Model training method based on multi-granularity risk label and sorting loss and related product

Abstract

The invention provides a model training method based on multi-granularity risk tags and sorting loss and related products. According to the method, a multi-granularity risk tag is converted into a corresponding risk level, then a risk correlation tag of each training sample is determined based on the risk level and a target tag gain mapping method which is adjustable in a hyper-parameter optimization process, and the risk correlation tag is used as a real tag of the training sample. The risk evolution process of the training sample in different life cycle stages can be more comprehensively described, the sorting loss is used as an optimization target, the risk assessment model is guided to focus on prediction of the risk relative sequence, the risk sorting capability of the risk assessment model is enhanced, and in addition, the risk evaluation efficiency is improved. According to the method, the training samples which are short in relation duration but show early-stage risk signs can be effectively included, the utilization efficiency of early-stage risk signals is effectively improved, and information loss caused by label flattening and rough storm deletion of a traditional risk assessment model is avoided.

Description

Technical Field

[0001] This disclosure relates to the field of artificial intelligence technology, specifically to a model training method and related products based on multi-granularity risk labels and ranking loss. Background Technology

[0002] In the field of financial risk control, risk assessment models are key tools for quantifying the risk level of entities. These models typically output a risk score of 300 to 1000 points to measure the risk status of entity samples (such as individuals, businesses, and transactions). Traditional risk assessment models mainly rely on supervised learning frameworks. In this approach, entity samples are labeled as low-risk or high-risk entities based on their behavior over a fixed period (e.g., 6 months), and then the risk assessment model is trained using cross-entropy loss based on these labels.

[0003] However, this traditional method has significant limitations: (1) the risk level of entity samples is simplified to a binary label (low-risk entity / high-risk entity), which cannot accurately reflect different levels of risk gradient; (2) for entity samples that have shown early signs of risk in a short period of relationship (such as slight abnormal behavior after only one month of observation), the traditional risk assessment model excludes them from the training samples because they do not conform to the labeling rules of fixed relationship duration, which means that these early signals that have potential value for long-term risk prediction cannot be used by the risk assessment model.

[0004] Therefore, it is necessary to propose a model training method and related products based on multi-granularity risk labels and ranking loss to solve at least one of the above-mentioned technical problems. Summary of the Invention

[0005] This disclosure presents a model training method and related products based on multi-granularity risk labels and ranking loss. By utilizing multi-granularity risk labels of training samples during different relationship durations and combining them with ranking loss to iteratively train the risk assessment model, it is possible to more precisely characterize the risk evolution process of training samples during different relationship durations, provide more discriminative supervisory information for the risk assessment model, and enhance the ability of the optimized risk assessment model to distinguish between high and low risk rankings.

[0006] In a first aspect, this disclosure provides a model training method based on multi-granularity risk labels and ranking loss, the method comprising: Obtain a training sample dataset, a risk assessment model, and a hyperparameter set for the risk assessment model. The training sample dataset includes training data for each training sample and multi-granularity risk labels. The multi-granularity risk labels are determined based on the performance data of the training samples during different relationship durations. The hyperparameter set includes label gain mapping hyperparameters. Based on the preset risk label level conversion method, the risk level corresponding to the multi-granularity risk label of each training sample is determined; The target label gain mapping method is determined based on the label gain mapping hyperparameters. Based on the risk level of each training sample and the target label gain mapping method, determine the risk correlation label of each training sample; The ranking loss for each training sample is determined based on the training data, the risk correlation labels, and the risk assessment model for each training sample. The risk assessment model is iteratively trained according to the ranking loss until the preset model training stop condition is met, thereby generating the target risk assessment model.

[0007] In some optional implementations, the label gain mapping hyperparameters include a target label gain mapping method identifier, and determining the target label gain mapping method based on the label gain mapping hyperparameters includes: Obtain a set of label gain mapping methods, wherein the set of label gain mapping methods includes at least one monotonically increasing label gain mapping method and a label gain mapping method identifier corresponding to each of the label gain mapping methods; The target label gain mapping method is determined based on the target label gain mapping method identifier and the label gain mapping method set.

[0008] In some optional implementations, the label gain mapping hyperparameters include the method type and method adjustment coefficient of the target label gain mapping method, wherein the method type includes at least one monotonically increasing label gain mapping method type, and determining the target label gain mapping method based on the label gain mapping hyperparameters includes: The target label gain mapping method is determined based on the method type and the method adjustment coefficient.

[0009] In some optional implementations, determining the risk relevance label of each training sample based on the risk level of each training sample and the target label gain mapping method includes: Based on the risk level of each training sample and the target label gain mapping method, determine the gain mapping value of each training sample; Based on the gain mapping values, the risk correlation labels of each training sample are determined.

[0010] In some optional implementations, the training sample dataset includes an omnichannel training sample dataset and / or a target channel training sample dataset, and the risk assessment model includes a general risk assessment model and / or a customized risk assessment model. The omnichannel training sample dataset includes training data and multi-granularity risk labels for each omnichannel training sample, and the target channel training sample dataset includes training data and multi-granularity risk labels or binary risk labels for each target channel training sample.

[0011] In some optional implementations, the step of iteratively training the risk assessment model based on each of the ranking losses until a preset model training stop condition is met to generate a target risk assessment model includes: The general risk assessment model or the first customized risk assessment model is iteratively trained based on the ranking loss of each of the omnichannel training samples or the ranking loss of each of the target channel training samples until the preset general model stop training condition or the first customized model stop training condition is met, thereby generating the target general risk assessment model or the first target customized risk assessment model.

[0012] In some optional implementations, after generating the target general risk assessment model, the method further includes: Based on the training data of each target channel training sample and the target general risk assessment model, generate a predicted risk correlation label for each target channel training sample; The predicted risk correlation labels are merged with the training data of the corresponding target channel training samples to generate merged training data of each target channel training sample. Based on the merged training data of each target channel training sample, the multi-granularity risk label, and the second customized risk assessment model, determine the ranking loss of each target channel training sample; The second customized risk assessment model is iteratively trained based on the ranking loss of the training samples of each target channel until the preset training stop condition of the second customized model is met, thereby generating the second target customized risk assessment model.

[0013] In some optional implementations, after generating the target general risk assessment model, the method further includes: Based on the training data of each target channel training sample and the target general risk assessment model, generate a predicted risk correlation label for each target channel training sample; The predicted risk correlation labels are merged with the training data of the corresponding target channel training samples to generate merged training data of each target channel training sample. Based on the merged training data of each target channel training sample, the binary risk label, and the third customized risk assessment model, determine the cross-entropy loss of each target channel training sample; The third customized risk assessment model is iteratively trained based on the cross-entropy loss of the training samples of each target channel until the preset training stop condition of the third customized model is met, thereby generating the third target customized risk assessment model.

[0014] In some optional implementations, determining the ranking loss for each training sample based on the training data, the risk correlation label, and the risk assessment model includes: The training data of each training sample is input into the risk assessment model, and the risk assessment model outputs the predicted risk correlation label of each training sample. The ranking loss of each training sample is determined based on the predicted risk correlation label and the risk correlation label of each training sample.

[0015] In some optional implementations, after determining the risk correlation label of each training sample based on the risk level of each training sample and the target label gain mapping method, the method further includes: The training sample dataset is divided into X training sample data subsets according to a preset sampling rule, where X is a positive integer greater than 1. The step of inputting the training data of each training sample into the risk assessment model, and outputting the predicted risk relevance label of each training sample through the risk assessment model, includes: For each subset of training sample data, the training data of each training sample in the subset of training sample data is input into the risk assessment model, and the risk assessment model outputs the predicted risk correlation label of each training sample in the subset of training sample data.

[0016] Secondly, this disclosure provides a risk assessment method based on multi-granularity risk labels and ranking losses, the method comprising: Obtain multi-dimensional data of the target object, wherein the multi-dimensional data of the target object includes data used to assess the risk of the target object; The multi-dimensional data is input into the target risk assessment model, and the target risk assessment model outputs the risk assessment result of the target object. The target risk assessment model is obtained through the following training method: determining the risk relevance label of each training sample based on the risk level and target label gain mapping method of each training sample in the training sample dataset; determining the ranking loss of each training sample based on the training data, the risk relevance label and the risk assessment model; and iteratively training the risk assessment model based on the ranking loss.

[0017] Thirdly, this disclosure provides a model training device based on multi-granularity risk labels and ranking loss, the device comprising: The training data acquisition unit is used to acquire the training sample dataset, the risk assessment model, and the hyperparameter set of the risk assessment model. The training sample dataset includes the training data of each training sample and multi-granularity risk labels. The multi-granularity risk labels are determined based on the performance data of the training samples during different relationship durations. The hyperparameter set includes label gain mapping hyperparameters. The risk label level conversion unit is used to determine the risk level corresponding to the multi-granularity risk label of each training sample according to the preset risk label level conversion method. The tag gain mapping method determination unit is used to determine the target tag gain mapping method based on the tag gain mapping hyperparameters. The risk-related label determination unit is used to determine the risk-related label of each training sample based on the risk level of each training sample and the target label gain mapping method. The ranking loss determination unit is used to determine the ranking loss of each training sample based on the training data of each training sample, the risk correlation label and the risk assessment model. The risk assessment model generation unit is used to iteratively train the risk assessment model according to each ranking loss until the preset model stopping training condition is met, thereby generating the target risk assessment model.

