Adaptive dynamic risk prediction method and apparatus

Abstract

The present disclosure relates to the technical field of artificial intelligence, and particularly provides a self-adaptive dynamic risk prediction method and device. The method comprises: obtaining a plurality of model parameters of a risk prediction model trained based on historical risk data and a parameter importance measure corresponding to each model parameter; performing incremental training on the risk prediction model based on new task data, determining a parameter constraint degree of each model parameter according to the parameter importance measure and a parameter offset degree of each model parameter; adjusting the parameter constraint degree of each model parameter according to the output difference of the risk prediction model for the historical risk data and the new task data, and updating each model parameter according to the adjusted parameter constraint degree and the new task data; and performing risk prediction on to-be-identified risk data by using the updated risk prediction model to obtain a risk prediction result. The present disclosure can improve the adaptability of the risk prediction model to the continuously changing risk pattern, thereby improving the accuracy of risk prediction.

Description

Technical Field

[0001] This disclosure relates to the field of artificial intelligence technology, and in particular to an adaptive dynamic risk prediction method and apparatus. Background Technology

[0002] With the development of internet finance, e-commerce, and other businesses, transaction data and user behavior data are growing rapidly, and the types and methods of fraud are becoming increasingly complex. Existing risk control systems typically collect transaction data, user behavior data, and device information based on big data platforms, and use historical data to train risk prediction models to identify fraudulent and other risky behaviors. Because fraud methods are constantly evolving and new risk patterns are continuously emerging, it is necessary to update risk prediction models using new data.

[0003] In related technologies, risk prediction models are typically updated using methods such as periodic offline training. However, these updates are somewhat delayed and struggle to adapt to rapidly evolving new fraudulent behaviors. Furthermore, when updating risk prediction models with new data, the model parameters change based on the new risk patterns, potentially leading to a decline in the model's ability to identify historical risk patterns and resulting in catastrophic amnesia. Summary of the Invention

[0004] This disclosure is made in view of the above-mentioned problems. This disclosure provides an adaptive dynamic risk prediction method and apparatus.

[0005] According to one aspect of this disclosure, an adaptive dynamic risk prediction method is provided, comprising: acquiring multiple model parameters of a risk prediction model trained based on historical risk data, and acquiring a parameter importance metric corresponding to each model parameter; wherein the parameter importance metric is used to characterize the importance of the corresponding model parameter for risk prediction based on historical risk data; acquiring new task data corresponding to a new risk pattern, incrementally training the risk prediction model based on the new task data, determining the parameter constraint degree of each model parameter according to the parameter importance metric of each model parameter and the parameter offset degree of each model parameter relative to the parameter before incremental training; adjusting the parameter constraint degree of each model parameter according to the output difference of the risk prediction model for historical risk data and new task data, and updating each model parameter according to the adjusted parameter constraint degree and the new task data to obtain an updated risk prediction model; and using the updated risk prediction model to perform risk prediction on the risk data to be identified to obtain a risk prediction result.

[0006] According to another aspect of this disclosure, an adaptive dynamic risk prediction device is provided, comprising: an acquisition module, configured to acquire multiple model parameters of a risk prediction model trained based on historical risk data, and acquire parameter importance metrics corresponding to each model parameter; wherein the parameter importance metrics are used to characterize the importance of the corresponding model parameter for risk prediction based on historical risk data; the acquisition module is further configured to acquire new task data corresponding to a new risk pattern; a training module, configured to incrementally train the risk prediction model based on the new task data, determine the parameter constraint degree of each model parameter according to the parameter importance metrics of each model parameter and the parameter offset degree of each model parameter relative to the previous incremental training; adjust the parameter constraint degree of each model parameter according to the output difference of the risk prediction model for historical risk data and new task data, and update each model parameter according to the adjusted parameter constraint degree and the new task data to obtain an updated risk prediction model; and a prediction module, configured to use the updated risk prediction model to perform risk prediction on the risk data to be identified, and obtain a risk prediction result.

[0007] In another aspect of exemplary embodiments of this disclosure, an electronic device is provided, including a memory, a processor, and a computer program stored in the memory, wherein the processor executes the computer program to implement the method of exemplary embodiments of this disclosure.

[0008] In another aspect of exemplary embodiments of the present disclosure, a computer-readable storage medium is provided having a computer program stored thereon that, when executed by a processor, implements the methods of exemplary embodiments of the present disclosure.

[0009] In another aspect of the exemplary embodiments of this disclosure, a computer program product is provided, including a computer program that, when executed by a processor, implements the method of the exemplary embodiments of this disclosure.

[0010] As will be described in detail below, the adaptive dynamic risk prediction method according to embodiments of this disclosure acquires multiple model parameters of a risk prediction model trained based on historical risk data, and acquires a parameter importance metric corresponding to each model parameter; wherein, the parameter importance metric is used to characterize the importance of the corresponding model parameter for risk prediction based on historical risk data; acquires new task data corresponding to new risk patterns, incrementally trains the risk prediction model based on the new task data, determines the parameter constraint degree of each model parameter according to the parameter importance metric of each model parameter and the parameter offset degree of each model parameter relative to the parameter before incremental training; adjusts the parameter constraint degree of each model parameter according to the output difference of the risk prediction model for historical risk data and new task data, and updates each model parameter according to the adjusted parameter constraint degree and new task data to obtain an updated risk prediction model; uses the updated risk prediction model to perform risk prediction on the risk data to be identified, and obtains the risk prediction result. This enables the risk prediction model to learn new risk patterns while reducing the degree of forgetting historical risk patterns, taking into account the risk prediction ability for both historical and new risk patterns, improving the adaptability of the risk prediction model to continuously changing risk patterns, and thus improving the accuracy of risk prediction.

[0011] It should be understood that both the foregoing general description and the following detailed description are exemplary and intended to provide further illustration of the claimed technology. Attached Figure Description

[0012] The above and other objects, features, and advantages of this disclosure will become more apparent from the more detailed description of the embodiments thereof in conjunction with the accompanying drawings. The drawings are provided to further illustrate the embodiments of this disclosure and form part of the specification. They are used together with the embodiments of this disclosure to explain the disclosure and do not constitute a limitation thereof. In the drawings, the same reference numerals generally represent the same components or steps.

[0013] Figure 1 A flowchart illustrating the adaptive dynamic risk prediction method provided in an exemplary embodiment of this disclosure is shown. Figure 2 A schematic diagram of the structure of the adaptive dynamic risk prediction device provided in an exemplary embodiment of this disclosure is shown. Figure 3 A schematic diagram of the structure of an electronic device provided in an exemplary embodiment of this disclosure is shown; Figure 4 A schematic diagram of the structure of a computer system provided in an exemplary embodiment of this disclosure is shown. Detailed Implementation

[0014] To make the objectives, technical solutions, and advantages of this disclosure more apparent, exemplary embodiments according to this disclosure will now be described in detail with reference to the accompanying drawings. Obviously, the described embodiments are merely some embodiments of this disclosure, and not all embodiments of this disclosure. It should be understood that this disclosure is not limited to the exemplary embodiments described herein.

[0015] It should be understood that the steps described in the method embodiments of this disclosure may be performed in different orders and / or in parallel. Furthermore, the method embodiments may include additional steps and / or omit the steps shown. The scope of this disclosure is not limited in this respect.

