Self-learning method, system and device of financial risk assessment model, and medium

Abstract

The invention belongs to the technical field of financial risk control, and particularly relates to a self-learning method, system and device for a financial risk assessment model, and a medium, and the method comprises the steps: generating a scheduling instruction for optimizing the financial risk assessment model, and updating the financial risk assessment model according to the scheduling instruction; based on the first instruction, adjusting parameters of a financial risk assessment model according to one or more pieces of financial transaction data updated within preset first time; on the basis of the second instruction, pushing abnormal transaction data accumulated in preset second time to an artificial expert for labeling, and retraining the financial risk assessment model according to the manually labeled abnormal transaction data; and based on the third instruction, simulating and generating virus transaction data, and retraining the financial risk assessment model according to the virus transaction data. According to the method, the accuracy and reliability of the financial risk assessment model are optimized by cooperatively operating three model optimization strategies of online learning, sample calibration and adversarial enhancement.

Description

Technical Field

[0001] This invention belongs to the field of financial risk control technology, specifically relating to a self-learning method, system, device, and medium for a financial risk assessment model. Background Technology

[0002] In the field of financial risk control, machine learning models and a large number of historical samples are used to build financial risk assessment models in order to quantify the risk status of financial transaction data, thereby effectively predicting the risk probability of financial transaction data.

[0003] However, most machine learning models suffer from problems such as model drift, over-reliance on historical samples, and insufficient ability to identify unknown risks when dealing with dynamic financial risks. Summary of the Invention

[0004] In view of the shortcomings of the prior art described above, the purpose of this invention is to provide a self-learning method for financial risk assessment models, which integrates three strategies: online real-time adaptation, active sample learning, and offline adversarial enhancement, in order to optimize the accuracy and reliability of financial risk assessment models and improve their predictive ability for new and abnormal risks.

[0005] This invention provides a self-learning method for a financial risk assessment model, comprising: generating a scheduling instruction for optimizing the financial risk assessment model, and updating the financial risk assessment model according to the scheduling instruction; adjusting the parameters of the financial risk assessment model based on a first instruction and one or more financial transaction data updated within a preset first time period; pushing abnormal transaction data accumulated within a preset second time period to human experts for annotation based on a second instruction, and retraining the financial risk assessment model based on the annotated abnormal transaction data; and simulating and generating virus transaction data based on a third instruction, and retraining the financial risk assessment model based on it.

[0006] In one embodiment of the present invention, the step of generating scheduling instructions for optimizing the financial risk assessment model includes: continuously generating the first instruction at a preset first frequency; continuously generating the second instruction at a preset second frequency; and continuously generating the third instruction at a preset third frequency; wherein the first frequency is higher than the second frequency, and the second frequency is higher than the third frequency.

[0007] In one embodiment of the present invention, the step of adjusting the parameters of the financial risk assessment model includes: incrementally updating the financial risk assessment model using the FTRL algorithm.

[0008] In one embodiment of the present invention, the financial risk assessment model is constructed based on the Wide & Deep model, and the step of incrementally updating the financial risk assessment model using the FTRL algorithm includes: sequentially inputting one or more financial transaction data updated within a preset first time period into the financial risk assessment model, and calculating a weighted cross-entropy loss function based on the output result of the financial risk assessment model and the true label of the corresponding financial transaction data; calculating gradients for the weight parameters of the Wide sub-model layer and the output layer parameters of the financial risk assessment model according to the weighted cross-entropy loss function, and accumulating historical gradients and corresponding regularization constraints using the FTRL algorithm to update the parameters of the Wide sub-model and the parameters of the output layer; and updating the parameters of the last layer of the Deep sub-model in the financial risk assessment model according to the weighted cross-entropy loss function.

[0009] In one embodiment of the present invention, the method further includes: updating the financial risk assessment model according to the cumulative number of abnormal transaction data, so that when the cumulative number of abnormal transaction data exceeds a preset first threshold, the accumulated abnormal transaction data is pushed to human experts for annotation, and the financial risk assessment model is retrained based on the human-annotated abnormal transaction data.

[0010] In one embodiment of the present invention, the step of identifying whether each piece of financial transaction data is abnormal transaction data includes: determining the uncertainty value corresponding to the financial transaction data according to an evaluation function, and classifying financial transaction data with uncertainty values ​​exceeding a preset threshold as abnormal transaction data; wherein, the formula of the evaluation function is as follows: ; This indicates the uncertainty value of the financial transaction data. This represents the risk assessment value of the financial transaction data by the financial risk assessment model, and the normalized risk assessment value is (0, 1).