[0018] Fourthly, this disclosure provides a model risk assessment device based on multi-granularity risk labels and ranking loss, the device comprising: A multi-dimensional data acquisition unit is used to acquire multi-dimensional data of a target object, wherein the multi-dimensional data of the target object includes data used to assess the risk of the target object; The risk assessment result output unit is used to input the multi-dimensional data into the target risk assessment model, and output the risk assessment result of the target object through the target risk assessment model. The target risk assessment model is obtained by determining the risk correlation label of each training sample based on the risk level and target label gain mapping method of each training sample in the training sample dataset; determining the ranking loss of each training sample based on the training data, the risk correlation label and the risk assessment model; and iteratively training the risk assessment model based on the ranking loss.

[0019] Fifthly, this disclosure provides an electronic device, including: One or more processors; Storage device, on which one or more programs are stored, When the above-described one or more programs are executed by the above-described one or more processors, the above-described one or more processors implement the methods as described in any embodiment of the first and / or second aspects of this disclosure.

[0020] In a sixth aspect, this disclosure provides a computer-readable storage medium having a computer program stored thereon, wherein the computer program, when executed by one or more processors, implements the method as described in any embodiment of the first and / or second aspects of this disclosure.

[0021] In a seventh aspect, this disclosure provides a computer program product including a computer program / instructions that, when executed by a processor, implement the method described in any embodiment of the first and / or second aspects of this disclosure.

[0022] This disclosure provides a model training method and related products based on multi-granularity risk labels and ranking losses. First, a training sample dataset, a risk assessment model, and a hyperparameter set for the risk assessment model are obtained. The training sample dataset includes training data for each training sample and multi-granularity risk labels. The multi-granularity risk labels are determined based on the performance data of the training samples during different relationship durations. The hyperparameter set includes label gain mapping hyperparameters. The risk level corresponding to the multi-granularity risk label for each training sample is determined according to a preset risk label level conversion method. Then, a target label gain mapping method is determined based on the label gain mapping hyperparameters. Next, the risk relevance label for each training sample is determined based on its risk level and the target label gain mapping method. Then, the ranking loss for each training sample is determined based on its training data, risk relevance label, and risk assessment model. Finally, the risk assessment model is iteratively trained based on each ranking loss until a preset model training stop condition is met, generating the target risk assessment model. This disclosure converts multi-granularity risk labels into corresponding risk levels, and then determines the risk relevance label of each training sample based on the risk level and target label gain mapping method. This risk relevance label is used as the true label of the training sample, which can more finely and comprehensively depict the risk evolution process of the training sample at different life cycle stages. This provides more discriminative supervision information for the risk assessment model. Furthermore, by using ranking loss as the optimization objective, the risk assessment model is guided to focus on predicting the relative order of risks among training samples, which enhances the ability of the optimized risk assessment model to distinguish between high and low risk rankings. In addition, this method can effectively include training samples with short relationship durations but early risk signs, effectively improving the utilization efficiency of early risk signals and avoiding information loss caused by label flattening and crude deletion in traditional risk assessment models. Attached Figure Description

[0023] Other features, objects, and advantages of this disclosure will become more apparent from the following detailed description of non-limiting embodiments with reference to the accompanying drawings. The drawings are for illustrative purposes only and are not intended to limit the invention. In the drawings: Figure 1 This is a system architecture diagram in which an embodiment of the model training and risk assessment method based on multi-granularity risk labeling and ranking loss of this disclosure can be applied; Figure 2 This is a flowchart of an embodiment of the model training method based on multi-granularity risk labels and ranking loss according to the present disclosure; Figure 3 This is an exploded flowchart of one embodiment of step 206 of this disclosure; Figure 4This is an exploded flowchart of another embodiment of step 206 of this disclosure; Figure 5 This is a flowchart of an embodiment of the risk assessment method based on multi-granularity risk labeling and ranking loss according to the present disclosure; Figure 6 This is a schematic diagram of the structure of an embodiment of a model training device based on multi-granularity risk labels and ranking loss according to the present disclosure; Figure 7 This is a schematic diagram of a structure of an embodiment of the model risk assessment device based on multi-granularity risk labeling and ranking loss according to the present disclosure; Figure 8 This is a schematic diagram of the structure of a computer system suitable for implementing embodiments of the present disclosure. Detailed Implementation

[0024] The present disclosure will now be described in further detail with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative of the invention and not intended to limit it. Furthermore, it should be noted that, for ease of description, only the parts relevant to the invention are shown in the accompanying drawings.

[0025] The user data involved in the embodiments of this disclosure shall be collected and / or used in strict accordance with the laws, regulations and industry standards of relevant countries and regions. The collection and acquisition of data involved in the embodiments of this disclosure shall be conducted in advance by actively prompting or prominently displaying information to inform users and obtaining authorization, or by obtaining full authorization from all parties. The processing, manipulation, forwarding and use of data involved in the embodiments of this disclosure shall be carried out with the full knowledge and authorization of the user or relevant party. In implementing the embodiments of this disclosure, the types of data or information, scope of use, and usage scenarios that may be involved shall be informed to users or relevant parties and authorization shall be obtained through appropriate means. The specific methods of notification and authorization may vary according to the actual situation, and this disclosure is not limited in this regard. The processing of personal information involved in the embodiments of this disclosure shall be carried out under the premise of having a legal basis (such as obtaining the consent of the personal information subject or being necessary for the performance of a contract), and shall only be processed within the prescribed or agreed scope. Sensitive personal information such as biometric information, medical and health information, financial account information, and precise location information involved in the embodiments of this disclosure shall be processed under the premise of having a specific purpose and sufficient necessity, and with the separate authorization and consent of the user or relevant party. In some embodiments of this disclosure, if a user or related party refuses to process personal information other than the information necessary for the basic functions, it will not affect the use of the basic functions of the embodiments of this disclosure.

[0026] It should be noted that, unless otherwise specified, the embodiments and features described in this disclosure can be combined with each other. This disclosure will now be described in detail with reference to the accompanying drawings and embodiments.

[0027] Figure 1 An exemplary system architecture 100 is shown, in which embodiments of the model training and risk assessment methods based on multi-granularity risk labels and ranking loss of this disclosure can be applied.

[0028] like Figure 1 As shown, system architecture 100 may include terminal devices 101, 102, and 103, network 104, and server 105. Network 104 is used to provide communication links between terminal devices 101, 102, and 103 and server 105. Network 104 may include various communication connection types, such as wired communication links, wireless communication links, etc.

[0029] Objects can use terminal devices 101, 102, and 103 to interact with server 105 via network 104 to receive or send messages, etc. Various communication client applications can be installed on terminal devices 101, 102, and 103, such as model building applications, risk assessment applications, voice interaction applications, video conferencing applications, short video social applications, web browser applications, shopping applications, search applications, instant messaging tools, email clients, social platform software, etc.

[0030] Terminal devices 101, 102, and 103 can be either hardware or software. When terminal devices 101, 102, and 103 are hardware, they can be various electronic devices with microphones and speakers, including but not limited to smartphones, tablets, e-book readers, MP3 players (Moving Picture Experts Group Audio Layer III), MP4 players (Moving Picture Experts Group Audio Layer IV), portable computers, and desktop computers, etc. When terminal devices 101, 102, and 103 are software, they can be installed in the aforementioned electronic devices. They can be implemented as multiple software programs or software modules (e.g., acquiring training sample datasets, risk assessment models, and hyperparameter sets of risk assessment models), or as a single software program or software module. No specific limitations are imposed here.

[0031] Server 105 can be a server that provides various services, such as a backend server that processes the training sample dataset, risk assessment model, and hyperparameter set of the risk assessment model obtained from terminal devices 101, 102, and 103. The backend server can perform corresponding processing based on the training sample dataset, risk assessment model, and hyperparameter set of the risk assessment model obtained from the terminal devices.

[0032] In some cases, the model training and risk assessment method based on multi-granularity risk labels and ranking loss provided in this disclosure can be jointly executed by terminal devices 101, 102, and 103 and server 105. For example, the step of "obtaining the training sample dataset, the risk assessment model, and the hyperparameter set of the risk assessment model" can be executed by terminal devices 101, 102, and 103, and the step of "determining the risk level corresponding to the multi-granularity risk label of each training sample according to the preset risk label level conversion method" can be executed by server 105. This disclosure does not limit this. Correspondingly, the model training and risk assessment device based on multi-granularity risk labels and ranking loss can also be respectively set in terminal devices 101, 102, and 103 and server 105.

[0033] In some cases, the model training and risk assessment method based on multi-granularity risk labels and ranking loss provided in this disclosure can be executed by server 105. Correspondingly, the model training and risk assessment device based on multi-granularity risk labels and ranking loss can also be set in server 105. In this case, the system architecture 100 may not include terminal devices 101, 102, and 103.

[0034] In some cases, the model training and risk assessment method based on multi-granularity risk labels and ranking loss provided in this disclosure can be executed by terminal devices 101, 102, and 103. Correspondingly, the model training and risk assessment device based on multi-granularity risk labels and ranking loss can also be set in terminal devices 101, 102, and 103. In this case, the system architecture 100 may not include server 105.

[0035] It should be noted that server 105 can be either hardware or software. When server 105 is hardware, it can be implemented as a distributed server cluster consisting of multiple servers, or as a single server. When server 105 is software, it can be implemented as multiple software programs or software modules (for example, used to provide distributed services), or as a single software program or software module. No specific limitations are made here.

[0036] It should be understood that Figure 1 The number of terminal devices, networks, and servers shown is merely illustrative. Depending on implementation needs, any number of terminal devices, networks, and servers can be included.

