Risk control model updating method and related device

Abstract

The embodiment of the invention discloses a risk control model updating method and a related device. The method comprises the following steps: acquiring a first model and newly added business data; inputting the newly added business data into the first model to obtain a first soft label output by the first model; newly establishing an initial risk control model and a target loss function, and training the initial risk control model through the target loss function to obtain a second model; and updating the first model into the second model. Therefore, the data processing pressure is relieved, meanwhile, the disastrous forgetting defect existing in the model updating process is effectively restrained, and the model updating efficiency is improved.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of model optimization, and particularly relates to a risk control model updating method and related device. BACKGROUND

[0002] In the financial market, in business scenarios such as enterprise credit investigation and credit approval, an accurate and real-time risk assessment model is the core competitiveness. These risk control models rely on learning a large amount of historical business data to identify risks. However, the financial market environment, industry policies and enterprise operating conditions are continuously dynamic, and new data is continuously generated, which puts higher requirements on the continuous learning and adaptive updating capabilities of the risk control model to maintain its prediction accuracy.

[0003] At present, the existing risk control model updating method is to update the node, merge new data with all past historical data to form a data set containing all known information, and then train a new risk control model from scratch based on the full data set. After training, the old model running online is replaced. Each update needs to process a large amount of full data, causing huge consumption of computing resources, and at the same time, it takes a long training time, resulting in a long model update cycle. SUMMARY

[0004] Therefore, the embodiments of the present application provide a risk control model updating method and related device to alleviate the data processing pressure by fusing the incremental learning means of knowledge distillation and parameter regularization, and effectively suppress the catastrophic forgetting defect existing in the model updating process, and improve the efficiency of model updating.

[0005] In a first aspect, the embodiments of the present application provide a risk control model updating method, which comprises:

[0006] obtaining a first model and new business data, wherein the first model refers to a deployed risk control model;

[0007] inputting the new business data into the first model to obtain first soft labels output by the first model;

[0008] creating an initial risk control model and a target loss function, training the initial risk control model through the target loss function to obtain a second model, and the target loss function contains the first soft labels;

[0009] updating the first model to the second model.

[0010] In a possible embodiment, the training of the initial risk control model according to the target loss function comprises: determining a loss result of the initial risk control model according to the new business data and the target loss function; and minimizing the loss result according to an optimization algorithm to train the initial risk control model.

[0011] In a possible embodiment, the determining of the loss result of the initial risk control model according to the new business data and the target loss function comprises: inputting the new business data into the initial risk control model, and determining a first sub-loss result, a second sub-loss result, and a third sub-loss result according to the target loss function; configuring a weight parameter for each of the first sub-loss result, the second sub-loss result, and the third sub-loss result, the weight parameter being used to reflect an influence degree of a sub-loss result on the loss result; and determining the loss result according to the first sub-loss result, the second sub-loss result, the third sub-loss result, and the corresponding weight parameters.

[0012] In a possible embodiment, the determining of the first sub-loss result according to the target loss function comprises: determining a predicted label of the new business data output by the initial risk control model, and a real label of the new business data; and calculating a difference between the predicted label and the real label as the first sub-loss result by using a cross-entropy loss, the first sub-loss result being used to represent a prediction error of the second model for the new business data.

[0013] In a possible embodiment, the determining of the second sub-loss result according to the target loss function comprises: calculating a second soft label of the new business data output by the initial risk control model; and determining a difference between the first soft label and the second soft label as the second sub-loss result.

[0014] In a possible embodiment, the determining of the third sub-loss result according to the target loss function comprises: determining a core task of the first model for the historical business data, and a core parameter in the first model, the core parameter being a parameter that has an influence on the core task; calculating an importance degree of the core task in the first model; and determining the third sub-loss result according to the importance degree of the core task, a preset penalty coefficient, and the core parameter by using a learning algorithm, the third sub-loss result being used to constrain a variation degree of the core parameter.

[0015] In a possible embodiment, the learning algorithm is an elastic weight consolidation algorithm.

[0016] In a possible embodiment, the method further includes: dynamically adjusting the weight parameter according to an increase in the number of times of training of the initial risk control model.

[0017] In a second aspect, an embodiment of the present application provides an updating device of a risk control model, the device comprising: an obtaining unit, an output unit, a training unit, and an updating unit; the obtaining unit is configured to obtain a first model and new business data, the first model being a deployed risk control model; the output unit is configured to input the new business data into the first model to obtain a first soft label output by the first model; the training unit is configured to newly create an initial risk control model and a target loss function, train the initial risk control model through the target loss function, and obtain a second model, the target loss function containing the first soft label; and the updating unit is configured to update the first model to the second model.

