature data of the first type of sample object carrying the preset label, and K is the number of preset feature subspaces; Wherein, the dynamic pairwise matrix is constructed according to the following formula based on the connected set and the unconnected set: where s i,j is the element in the i-th row and j-th column of the dynamic pair matrix S, which is used to represent the interaction between the sample object numbered i and the sample object numbered j. ML is the connection set, and CL is the non-connection set; where, ML = {(x i , x j ) | l i = l j}, CL = {(x i , x j ) | l i ≠ l j}, x i represents the feature data of the sample object numbered i, x j represents the feature data of the sample object numbered j, l i represents the label of the sample object numbered i, l j represents the label of the sample object numbered j.

2. The method according to claim 1, wherein The preset processing model further includes a discriminant structure; wherein, the discriminant structure is connected to the multiple preset classifiers; the discriminant structure is used to generate the processing results according to the classification results output by the multiple preset classifiers.

3. The method according to claim 2, wherein The preset processing model is established in the following manner: Obtain multiple sample data; wherein, the sample data corresponds to a sample object, and the sample data includes multiple feature data of the corresponding sample object; the sample object includes a first type of sample object with a preset label and a second type of sample object without a preset label; According to a preset mapping rule, map the multiple feature data of the first type of sample object into multiple preset feature subspaces to obtain first type of training data; map the multiple feature data of the second type of sample object into multiple preset feature subspaces to obtain second type of training data; Use the first type of training data to train multiple initial classifiers to obtain corresponding multiple intermediate classifiers; wherein, each initial classifier corresponds to a preset feature subspace; Use the multiple intermediate classifiers, the first type of training data, and the second type of training data to construct an objective function for the multiple intermediate classifiers; wherein, the objective function includes a subspace consistency constraint formula and a dynamic pairwise constraint formula; According to the objective function, determine multiple preset classifiers that meet the requirements to construct a preset processing model.

4. The method according to claim 3, wherein Using the multiple intermediate classifiers, the first type of training data, and the second type of training data to construct an objective function for the multiple intermediate classifiers includes: Call the multiple intermediate classifiers to process the second type of training data to obtain dynamic labels for the second type of sample objects; Construct a dynamic paired constraint equation according to the preset labels of the first - type sample objects and the dynamic labels of the second - type sample objects; Call multiple intermediate classifiers to process the first - type training data to obtain multiple classification results of the first - type sample objects; Construct a subspace consistency constraint equation according to the multiple classification results of the first - type sample objects; Construct the objective function according to the dynamic paired constraint equation and the subspace consistency constraint equation.

5. The method according to claim 4, characterized in that, Construct a dynamic paired constraint equation according to the preset labels of the first - type sample objects and the dynamic labels of the second - type sample objects, including: Establish a connection set and a non - connection set according to the preset labels of the first - type sample objects and the dynamic labels of the second - type sample objects; wherein, the connection set contains multiple connection groups, and each connection group contains two sample objects with the same label; the non - connection set contains multiple non - connection groups, and each non - connection group contains two sample objects with different labels; Construct a dynamic paired matrix according to the connection set and the non - connection set; Construct the dynamic paired constraint equation according to the dynamic paired matrix and the multiple intermediate classifiers.

6. The method according to claim 4, characterized in that, Construct a subspace consistency constraint equation according to the multiple classification results of the first - type sample objects, including: Divide multiple comparison groups according to the multiple classification results of the first - type sample objects; wherein, each comparison group contains two classification results of the same first - type sample object; Construct the subspace consistency constraint equation according to the multiple comparison groups.

7. The method according to claim 3, wherein Determine multiple preset classifiers that meet the requirements according to the objective function to construct a preset processing model, including: Use the gradient descent algorithm to solve the optimal solution of the objective function to determine multiple preset classifiers that meet the requirements; Connect the multiple preset classifiers to a discriminant structure to obtain a preset processing model.

8. The method according to claim 3, characterized in that, After obtaining multiple sample data, the method further includes: Detect whether the missing quantity of the feature data of the sample data is less than a preset missing quantity threshold; In the case of determining that the missing quantity of the feature data of the sample data is less than the preset missing quantity threshold, perform zero - filling processing on the missing feature data.

9. A risk determination device, characterized in that, Include: An acquisition module for acquiring multiple feature data of a target object; wherein, the target object includes a trading account; the feature data includes: identity - type feature data of the trading account, historical trading behavior - type feature data of the trading account, and current trading behavior - type feature data of the trading account; A mapping module for mapping the multiple feature data into multiple preset feature sub - spaces according to a preset mapping rule to obtain multiple feature groups of the target object; wherein, each feature group corresponds to a preset feature sub - space; A call module for calling a preset processing model to process the multiple feature groups to obtain corresponding processing results; wherein, the preset processing model at least includes multiple preset classifiers, and each preset classifier corresponds to a preset feature sub - space; the preset processing model is pre - trained based on subspace consistency constraints and dynamic paired constraints. A determination module, configured to determine whether a target object has a preset risk according to the processing result; the preset risk includes a risk of overdue transaction data; Wherein, the preset processing model is trained based on an objective function; the objective function is constructed by using subspace consistency constraints and dynamic pairwise constraints; the objective function is expressed as the following formula: Among them, L represents the function value of the objective function, Y represents the preset label of the first type of sample object X L , S represents the dynamic pairwise matrix, and f k (·) is the intermediate classifier corresponding to the preset feature subspace numbered k, X and X' represent the feature data of any two sample objects, and f p (·), f q (·) represent two intermediate classifiers corresponding to different feature subspaces, and X L represents the feature data of the first type of sample object carrying the preset label, and K is the number of preset feature subspaces; Wherein, the dynamic pairwise matrix is constructed according to the following formula based on the connection set and the non-connection set: where s i,j is the element in the i-th row and j-th column of the dynamic pair matrix S, used to represent the interaction between the sample object numbered i and the sample object numbered j, ML is the connection set, and CL is the non-connection set; where, ML = {(x i , x j ) | l i = l j}, CL = {(x i , x j ) | l i ≠ l j}, x i represents the feature data of the sample object numbered i, x j represents the feature data of the sample object numbered j, l i represents the label of the sample object numbered i, l j represents the label of the sample object numbered j.

10. A server, characterized in that, It includes a processor and a memory for storing processor-executable instructions, and when the processor executes the instructions, the steps of the method according to any one of claims 1 to 8 are implemented.

11. A computer-readable storage medium, characterized in that, Computer instructions are stored thereon, and when the instructions are executed, the steps of the method according to any one of claims 1 to 8 are implemented.

Images