[0037] Continue to refer to Figure 2 , Figure 2 A flowchart 200 illustrates an embodiment of a model training method based on multi-granularity risk labels and ranking loss according to this disclosure. Figure 2 The model training method shown, based on multi-granularity risk labels and ranking loss, can be applied to... Figure 1The terminal device or server shown. This process 200 includes at least the following steps 201-203.

[0038] Step 201: Obtain the training sample dataset, the risk assessment model, and the hyperparameter set of the risk assessment model.

[0039] In this embodiment, the training sample dataset may include training data for each training sample and multi-granularity risk labels. The multi-granularity risk labels are determined based on the performance data of the training samples during different relationship durations. The hyperparameter set includes label gain mapping hyperparameters.

[0040] The training data of the training samples can refer to the multi-dimensional data of the training samples. The multi-dimensional data can include data used to assess the risk of the target object. Specifically, the multi-dimensional data can include data collected from different angles, sources, and scenarios to assess the risk of the training samples.

[0041] For example, multi-dimensional data can include training sample consumption behavior, historical risk scores, social behavior data, resource data, etc., without specific limitations.

[0042] The training data of each training sample in the training sample dataset can be used to iteratively train the risk assessment model.

[0043] Multi-granularity risk labels are determined based on the performance data of each training sample during different relationship durations. Here, the relationship duration can refer to the time span from the date of resource allocation to the current observation date, usually measured in months.

[0044] For example, if the resources corresponding to training sample 1 were distributed on January 1, 2024, and the current observation date is July 15, 2024, then the duration of the relationship for training sample 1 is 6 months (denoted as MOB6).

[0045] Multi-granularity risk labels support a refined characterization of the risk status for different relationship durations.

[0046] For example, multi-granularity risk labels may include MOB1 7+, MOB3 15+, MOB6 30+, MOB6 60+, etc., without any restrictions.

[0047] Among them, "MOB1 7+ indicates that a risk event occurred within the first month of the relationship's duration and lasted for ≥7 days."

[0048] "MOB3 15+ means that during the first 3 months of the relationship, the duration of the risk event is ≥15 days."

[0049] "MOB6 30+" means that during the first 6 months of the relationship, the duration of the risk event is ≥30 days.

[0050] "MOB6 60+" means that during the first 6 months of the relationship, the duration of the risk event is ≥60 days.

[0051] Traditional risk assessment models simplify the risk level of entity samples into a binary label (low-risk entity / high-risk entity), failing to accurately reflect different risk gradients. For example, entity samples with "MOB (Month on Book) 6 90+" (meaning the risk status persists for more than 90 days within 6 months of the establishment of the relationship) have a higher default risk than entity samples with "MOB 6 30+" (meaning the risk status persists for more than 30 days within 6 months of the establishment of the relationship). This difference is not adequately expressed in traditional risk assessment models. Furthermore, traditional risk assessment models typically require the relationship duration of training samples to reach a specific threshold (e.g., 6 months) before inclusion in training, resulting in the exclusion of many samples with short relationship durations but already showing early risk signals. In some examples, the relationship duration is referred to as the aging of accounts receivable.

[0052] For example, if a sample that has only existed for 2 months has already exhibited overdue behavior in the MOB1 stage, although it does not meet the "MOB6 30+" requirement, it still contains valuable early warning information.

[0053] This embodiment constructs a multi-level, multi-granularity risk labeling system that can accurately reflect different levels of risk gradients. Furthermore, it allows short-term training samples to participate in risk assessment at the matching granularity based on the duration of the relationship (such as MOB1 7+), thereby effectively utilizing the training data of short-term training samples and significantly improving the information utilization rate of training samples and the risk assessment model's ability to identify early risks.

[0054] A hyperparameter set refers to a set of configuration parameters used to control the training process of a risk assessment model. The combination of these parameters directly affects the model's convergence, generalization ability, and final performance. This hyperparameter set can be determined through various hyperparameter optimization methods, including but not limited to grid search, random search, or Bayesian optimization.

[0055] The hyperparameter set in this step can refer to the hyperparameter set used for this round of training of the risk assessment model, determined by any of the methods mentioned above for finding the optimal hyperparameter combination. The hyperparameter set in this embodiment may include key hyperparameters used to determine the target label gain mapping method, that is, label gain mapping hyperparameters.

[0056] Step 202: Determine the risk level corresponding to the multi-granularity risk label of each training sample according to the preset risk label level conversion method.

[0057] It is important to understand that, since each training sample has a different relationship duration (such as MOB1, MOB3, MOB6, etc.), each training sample may be associated with multiple multi-granularity risk labels at the same time. For example, the same training sample may simultaneously satisfy MOB1 7+, MOB3 15+ and MOB6 30+, that is, it exhibits different levels of risk events under different relationship durations.

[0058] If these multi-granularity risk labels are used in parallel for training a risk assessment model, it may lead to problems such as label overlap and blurred risk gradients.

[0059] To address the aforementioned issues, this embodiment introduces a pre-defined risk label level conversion method, which maps multiple multi-granularity risk labels of each training sample into a single risk level, thereby constructing a clear risk monitoring signal.

[0060] Specifically, the risk label level conversion method can adopt the following strategies: Initialization: Set the initial risk level of all training samples to 0, indicating that no risk events have been observed, or that all training samples are in a risk-free state.

[0061] Level 1 determination: For all training samples, if the multi-granularity risk label of each training sample contains MOB1 7+, then the risk level of the corresponding training sample is set to 1.

[0062] Second-level judgment: For all training samples, if the multi-granularity risk label of each training sample contains MOB3 15+, the risk level of the corresponding training sample is set to 2.

[0063] Level 3 determination: For all training samples, if the multi-granularity risk label of each training sample contains MOB6 30+, the risk level of the corresponding training sample is set to 3.

[0064] Level 4 Judgment: For all training samples, if the multi-granularity risk label of each training sample contains MOB6 60+, the risk level of the corresponding training sample is set to 4.

[0065] Similarly, higher levels can be defined, such as MOB6 90+ corresponding to risk level 5, to support more granular risk stratification, without specific restrictions here.

[0066] This risk level not only reflects the severity of the risk event but also implies the timing and duration of its occurrence, providing a structured and comparable supervisory basis for subsequent model optimization based on ranking loss. Especially for training samples that show early signs of risk but have not yet entered the long-term observation window, this method can still assign them a reasonable risk level, thereby ensuring the effective participation of these training samples in the training process and improving the sensitivity and discrimination ability of the risk assessment model to early risks.

[0067] Step 203: Determine the target label gain mapping method based on the label gain mapping hyperparameters.

[0068] In some alternative implementations, the label gain mapping hyperparameters include the method type and method adjustment coefficients of the target label gain mapping method.

[0069] Among them, the method type includes at least one monotonically increasing label gain mapping method type.

[0070] For example, label gain mapping methods may include linear mapping functions g(y)=ky and power-law mapping functions g(y)=y m Where y is the risk level of the training sample, g(y) is the gain value corresponding to the risk level of the training sample, and k and m are the function adjustment coefficients of the gain mapping function, satisfying k > 0 and m > 0. Here, the function adjustment coefficients are an example of a method adjustment coefficient implementation.

[0071] It should be noted that the above-described label gain mapping method is merely an illustrative example. Those skilled in the art can employ other forms of label gain mapping methods, such as logarithmic functions, exponential functions, piecewise functions, etc., based on actual business needs and risk definitions, provided that the monotonically increasing constraint is satisfied. All monotonically increasing functions that can map discrete initial risk levels to non-negative gain values ​​for model training should be included within the scope of this disclosure and are not limited herein.

[0072] The method adjustment coefficient can be used to specify the parameters of the method type of the selected target label gain mapping method, thereby precisely controlling the rate and curvature of the gain growth of the target label gain mapping method.

[0073] The method adjustment coefficient directly affects the weighting of high-risk samples relative to low-risk samples in the ranking loss. For example, a larger method adjustment coefficient will significantly amplify the gain of high-risk training samples, thus giving them higher priority during training; while a smaller method adjustment coefficient will make the gain distribution more even, emphasizing overall ranking consistency rather than highlighting extreme risks.

[0074] When the label gain mapping hyperparameters include the method type and method adjustment coefficient of the target label gain mapping method, step 203 can be implemented as follows.

[0075] The target label gain mapping method is determined based on the method type and method adjustment coefficient.

[0076] Specifically, the method type specified in the hyperparameter set can be parsed, the corresponding label gain mapping method template can be loaded, and then the method adjustment coefficients can be substituted into the label gain mapping method template to generate a specific, computable mapping function instance. The mapping function instance is the target label gain mapping method used in this training, which is used to map the risk level of each sample to a risk correlation label.

[0077] For example, the label gain mapping hyperparameters may include a power-law mapping function, and the corresponding method adjustment coefficients may include k1 = 2 and k2 = 1, that is, y = x k1 +k2, then the determined target label gain mapping method is y=x 2 +1.

[0078] Here, the specific number of method adjustment coefficients can be greater than or equal to 1. For example, the number of method adjustment coefficients can also be 3 or more, and there is no restriction here.

[0079] This target label gain mapping method will be used in subsequent steps to convert the risk level of each training sample into a risk relevance label, providing a key supervisory signal for calculating the ranking loss, and thus guiding the risk assessment model to learn the relative risk order between samples more accurately.

[0080] In some alternative implementations, the label gain mapping hyperparameters include a target label gain mapping method identifier.

[0081] The target label gain mapping method identifier can be used to uniquely identify a predefined label gain mapping method.

[0082] When the label gain mapping hyperparameters include the target label gain mapping method identifier, step 203 can be implemented as follows.