[0018] In a third aspect, an embodiment of the present application provides an electronic device, comprising a processor, a memory, a communication interface, and one or more programs, wherein the one or more programs are stored in the memory and configured to be executed by the processor, and the programs comprise instructions for performing steps in any method of the first aspect of the embodiments of the present application.

[0019] In a fourth aspect, an embodiment of the present application provides a computer readable storage medium, wherein the computer readable storage medium stores a computer program for electronic data exchange, and the computer program causes a computer to perform some or all of the steps described in any method of the first aspect of the embodiments of the present application.

[0020] In a fifth aspect, an embodiment of the present application provides a computer program product, comprising a non-transitory computer readable storage medium storing a computer program, the computer program operable to cause a computer to perform some or all of the steps described in any method of the first aspect of the embodiments of the present application. The computer program product can be a software installation package.

[0021] It can be seen that, by the updating method of the risk control model and the related device provided in the present application, first, a first model and new business data are obtained; then the new business data is input into the first model to obtain a first soft label output by the first model; then, an initial risk control model and a target loss function are newly created, the initial risk control model is trained through the target loss function, and a second model is obtained; finally, the first model is updated to the second model. In this way, by fusing the incremental learning means of knowledge distillation and parameter regularization, the data processing pressure is relieved, and the efficiency of model updating is improved while effectively inhibiting the catastrophic forgetting defect existing in the model updating process. BRIEF DESCRIPTION OF DRAWINGS

[0022] In order to more clearly illustrate the technical solutions in the embodiments of the present application or the prior art, the accompanying drawings needed to be used in the description of the embodiments or the prior art will be briefly introduced. Obviously, the accompanying drawings in the following description only represent some of the embodiments of the present application, and for those skilled in the art, other drawings can also be obtained from these drawings without any creative effort.

[0023] Figure 1 is a schematic diagram of a risk control model training architecture provided by an embodiment of the present application;

[0024] Figure 2 is a flowchart of a risk control model updating method provided by an embodiment of the present application;

[0025] Figure 3 is a flowchart of a risk control model training method provided by an embodiment of the present application;

[0026] Figure 4 is a flowchart of another risk control model updating method provided by an embodiment of the present application;

[0027] Figure 5 is a functional unit composition block diagram of a risk control model updating device provided by an embodiment of the present application;

[0028] Figure 6 is a structural block diagram of an electronic device provided by an embodiment of the present application. DETAILED DESCRIPTION

[0029] In order to make the person skilled in the art better understand the present application, the technical solutions in the embodiments of the present application will be described clearly and completely below with reference to the accompanying drawings in the embodiments of the present application. Obviously, the described embodiments are only some of the embodiments of the present application, not all the embodiments. Based on the embodiments in the present application, all other embodiments obtained by those skilled in the art without any creative effort fall within the scope of the present application.

[0030] The terms "first", "second", etc. in the specification and claims of the present application and the above-mentioned accompanying drawings are used to distinguish different objects, not to describe a specific order. In addition, the terms "include" and "have" and any variations thereof are intended to cover non-exclusive inclusion. For example, a process, method, system, product or device including a series of steps or units is not limited to the listed steps or units, but can optionally include steps or units not listed or can optionally include other steps or units inherent to the process, method, product or device.

[0031] It should be understood that the term "and / or" in this document merely describes an associated relationship between associated objects, and means that three relationships can exist, for example, A and / or B can mean that A exists alone, A and B exist together, and B exists alone. In addition, the character " / " in this document means that the front and rear associated objects are in an "or" relationship.

[0032] The "multiple" appearing in the embodiments of the present application means two or more. The "connection" appearing in the embodiments of the present application means direct connection or indirect connection and various connection modes to realize communication between devices, which is not limited by the embodiments of the present application.

[0033] Reference to "embodiments" in this document means that the specific features, structures or characteristics described in conjunction with the embodiments can be included in at least one embodiment of the present application. The phrase appears at various places in the specification does not necessarily mean the same embodiment, nor is it an independent or alternative embodiment to other embodiments. The skilled person explicitly and implicitly understands that the embodiments described herein can be combined with other embodiments.

[0034] The related content, concepts, meanings, technical problems, technical solutions, beneficial effects and the like involved in the embodiments of the present application are described below.