[0016] The term "comprising" and its variations as used herein are open-ended, meaning "including but not limited to". The term "based on" means "at least partially based on". The term "one embodiment" means "at least one embodiment"; the term "another embodiment" means "at least one additional embodiment"; the term "some embodiments" means "at least some embodiments". Definitions of other terms will be given in the description below. It should be noted that the concepts of "first", "second", etc., used in this disclosure are only used to distinguish different devices, modules, or units, and are not intended to limit the order of functions performed by these devices, modules, or units or their interdependencies.

[0017] It should be noted that the terms "a" and "a plurality of" used in this disclosure are illustrative rather than restrictive, and those skilled in the art should understand that, unless otherwise expressly indicated in the context, they should be understood as "one or more".

[0018] The names of messages or information exchanged between multiple devices in the embodiments of this disclosure are for illustrative purposes only and are not intended to limit the scope of such messages or information.

[0019] For users with abundant data, a multi-layered architecture is built based on big data platforms such as Hadoop and Spark, covering the entire process of data collection, real-time computing, model training, and rule-based decision-making. Feature engineering is established through user profiles (such as consumption levels, credit scores, and device fingerprints) and behavioral time-series data (such as transaction frequency and IP changes), and risk probabilities are predicted by combining models such as Long Short-Term Memory (LSTM) and Graph Neural Networks (GNN).

[0020] Existing technical solutions generally face two major challenges: First, the lag in model updates: mainstream risk control systems, such as International Business Machines Safer Payments (IBM Safer Payments) and SAS Fraud Framework (Statistical Analysis System Fraud Framework), primarily train their models on periodic offline tasks (usually daily or weekly), making it difficult to cope with new types of online fraud that evolve at the minute or even second level, such as deepfake voice scams and metaverse virtual asset theft. Second, the problem of catastrophic forgetting: when a model is fully retrained using newly collected fraud data (new tasks), it inevitably overwrites the memory of historical fraud patterns (old tasks), leading to a decrease in the ability to defend against known risks. Although the academic community has proposed algorithms such as Elastic Weight Consolidation (EWC) in the field of incremental learning in recent years, how to balance the speed and accuracy of retaining historical knowledge and learning new knowledge in industrial-grade big data risk control scenarios remains an open question.

[0021] Therefore, in order to solve the above problems, this disclosure provides an adaptive dynamic risk prediction method. In the process of incrementally training the risk prediction model using new task data corresponding to new risk patterns, the model parameters are constrained based on the parameter importance measure and parameter offset degree of the model parameters. The degree of parameter constraint is adjusted according to the output difference of the risk prediction model for historical risk data and new task data. This enables the risk prediction model to learn new risk patterns while reducing the degree of forgetting historical risk patterns and improving the adaptability of the risk prediction model to continuously changing risk patterns.

[0022] The adaptive dynamic risk prediction method provided in this disclosure can be executed by a server or by a chip applied to the server.

[0023] For example, the server described above can be an independent physical server, a server cluster or distributed system composed of multiple physical servers, or a cloud server that provides basic cloud computing services such as cloud services, cloud databases, cloud computing, cloud functions, cloud storage, network services, cloud communication, middleware services, domain name services, security services, content delivery networks (CDN), and big data and artificial intelligence platforms. This exemplary embodiment does not limit the scope of the disclosure.

[0024] Figure 1 A flowchart illustrating the adaptive dynamic risk prediction method provided in an exemplary embodiment of this disclosure is shown. Figure 1 As shown, the adaptive dynamic risk prediction method includes: S101, obtain multiple model parameters of the risk prediction model trained based on historical risk data, and obtain the parameter importance metric corresponding to each model parameter; wherein, the parameter importance metric is used to characterize the importance of the corresponding model parameter for risk prediction based on historical risk data; S102, acquire new task data corresponding to the new risk pattern, incrementally train the risk prediction model based on the new task data, determine the parameter constraint degree of each model parameter according to the parameter importance metric of each model parameter and the parameter offset degree of each model parameter relative to the previous incremental training; adjust the parameter constraint degree of each model parameter according to the output difference of the risk prediction model for historical risk data and new task data, and update each model parameter according to the adjusted parameter constraint degree and the new task data to obtain the updated risk prediction model; S103, use the updated risk prediction model to perform risk prediction on the risk data to be identified, and obtain the risk prediction results.

[0025] Specifically, the aforementioned historical risk data, also known as old task data or old task samples, can be historical business data for which risk categories or outcomes have been determined. For example, in a transaction risk prediction scenario, historical risk data can include historical transaction data, user behavior data, device information, and corresponding risk tags. Among these, risk tags can be used to characterize whether the corresponding historical business data carries risk; for example, they can include tags for normal transactions and tags for fraudulent transactions.

[0026] Before training a risk prediction model using historical risk data, the historical risk data can be preprocessed. For example, historical risk data can include features processed by principal component analysis (such as V1-V28), time features, amount features, and risk labels.

[0027] For numerical features such as monetary amounts, Z-score standardization can be used. The formula is as follows:

[0028] in, This represents the standardized numerical features; This represents numerical characteristics before standardization, such as transaction amount; This represents the mean of the corresponding feature; This represents the standard deviation of the corresponding feature.

[0029] For time features, second-level time can be converted to hour-level time to reduce the impact of second-level time fluctuations on model training stability.

[0030] When the proportion of risky samples is low, such as when the proportion of fraudulent samples is below a preset threshold (e.g., 0.5%, which can be set according to the actual scenario), downsampling and other methods can be used to adjust the ratio of normal samples to risky samples in order to reduce the impact of sample class imbalance on the training of risk prediction models.

[0031] After preprocessing historical risk data, a risk prediction model can be trained using the processed historical risk data. The risk prediction model can be a model used to determine the corresponding risk prediction result based on the input risk data. For example, the risk prediction model can employ a logistic regression model, a random forest model, or a neural network model; this disclosure does not specifically limit the specific model to these methods.

[0032] During the training of a risk prediction model, improving the recall rate and reducing the false positive rate can be used as optimization objectives, and model parameters can be fine-tuned using methods such as grid search. For example, a recall rate no lower than a preset recall rate (e.g., 95%, which can be set according to the actual scenario) and a false positive rate no higher than a preset false positive rate (e.g., 20%, which can be set according to the actual scenario) can be used as model performance validation metrics. Once the risk prediction model meets the corresponding performance requirements, the trained risk prediction model and its multiple model parameters can be saved as the basis for subsequent incremental training.

[0033] A risk prediction model comprises multiple model parameters, and different model parameters may have varying impacts on the model's ability to predict risk based on historical risk data. Therefore, it is possible to obtain the parameter importance metric for each model parameter. A higher parameter importance metric indicates a greater importance of the corresponding model parameter in risk prediction based on historical risk data; conversely, a lower parameter importance metric indicates a lower importance of the corresponding model parameter in risk prediction based on historical risk data.

[0034] It should be noted that the specific method for determining the importance of parameters is not limited in the embodiments disclosed herein, as long as it can characterize the importance of the corresponding model parameters for risk prediction based on historical risk data.