[0083] Obtain the label gain mapping method set, and determine the target label gain mapping method based on the target label gain mapping method identifier and the label gain mapping method set.

[0084] The tag gain mapping method set includes at least one monotonically increasing tag gain mapping method and tag gain mapping method identifiers corresponding to each tag gain mapping method.

[0085] In this embodiment, each label gain mapping method can be used to map discrete risk levels (such as 1, 2, 3, 4) to a non-negative gain value, which is the risk correlation label of the subsequent training samples. The specific meaning of the risk correlation label will be explained later.

[0086] Label gain mapping methods can include, for example, linear mapping functions g(y) = ky and power-law mapping functions g(y) = y m Where y is the risk level of the training sample, g(y) is the gain value corresponding to the risk level of the training sample, and k and m are the method adjustment coefficients of the label gain mapping method, satisfying k > 0 and m > 0.

[0087] It should be noted that the above-described label gain mapping method is merely an illustrative example. Those skilled in the art can, based on actual business needs and risk definitions, employ other forms of mapping functions, such as logarithmic functions, exponential functions, piecewise functions, etc., provided that the monotonically increasing constraint is satisfied. All monotonically increasing functions that can map discrete initial risk levels to non-negative gain values ​​for model training should be included within the scope of this disclosure and are not restricted herein.

[0088] To facilitate management and invocation, each label gain mapping method is associated with a unique label gain mapping method identifier, which is used to quickly locate the target label gain mapping method.

[0089] Based on the obtained label gain mapping method set, the target label gain mapping method used for this model training is retrieved and determined from the label gain mapping method set according to the target label gain mapping method identifier specified in the hyperparameter set.

[0090] Specifically, the hyperparameter set configuration used in the current training round can be parsed, the target label gain mapping method identifier can be extracted, and then a matching search can be performed in the label gain mapping method set.

[0091] If the target label gain mapping method identifier is found, then the label gain mapping method corresponding to the target label gain mapping method identifier is determined as the target label gain mapping method.

[0092] The finalized target label gain mapping method will be used in subsequent steps to convert the risk level of each training sample into a risk-related label, providing a key supervisory signal for calculating the ranking loss, and thus guiding the risk assessment model to learn the relative risk order between samples more accurately.

[0093] Step 204: Determine the risk correlation label for each training sample based on the risk level of each training sample and the target label gain mapping method.

[0094] In this embodiment, after completing the risk level conversion (step 202) and the target label gain mapping method determination (step 203), this step is entered, which aims to transform discrete and ordered risk levels into continuous and quantifiable risk correlation labels.

[0095] Risk-related labels can be used to assess the risk severity level of training samples during the risk evolution process and serve as a key supervisory signal in subsequent ranking loss calculations.

[0096] In some alternative implementations, step 204 may include the following A1-A2.

[0097] A1. Determine the gain mapping value for each training sample based on the risk level of each training sample and the target label gain mapping method.

[0098] In this embodiment, for each training sample, the risk level of the training sample can be used as input, substituted into the target label gain mapping method, and the corresponding output value can be calculated, which is the gain mapping value of the training sample.

[0099] The higher the risk level of the training samples, the larger their corresponding gain mapping value. The gain mapping value is essentially a non-linear weighting or amplification of the risk level, which can reflect the relative difference or importance difference between different risk levels.

[0100] For example, the target label gain mapping method is g(y)=y m y is the risk level of the training sample (e.g., y is 1, 2, 3, 4), and m is used to control the rate at which the gain increases with the risk level (e.g., m is 2.5).

[0101] Correspondingly, the gain mapping values ​​for each risk level are 1, 5.66, 15.59 and 32.00, showing an accelerating growth trend, which significantly enhances the influence of high-risk samples in subsequent training.

[0102] A2, based on each gain mapping value, determine the risk correlation label for each training sample.

[0103] In this embodiment, the gain mapping value itself is the risk correlation label, and each gain mapping value can be directly determined as the risk correlation label for each training sample.

[0104] Here, risk-related labels can be used as true labels to drive the risk assessment model to learn the correct risk ranking structure.

[0105] It is important to note that in the subsequent calculation of ranking loss, risk-related labels (such as mapped true label values) participate in the calculation of first and second derivatives, thereby affecting the gradient distribution of the risk assessment model, making it more reasonable, and directly influencing the model training process (e.g., the core splitting process and decision tree construction method of the LightGBM model). This mechanism essentially exerts technical intervention on the learning optimization path of the risk assessment model, helping to improve the model's discriminative ability and ranking performance.

[0106] Step 205: Determine the ranking loss for each training sample based on the training data, risk correlation labels, and risk assessment model of each training sample.

[0107] In this embodiment, the aim is to calculate the ranking loss to guide model optimization based on the training data of each training sample, the constructed risk correlation labels, and the risk assessment model.

[0108] The core objective of this ranking loss is to ensure that the predicted risk correlation labels output by the risk assessment model can reflect the true relative order of risk among the training samples as accurately as possible.

[0109] In some alternative implementations, step 205 may include the following B1-B2.

[0110] B1 inputs the training data of each training sample into the risk assessment model, and outputs the predicted risk correlation label of each training sample through the risk assessment model.

[0111] In this step, the training data for each training sample can be input into the current risk assessment model, which could be a gradient boosting tree.

[0112] The risk assessment model can output a predicted risk relevance label for each training sample based on the training data of each training sample. This predicted risk relevance label is used to characterize the current risk assessment model's estimate of the risk level of each training sample. The higher the predicted risk relevance label (e.g., the predicted risk relevance score), the higher the risk assessment model judges the training sample to be.

[0113] In some alternative implementations, given the massive amount of sample data in the risk control field, typically ranging from tens of thousands to millions, if the entire sample is forcibly treated as a single group and ranking loss is used to train the risk assessment model, the computational complexity will increase exponentially. For example, the computational complexity of the pairwise ranking loss algorithm is O(n^2) * ... 2 When the sample size reaches millions, the computation time of the ranking loss algorithm becomes unacceptable.

[0114] Therefore, after step 204, the training sample dataset can be divided into X training sample data subsets according to preset sampling rules, where X is a positive integer greater than 1.

[0115] In other words, by converting a large dataset into multiple smaller datasets and then applying ranking loss to each smaller dataset (a subset of training sample data), it is possible to transfer the ranking loss to risk assessment models in the risk control field, thereby improving the risk assessment model's performance on the first few bins.

[0116] For example, we have a training sample dataset containing 10,000 records, and we plan to divide it into 100 training sample data subsets (i.e., X=100), each of which can include 100 training samples.

[0117] In the example above, if the ranking loss is directly applied to a training sample dataset of 10,000 records, the time complexity is enormous. However, after transforming it into 100 training sample data subsets, each subset contains only 100 training samples. Applying the ranking loss to each training sample data subset significantly reduces the time complexity. Thus, the ranking loss method can be transferred to risk assessment models in the risk control field.

[0118] In some alternative implementations, the training sample dataset can be stratified and sampled according to at least one preset stratification method.

[0119] For example, the training sample dataset can be stratified based on the proportion of positive and / or negative training samples in the training sample dataset. It can also be stratified based on the proportion of training samples in each month / quarter / year. Furthermore, it can be stratified based on the proportion of training samples from different data sources in the training sample dataset.

[0120] Based on this, B1 can also be implemented in the following way: for each subset of training sample data, the training data of each training sample in the subset of training sample data is input into the risk assessment model, and the risk assessment model outputs the predicted risk correlation label of each training sample in the subset of training sample data.

[0121] In this way, the ranking loss can be calculated independently within each subset of training sample data, thereby significantly reducing the scale and complexity of a single loss calculation.

[0122] B2, based on the predicted risk correlation label and risk correlation label of each training sample, determine the ranking loss of each training sample.

[0123] In this step, the risk relevance label can serve as the true label (true risk relevance score) for each training sample, reflecting the ideal position of the training sample in the risk ranking.

[0124] After obtaining the predicted risk relevance labels, the ranking loss of each training sample can be calculated using the ranking learning loss function based on the predicted risk relevance labels and the risk relevance labels.

[0125] The ranking learning loss function can be any of the known ranking loss algorithms, including but not limited to: Listwise Ranking Loss, XENDCG (eXtended Normalized Discounted Cumulative Gain), and LambdaRank.

[0126] Specifically, a suitable sorting loss algorithm can be selected according to actual needs, and no restrictions are imposed here.

[0127] Thus, by combining risk assessment model prediction, real risk correlation labels, and ranking loss function, a training mechanism with the relative order of risks as the core optimization objective is constructed, providing technical support for generating a target risk assessment model with high discriminative power and strong generalization.

[0128] Step 206: Iteratively train the risk assessment model based on each ranking loss until the preset model training stop condition is met, and generate the target risk assessment model.

[0129] In this embodiment, based on the ranking loss calculated in step 205, this step performs iterative optimization training on the risk assessment model. By repeatedly adjusting the model parameters to minimize the ranking loss, the ability of the risk assessment model to discriminate the relative order of risks among training samples is continuously improved. The training process continues until a pre-set model training stop condition is met, at which point the current state of the risk assessment model is determined as the final target risk assessment model.

[0130] Training stopping conditions can refer to one or more of the following combinations, including but not limited to: reaching the preset maximum number of iterations, triggering an early stopping mechanism, and the ranking performance index not improving for several consecutive iterations, such as normalized loss cumulative gain, KS value (Kolmogorov–Smirnov Statistic) not showing significant growth in the last N iterations.