[0011] In one embodiment of the present invention, the step of pushing accumulated abnormal transaction data to human experts for annotation and retraining the financial risk assessment model based on the annotated abnormal transaction data includes: for each abnormal transaction data, identifying whether it has been manually annotated; if yes, stopping the push; if no, pushing it to human experts and waiting for their feedback; integrating all manually annotated abnormal transaction data and retraining the financial risk assessment model based on it.

[0012] In one embodiment of the present invention, the financial risk assessment model is constructed based on the Wide & Deep model, and the step of retraining the financial risk assessment model based on manually labeled abnormal transaction data includes: sequentially inputting the manually labeled abnormal transaction data into the financial risk assessment model, and calculating a weighted cross-entropy loss function based on the output of the financial risk assessment model and the true labels of the corresponding abnormal transaction data; updating the weight parameters of each level of the Wide sub-model in the financial risk assessment model according to the weighted cross-entropy loss function; and updating the parameters of the last layer or the last two layers of the Deep sub-model in the financial risk assessment model according to the weighted cross-entropy loss function.

[0013] In one embodiment of the present invention, the method further includes: updating the financial risk assessment model according to the performance evaluation index of the financial risk assessment model, so as to simulate and generate virus transaction data when the performance evaluation index of the financial risk assessment model is lower than a preset second threshold, and retraining the financial risk assessment model based on the data.

[0014] In one embodiment of the present invention, the step of simulating and generating virus transaction data includes: generating initial virus transaction data by simulating random noise based on abnormal transaction data; and obtaining final virus transaction data by conducting multiple rounds of adversarial game on the initial virus transaction data.

[0015] In one embodiment of the present invention, the financial risk assessment model is constructed based on the Wide & Deep model, and the step of retraining the financial risk assessment model according to the virus transaction data includes: sequentially inputting the simulated virus transaction data into the financial risk assessment model, and calculating a first loss function based on the output of the financial risk assessment model and the real labels of the corresponding virus transaction data; updating the last layer parameters of the Wide sub-model and the last layer parameters of the Deep sub-model in the financial risk assessment model according to the first loss function; re-inputting the virus transaction data into the updated financial risk assessment model, and calculating a second loss function based on the output of the financial risk assessment model and the real labels of the virus transaction data; and updating the penultimate and penultimate layer parameters of the Deep sub-model in the financial risk assessment model according to the second loss function.

[0016] The present invention also provides a self-learning system for a financial risk assessment model, comprising: a central scheduling module, used to generate scheduling instructions for optimizing the financial risk assessment model, and to update the financial risk assessment model according to the scheduling instructions; a model update module, used to adjust the parameters of the financial risk assessment model based on a first instruction and one or more financial transaction data updated within a preset first time period; based on a second instruction, to push abnormal transaction data accumulated within a preset second time period to human experts for annotation, and to retrain the financial risk assessment model based on the annotated abnormal transaction data; and based on a third instruction, to simulate and generate virus transaction data, and to retrain the financial risk assessment model based on it.

[0017] The present invention also provides an electronic device, including a processor coupled to a memory, the memory storing program instructions, which, when executed by the processor, implement the method described above.

[0018] The present invention also provides a computer-readable storage medium including a program that, when run on a computer, causes the computer to perform the methods described above.

[0019] The beneficial effects of this invention are: by using three model optimization strategies—cooperative online learning, sample calibration, and adversarial enhancement—to optimize the accuracy and reliability of the financial risk assessment model and improve its predictive ability for novel and abnormal risks. Attached Figure Description

[0020] The accompanying drawings, which are incorporated in and form part of this specification, illustrate embodiments consistent with this application and, together with the description, serve to explain the principles of this application. It is obvious that the drawings described below are merely some embodiments of this application, and those skilled in the art can obtain other drawings based on these drawings without any inventive effort.