[0035] A new risk pattern can be a risk pattern that emerges or changes compared to the historical risk patterns represented by historical risk data. For example, in a transaction fraud risk prediction scenario, a new risk pattern could be a newly emerging fraud method, or a fraud pattern formed after existing fraud methods have changed. Accordingly, risk data corresponding to the aforementioned new risk patterns can be obtained, and the obtained risk data can be used as new task data.

[0036] For example, new task data may include actual collected and confirmed risk data, which can serve as the raw data layer. For instance, real-time transaction data, user behavior events, and device information generated from different business channels can be acquired, and risk labels can be determined for the corresponding data through risk identification or manual review. The confirmed risk data can then be used as new task data.

[0037] New task data can also include simulated risk data, which can be highly realistic fraud data. For example, user behavior logs, transaction data, and external data can be integrated, and simulated fraud data can be generated based on random noise and fraud characteristics using a data generation model. The generated simulated fraud data can be mixed with actually collected and confirmed fraud data to build a dynamically updated sample library, and the data in the dynamic sample library can be used as new task data.

[0038] One implementation approach involves introducing real-time adversarial training. Dynamic game theory enhances the risk prediction model's ability to predict novel fraudulent activities while simultaneously increasing the model's iteration rate. Real-time collection and fusion of multi-source data are possible, such as dynamically integrating structured and unstructured data like transaction logs, user behavior, and device fingerprints. Random noise and fraudulent features are then input to generate highly realistic fraud data for simulation training, thereby improving the model's generalization ability.

[0039] For example, generative adversarial networks can be used to generate simulated fraud data corresponding to new fraud methods. In one example, the generation rate of simulated fraud data can reach 5,000 to 10,000 records per second, thereby expanding the new task data corresponding to new risk patterns.

[0040] After acquiring new task data, an online learning mechanism can be used to incrementally train the risk prediction model based on the new task data, retaining historical information of important parameters. This allows the risk prediction model to quickly learn the new risk patterns represented by the new task data while retaining the existing model parameters.

[0041] In one alternative implementation, a dynamic model update mechanism based on reinforcement learning and transfer learning can be adopted: after obtaining the corresponding parameter model through training on historical risk data, the model weights are automatically adjusted according to the new task data corresponding to the new attack mode, and the model parameters are continuously updated by combining parameter importance measurement, parameter constraints and dynamic tuning process.

[0042] During incremental training, the parameters of the risk prediction model can change based on new task data. For any given model parameter, the degree of its shift relative to the parameter before incremental training can be determined, and the degree of parameter constraint can be determined by combining the corresponding parameter importance metric.

[0043] For model parameters that are important for risk prediction based on historical risk data, appropriate parameter constraints can be used to limit excessive changes in these parameters; for model parameters that are less important, relatively large adjustments can be made based on new task data.

[0044] Furthermore, the parameter constraints of each model can be adjusted based on the differences in the output of the risk prediction model for historical risk data and new task data. The adjusted parameter constraints and the new task data are then used to update the model parameters, resulting in an updated risk prediction model. This allows the risk prediction model to learn new risk patterns using new task data, while reducing the likelihood of significant changes in model parameters that are crucial for historical risk prediction.

[0045] The aforementioned risk data to be identified can be business data that currently requires risk identification.

[0046] In practical applications, the risk prediction model can be deployed as an online risk prediction service. When risk monitoring is required, during the initialization and cold start phase, the risk prediction model and its corresponding model parameters can be loaded from persistent storage, and the risk prediction model can be published as an online application programming interface service to access real-time data streams and officially begin processing risk control requests.

[0047] Online risk prediction services can continuously monitor real-time business data generated from different business channels (data access process), such as real-time transaction logs, user behavior events, and device information. They also perform data standardization, feature extraction, and feature transformation on this real-time business data, consistent with the model training phase (streaming feature computation process). For example, based on real-time business data, dynamic risk characteristics such as the number of logins from the same device in the last minute and the geographical distance between the current transaction location and the user's frequently used locations can be determined. Multiple risk characteristics are then assembled into risk data to be identified and input into an updated risk prediction model. Processing this updated model yields the corresponding risk prediction results (request assembly and online service process).

[0048] For example, in a transaction fraud risk prediction scenario, the assembled risk data to be identified is input into the updated risk prediction model, and forward propagation calculation is performed to output the fraud probability P corresponding to the transaction to be identified. fraud The probability of fraud, P, can be... fraud The transaction is compared with a preset dynamic risk threshold. When the probability of fraud is lower than the preset dynamic risk threshold, the corresponding transaction can be determined as a low-risk transaction and the transaction can be released. When the probability of fraud is higher than the preset dynamic risk threshold, risk intervention can be triggered, such as requiring the user to perform secondary authentication, suspending the current transaction, or sending the current transaction to manual review.

[0049] Here, the preset dynamic risk threshold can also be dynamically adjusted based on real-time false alarms to balance risk identification capabilities and normal business pass rates.

[0050] For business data identified as high-risk by the risk prediction model and confirmed by manual review, corresponding risk labels can be generated based on the manual review results, and the corresponding business data can be fed back to the dynamic sample library as new task data. This allows new risk data acquired during online risk prediction to be further used for incremental training of the risk prediction model, forming a data closed loop between risk prediction, risk confirmation, and incremental model training.

[0051] Once the updated risk prediction model obtained from incremental training passes the preset model performance verification, the currently running risk prediction model can be replaced with the updated risk prediction model using hot swapping. This allows the model update to be completed without interrupting the online risk prediction service, achieving a "seamless upgrade" of the business.

[0052] Through the above-mentioned online inference, data feedback, incremental training, and hot switching processes, training and inference can form an automated closed loop, enabling rapid response to real-time risks and continuous updates to risk prediction capabilities as attack methods evolve, achieving millisecond-level online risk response.

[0053] According to the technical solution of the exemplary embodiments of this disclosure, multiple model parameters of a risk prediction model trained based on historical risk data are obtained, and the parameter importance metric corresponding to each model parameter is obtained. The parameter importance metric characterizes the importance of the corresponding model parameter for risk prediction based on historical risk data. New task data corresponding to new risk patterns is obtained, and the risk prediction model is incrementally trained based on the new task data. The parameter constraint degree of each model parameter is determined according to the parameter importance metric and the parameter offset degree of each model parameter relative to the parameter before incremental training. The parameter constraint degree of each model parameter is adjusted according to the output difference of the risk prediction model for historical risk data and new task data, and the model parameters are updated according to the adjusted parameter constraint degree and the new task data to obtain an updated risk prediction model. The updated risk prediction model is used to predict the risk data to be identified, and the risk prediction result is obtained. This enables the risk prediction model to learn new risk patterns while reducing the forgetting of historical risk patterns, balancing the risk prediction ability for both historical and new risk patterns, improving the adaptability of the risk prediction model to continuously changing risk patterns, and thus improving the accuracy of risk prediction.

[0054] Based on the above embodiments, in another embodiment provided in this disclosure, obtaining the parameter importance metric corresponding to each model parameter may include: Input historical risk data into the risk prediction model to obtain the output probability distribution corresponding to the historical risk data; Based on the output probability distribution, determine the gradient corresponding to each model parameter; The gradients corresponding to each model parameter are squared, and the parameter importance metric corresponding to each model parameter is determined based on the average of the squared gradients corresponding to multiple historical risk data.