[0131] Thus, this disclosure converts multi-granularity risk labels into corresponding risk levels, and then determines the risk relevance label for each training sample based on the risk level and target label gain mapping method. Using this risk relevance label as the true label for the training sample allows for a more refined and comprehensive characterization of the risk evolution process of the training sample at different lifecycle stages. This provides more discriminative supervisory information for the risk assessment model. Furthermore, by using ranking loss as the optimization objective, the method guides the risk assessment model to focus on predicting the relative risk order among training samples, enhancing the model's ability to distinguish between high and low risk rankings. In addition, this method can effectively incorporate training samples with short relationship durations but already showing early signs of risk, effectively improving the utilization efficiency of early risk signals and avoiding information loss caused by label flattening and crude censoring in traditional risk assessment models. The target risk assessment model can be subsequently applied to scenarios such as risk assessment of new samples and post-resource allocation monitoring, providing resource application institutions with high-precision, interpretable, and forward-looking risk identification capabilities.

[0132] In some alternative implementations, the training sample dataset includes an omnichannel training sample dataset and / or a target channel training sample dataset, and the risk assessment model includes a general risk assessment model and / or a customized risk assessment model.

[0133] The omnichannel training sample dataset includes training data and multi-granularity risk labels for each omnichannel training sample, while the target channel training sample dataset includes training data and multi-granularity risk labels or binary classification risk labels for each target channel training sample.

[0134] In this embodiment, a channel can refer to a business unit with an independent source of training samples.

[0135] A multi-channel training sample dataset can refer to a training sample dataset that includes training data from multiple different channels and corresponding multi-granularity risk labels for the training samples. A multi-channel training sample dataset can be used to train a general risk assessment model.

[0136] The risk levels of training samples covered by different channels often differ significantly.

[0137] For example, training samples from channel A have better creditworthiness and a relatively lower probability of risk events, while training samples from channel B have poorer creditworthiness and a relatively higher probability of risk events.

[0138] Due to the heterogeneity of training samples across different channels, modeling based solely on data from a single channel can easily lead to overfitting of the risk assessment model to a specific group, limiting its generalization ability. Therefore, this embodiment introduces a dataset of training samples from all channels, and by fusing training samples from multiple channels, constructs a general risk assessment model with stronger robustness and adaptability.

[0139] The target channel training sample dataset can refer to a training sample dataset that includes training data of training samples from the same target channel and corresponding multi-granularity risk labels or binary risk labels. The target channel training sample dataset can be used to train a customized risk assessment model.

[0140] Here, the target channel can refer to a specific channel for which a dedicated credit scoring model needs to be built. This target channel has relatively independent training samples and risk characteristics. In such scenarios, general risk assessment models may lose accuracy by ignoring channel specificity. Therefore, this embodiment can support training a customized risk assessment model based on the target channel's training sample dataset. This customized risk assessment model can be deeply adapted to the distribution of training samples and risk evolution patterns of the target channel, thereby achieving higher risk identification accuracy in local scenarios.

[0141] Thus, by distinguishing between the training sample datasets of all channels and the training sample datasets of the target channels, this embodiment achieves a combination of broad coverage and deep focus, which not only improves the overall robustness of the risk assessment model in a complex multi-channel environment, but also ensures the refined risk control capabilities under key target channels.

[0142] In some alternative implementations, the construction of a general risk assessment model for the target or a customized risk assessment model for the first target can be achieved through the following methods.

[0143] Specifically, the general risk assessment model or the first customized risk assessment model is iteratively trained based on the ranking loss of each training sample from all channels or the ranking loss of each training sample from the target channel until the preset general model stop training condition or the first customized model stop training condition is met, thereby generating the target general risk assessment model or the first target customized risk assessment model.

[0144] In some optional implementations, the method described in steps 201-205 can be used to calculate the corresponding ranking loss for each omnichannel training sample. Subsequently, the general risk assessment model is iteratively updated with the ranking loss of all omnichannel training samples as the optimization objective. The training process continues until a preset general model training stop condition is met. The general model training stop condition can refer to the target risk assessment model's training stop condition, which will not be elaborated here.

[0145] In some optional implementations, the method described in steps 201-205 can be used to calculate the corresponding ranking loss for each target channel training sample. Subsequently, the ranking loss of all target channel training samples is used as the optimization objective to iteratively update the first customized risk assessment model. The training process continues until the preset training stop condition of the first customized model is met. The training stop condition of the first customized model can refer to the training stop condition of the target risk assessment model, and will not be elaborated here.

[0146] Thus, through the aforementioned training mechanism, this embodiment achieves a combination of broad coverage and deep focus, enhancing the overall robustness of the risk assessment model in complex multi-channel environments while ensuring refined risk control capabilities for key target channels. Furthermore, the two types of risk assessment models can be trained in parallel without interference, improving development and deployment efficiency.

[0147] In some optional implementations, after generating the target general risk assessment model, to further improve the risk identification accuracy of the target channel, this disclosure also provides a customized modeling enhancement mechanism based on knowledge transfer from the general model. This mechanism injects the predictive capability of the general model as a feature into the training process of the customized risk assessment model for the target channel, achieving synergistic optimization of general capability reuse and channel characteristic fine-tuning.

[0148] Specifically, step 206 may also include steps 2061-2064.

[0149] Step 2061: Based on the training data of the training samples of each target channel and the target general risk assessment model, generate the predicted risk correlation label of the training samples of each target channel.

[0150] In this step, the original training data of each training sample in the target channel training sample dataset can be input into the trained target general risk assessment model. The target general risk assessment model outputs the predicted risk correlation label (e.g., predicted risk correlation score) of each target channel training sample. The predicted risk correlation label can reflect the target general risk assessment model's estimate of the risk level of the target channel training sample.

[0151] The predicted risk correlation label is the risk ranking judgment of the target channel sample by the target general risk assessment model. Although it is not optimized for the specificity of the target channel, it still contains risk signals common to cross channels and has high risk information value.

[0152] Step 2062: Merge each predicted risk correlation label with the training data of the corresponding target channel training sample to generate merged training data of each target channel training sample.

[0153] After obtaining the predicted risk correlation label for each target channel sample, this step adds the predicted risk correlation label as a new feature dimension, which can be concatenated with the original training data to form the merged training data of each target channel training sample.

[0154] Step 2063: Determine the ranking loss of the training samples for each target channel based on the merged training data, multi-granularity risk labels, and the second customized risk assessment model.

[0155] In this step, based on the merged training data described above, and combined with the multi-granularity risk labels of the target channel samples, the ranking loss can be calculated for an independent second customized risk assessment model. The specific calculation process of the ranking loss can be referred to steps 2051-2052, and will not be elaborated here.

[0156] Step 2064: Iteratively train the second customized risk assessment model based on the ranking loss of the training samples of each target channel until the preset training stop condition of the second customized model is met, and generate the second target customized risk assessment model.

[0157] In this step, the ranking loss calculated in step 2063 is used as the optimization objective to iteratively update the second customized risk assessment model. The training process continues until the preset stopping condition for the second customized model is met. The stopping condition for the second customized model can be referred to the stopping condition for the target risk assessment model, and will not be elaborated here.

[0158] Thus, through steps 2061-2064, not only is the performance ceiling of the second customized risk assessment model improved, but a scalable and reusable general and customized risk control model paradigm is also constructed, which is suitable for complex multi-channel businesses.

[0159] In some optional implementations, after training the target general risk assessment model, to further adapt to the target channel's needs for traditional modeling paradigms, this disclosure also provides a modeling path for a binary classification customized risk assessment model based on general model feature enhancement. This path is suitable for scenarios where the target channel only has binary risk labels (such as 0 and 1), while still effectively utilizing the cross-channel risk discrimination capability inherent in the general risk assessment model.

[0160] Specifically, step 206 may further include the following sub-steps 2061' to 2064'.

[0161] In some alternative implementations, after generating the target general risk assessment model, step 206 may also include steps 2061'-2064'.

[0162] Step 2061': Based on the training data of the training samples of each target channel and the target general risk assessment model, generate the predicted risk correlation label of each target channel training sample.

[0163] In this step, the original training data of each training sample in the target channel training sample dataset can be input into the already trained target general risk assessment model. The target general risk assessment model outputs the predicted risk correlation label of each target channel training sample. The predicted risk correlation label can reflect the target general risk assessment model's estimate of the risk level of the target channel training sample.

[0164] The predicted risk correlation label is the risk ranking judgment of the target channel sample by the target general risk assessment model. Although it is not optimized for the specificity of the target channel, it still contains risk signals common to cross channels and has high risk information value.

[0165] Step 2062': Merge each predicted risk correlation label with the training data of the corresponding target channel training sample to generate merged training data of each target channel training sample.

[0166] After obtaining the predicted risk correlation label for each target channel sample, this step adds the predicted risk correlation label as a new feature dimension, which can be concatenated with the original training data to form the merged training data of each target channel training sample.

[0167] Step 2063': Based on the training data after merging the training samples of each target channel, the binary risk labels, and the third customized risk assessment model, determine the cross-entropy loss of the training samples of each target channel.

[0168] Unlike step 2063, step 2063' can calculate the cross-entropy loss for an independent third-custom risk assessment model based on the above-mentioned merged training data and the binary risk labels of the target channel samples.

[0169] Step 2064': Iteratively train the third customized risk assessment model based on the cross-entropy loss of the training samples of each target channel until the preset training stop condition of the third customized model is met, and generate the third target customized risk assessment model.