[0035] EWC (Elastic Weight Consolidation): It is an algorithm for preventing deep learning models from suffering catastrophic forgetting when learning new tasks continuously. EWC adds a regularization term on weights to make key weights change less, keeping in the low error area of old tasks, thereby realizing continuous learning.

[0036] Synaptic Intelligence (SI): It is a continuous learning algorithm inspired by the synaptic mechanism of biological neural networks, the core of which is to assign a dynamic importance measure to each synapse, and to punish the modification of key synapses for old tasks to retain historical business data when learning new tasks.

[0037] KL divergence (Kullback-Leibler Divergence): It is an index in information theory for quantifying the difference between two probability distributions, which measures the information entropy lost when one probability distribution is used to approximate another probability distribution.

[0038] In the financial market, in the business scenarios such as enterprise credit investigation and credit approval, an accurate and real-time risk assessment model is the core competitiveness. These risk control models rely on learning a large amount of historical business data to identify risks. However, the financial market environment, industry policy and enterprise operating conditions are continuously and dynamically changing, and new data is continuously generated, which puts higher requirements on the continuous learning and adaptive updating capability of the risk control model to maintain the accuracy of its prediction. At present, the existing risk control model updating method is to update the node, merge the new data with all the historical data in the past to form a data set containing all known information, and then train a new risk control model from scratch based on the full data set, and after the training is completed, the old model running online is replaced. Each update needs to process a large amount of full data, causing huge consumption of computing resources, and at the same time, it needs to consume a long training time, resulting in a too long model update cycle.

[0039] To solve the above problems, the embodiments of the present application provide a risk control model updating method and related device, in order to alleviate the data processing pressure by fusing the incremental learning means of knowledge distillation and parameter regularization, and effectively inhibit the catastrophic forgetting defect existing in the model updating process, and improve the efficiency of model updating.

[0040] Firstly, the method in the embodiments of the present application is applied to a general distributed architecture of machine learning model training and deployment, mainly involving the risk control model training architecture in the financial business field. Please refer to Figure 1 , Figure 1 is a schematic diagram of the risk control model training architecture provided by the embodiments of the present application, as shown in Figure 1 , the risk control model training architecture 10 includes a data storage layer 110, a computing layer 120, a network layer 130, and a deployment and monitoring layer 140; wherein the data storage layer 110 is used to store new business data, parameters of the first model and historical business data; the computing layer 120 is used to perform incremental training of the initial risk control model, and the core is to optimize the parameters through a composite loss function to update the initial risk control model to a second model; the network layer 130 is used to realize data transmission between modules, such as new data transmission from the storage layer to the computing layer 120, soft label transmission from the teacher model node to the training node; the deployment and monitoring layer 140 is used to deploy the second model after training, and monitor its online performance, such as accuracy and stability and other indicators.

[0041] Specifically, the data storage layer 110 can be a distributed file system such as HDFS, Ceph, which stores structured enterprise credit data and unstructured data, supports high-throughput read and write; a relational database such as a MySQL cluster, which stores model training logs, parameter version information, and ensures traceability; a cache system such as Redis, which temporarily caches new data batches and teacher model soft labels to speed up data access during training. The computing layer 120 can be a GPU cluster that uses the parallel computing capabilities of GPUs to accelerate matrix operations of loss functions, suitable for neural network models; a CPU server that can train small batches of data in parallel through multi-core CPUs if the model is lightweight, reducing costs. The network layer 130 can be a high-speed Ethernet or InfiniBand network that ensures low latency for parameter synchronization and data transmission during distributed training. The deployment and monitoring layer 140 can be an application server that deploys the inference interface of the student model; a monitoring server that runs hardware monitoring tools to collect CPU / GPU utilization, memory usage, and other indicators in real time, ensuring the stability of the deployed second model.

[0042] The following will be described in combination with Figure 2 The updating method of a risk control model in an embodiment of the present application is described below, Figure 2 is a flowchart of an updating method of a risk control model provided by an embodiment of the present application. The method is applied to an architecture as shown in Figure 1 and specifically includes the following steps:

[0043] In step S210, a first model and new business data are obtained.

[0044] The first model refers to a deployed risk control model. The first model is an old version of the risk control model currently running in the actual online business, which has been deployed and is in use, and serves as a baseline risk control model M old (teacher model); the new business data refers to the latest business data generated within the model update cycle, such as enterprise credit application records and credit data in the past month, which contains new risk features that the old model has not learned. It should be understood that the model architecture selection of the present scheme is independent of the implementation of the scheme, and can be adapted to neural network or logistic regression models, as well as other mainstream risk control models such as gradient boosting decision tree (GBDT).