[0055] Specifically, historical risk data can be input into the risk prediction model, and the corresponding output probability distribution can be obtained through forward propagation, which provides a basis for subsequently determining the gradients corresponding to each model parameter.

[0056] Before forward propagation, historical risk data can be preprocessed using the preprocessing methods described above, converting it into the input format required by the risk prediction model, such as standardizing features and imputing missing values.

[0057] The input features corresponding to historical risk data can be represented as an input vector:

[0058] Where x represents the input vector, such as the feature vector of telecommunications fraud; d represents the dimension of the input vector.

[0059] After inputting the input vector into the risk prediction model, forward propagation can be performed layer by layer. The linear transformation of the u-th layer is:

[0060] Among them, z u W represents the linear transformation result of the u-th layer; u Let a represent the weight matrix of the u-th layer; u 1 b represents the activation value of the previous layer (the (u-1)th layer); u This represents the bias term of the u-th layer. Here, the weights and bias terms in the weight matrix can belong to the model parameters of the aforementioned risk prediction model.

[0061] Activation processing is performed on the linear transformation result:

[0062] Among them, a u σ represents the activation value of the u-th layer; σ represents the activation function. For example, the activation function can be the linear rectified function ReLU, used to introduce nonlinearity.

[0063] After layer-by-layer forward propagation, the output probability distribution corresponding to the historical risk data can be obtained through the output layer. The output layer function is:

[0064] in, Indicates in model parameters Below, the input vector x corresponds to the output probability of the risk label y; z(x) y This represents the output value of the risk prediction model for the risk label y; z(x) v This represents the output value of the risk prediction model for the v-th risk category; k represents the number of risk categories, such as two categories: normal transactions and fraudulent transactions.

[0065] The output probabilities corresponding to different risk categories together constitute the output probability distribution of historical risk data.

[0066] During forward propagation, the linear transformation results and activation values ​​of each layer can be saved for use in subsequent backpropagation.

[0067] Furthermore, the gradients corresponding to each model parameter can be determined based on the aforementioned output probability distribution. For the i-th historical risk data point, its input feature is represented as x. i The corresponding risk label is represented as y. i The output probability distribution corresponds to the risk label y. iTaking the logarithm of the corresponding output probability yields the log-likelihood function:

[0068] Calculate the gradient of the above log-likelihood with respect to the j-th model parameter θj:

[0069] Where gji represents the gradient of the j-th model parameter corresponding to the i-th historical risk data; θj represents the j-th model parameter.

[0070] Backpropagation can be performed layer by layer using the chain rule. For example, when the model parameters are the weights in the weight matrix of the u-th layer... hour:

[0071] in, This represents the weight corresponding to the u-th layer; This represents the input corresponding to the m-th neuron in the u-th layer; This represents the activation value of the nth neuron in the (u-1)th layer.

[0072] Furthermore, the gradient corresponding to each historical risk data point is squared:

[0073] And average the squared gradients corresponding to multiple historical risk data:

[0074] Among them, F j This represents the parameter importance metric corresponding to the j-th model parameter; N represents the number of historical risk data points. This represents the squared gradient of the j-th model parameter corresponding to the i-th historical risk data.

[0075] In one implementation, the diagonal elements of the Fisher information matrix can be used to determine the Fisher information values ​​corresponding to the model parameters, and these Fisher information values ​​can be used as a measure of parameter importance for each model parameter. A higher parameter importance measure indicates a greater degree of importance for the corresponding model parameter in risk prediction based on historical risk data.

[0076] This approach can quantify the importance of each model parameter for historical risk prediction, providing a basis for subsequently implementing differentiated constraints on different model parameters.

[0077] Based on the above embodiments, in another embodiment provided in this disclosure, the triggering conditions for incremental training include at least one of the following: reaching a preset training period, the amount of new task data reaching a preset data amount threshold, and the risk prediction results of the risk prediction model experiencing a preset degree of distribution shift.

[0078] Specifically, embodiments of this disclosure can set various triggering conditions for incremental training.

[0079] For example, incremental training can be triggered according to a preset training cycle. The preset training cycle can be set according to actual needs, and this embodiment of the disclosure does not specifically limit it. For example, 30 minutes can be set as the preset training cycle, and incremental training can be triggered once each preset training cycle is reached.

[0080] For example, incremental training can be triggered based on a preset data volume threshold for new task data. The preset data volume threshold can be set according to actual needs, and this embodiment does not specifically limit it. For example, 1000 data entries can be set as the preset data volume threshold, and incremental training is triggered when the accumulated new task data reaches 1000 data entries.

[0081] For example, incremental training can be triggered based on whether the risk prediction results of the risk prediction model have undergone a preset degree of distribution shift. This can be configured according to actual needs, and this disclosure does not impose specific limitations. For instance, the risk score distribution obtained by the risk prediction model based on recent business data can be monitored in real time, and incremental training can be triggered when a significant shift in the risk score distribution occurs. For example, the KS test can be used to determine whether the risk prediction results have undergone a preset degree of distribution shift.

[0082] The aforementioned triggering conditions, such as the preset training period, the amount of new task data, and the distribution offset of risk prediction results, can be used individually or in combination.

[0083] This solution can trigger incremental training in a timely manner based on time, the amount of new task data, or changes in model performance, continuously optimizing the fraud detection rate and false positive rate, and improving the timeliness of the risk prediction model's response to new risk patterns.

[0084] Based on the above embodiments, in another embodiment provided in this disclosure, determining the parameter constraint degree of each model parameter according to the parameter importance metric of each model parameter and the parameter offset of each model parameter relative to before incremental training may include: Obtain the current parameter values ​​of each model during incremental training and the initial parameter values ​​before incremental training; Based on the current parameter values ​​and the initial parameter values, determine the degree of parameter shift of each model relative to the parameters before incremental training; The degree of parameter constraint for each model parameter is determined based on the parameter importance metric and the parameter offset; the greater the parameter importance metric and / or the greater the parameter offset, the greater the corresponding degree of parameter constraint.

[0085] Specifically, when performing incremental training based on new task data, the model parameters θ of the risk prediction model before incremental training can be retained. old The initial state is used as the model parameters, and this initial state is used to calculate the model output corresponding to the new task data. In one implementation, the key layer parameters can also be frozen to preserve the existing risk prediction knowledge of the risk prediction model.

[0086] This disclosure presents a two-layer loss function that adds a resilient constraint term based on Fisher information to the new task loss determined based on new task data. The resilient constraint term is based on the Fisher matrix F. j The regularization term penalizes behaviors that deviate from the important parameters of the old task. The total loss function is:

[0087]

[0088] Among them, L total L represents the total loss corresponding to incremental training; new This represents the loss of the new task determined based on the new task data; λ represents the regularization coefficient, which acts as an important control valve; F j θ represents the parameter importance metric corresponding to the j-th model parameter; j θ represents the current parameter value of the j-th model parameter during incremental training. old,j This represents the initial parameter value of the j-th model parameter before incremental training; (θ) j θ old,j ) 2 Used to characterize the degree of shift of the j-th model parameter relative to the parameter before incremental training.