[0170] In this step, the cross-entropy loss calculated in step 2063' is used as the optimization objective to iteratively update the third customized risk assessment model. The training process continues until the preset training stop condition for the third customized model is met. The training stop condition for the third customized model can be referred to the training stop condition for the target risk assessment model, and will not be elaborated here.

[0171] Thus, by using the ranking ability of the target general risk assessment model as a feature and combining it with traditional cross-entropy loss, the results of multi-granularity modeling can be effectively transferred to binary classification scenarios.

[0172] In some alternative implementations, after obtaining a target risk assessment model, the hyperparameter set of the binary classification model can be changed, and the binary classification model can be trained multiple times to obtain multiple target risk assessment models.

[0173] Specifically, it is determined whether the number of obtained target risk assessment models has reached a preset number R. In this embodiment, R is a positive integer, for example, R is 1000.

[0174] If not, update the hyperparameter set based on the performance evaluation results of the current target risk scoring model and the preset hyperparameter optimization method to obtain the updated hyperparameter set (the updated hyperparameter set is the hyperparameter set for the next round of risk assessment model training). The hyperparameter set before the update may include model training hyperparameters and label gain mapping hyperparameters. Similarly, the updated hyperparameter set may also include model training hyperparameters and label gain mapping hyperparameters.

[0175] In this embodiment, the performance of the current target risk assessment model (such as NDCG, KS value, etc.) can be used as a feedback signal to drive the hyperparameter optimization algorithm to generate a better or more exploratory set of hyperparameters. Here, jointly optimizing training hyperparameters (such as learning rate, number of leaf nodes, regularization term, etc.) and label gain mapping hyperparameters can shorten the overall number of rounds of hyperparameter set optimization, thereby shortening the training time.

[0176] Specifically, the traditional approach typically involves first fixing the label gain mapping method and conducting extensive searches (e.g., 1000 rounds of experiments) on the training hyperparameters to find the optimal configuration; then, a different label gain mapping method is used, and the entire process is repeated. If four different label gain mapping methods are preset, the total number of experiments can reach as high as 4000 rounds.

[0177] By employing a strategy of jointly optimizing training hyperparameters and label gain mapping methods, the synergistic effect between the two during training—for example, certain forms of label gain mapping methods perform significantly better at specific learning rates or regularization strengths—allows for efficient exploration of the combinatorial space by simultaneously adjusting training hyperparameters and selecting superior label gain mapping methods in each experiment. Experiments demonstrate that this method can approximate the global optimum in approximately 1200 trials, saving over 50% of training time.

[0178] The label gain mapping method included in the updated hyperparameter set may be the same as or different from the label gain mapping method included in the original hyperparameter set.

[0179] The method adjustment coefficients can also be resampled, and this flexibility allows for the automatic exploration of the risk gain scale that best suits the current data distribution.

[0180] After obtaining the updated hyperparameter set, the target risk assessment model for this round of training can be obtained again according to the methods in steps 201-206 above, until the Rth target risk assessment model is generated.

[0181] Thus, after obtaining multiple target risk assessment models, one can select the best-performing risk assessment model based on the performance indicators of each model, or select multiple high-performing risk assessment models for use. Alternatively, one can choose the appropriate risk assessment model for application based on the needs of the actual scenario.

[0182] Furthermore, it should be noted that after 1000 rounds of training, the evaluation results of 1000 target risk scoring models show that the best-performing label gain mapping method is the power-law mapping function, with an adjustment coefficient of 1.5, specifically in the form y=X. 1.5 .

[0183] The following examples illustrate the effectiveness of the general risk assessment model and the third-objective customized risk assessment model disclosed herein.

[0184] Example: Training dataset 1 (approximately 200,000 training samples) and training dataset 2 (approximately 300,000 samples) both contain multiple risk assessment labels (such as MOB3 30+, MOB3 15+, MOB1 7+, etc.) for jointly training a general risk assessment model for the objective. Training dataset 3 (approximately 50,000 samples) contains only one MOB3 15+ binary label, used to construct a third-objective customized risk assessment model, and the prediction results of the general risk assessment model for the objective are used as one of the input features.

[0185] Comparing the performance of different modeling methods on training samples from all channels in training datasets 1 and 2, it can be seen that the target general risk assessment model constructed using ranking learning (with the same GBDT decision tree algorithm for full parameter tuning) is superior to the traditional cross-entropy binary classification model construction method in terms of KS index.

[0186] Specifically, KS can be represented by the three key risk indicators MOB3_30, MOB3_15, and FPD7 as shown in Table 1.

[0187] Table 1: Three Key Risk Indicators: MOB3_30, MOB3_15, and FPD1_7 Among them, the target label gain mapping method determined by the target general risk assessment model based on ranking loss is y=x^k, k is 1.5, and the original 0, 1, 2, 3 are mapped to 0, 1, 2.8, 5.2.

[0188] It can be seen that the KS value of the target general risk assessment model based on ranking loss is significantly higher than that of the cross-entropy loss binary classification model 1 and the cross-entropy loss binary classification model 2.

[0189] It is important to note that during the calculation of ranking loss, the risk-related labels (i.e., the mapped true label values ​​of 0, 1, 2.8, and 5.2) participate in the calculation of the first and second derivatives, thus affecting the gradient distribution of the target general risk assessment model, making it more reasonable, and directly influencing the model training process (e.g., the core splitting process and decision tree construction method of the LightGBM model). This mechanism essentially exerts technical intervention on the learning optimization path of the target general risk assessment model, helping to improve the model's discriminative ability and ranking performance.

[0190] Furthermore, in the modeling of the third customized risk assessment model in training sample dataset 1, when the traditional cross-entropy binary classification method is adopted and the predicted output of the target general risk assessment model is used as the feature of the third customized risk assessment model, it is shown that the target general risk assessment model with ranking loss not only has a stronger risk discrimination ability, but can also more effectively improve the performance of the downstream third customized risk assessment model.

[0191] Specifically, KS's performance in the third customized risk assessment model of the target general risk assessment model can be shown in Table 2.

[0192] Table 2: Performance of KS in the Third Customized Risk Assessment Model of the Target General Risk Assessment Model Continue to refer to Figure 5 , Figure 5 A flowchart 500 shows an embodiment of the risk assessment method based on multi-granularity risk labeling and ranking loss according to this disclosure. Figure 5 The risk assessment method based on multi-granularity risk labeling and ranking loss shown can be applied to... Figure 1 The terminal device or server shown. This process 500 includes at least the following steps 501-502.

[0193] Step 501: Obtain multi-dimensional data of the target object.

[0194] In this embodiment, the target object can refer to the object whose credit risk is to be assessed.

[0195] The multidimensional data of the target object includes data used to assess the risk of the target object.

[0196] Step 502: Input multi-dimensional data into the target risk assessment model, and output the risk assessment results of the target object through the target risk assessment model.

[0197] In this embodiment, the target risk assessment model is trained through steps 201–206 of the aforementioned process 200.

[0198] Specifically, the target risk assessment model is obtained through the following training method: The risk relevance label of each training sample is determined based on the risk level and target label gain mapping method of each training sample in the training sample dataset. The ranking loss of each training sample is determined based on the training data, risk relevance labels, and risk assessment model. The risk assessment model is then iteratively trained using each ranking loss.

[0199] Here, the risk score is usually a continuous value of 300-1000 points. The lower the score, the higher the probability of a risk event occurring (or vice versa depending on the business definition).

[0200] Further reference Figure 6 As an implementation of the methods shown in the above figures, this disclosure provides an embodiment of a model training device based on multi-granularity risk labels and ranking loss. This device embodiment is similar to... Figure 2 Corresponding to the method embodiments shown, this device can be specifically applied to various terminal devices.

[0201] like Figure 6As shown, this embodiment of the model training device based on multi-granularity risk labels and ranking loss includes a training data acquisition unit 601, a risk label level conversion unit 602, a label gain mapping method determination unit 603, a risk correlation label determination unit 604, a ranking loss determination unit 605, and a risk assessment model generation unit 606. The training data acquisition unit 601 acquires a training sample dataset, a risk assessment model, and a hyperparameter set for the risk assessment model. The training sample dataset includes training data and multi-granularity risk labels for each training sample. The multi-granularity risk labels are determined based on the performance data of the training samples during different relationship durations. The hyperparameter set includes label gain mapping hyperparameters. The risk label level conversion unit 602 determines the risk level corresponding to the multi-granularity risk label of each training sample according to a preset risk label level conversion method. The label gain mapping method determination unit 603 determines a target label gain mapping method based on the label gain mapping hyperparameters. The risk correlation label determination unit 604 determines the risk correlation label of each training sample based on its risk level and the target label gain mapping method. The ranking loss determination unit 605 is used to determine the ranking loss of each training sample based on the training data, risk correlation labels, and risk assessment model of each training sample. The risk assessment model generation unit 606 is used to iteratively train the risk assessment model based on each ranking loss until a preset model stopping condition is met, thereby generating the target risk assessment model.

[0202] In this embodiment, the specific processing of the training data acquisition unit 601, the risk label level conversion unit 602, the label gain mapping method determination unit 603, the risk correlation label determination unit 604, the ranking loss determination unit 605, and the risk assessment model generation unit 606, and the resulting technical effects, can be found in reference to [the relevant documentation]. Figure 2 The relevant descriptions of steps 201 to 206 in the corresponding embodiments will not be repeated here.

[0203] In some optional implementations, the tag gain mapping hyperparameters include a target tag gain mapping method identifier, and the tag gain mapping method determination unit 603 can be further used for: Obtain a set of label gain mapping methods, wherein the set of label gain mapping methods includes at least one monotonically increasing label gain mapping method and a label gain mapping method identifier corresponding to each label gain mapping method; The target label gain mapping method is determined based on the target label gain mapping method identifier and the label gain mapping method set.