[0045] In step S220, the new business data is input into the first model to obtain first soft labels output by the first model.

[0046] The first soft label refers to a probability distribution prediction of the first model on the new business data, rather than a hard label such as default or non-default. Specifically, the first soft label of a certain enterprise can be “default probability 8%, normal probability 92%”, which contains the implicit cognition of the first model on the risk of the enterprise and is a compressed form of the knowledge of the first model.

[0047] In step S230, an initial risk control model and a target loss function are newly created, the initial risk control model is trained through the target loss function, and a second model is obtained.

[0048] The target loss function contains the first soft label, and the target loss function contains a distillation loss related to the first soft label, such as a KL divergence loss. In addition, the target loss function can also contain a task loss based on the real label of the new business data, a regularization loss to protect the key parameters of the first model, and an avoidance of drastic changes. The first soft label is used to make the new model (the second model) imitate the decision distribution of the old model (the first model) and achieve knowledge distillation. The initial risk control model refers to a model (student model) that is newly created and not trained, and has the same architecture as the first model, such as the same number of neural network layers and feature processing logic. new The initial risk control model refers to a model (student model) that is newly created and not trained, and has the same architecture as the first model, such as the same number of neural network layers and feature processing logic.

[0049] In step S240, the first model is updated to the second model.

[0050] The first model is replaced by the trained second model, which is directly applied to actual risk control business such as enterprise credit approval and risk assessment. The online iteration of the second model is completed, and the updated second model plays a role in actual business. After replacement, the second model can identify new risk patterns in the new business data and retain the judgment ability of the first model on historical risk characteristics, solving the defects of high cost and catastrophic forgetting of the traditional updating method.

[0051] It can be seen that, by using the risk control model updating method and related device provided in the present application, the first model and the new business data are first obtained; then the new business data is input into the first model to obtain the first soft label output by the first model; then an initial risk control model and a target loss function are newly created, the initial risk control model is trained through the target loss function, and a second model is obtained; finally, the first model is updated to the second model. In this way, by using the incremental learning method of fusing knowledge distillation and parameter regularization, the data processing pressure is relieved, and the catastrophic forgetting defect in the model updating process is effectively inhibited, and the efficiency of model updating is improved.

[0052] In a possible embodiment, the initial risk control model is trained by a target loss function, including: determining a loss result of the initial risk control model according to the new business data and the target loss function; and minimizing the loss result according to an optimization algorithm to train the initial risk control model.

[0053] For details, please refer to Figure 3 , Figure 3 is a flowchart of training a risk control model provided by an embodiment of the present application, as shown in Figure 3 , the loss result of the initial risk control model is determined according to the new business data and the target loss function, including the following steps:

[0054] Step S301, input the new business data into the initial risk control model, and determine a first sub-loss result, a second sub-loss result and a third sub-loss result according to the target loss function.

[0055] Specifically, the target loss function is shown in the following formula:

[0056] L total =αL task +βL distill +γL reg ;

[0057] Wherein, L total is the total loss result, L task (task loss) as the first sub-loss result, is used to force learning the real law of the new business data, such as the new features of credit default, reflecting the prediction accuracy of the real label of the new business data; L distill (distillation loss) as the second sub-loss result, in order to inherit the historical decision logic of the old model, such as the risk judgment preference of the first model to the enterprise scale, reflecting the imitation of the decision logic of the first model; L reg (regularization loss) as the third sub-loss result, is used to protect the key parameters of the historical knowledge logic, such as the neuron weight sensitive to the historical repayment record in the first model, reflecting the stability of the key parameters.

[0058] Step S302, configure weight parameters for the first sub-loss result, the second sub-loss result and the third sub-loss result, respectively.

[0059] Wherein, the weight parameters are used to reflect the influence degree of the sub-loss result on the loss result, such as alpha, beta and gamma shown in the above formula.

[0060] Specifically, in a possible embodiment, the method further includes: dynamically adjusting the weight parameters according to the increase of the training times of the initial risk control model.

[0061] For example, initially, alpha = 0.7, which means that new data is prioritized; beta = 0.2, which means that old logic is inherited; gamma = 0.1, which means that core parameters are protected. As the training iteration progresses, gamma can be dynamically reduced, for example, by 10% every 100 times, to avoid over-restricting the adaptability of the new model.

[0062] As can be seen, in the embodiment, through the above dynamic adjustment scheme and the decay mechanism of the weight parameter, the variation range of the related parameters of the old logic is controlled within a safe range, and the catastrophic forgetting problem is inhibited; only the newly added business data needs to be processed, without loading the full amount of historical data, thereby reducing the waste of computing resources and improving the efficiency and stability of model training.