[0089] Therefore, the greater the parameter importance metric and / or the greater the parameter offset, the stronger the parameter constraints on the corresponding model parameters.

[0090] In actual training, for highly important model parameters, their Fi is larger, which amplifies the offset penalty and thus strictly limits their update magnitude; for example, it can limit their deviation from the initial parameter values ​​to no more than 5%. For less important model parameters, their Fi is smaller, allowing for larger adjustments to quickly fit the data patterns of new tasks. The initial value of the regularization coefficient λ can be set to 500 / N, where N is the number of historical risk data points.

[0091] This scheme can determine parameter constraints by combining the importance of model parameters and the degree of parameter deviation, so that important model parameters can be more effectively protected and the historical risk knowledge can be forgotten.

[0092] Based on the above embodiments, in another embodiment provided in this disclosure, adjusting the parameter constraint degree of each model parameter according to the output difference of the risk prediction model for historical risk data and new task data may include: Obtain the first output probability distribution of the risk prediction model for historical risk data, and the second output probability distribution for new task data; The output difference is determined based on the first output probability distribution and the second output probability distribution; The degree of parameter constraint for each model parameter is adjusted based on the output difference; the greater the output difference, the smaller the degree of parameter constraint for each model parameter.

[0093] Specifically, the output differences can be characterized using KL divergence:

[0094] Where, p old This represents the first output probability distribution of the risk prediction model based on historical risk data; p new This represents the second output probability distribution of the risk prediction model based on historical risk data; D old Represents historical risk data; θ old θ represents the model parameters of the risk prediction model before incremental training. new This represents the model parameters of the risk prediction model after incremental training. Here, the model parameters θ of the risk prediction model before incremental training can be used. new Historical risk data D old Perform inference and calculate the second output probability distribution p. new .

[0095] Furthermore, the regularization coefficient λ, used to control the degree of constraint on model parameters, can be adjusted based on the output differences, thereby adjusting the parameter constraint degree of each model parameter. The formula for adjusting the regularization coefficient can be:

[0096] Where, λ new This represents the adjusted regularization coefficient, used to update the unadjusted regularization coefficient λ; λ init λ represents the initial value of the regularization coefficient, i.e., the unadjusted regularization coefficient; ε represents the preset adjustment coefficient threshold, which can be, for example, 0.1. The constraint is reduced when KL is large or the task differences are significant, and vice versa.

[0097] By adjusting the regularization coefficient, the constraint effect of the elastic constraint term on the update of each model parameter during incremental training can be changed, thereby achieving dynamic adjustment of the degree of constraint on each model parameter. Specifically, when the output difference between historical risk data and new task data is large, it indicates that the new risk pattern corresponding to the new task data is significantly different from the historical risk pattern. In this case, the constraint on the model parameters can be reduced, allowing for greater adjustment space and improving the risk prediction model's ability to learn new risk patterns. Conversely, when the output difference is small, the constraint on the model parameters can be strengthened to reduce changes in important model parameters during incremental training and maintain the risk prediction capability of the risk prediction model based on historical risk data.

[0098] Furthermore, after adjusting the parameter constraint degree of each model parameter according to the output difference, the model parameters can be updated according to the adjusted parameter constraint degree.

[0099] Specifically, the regularization coefficient λ is used to adjust the constraint strength of the elastic constraint term on the model parameter update. Therefore, adjusting the regularization coefficient will correspondingly change the constraint effect of the elastic constraint term in the model parameter update process.

[0100] For the j-th model parameter, the gradient of the elastic constraint term with respect to that model parameter is:

[0101] Here, λ represents the regularization coefficient adjusted based on the output differences.

[0102] Therefore, the adjusted regularization coefficient λ and the parameter importance metric F of the model parameters are... j and the degree of parameter shift of the model relative to the parameters before incremental training (θ) j θ old,j ) 2 These factors collectively influence the constraints on the model parameters. Given a fixed measure of parameter importance and parameter bias, adjusting the regularization coefficient can alter the degree of parameter constraint on the corresponding model parameters during this incremental training.

[0103] Furthermore, the gradient corresponding to the new task loss determined based on the new task data can be synthesized with the gradient corresponding to the elastic constraint term mentioned above to obtain the total gradient corresponding to incremental training:

[0104] in, θ L new This represents the gradient of the new task loss determined based on the new task data, relative to each model parameter; λF⊙(θ) θold ) represents the constraint gradient generated by the adjusted parameter constraints on the update of each model parameter; ⊙ represents element-wise multiplication.

[0105] Based on the total gradient, the parameters of each model can be iteratively updated:

[0106] Where, θ t θ represents the model parameters corresponding to the t-th iteration; t+1 η represents the model parameters corresponding to the (t+1)th iteration; η represents the learning rate.

[0107] Therefore, the impact of output differences on model parameter updates can be reflected by the degree of parameter constraint: adjusting the regularization coefficient based on the output difference between the first and second output probability distributions, thereby adjusting the constraint effect of the elastic constraint term on model parameter updates, and updating each model parameter under the gradient corresponding to the new task loss and the adjusted constraint effect. In this way, the update magnitude of model parameters during incremental training can be adaptively controlled according to the degree of change of the new risk pattern relative to the historical risk pattern. (Risk prediction model, Fisher information matrix F, and first output probability distribution P are mentioned.) old The initial "memory" formed by these two elements serves as the foundation for subsequent incremental learning.

[0108] For example, the learning rate can be set to 1 / 10 of the normal learning rate to mitigate the forgetting of historical risk prediction capabilities. The changes in model parameters and the corresponding parameter importance metrics can satisfy the following relationship:

[0109] Therefore, for model parameters with higher parameter importance, relatively strong constraints can be imposed to reduce their changes during incremental training; for model parameters with lower parameter importance, relatively large adjustment space can be reserved to learn new risk patterns corresponding to new task data.

[0110] This approach can dynamically adjust parameter constraints based on the output differences between historical risk data and new task data, thereby improving the model's adaptability in maintaining historical knowledge and learning new risk patterns.

[0111] Based on the above embodiments, in another embodiment provided in this disclosure, the method may further include: Based on the new task data, determine the current parameter importance measure for each model parameter in the updated risk prediction model; The current parameter importance measure and the parameter importance measure before incremental training of each model parameter are weighted to obtain the updated parameter importance measure of each model parameter in the updated risk prediction model.

[0112] Specifically, after completing one round of incremental training based on the new task data, the model parameters in the risk prediction model have been updated according to the new task data. In order to ensure that the parameter importance measures of each model parameter can reflect the new risk patterns corresponding to the new task data, the current parameter importance measures of each model parameter in the updated risk prediction model can be redefined based on the new task data.

[0113] New task data can be input into the updated risk prediction model. The output probability distribution corresponding to the new task data is obtained through forward propagation. Based on this output probability distribution, the gradients corresponding to each model parameter are determined through backpropagation. Furthermore, the gradients corresponding to multiple new task data points are squared and averaged. The resulting average squared gradient is used as a measure of the current parameter importance of the corresponding model parameter.

[0114] The current parameter importance metric characterizes the significance of the corresponding model parameters for risk prediction based on new task data after completing this round of incremental training. A higher current parameter importance metric indicates greater importance of the corresponding model parameters to the new risk patterns represented by the new task data.