[0204] In some optional implementations, the tag gain mapping hyperparameters include the method type and method adjustment coefficient of the target tag gain mapping method, wherein the method type includes at least one monotonically increasing tag gain mapping method type, and the tag gain mapping method determination unit 603 can be further used for: The target label gain mapping method is determined based on the method type and method adjustment coefficient.

[0205] In some optional implementations, the risk-related label determination unit 604 may be further used for: Based on the risk level of each training sample and the target label gain mapping method, determine the gain mapping value of each training sample; Based on each gain mapping value, determine the risk relevance label for each training sample.

[0206] In some optional implementations, the training sample dataset includes an omnichannel training sample dataset and / or a target channel training sample dataset, and the risk assessment model includes a general risk assessment model and / or a customized risk assessment model. The omnichannel training sample dataset includes training data and multi-granularity risk labels for each omnichannel training sample, and the target channel training sample dataset includes training data and multi-granularity risk labels or binary risk labels for each target channel training sample.

[0207] In some alternative implementations, the risk assessment model generation unit 606 may be further used for: The general risk assessment model or the first customized risk assessment model is iteratively trained based on the ranking loss of training samples from each omnichannel or each target channel until the preset training stop conditions for the general model or the first customized model are met, thereby generating the target general risk assessment model or the first target customized risk assessment model.

[0208] In some alternative implementations, the risk assessment model generation unit 606 may be further used for: Based on the training data of the training samples of each target channel and the target general risk assessment model, generate the predicted risk correlation label of the training samples of each target channel; The training data of each predicted risk correlation label and the corresponding target channel training sample are merged to generate the merged training data of each target channel training sample. Based on the training data after merging the training samples of each target channel, multi-granularity risk labels, and the second customized risk assessment model, the ranking loss of the training samples of each target channel is determined. The second customized risk assessment model is iteratively trained based on the ranking loss of the training samples of each target channel until the preset training stop condition of the second customized model is met, thus generating the second target customized risk assessment model.

[0209] In some alternative implementations, the risk assessment model generation unit 606 may be further used for: Based on the training data of the training samples of each target channel and the target general risk assessment model, generate the predicted risk correlation label of the training samples of each target channel; The training data of each predicted risk correlation label and the corresponding target channel training sample are merged to generate the merged training data of each target channel training sample. Based on the training data after merging the training samples of each target channel, the binary risk labels, and the third customized risk assessment model, the cross-entropy loss of the training samples of each target channel is determined. The third customized risk assessment model is iteratively trained based on the cross-entropy loss of the training samples of each target channel until the preset training stop condition of the third customized model is met, thus generating the third target customized risk assessment model.

[0210] In some optional implementations, the sorting loss determination unit 605 may be further used to: The training data of each training sample is input into the risk assessment model, and the risk assessment model outputs the predicted risk correlation label of each training sample. The ranking loss for each training sample is determined based on the predicted risk correlation label and the risk correlation label for each training sample.

[0211] In some alternative embodiments, device 600 further includes a dataset partitioning unit 607. Figure 6 (not shown), used to divide the training sample dataset into X training sample data subsets according to a preset sampling rule, where X is a positive integer greater than 1; Risk-related label determination unit 604 can be further used for: For each subset of training sample data, the training data of each training sample in the subset is input into the risk assessment model, and the risk assessment model outputs the predicted risk correlation label of each training sample in the subset.

[0212] It should be noted that the implementation details and technical effects of each unit in the model risk assessment device based on multi-granularity risk labels and ranking loss provided in the embodiments of this disclosure can be referred to the descriptions of other embodiments in this disclosure, and will not be repeated here.

[0213] Further reference Figure 7 As an implementation of the methods shown in the above figures, this disclosure provides an embodiment of a model risk assessment device based on multi-granularity risk labels and ranking loss. This device embodiment is similar to... Figure 5Corresponding to the method embodiments shown, this device can be specifically applied to various terminal devices.

[0214] like Figure 7 As shown, the model risk assessment device based on multi-granularity risk labels and ranking loss in this embodiment includes: Device 700. The system includes a multi-dimensional data acquisition unit 701 and a risk assessment result output unit 702. The multi-dimensional data acquisition unit 701 acquires multi-dimensional data of the target object, including data used to assess the risk of the target object. The risk assessment result output unit 702 inputs the multi-dimensional data into a target risk assessment model and outputs the risk assessment result of the target object. The target risk assessment model is obtained through the following training method: determining the risk correlation label of each training sample based on the risk level and target label gain mapping method of each training sample in the training sample dataset; determining the ranking loss of each training sample based on the training data, risk correlation labels, and risk assessment model; and iteratively training the risk assessment model based on the ranking losses.

[0215] In this embodiment, the specific processing of the multi-dimensional data acquisition unit 701 and the risk assessment result output unit 702, and the resulting technical effects, can be found in reference to [the relevant documentation]. Figure 5 The relevant descriptions of steps 501 to 502 in the corresponding embodiments will not be repeated here.

[0216] It should be noted that the implementation details and technical effects of each unit in the model risk assessment device based on multi-granularity risk labels and ranking loss provided in the embodiments of this disclosure can be referred to the descriptions of other embodiments in this disclosure, and will not be repeated here.

[0217] The following is for reference. Figure 8 It shows a schematic diagram of the structure of a computer system 800 suitable for implementing the terminal device of this disclosure. Figure 8 The computer system 800 shown is merely an example and should not be construed as limiting the functionality and scope of the embodiments of this disclosure.

[0218] like Figure 8As shown, the computer system 800 may include a processing device (e.g., a central processing unit, a graphics processing unit, etc.) 801, which can perform various appropriate actions and processes according to a program stored in a read-only memory (ROM) 802 or a program loaded from a storage device 808 into a random access memory (RAM) 803. The RAM 803 also stores various programs and data required for the operation of the computer system 800. The processing device 801, ROM 802, and RAM 803 are interconnected via a bus 804. An input / output (I / O) interface 805 is also connected to the bus 804.

[0219] Typically, the following devices can be connected to I / O interface 805: input devices 806 including, for example, touchscreens, touchpads, keyboards, mice, cameras, microphones, etc.; output devices 807 including, for example, liquid crystal displays (LCDs), speakers, vibrators, etc.; storage devices 808 including, for example, magnetic tapes, hard disks, etc.; and communication devices 809. Communication device 809 allows computer system 800 to communicate wirelessly or wiredly with other devices to exchange data. Although Figure 8 A computer system 800 with various electronic devices is shown; however, it should be understood that it is not required to implement or possess all of the devices shown. More or fewer devices may be implemented or possessed alternatively.

[0220] In particular, according to embodiments of this disclosure, the processes described above with reference to the flowcharts can be implemented as computer software programs. For example, embodiments of this disclosure include a computer program product comprising a computer program carried on a computer-readable medium, the computer program containing program code for performing the methods shown in the flowcharts. In such embodiments, the computer program can be downloaded and installed from a network via a communication device 809, or installed from a storage device 808, or installed from a ROM 802. When the computer program is executed by a processing device 801, it performs the functions defined in the methods of embodiments of this disclosure.

[0221] It should be noted that the computer-readable medium described in this disclosure can be a computer-readable signal medium or a computer-readable storage medium, or any combination thereof. A computer-readable storage medium can be, for example,—but not limited to—an electrical, magnetic, optical, electromagnetic, infrared, or semiconductor system, apparatus, or device, or any combination thereof. More specific examples of a computer-readable storage medium may include, but are not limited to: an electrical connection having one or more wires, a portable computer disk, a hard disk, random access memory (RAM), read-only memory (ROM), erasable programmable read-only memory (EPROM or flash memory), optical fiber, portable compact disk read-only memory (CD-ROM), optical storage device, magnetic storage device, or any suitable combination thereof. In this disclosure, a computer-readable storage medium can be any tangible medium containing or storing a program that can be used by or in conjunction with an instruction execution system, apparatus, or device. In this disclosure, a computer-readable signal medium can include a data signal propagated in baseband or as part of a carrier wave, carrying computer-readable program code. Such propagated data signals can take various forms, including but not limited to electromagnetic signals, optical signals, or any suitable combination thereof. A computer-readable signal medium can be any computer-readable medium other than a computer-readable storage medium, which can send, propagate, or transmit a program for use by or in connection with an instruction execution system, apparatus, or device. The program code contained on the computer-readable medium can be transmitted using any suitable medium, including but not limited to: wires, optical fibers, RF (radio frequency), etc., or any suitable combination thereof.

[0222] The aforementioned computer-readable medium may be included in the aforementioned electronic device; or it may exist independently and not assembled into the electronic device.

[0223] The aforementioned computer-readable medium carries one or more programs, which, when executed by the electronic device, cause the electronic device to perform the following functions: Figure 2 or Figure 5 The embodiments shown and their alternative implementations illustrate a model training and risk assessment method based on multi-granularity risk labels and ranking loss.

[0224] Computer program code for performing the operations of this disclosure can be written in one or more programming languages ​​or a combination thereof, including object-oriented programming languages ​​such as Java, C++, and Python, and conventional procedural programming languages ​​such as the "C" language or similar programming languages. The program code can be executed entirely on the target computer, partially on the target computer, as a standalone software package, partially on the target computer and partially on a remote computer, or entirely on a remote computer or server. In cases involving remote computers, the remote computer can be connected to the target computer via any type of network—including a local area network (LAN) or a wide area network (WAN)—or can be connected to an external computer (e.g., via the Internet using an Internet service provider).