[0063] In step S303, the loss result is determined according to the first sub-loss result, the second sub-loss result, the third sub-loss result, and the corresponding weight parameter.

[0064] The double mechanism ensures the retention of the knowledge logic of the model, because the knowledge distillation retains the decision-making behavior of the first model at the macro output level, and the EWC regularization protects the model structure important to the historical knowledge logic at the micro parameter level. The double protection enables the model to learn new knowledge while retaining historical experience, thereby ensuring the stability of the updated risk control model on new and old data.

[0065] In one possible embodiment, the first sub-loss result is determined according to the target loss function, including: determining the predicted label of the newly added business data output by the initial risk control model, and the real label of the newly added business data; calculating the difference between the predicted label and the real label by cross-entropy loss as the first sub-loss result, which is used to represent the prediction error of the second model on the newly added business data.

[0066] Specifically, for example: for enterprise A participating in the green energy project and having an asset-liability ratio of 65%, the first model outputs a soft label of "default probability 6%, normal probability 94%"; when training using the initial risk control model, based on the real label of the newly added business data, such as the actual non-default of enterprise A, the real label is normal, and the difference between the prediction of the initial risk control model and the real label is calculated by cross-entropy loss; if the initial risk control model initially predicts a default probability of 30% for enterprise A, the first sub-loss result is large, which forces it to optimize in the direction of the real normal and reduce the default probability prediction. The goal of the first sub-loss result is to learn the pattern of new data.

[0067] In one possible embodiment, the second sub-loss result is determined according to the target loss function, including: calculating the second soft label of the newly added business data output by the initial risk control model; determining the difference between the first soft label and the second soft label as the second sub-loss result.

[0068] Specifically, as in the above example, the KL divergence between the initial risk control model prediction and the first model soft label, enterprise A "default probability 6%" is calculated; if the initial risk control model predicts "default probability 15%" after optimization, although it is closer to the true label than the initial 30%, it still has a gap with the first model's 6%, and the second sub-loss result will further reduce it to close to 6%, ensuring the inheritance of the first model's judgment on "asset-liability ratio 65%" such historical key features, avoiding ignoring traditional risk indicators due to learning new policies. The goal of the second sub-loss result is to mimic and inherit the decision logic of the old model.

[0069] In one possible embodiment, according to the target loss function, the third sub-loss result is determined, including: determining a core task of the first model for historical business data, and core parameters in the first model, the core parameters referring to parameters that have an impact on the core task; calculating the importance of the core task; according to the importance of the core task, a preset penalty coefficient and the core parameters, determining the third sub-loss result by a learning algorithm, the third sub-loss result being used to constrain the degree of change of the core parameters.

[0070] Specifically, as in the above example, the EWC algorithm is used to impose a penalty on the change of "parameters critical to old knowledge" in the first model, such as the weight parameter W of the historical overdue record feature; if the initial risk control model tries to significantly reduce W during training, such as from 0.8 to 0.3, the third sub-loss result will significantly increase, limiting the above change and ensuring that historical overdue records as an important risk indicator are not forgotten.

[0071] Specifically, please refer to Figure 4 , Figure 4 is a flowchart of another method for updating a risk control model provided by the embodiments of the present application, as shown in Figure 4 , the updating method comprises:

[0072] The preparation stage 410 starts; the new business data is input into the first model M old , knowledge distillation is implemented, and the soft label of the new business data is generated.

[0073] The incremental learning stage 420 determines the initial risk control model; the task loss L task , the distillation loss L distill , and the regularization loss L reg are taken as factors to calculate the composite loss L total ; the initial risk control model is optimized by minimizing L total ; and the updated second model M new is generated.

[0074] The deployment stage 430 deploys the updated second model M new to replace the first model Mold ; end.

[0075] In other possible embodiments, the learning algorithm can also employ other continual learning algorithms for alleviating catastrophic forgetting, such as Synaptic Intelligence (SI), Memory Aware Synapses (MAS), etc.

[0076] Specifically, Synaptic Intelligence measures the importance of parameters by tracking the contribution of model parameters in the process of learning old tasks (i.e., the impact of parameter changes on the performance of old tasks); specifically, SI records the cumulative amount of change of each parameter in previous task training and the contribution to loss reduction, and converts this information into the importance weight of the parameter. When learning a new task incrementally, if the update direction of the parameter conflicts with the contribution direction in the previous task, SI will impose a penalty on it, and the penalty is proportional to the importance weight of the parameter. Synaptic Intelligence replaces EWC as an implementation way of calculating the compound loss, protects key parameters from the perspective of dynamic contribution of parameters to previous tasks, and avoids excessive modification of important parameters for historical business data by the new model to adapt to new data.