[0115] Furthermore, to avoid directly overwriting the parameter importance information that existed before incremental training by determining the parameter importance metric solely based on the data from the new task in this round, the current parameter importance metric of each model parameter can be weighted with the parameter importance metric corresponding to the previous parameter importance metric before incremental training, so as to retain the parameter importance information corresponding to both historical risk patterns and new risk patterns.

[0116] The formula for incremental updates of Fisher information is:

[0117] Among them, F new F represents the updated parameter importance measure for the corresponding model parameters in the updated risk prediction model; old This represents the parameter importance measure of the model parameters before this incremental training; the average of the squared gradients represents the current parameter importance measure determined based on the new task data in this round; γ represents the decay factor, used to adjust the influence of the parameter importance measures before incremental training and the current parameter importance measures on the updated parameter importance measures; N represents the amount of new task data used to determine the current parameter importance measure; x i Let y represent the input feature corresponding to the i-th new task data. i This represents the risk label corresponding to the i-th new task data.

[0118] For example, the decay factor γ can be set to 0.9. By setting the decay factor, the parameter importance information reflected by the new task data in this round can be integrated into the updated parameter importance metric while retaining the parameter importance information before incremental training.

[0119] Therefore, the current parameter importance metric is different from the updated parameter importance metric. The current parameter importance metric mainly reflects the importance of the corresponding model parameters to the new task data in this round; the updated parameter importance metric is the result of weighting the parameter importance metric before incremental training and the current parameter importance metric, and is used to comprehensively reflect the importance of the corresponding model parameters in the continuous incremental training process.

[0120] Furthermore, after obtaining the updated parameter importance metrics for each model parameter, these updated metrics can be used as the parameter importance metrics before the next round of incremental training, in order to determine the degree of parameter constraint for the corresponding model parameters during the next round of incremental training. Thus, with the continuous acquisition of new task data and the continuous incremental training of the risk prediction model, the parameter importance metrics for each model parameter can be dynamically updated according to changes in new risk patterns.

[0121] In one implementation, multiple model parameters can be filtered based on the updated parameter importance metric to reduce the computational cost of parameter constraints in subsequent incremental training. For example, the parameters can be filtered according to their Fisher information values ​​(i.e., parameter importance metrics), retaining only the parameter constraints corresponding to the top k% of model parameters with Fisher information values, while setting the Fisher information values ​​of the remaining model parameters to 0.

[0122] For example, k can be 30%, and the filtering method for parameter importance metrics is as follows:

[0123] Among them, F j Let F represent the parameter importance metric corresponding to the j-th model parameter; Percentile(F,70%) represents the 70th percentile of the parameter importance metrics for multiple model parameters. Therefore, we can retain the model parameters whose parameter importance metrics rank in the top 30% and constrain their parameter updates in subsequent incremental training, while reducing the computation of corresponding parameter constraints for the remaining model parameters.

[0124] By using the above method, the parameter importance metrics before incremental training and the parameter importance metrics corresponding to the new task data in this round can be integrated, so that the parameter importance metrics can be dynamically updated with the new risk model, and the parameter constraints in subsequent incremental training can adapt to changes as the model continues to train.

[0125] In one alternative implementation, embodiments of this disclosure can further protect gradients and Fisher information through differential privacy to reduce the risk of sensitive information leakage during cross-agency or multi-source data processing. Furthermore, interpretability analysis can be performed on the risk prediction results. For example, the Shapley Additive (SHAP) value can be used to determine the contribution of different risk characteristics to the risk prediction results, thereby improving the interpretability of the risk prediction results and meeting regulatory audit requirements.

[0126] The adaptive dynamic risk prediction method provided in this disclosure has the following beneficial effects: 1. Balancing the preservation of historical risk patterns with the learning of new risk patterns enhances the continuous adaptability of risk prediction models. By obtaining parameter importance metrics for each model parameter and determining the degree of parameter constraint based on the parameter shift relative to the pre-incremental training state, model parameters that are more important for historical risk prediction are subject to stronger constraints, while less important model parameters have greater adjustment space. This approach reduces the forgetting of historical risk patterns while learning new risk patterns using new task data, thus balancing the risk prediction model's ability to predict both historical and new risk patterns.

[0127] 2. It can adaptively adjust the update constraints of model parameters based on the differences between old and new risk patterns, improving the flexibility of incremental model training. By obtaining the output probability distribution of the risk prediction model for historical risk data and new task data, the parameter constraints of each model are dynamically adjusted according to the differences between the two outputs. When the difference between the new risk pattern and the historical risk pattern is large, the parameter constraints are appropriately reduced to improve the model's learning ability for the new risk pattern; when the difference is small, the parameter constraints are appropriately strengthened to further maintain the model's existing risk prediction ability, thereby achieving adaptive adjustment of the model parameter constraint degree.

[0128] 3. It enables continuous updates to parameter importance metrics as risk patterns change, improving the accuracy of subsequent parameter constraints. By determining the current parameter importance metrics of each model parameter in the updated risk prediction model based on new task data, and weighting these current metrics with those before incremental training, it can retain parameter importance information corresponding to historical risk patterns while integrating parameter importance information corresponding to new risk patterns. This allows the parameter importance metrics of each model parameter to be dynamically updated as the model continues to undergo incremental training, thus providing a more accurate basis for parameter constraints in subsequent incremental training.

[0129] 4. It can reduce the constraint calculations for non-critical model parameters, thereby lowering the computational cost during incremental training. By filtering parameters based on their importance metrics, only the parameter constraints for model parameters with higher importance metrics are retained, while reducing or eliminating the parameter constraints for model parameters with lower importance metrics. This prioritizes the protection of model parameters that are more important for historical risk prediction while reducing unnecessary parameter constraint calculations, thus lowering the computational cost during incremental training and improving model update efficiency.

[0130] 5. The system can promptly initiate incremental training based on the model's operational status, improving its responsiveness to continuously changing risk patterns. By using at least one of the following as trigger conditions for incremental training—reaching a preset training period, the amount of new task data reaching a preset data volume threshold, or a preset degree of distribution shift in risk prediction results—incremental training can be initiated promptly based on time, the accumulation of new task data, or changes in the model's risk prediction results. This allows the risk prediction model to learn newly emerging risk patterns in a timely manner, improving its responsiveness to continuously changing risk environments.

[0131] The foregoing mainly describes the solutions provided by the embodiments of this disclosure. It is understood that, in order to achieve the above functions, the electronic device includes hardware structures and / or software modules corresponding to the execution of each function. Those skilled in the art should readily recognize that, based on the units and algorithm steps of the various examples described in conjunction with the embodiments disclosed herein, this disclosure can be implemented in hardware or a combination of hardware and computer software. Whether a function is executed in hardware or by computer software driving hardware depends on the specific application and design constraints of the technical solution. Those skilled in the art can use different methods to implement the described functions for each specific application, but such implementation should not be considered beyond the scope of this disclosure.

[0132] This disclosure embodiment can divide the electronic device into functional units according to the above method example. For example, each function can be divided into a separate functional module, or two or more functions can be integrated into one processing module. The integrated module can be implemented in hardware or as a software functional module. It should be noted that the module division in this disclosure embodiment is illustrative and only represents one logical functional division; other division methods may be used in actual implementation.