[0225] The flowcharts and block diagrams in the accompanying drawings illustrate the architecture, functionality, and operation of possible implementations of systems, methods, and computer program products according to various embodiments of this disclosure. In this regard, each block in a flowchart or block diagram may represent a module, segment, or portion of code containing one or more executable instructions for implementing a specified logical function. It should also be noted that in some alternative implementations, the functions indicated in the blocks may occur in a different order than those indicated in the drawings. For example, two consecutively indicated blocks may actually be executed substantially in parallel, and they may sometimes be executed in reverse order, depending on the functions involved. It should also be noted that each block in the block diagrams and / or flowcharts, and combinations of blocks in the block diagrams and / or flowcharts, can be implemented using a dedicated hardware-based system that performs the specified function or operation, or using a combination of dedicated hardware and computer instructions.

[0226] The units described in the embodiments of this disclosure can be implemented in software or in hardware. The name of a unit does not necessarily limit the unit itself; for example, a multi-dimensional data acquisition unit can also be described as "a unit for acquiring multi-dimensional data".

[0227] The above description is merely a preferred embodiment of this disclosure and an explanation of the technical principles employed. Those skilled in the art should understand that the scope of this disclosure is not limited to technical solutions formed by specific combinations of the above-described technical features, but should also cover other technical solutions formed by arbitrary combinations of the above-described technical features or their equivalents without departing from the above-described concept. For example, technical solutions formed by substituting the above features with (but not limited to) technical features disclosed in this disclosure that have similar functions.

Claims

1. A model training method based on multi-granularity risk labels and ranking loss, characterized in that, The method includes: Obtain a training sample dataset, a risk assessment model, and a hyperparameter set for the risk assessment model. The training sample dataset includes training data for each training sample and multi-granularity risk labels. The multi-granularity risk labels are determined based on the performance data of the training samples during different relationship durations. The hyperparameter set includes label gain mapping hyperparameters. Based on the preset risk label level conversion method, the risk level corresponding to the multi-granularity risk label of each training sample is determined; The target label gain mapping method is determined based on the label gain mapping hyperparameters. Based on the risk level of each training sample and the target label gain mapping method, determine the risk correlation label of each training sample; The ranking loss for each training sample is determined based on the training data, the risk correlation labels, and the risk assessment model for each training sample. The risk assessment model is iteratively trained according to the ranking loss until the preset model training stop condition is met, thereby generating the target risk assessment model.

2. The method according to claim 1, characterized in that, The label gain mapping hyperparameters include a target label gain mapping method identifier, and determining the target label gain mapping method based on the label gain mapping hyperparameters includes: Obtain a set of label gain mapping methods, wherein the set of label gain mapping methods includes at least one monotonically increasing label gain mapping method and a label gain mapping method identifier corresponding to each of the label gain mapping methods; The target label gain mapping method is determined based on the target label gain mapping method identifier and the label gain mapping method set.

3. The method according to claim 1, characterized in that, The label gain mapping hyperparameters include the method type and method adjustment coefficient of the target label gain mapping method, wherein the method type includes at least one monotonically increasing label gain mapping method type, and determining the target label gain mapping method based on the label gain mapping hyperparameters includes: The target label gain mapping method is determined based on the method type and the method adjustment coefficient.

4. The method according to claim 1, characterized in that, The step of determining the risk-related label of each training sample based on the risk level of each training sample and the target label gain mapping method includes: Based on the risk level of each training sample and the target label gain mapping method, determine the gain mapping value of each training sample; Based on the gain mapping values, the risk correlation labels of each training sample are determined.

5. The method according to claim 1, characterized in that, The training sample dataset includes an omnichannel training sample dataset and / or a target channel training sample dataset. The risk assessment model includes a general risk assessment model and / or a customized risk assessment model. The omnichannel training sample dataset includes training data and multi-granularity risk labels for each omnichannel training sample. The target channel training sample dataset includes training data and multi-granularity risk labels or binary risk labels for each target channel training sample.

6. The method according to claim 5, characterized in that, The step of iteratively training the risk assessment model based on each of the ranking losses until a preset model training stop condition is met, thereby generating a target risk assessment model, includes: The general risk assessment model or the first customized risk assessment model is iteratively trained based on the ranking loss of each of the omnichannel training samples or the ranking loss of each of the target channel training samples until the preset general model stop training condition or the first customized model stop training condition is met, thereby generating the target general risk assessment model or the first target customized risk assessment model.

7. The method according to claim 6, characterized in that, After generating the target general risk assessment model, the method further includes: Based on the training data of each target channel training sample and the target general risk assessment model, generate a predicted risk correlation label for each target channel training sample; The predicted risk correlation labels are merged with the training data of the corresponding target channel training samples to generate merged training data of each target channel training sample. Based on the merged training data of each target channel training sample, the multi-granularity risk label, and the second customized risk assessment model, determine the ranking loss of each target channel training sample; The second customized risk assessment model is iteratively trained based on the ranking loss of the training samples of each target channel until the preset training stop condition of the second customized model is met, thereby generating the second target customized risk assessment model.

8. The method according to claim 6, characterized in that, After generating the target general risk assessment model, the method further includes: Based on the training data of each target channel training sample and the target general risk assessment model, generate a predicted risk correlation label for each target channel training sample; The predicted risk correlation labels are merged with the training data of the corresponding target channel training samples to generate merged training data of each target channel training sample. Based on the merged training data of each target channel training sample, the binary risk label, and the third customized risk assessment model, determine the cross-entropy loss of each target channel training sample; The third customized risk assessment model is iteratively trained based on the cross-entropy loss of the training samples of each target channel until the preset training stop condition of the third customized model is met, thereby generating the third target customized risk assessment model.

9. The method according to claim 1, characterized in that, The step of determining the ranking loss for each training sample based on the training data, the risk correlation label, and the risk assessment model includes: The training data of each training sample is input into the risk assessment model, and the risk assessment model outputs the predicted risk correlation label of each training sample. The ranking loss of each training sample is determined based on the predicted risk correlation label and the risk correlation label of each training sample.

10. The method according to claim 9, characterized in that, After determining the risk-related label of each training sample based on the risk level of each training sample and the target label gain mapping method, the method further includes: The training sample dataset is divided into X training sample data subsets according to a preset sampling rule, where X is a positive integer greater than 1. The step of inputting the training data of each training sample into the risk assessment model, and outputting the predicted risk relevance label of each training sample through the risk assessment model, includes: For each subset of training sample data, the training data of each training sample in the subset of training sample data is input into the risk assessment model, and the risk assessment model outputs the predicted risk correlation label of each training sample in the subset of training sample data.

11. A risk assessment method based on multi-granularity risk labeling and ranking loss, characterized in that, The method includes: Obtain multi-dimensional data of the target object, wherein the multi-dimensional data of the target object includes data used to assess the risk of the target object; The multi-dimensional data is input into the target risk assessment model, and the target risk assessment model outputs the risk assessment result of the target object. The target risk assessment model is obtained through the following training method: determining the risk relevance label of each training sample based on the risk level and target label gain mapping method of each training sample in the training sample dataset; determining the ranking loss of each training sample based on the training data, the risk relevance label and the risk assessment model; and iteratively training the risk assessment model based on the ranking loss.

12. A model training device based on multi-granularity risk labels and ranking loss, characterized in that, The device includes: The training data acquisition unit is used to acquire the training sample dataset, the risk assessment model, and the hyperparameter set of the risk assessment model. The training sample dataset includes the training data of each training sample and multi-granularity risk labels. The multi-granularity risk labels are determined based on the performance data of the training samples during different relationship durations. The hyperparameter set includes label gain mapping hyperparameters. The risk label level conversion unit is used to determine the risk level corresponding to the multi-granularity risk label of each training sample according to the preset risk label level conversion method. The tag gain mapping method determination unit is used to determine the target tag gain mapping method based on the tag gain mapping hyperparameters. The risk-related label determination unit is used to determine the risk-related label of each training sample based on the risk level of each training sample and the target label gain mapping method. The ranking loss determination unit is used to determine the ranking loss of each training sample based on the training data of each training sample, the risk correlation label and the risk assessment model. The risk assessment model generation unit is used to iteratively train the risk assessment model according to each ranking loss until the preset model stopping training condition is met, thereby generating the target risk assessment model.

13. A model risk assessment device based on multi-granularity risk labeling and ranking loss, characterized in that, The device includes: A multi-dimensional data acquisition unit is used to acquire multi-dimensional data of a target object, wherein the multi-dimensional data of the target object includes data used to assess the risk of the target object; A risk assessment result output unit is used to input the multi-dimensional data into a target risk assessment model and output the risk assessment result of the target object through the target risk assessment model. The target risk assessment model is obtained through the following training method: determining the risk correlation label of each training sample based on the risk level and target label gain mapping method of each training sample in the training sample dataset; determining the ranking loss of each training sample based on the training data, the risk correlation label, and the risk assessment model; and iteratively training the risk assessment model based on the ranking loss.

14. An electronic device, characterized in that, include: One or more processors; Storage device, on which one or more programs are stored, When one or more programs are executed by one or more processors, the one or more processors implement the method as described in any one of claims 1-10 or 11.

15. A computer-readable storage medium, characterized in that, It stores a computer program thereon, wherein the computer program, when executed by one or more processors, implements the method as described in any one of claims 1-10 or 11.

16. A computer program product, characterized in that, Includes a computer program / instruction that, when executed by a processor, implements the method as described in any one of claims 1-10 or 11.

Images