[0077] Specifically, Memory Aware Synapses determines the importance of parameters by evaluating the impact of parameters on the prediction performance of representative samples in old tasks. Specifically, MAS selects a small number of key samples from old task data and calculates the sensitivity of each parameter to the prediction results of these samples. The higher the sensitivity of the parameter, the more important it is considered to be for historical business data; when learning a new task, the update amplitude of such parameters will be subject to greater penalties. By indirectly measuring the importance of parameters through key samples, Memory Aware Synapses replaces the Fisher information matrix calculation method of EWC, and more lightweightly protects the parameters related to historical business data.

[0078] In one possible embodiment, the learning algorithm is an Elastic Weight Consolidation algorithm.

[0079] The Elastic Weight Consolidation algorithm measures the importance of each parameter to the old task through the Fisher information matrix, and imposes greater penalties on changes to important parameters to prevent their results from deviating dramatically from the parameter values important to the old task.

[0080] The EWC Elastic Weight Consolidation algorithm, as a core means of regularization loss term, protects the risk features proven effective in historical data by parameter importance quantification + differentiated penalties, preventing new data noise from destroying old knowledge; allows non-critical parameters to be adjusted flexibly to adapt to new risk patterns; only uses new data for training, reducing costs while maintaining model stability. Ultimately, the synergy of EWC and knowledge distillation makes the update of the risk control model more accurate and stable, and improves the efficiency of incremental training, solving the technical defects of traditional solutions.

[0081] The embodiments of the present application can divide the functional units of the electronic device according to the above method examples. For example, each functional unit can be divided according to each function, or two or more functions can be integrated in one processing unit. The integrated unit can be realized in the form of hardware or in the form of a software functional unit. It should be noted that the division of the units in the embodiments of the present application is illustrative, and is only a logical functional division. When actually implemented, there can be another division manner.

[0082] Consistent with embodiments of the present application, please refer to Figure 2 Figure 5 , Figure 5 is a functional unit composition block diagram of a risk control model updating device provided by the embodiments of the present application, as shown in Figure 5 The risk control model updating device 500 includes an acquisition unit 510, an output unit 520, a training unit 530, and an updating unit 540. The acquisition unit 510 is configured to acquire a first model and new business data. The first model refers to a deployed risk control model. The output unit 520 is configured to input the new business data into the first model to obtain a first soft label output by the first model. The training unit 530 is configured to newly create an initial risk control model and a target loss function, train the initial risk control model through the target loss function to obtain a second model, and the target loss function contains the first soft label. The updating unit 540 is configured to update the first model to the second model.

[0083] In one possible embodiment, in terms of training the initial risk control model through the target loss function, the training unit 530 is specifically configured to: determine a loss result of the initial risk control model according to the new business data and the target loss function; and minimize the loss result according to an optimization algorithm to train the initial risk control model.

[0084] In one possible embodiment, in terms of determining the loss result of the initial risk control model according to the new business data and the target loss function, the training unit 530 is specifically configured to: input the new business data into the initial risk control model, and determine a first sub-loss result, a second sub-loss result, and a third sub-loss result according to the target loss function; configure a weight parameter for the first sub-loss result, the second sub-loss result, and the third sub-loss result, respectively, and the weight parameter is used to reflect the influence degree of the sub-loss result on the loss result; and determine the loss result according to the first sub-loss result, the second sub-loss result, the third sub-loss result, and the corresponding weight parameter.

[0085] ​In a possible implementation, in the aspect of determining the first sub-loss result according to the target loss function, the training unit 530 is specifically configured to: determine a predicted label of the initial risk control model output for the new business data, and a real label of the new business data; and calculate a difference between the predicted label and the real label as the first sub-loss result through a cross-entropy loss, where the first sub-loss result is used to represent a prediction error of the second model for the new business data.

[0086] In a possible implementation, in the aspect of determining the second sub-loss result according to the target loss function, the training unit 530 is specifically configured to: calculate a second soft label of the initial risk control model output for the new business data; and determine a difference between the first soft label and the second soft label as the second sub-loss result.