[0133] By dividing each functional module according to its corresponding function, an exemplary embodiment of this disclosure provides an adaptive dynamic risk prediction device, which can be a server or a chip applied to a server. Figure 2 A schematic diagram of the structure of an adaptive dynamic risk prediction device provided in an exemplary embodiment of this disclosure is shown. Figure 2 As shown, the device 200 includes: The acquisition module 201 is used to acquire multiple model parameters of the risk prediction model trained based on historical risk data, and to acquire the parameter importance metric corresponding to each model parameter; wherein, the parameter importance metric is used to characterize the importance of the corresponding model parameter for risk prediction based on historical risk data; The acquisition module 201 is also used to acquire new task data corresponding to the new risk pattern; Training module 202 is used to incrementally train the risk prediction model based on new task data. It determines the parameter constraint degree of each model parameter according to the parameter importance metric and the parameter offset degree of each model parameter relative to the parameter before incremental training. Based on the output difference of the risk prediction model for historical risk data and new task data, it adjusts the parameter constraint degree of each model parameter and updates each model parameter according to the adjusted parameter constraint degree and the new task data to obtain the updated risk prediction model. The prediction module 203 is used to perform risk prediction on the risk data to be identified using the updated risk prediction model, and obtain the risk prediction result.

[0134] In some embodiments, the acquisition module 201 is further configured to input historical risk data into the risk prediction model to obtain the output probability distribution corresponding to the historical risk data; determine the gradient corresponding to each model parameter based on the output probability distribution; square the gradient corresponding to each model parameter; and determine the parameter importance metric corresponding to each model parameter based on the average of the squared gradients corresponding to multiple historical risk data.

[0135] In some embodiments, the triggering conditions for incremental training include at least one of the following: reaching a preset training period, the amount of new task data reaching a preset data amount threshold, and the risk prediction results of the risk prediction model experiencing a preset degree of distribution shift.

[0136] In some embodiments, the training module 202 is further configured to obtain the current parameter values ​​of each model parameter during incremental training and the initial parameter values ​​before incremental training; determine the degree of parameter offset of each model parameter relative to before incremental training based on the current parameter values ​​and the initial parameter values; and determine the degree of parameter constraint of each model parameter based on the parameter importance metric and the degree of parameter offset; wherein, the greater the parameter importance metric and / or the greater the degree of parameter offset, the greater the corresponding degree of parameter constraint.

[0137] In some embodiments, the training module 202 is further configured to obtain a first output probability distribution of the risk prediction model for historical risk data and a second output probability distribution for new task data; determine the output difference based on the first output probability distribution and the second output probability distribution; and adjust the parameter constraint degree of each model parameter based on the output difference; wherein, the greater the output difference, the smaller the parameter constraint degree of each model parameter.

[0138] In some embodiments, the training module 202 is further configured to determine the current parameter importance measure of each model parameter in the updated risk prediction model based on the new task data; and to perform weighted processing on the current parameter importance measure of each model parameter and the parameter importance measure before incremental training to obtain the updated parameter importance measure corresponding to each model parameter in the updated risk prediction model.

[0139] This disclosure also provides an electronic device, including a memory, a processor, and a computer program stored in the memory, wherein the processor executes the computer program to implement the steps of the methods disclosed in this disclosure.

[0140] Figure 3 A schematic diagram of the structure of an electronic device provided in an exemplary embodiment of this disclosure is shown. For example... Figure 3 As shown, the electronic device 300 includes at least one processor 301 and a memory 302 coupled to the processor 301, which can perform the corresponding steps in the methods disclosed in the embodiments of this disclosure.

[0141] The processor 301 described above can also be called a Central Processing Unit (CPU), which can be an integrated circuit chip with signal processing capabilities. Each step in the method disclosed in this embodiment can be implemented by the integrated logic circuitry in the processor 301 or by software instructions. The processor 301 can be a general-purpose processor, a digital signal processor (DSP), an ASIC, a field-programmable gate array (FPGA), or other programmable logic devices, discrete gate or transistor logic devices, or discrete hardware components. The general-purpose processor can be a microprocessor or any conventional processor. The steps of the method disclosed in this embodiment can be directly implemented by a hardware decoding processor, or by a combination of hardware and software modules in the decoding processor. The software modules can be located in the memory 302, such as random access memory, flash memory, read-only memory, programmable read-only memory, electrically erasable programmable memory, registers, or other mature storage media in the art. The processor 301 reads information from the memory 302 and, in conjunction with its hardware, completes the steps of the method described above.

[0142] Furthermore, various operations / processes according to this disclosure, implemented via software and / or firmware, can be transmitted from a storage medium or network to a computer system with a dedicated hardware architecture, for example, Figure 4 The computer system 400 shown is equipped with the programs that constitute the software. When various programs are installed, the computer system is able to perform various functions, including functions such as those described above. Figure 4 A schematic diagram of the structure of a computer system provided in an exemplary embodiment of this disclosure is shown.

[0143] Computer system 400 is intended to represent various forms of digital electronic computer devices, such as laptop computers, desktop computers, workstations, personal digital assistants, servers, blade servers, mainframe computers, and other suitable computers. Electronic devices may also represent various forms of mobile devices, such as personal digital assistants, cellular phones, smartphones, wearable devices, and other similar computing devices. The components shown herein, their connections and relationships, and their functions are merely illustrative and are not intended to limit the implementation of the present disclosure described and / or claimed herein.

[0144] like Figure 4As shown, the computer system 400 includes a computing unit 401, which can perform various appropriate actions and processes based on a computer program stored in a read-only memory (ROM) 402 or a computer program loaded from a storage unit 408 into a random access memory (RAM) 403. The RAM 403 may also store various programs and data required for the operation of the computer system 400. The computing unit 401, ROM 402, and RAM 403 are interconnected via a bus 404. An input / output (I / O) interface 405 is also connected to the bus 404.

[0145] Multiple components in the computer system 400 are connected to the I / O interface 405, including: an input unit 406, an output unit 407, a storage unit 408, and a communication unit 409. The input unit 406 can be any type of device capable of inputting information into the computer system 400. The input unit 406 can receive input numerical or character information and generate key signal inputs related to user settings and / or function control of the electronic device. The output unit 407 can be any type of device capable of presenting information and may include, but is not limited to, a monitor, speaker, video / audio output terminal, vibrator, and / or printer. The storage unit 408 may include, but is not limited to, a hard disk and an optical disk. The communication unit 409 allows the computer system 400 to exchange information / data with other devices via a network such as the Internet, and may include, but is not limited to, a modem, network card, infrared communication device, wireless communication transceiver, and / or chipset, such as Bluetooth™ device, WiFi device, WiMax device, cellular communication device, and / or the like.

[0146] The computing unit 401 can be a variety of general-purpose and / or special-purpose processing components with processing and computing capabilities. Some examples of the computing unit 401 include, but are not limited to, a central processing unit (CPU), a graphics processing unit (GPU), various special-purpose artificial intelligence (AI) computing chips, various computing units running machine learning model algorithms, a digital signal processor (DSP), and any suitable processor, controller, microcontroller, etc. The computing unit 401 performs the various methods and processes described above. For example, in some embodiments, the methods disclosed in this disclosure can be implemented as a computer software program tangibly contained in a machine-readable medium, such as storage unit 408. In some embodiments, part or all of the computer program can be loaded and / or installed on an electronic device via ROM 402 and / or communication unit 409. In some embodiments, the computing unit 401 can be configured to perform the methods disclosed in this disclosure by any other suitable means (e.g., by means of firmware).