[0087] In a possible implementation, in the aspect of determining the third sub-loss result according to the target loss function, the training unit 530 is specifically configured to: determine a core task of the first model for the historical business data, and a core parameter in the first model, where the core parameter refers to a parameter that has an influence on the core task; calculate an importance degree of the core task in the first model; and determine the third sub-loss result through a learning algorithm according to the importance degree of the core task, a preset penalty coefficient, and the core parameter, where the third sub-loss result is used to constrain a variation degree of the core parameter.

[0088] In a possible implementation, the learning algorithm is an elastic weight consolidation algorithm.

[0089] In a possible implementation, the risk control model updating apparatus 500 is specifically further configured to: dynamically adjust the weight parameter according to an increase in a training number of the initial risk control model.

[0090] It can be understood that, since the method embodiments and the apparatus embodiments are different presentation forms of the same technical concept, the content of the method embodiment part in the present application should be synchronously adapted to the apparatus embodiment part, which will not be repeated here.

[0091] Figure 6 is a structural block diagram of an electronic device provided by an embodiment of the present application. As shown in Figure 6 , the electronic device 600 can include one or more of the following components: a processor 601, a memory 602 coupled with the processor 601, wherein the memory 602 can store one or more computer programs, and the one or more computer programs can be configured to be executed by the one or more processors 601 to implement the method described in the above examples. Wherein, the electronic device 600 includes an architecture as shown in Figure 1 .

[0092] The processor 601 can include one or more processing cores. The processor 601 connects various parts within the entire electronic device 600 by running or executing instructions, programs, code sets or instruction sets stored in the memory 602, and calling data stored in the memory 602, to perform various functions and process data of the electronic device 600. Optionally, the processor 601 can be implemented in at least one of a hardware form of a digital signal processing (DSP), a field-programmable gate array (FPGA), a programmable logic array (PLA). The processor 601 can integrate a combination of one or several of a central processing unit (CPU), a graphics processor (GPU), and a modem. It can be understood that the above-mentioned modem can also not be integrated into the processor 601, but be implemented by a separate communication chip.

[0093] The memory 602 can include a random access memory (RAM) and can also include a read-only memory (ROM). The memory 602 can be used to store instructions, programs, codes, code sets or instruction sets. The memory 602 can include a program storage area and a data storage area, wherein the program storage area can store instructions for implementing an operating system, instructions for implementing at least one function (such as a touch function, a sound playing function, an image playing function, etc.), instructions for implementing the above-mentioned various method examples, etc. The data storage area can also store data created by the electronic device 600 in use, etc.

[0094] It can be understood that the electronic device 600 can include more or fewer structural elements than those in the above structural block diagram, for example, a power module, a physical key, a WiFi (Wireless Fidelity) module, a speaker, a Bluetooth module, a sensor, etc., which are not limited herein.

[0095] The embodiments of the present application also provide a computer storage medium, wherein a computer program / instructions are stored on the computer storage medium, and the computer program / instructions are executed by a processor to implement some or all steps of any method described in the above method embodiments.

[0096] The embodiment of the present application further provides a computer program product, which comprises a non-transitory computer-readable storage medium storing a computer program, and the computer program is operable to cause a computer to perform some or all of the steps of any method described in the above method embodiments.

[0097] It should be understood that the size of the sequence number of the above processes does not mean the order of execution in various embodiments of the present application. The execution order of each process should be determined according to its function and inherent logic, and should not constitute any limitation on the implementation process of the embodiments of the present application.

[0098] In several embodiments provided in the present application, it should be understood that the disclosed methods, devices and systems can be implemented in other ways. For example, the above-described device embodiments are only illustrative; for example, the division of units is only a logical function division, and actual implementation can have another division manner; for example, a plurality of units or components can be combined or integrated into another system, or some features can be omitted or not executed. In addition, the coupling or direct coupling or communication connection between the displayed or discussed units can be indirect coupling or communication connection through some interfaces, devices or units, which can be electrical, mechanical or other forms.

[0099] The units described as separate components can or can not be physically separated, and the components shown as units can or can not be physical units, that is, they can be located in one place, or can be distributed on multiple network units. According to actual needs, some or all of the units can be selected to achieve the purpose of the present embodiment.

[0100] In addition, each functional unit in each embodiment of the present application can be integrated in one processing unit, or each unit can be physically included separately, or two or more units can be integrated in one unit. The integrated unit can be realized in the form of hardware, or in the form of hardware plus software functional unit.

[0101] The integrated unit in the form of software function unit can be stored in a computer readable storage medium. The software function unit is stored in a storage medium, including a plurality of instructions for enabling a computer device (which can be a personal computer, a server, or a network device, etc.) to execute part of the steps of the method of various embodiments of the present application. The aforementioned storage medium includes a U disk, a mobile hard disk, a magnetic disk, an optical disk, a volatile memory or a non-volatile memory. The non-volatile memory can be a read-only memory (ROM), a programmable read-only memory (PROM), an erasable programmable read-only memory (EPROM), an electrically erasable programmable read-only memory (EEPROM) or a flash memory. The volatile memory can be a random access memory (RAM) used as an external cache. By way of example, but not limitation, many forms of random access memory (RAM) can be used, such as static random access memory (SRAM), dynamic random access memory (DRAM), synchronous dynamic random access memory (SDRAM), double data rate synchronous dynamic random access memory (DDR SDRAM), enhanced synchronous dynamic random access memory (ESDRAM), synchlink dynamic random access memory (SLDRAM) and direct rambus random access memory (DRAM) and various other media which can store program codes.

[0102] Although the present application is disclosed as above, the present application is not limited thereto. Any person skilled in the art can easily conceive of changes or substitutions without departing from the spirit and scope of the present application, and various modifications can be made, including combinations of different functions and implementation steps, including software and hardware implementations, which are all within the scope of the present application.

Claims

1. A method for updating a risk control model, characterized in that, The method includes: Obtain the first model and new business data, where the first model refers to the risk control model that has been deployed; The newly added business data is input into the first model to obtain the first soft label output by the first model. A new initial risk control model and a target loss function are created. The initial risk control model is trained using the target loss function to obtain a second model. The target loss function includes a first soft label. Update the first model to the second model.

2. The method according to claim 1, characterized in that, The step of training the initial risk control model using the target loss function includes: Based on the newly added business data and the target loss function, determine the loss result of the initial risk control model; The initial risk control model is trained by minimizing the loss result according to the optimization algorithm.

3. The method according to claim 2, characterized in that, The step of determining the loss result of the initial risk control model based on the new business data and the target loss function includes: The newly added business data is input into the initial risk control model, and the first sub-loss result, the second sub-loss result, and the third sub-loss result are determined according to the target loss function. Weight parameters are configured for the first sub-loss result, the second sub-loss result, and the third sub-loss result, respectively. The weight parameters are used to reflect the degree of influence of the sub-loss result on the loss result. The loss result is determined based on the first sub-loss result, the second sub-loss result, the third sub-loss result, and the corresponding weight parameters.

4. The method according to claim 3, characterized in that, Determining the first sub-loss result based on the target loss function includes: Determine the predicted labels for the new business data output by the initial risk control model, and the actual labels for the new business data; The difference between the predicted label and the true label is calculated by cross-entropy loss and used as the first sub-loss result. The first sub-loss result is used to characterize the prediction error of the second model for the new business data.

5. The method according to claim 4, characterized in that, Determining the second sub-loss result based on the target loss function includes: Calculate the second soft label for the newly added business data output by the initial risk control model; The difference between the first soft label and the second soft label is determined as the second sub-loss result.

6. The method according to claim 5, characterized in that, The step of determining the third sub-loss result based on the target loss function includes: The core tasks of the first model in response to historical business data are determined, as well as the core parameters in the first model, wherein the core parameters refer to the parameters that have an impact on the core tasks; The importance of the core tasks described in the first model is determined through calculation. Based on the importance of the core task, the preset penalty coefficient, and the core parameters, the third sub-loss result is determined through a learning algorithm. The third sub-loss result is used to constrain the drastic changes in the core parameters.

7. The method according to claim 6, characterized in that, The learning algorithm is an elastic weight consolidation algorithm.

8. The method according to any one of claims 3-7, characterized in that, The method further includes: The weight parameters are dynamically adjusted as the number of training iterations of the initial risk control model increases.

9. A risk control model updating device, characterized in that, The device includes: an acquisition unit, an output unit, a training unit, and an update unit; wherein, The acquisition unit is used to acquire the first model and the newly added business data, wherein the first model refers to the risk control model that has been deployed. The output unit is used to input the new business data into the first model to obtain the first soft label output by the first model; The training unit is used to create an initial risk control model and a target loss function, and to train the initial risk control model using the target loss function to obtain a second model. The target loss function includes a first soft label. The update unit is used to update the first model to the second model.

10. An electronic device, characterized in that, It includes a processor and a memory, the memory being used to store one or more programs and configured to be executed by the processor, the programs including instructions for performing the steps of the method as described in any one of claims 1-8.

Images