[0147] This disclosure also provides a computer-readable storage medium, wherein when the instructions in the computer-readable storage medium are executed by a processor of an electronic device, the electronic device is able to perform the methods disclosed in this disclosure.

[0148] The computer-readable storage medium in this disclosure can be a tangible medium that may contain or store a program for use by or in conjunction with an instruction execution system, apparatus, or device. The aforementioned computer-readable storage medium may include, but is not limited to, electronic, magnetic, optical, electromagnetic, infrared, or semiconductor systems, apparatus, or devices, or any suitable combination of the foregoing. More specifically, the aforementioned computer-readable storage medium may include electrical connections based on 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 devices, magnetic storage devices, or any suitable combination of the foregoing.

[0149] 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.

[0150] This disclosure also provides a computer program product, including a computer program, wherein when the computer program is executed by a processor, it implements the methods disclosed in the embodiments of this disclosure.

[0151] In embodiments of this disclosure, computer program code for performing the operations of this disclosure can be written in one or more programming languages ​​or a combination thereof. These programming languages ​​include, but are not limited to, object-oriented programming languages ​​such as Java, Smalltalk, and C++, as well as conventional procedural programming languages ​​such as the "C" language or similar programming languages. The program code can be executed entirely on the user's computer, partially on the user's computer, as a standalone software package, partially on the user's 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 user's computer via any type of network (including a local area network (LAN) or a wide area network (WAN)), or it can be connected to an external computer.

[0152] 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.

[0153] The modules, components, or units described in the embodiments of this disclosure can be implemented in software or hardware. The names of the modules, components, or units do not necessarily constitute a limitation on the module, component, or unit itself.

[0154] The functions described above in this document can be performed at least in part by one or more hardware logic components. For example, without limitation, exemplary hardware logic components that can be used include: field-programmable gate arrays (FPGAs), application-specific integrated circuits (ASICs), application-specific standard products (ASSPs), system-on-a-chip (SoCs), complex programmable logic devices (CPLDs), and so on.

[0155] The above description is merely an illustration of some embodiments of this disclosure and 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.

[0156] While specific embodiments of this disclosure have been described in detail by way of example, those skilled in the art should understand that the examples are for illustrative purposes only and not intended to limit the scope of this disclosure. Those skilled in the art should understand that modifications can be made to the above embodiments without departing from the scope and spirit of this disclosure. The scope of this disclosure is defined by the appended claims.

Claims

1. An adaptive dynamic risk prediction method, characterized in that, include: Obtain multiple model parameters of a risk prediction model trained based on historical risk data, and obtain the parameter importance metric corresponding to each model parameter; wherein, the parameter importance metric is used to characterize the importance of the corresponding model parameter for risk prediction based on the historical risk data; Acquire new task data corresponding to the new risk pattern, perform incremental training on the risk prediction model based on the new task data, determine the parameter constraint degree of each model parameter according to the parameter importance metric of each model parameter and the parameter offset degree of each model parameter relative to the previous incremental training; adjust the parameter constraint degree of each model parameter according to the output difference of the risk prediction model for the historical risk data and the new task data, and update each model parameter according to the adjusted parameter constraint degree and the new task data to obtain the updated risk prediction model; The updated risk prediction model is used to predict the risk data to be identified, and the risk prediction results are obtained.

2. The method as described in claim 1, characterized in that, The step of obtaining the parameter importance metric corresponding to each of the model parameters includes: The historical risk data is input into the risk prediction model to obtain the output probability distribution corresponding to the historical risk data; Based on the output probability distribution, determine the gradient corresponding to each of the model parameters; The gradients corresponding to each of the model parameters are squared, and the parameter importance metric corresponding to each of the model parameters is determined based on the average of the squared gradients corresponding to multiple historical risk data.

3. The method as described in claim 1, characterized in that, The triggering conditions for incremental training include at least one of the following: reaching a preset training period, the amount of new task data reaching a preset data amount threshold, and the risk prediction results of the risk prediction model experiencing a preset degree of distribution shift.

4. The method of claim 1, wherein determining the parameter constraint degree of each model parameter based on the parameter importance metric of each model parameter and the parameter offset of each model parameter relative to the period before incremental training includes: Obtain the current parameter values ​​of each model parameter during the incremental training and the initial parameter values ​​before the incremental training; Based on the current parameter values ​​and the initial parameter values, determine the degree of parameter shift of each model parameter relative to the parameters before the incremental training; The degree of parameter constraint for each model parameter is determined based on the parameter importance metric and the parameter offset; wherein, the greater the parameter importance metric and / or the greater the parameter offset, the greater the corresponding degree of parameter constraint.

5. The method as described in claim 1, characterized in that, The step of adjusting the parameter constraint degree of each model parameter based on the output difference of the risk prediction model for the historical risk data and the new task data includes: Obtain the first output probability distribution of the risk prediction model for the historical risk data, and the second output probability distribution for the new task data; The output difference is determined based on the first output probability distribution and the second output probability distribution; The degree of parameter constraint of each model parameter is adjusted according to the output difference; wherein, the greater the output difference, the smaller the degree of parameter constraint of each model parameter.

6. The method according to any one of claims 1 to 5, characterized in that, The method further includes: Based on the new task data, determine the current parameter importance measure for each model parameter in the updated risk prediction model; The current parameter importance measure of each model parameter and the parameter importance measure before the incremental training are weighted to obtain the updated parameter importance measure corresponding to each model parameter in the updated risk prediction model.

7. An adaptive dynamic risk prediction device, characterized in that, include: The acquisition module is used to acquire multiple model parameters of a risk prediction model trained based on historical risk data, and to acquire the parameter importance metric corresponding to each model parameter; wherein, the parameter importance metric is used to characterize the importance of the corresponding model parameter for risk prediction based on the historical risk data; The acquisition module is also used to acquire new task data corresponding to the new risk model; The training module is used to incrementally train the risk prediction model based on the new task data. It determines the parameter constraint degree of each model parameter based on the parameter importance metric and the parameter offset degree of each model parameter relative to the period before incremental training. Based on the output difference of the risk prediction model for the historical risk data and the new task data, it adjusts the parameter constraint degree of each model parameter and updates the model parameters based on the adjusted parameter constraint degree and the new task data to obtain the updated risk prediction model. The prediction module is used to perform risk prediction on the risk data to be identified using the updated risk prediction model, and obtain the risk prediction result.

8. An electronic device comprising a memory, a processor, and a computer program stored in the memory, characterized in that, The processor executes the computer program to implement the steps of the method according to any one of claims 1 to 6.

9. A computer-readable storage medium having a computer program stored thereon, characterized in that, When executed by a processor, the computer program implements the steps of the method described in any one of claims 1 to 6.

10. A computer program product, comprising a computer program, characterized in that, When executed by a processor, the computer program implements the steps of the method described in any one of claims 1 to 6.

Images