[0021] Figure 1 This is a flowchart illustrating a specific embodiment of a self-learning method for a financial risk assessment model provided in one embodiment of the present invention. Figure 2 This is a self-learning method for a financial risk assessment model provided in one embodiment of the present invention; a flowchart of another specific embodiment; Figure 3 This is a flowchart illustrating a specific embodiment of simulating and generating virus transaction data provided in one embodiment of the present invention; Figure 4 This is a schematic diagram of a specific embodiment of a self-learning system for a financial risk assessment model provided in one embodiment of the present invention; Figure 5This is a structural block diagram of a specific embodiment of an electronic device provided in one embodiment of the present invention. Detailed Implementation

[0022] To facilitate understanding of this application, a more complete description will be provided below with reference to the accompanying drawings, which illustrate embodiments of the present application. However, the present application can be implemented in many different forms and is not limited to the embodiments described herein. Rather, these embodiments are provided so that the disclosure of this application will be thorough and complete.

[0023] Unless otherwise defined, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this application belongs. The terminology used herein is for the purpose of describing particular embodiments only and is not intended to be limiting of the application.

[0024] The following specific examples illustrate the implementation of the present invention. Those skilled in the art can easily understand other advantages and effects of the present invention from the content disclosed in this specification. The present invention can also be implemented or applied through other different specific embodiments, and various details in this specification can also be modified or changed based on different viewpoints and applications without departing from the spirit of the present invention. It should be noted that, unless otherwise specified, the following embodiments and features described therein can be combined with each other.

[0025] In the following description, numerous details are explored to provide a more thorough explanation of embodiments of the invention. However, it will be apparent to those skilled in the art that embodiments of the invention may be practiced without these specific details. In other embodiments, well-known structures and devices are shown in block diagram form rather than in detail to avoid obscuring embodiments of the invention.

[0026] Example 1 Please see Figure 1 , 2 As shown, a self-learning method for a financial risk assessment model includes: Step S100: Generate a scheduling instruction for optimizing the financial risk assessment model, and update the financial risk assessment model according to the scheduling instruction.

[0027] Step S210: Based on the first instruction, adjust the parameters of the financial risk assessment model according to one or more financial transaction data updated within a preset first time period.

[0028] Step S220: Based on the second instruction, the abnormal transaction data accumulated within a preset second time period is pushed to human experts for annotation, and the financial risk assessment model is retrained based on the annotated abnormal transaction data.

[0029] Step S230: Based on the third instruction, simulate and generate virus transaction data, and retrain the financial risk assessment model accordingly.

[0030] First, it should be noted that the following three strategies will be used in this embodiment to update and optimize the financial risk assessment model.

[0031] Strategy 1, High-Frequency Online Adaptation Loop, is an optimization strategy that combines real-time data processing, dynamic model updates, and closed-loop feedback mechanisms. It aims to achieve rapid response to environmental changes and continuous performance optimization. Through high-frequency online learning and incremental model adjustments, it can effectively maintain a certain degree of adaptability and robustness in scenarios with continuous data inflow.

[0032] Therefore, in this embodiment, the financial risk assessment model is kept agile by continuously tracking short-term market changes in the ever-growing financial transaction data. Specifically, in practical applications, this can be performed on each individual or batch (e.g., N transactions) of financial transaction data to update the financial risk assessment model in real time.

[0033] In this embodiment, the financial risk assessment model is updated in real-time with incremental, minute changes using the FTRL (Follow The Regularized Leader) online learning optimization algorithm. Specifically, the FTRL algorithm determines the current parameter update direction by accumulating historical gradient information and finds the optimal parameters through regularization constraints, thereby enabling the optimization and adjustment of model parameters.

[0034] In one specific embodiment, the financial risk assessment model is constructed using the Wide & Deep model. The corresponding financial risk assessment model includes a Wide sub-model, a Deep sub-model, and a final output layer. The high-frequency online incremental adaptation loop focuses more on optimizing and updating the Wide sub-model and the output layer of the financial risk assessment model.

[0035] Specifically, one or more real-time financial transaction data are sequentially input into the financial risk assessment model. The weighted cross-entropy loss function is calculated using the model's output and the corresponding real labels of the financial transaction data. Furthermore, gradients can be calculated for the weight parameters of some layers of the Wide sub-model and the parameters of the output layer based on this loss function. During this process, the FTRL algorithm is used to accumulate historical gradients and corresponding regularization constraints, thereby updating the parameters of the Wide sub-model and the output layer.

[0036] At the same time, this loss function can be used to make small updates to the parameters of the last layer of the Deep sub-model, while freezing the remaining layers to ensure stability.

[0037] Among them, regarding the learning rate of the FTRL algorithm The L1 regularization coefficient can be set within the range of [0.001, 0.05]. It can be set in the range of [0.1, 1.0] to ensure the stability of model optimization and updates.

[0038] Strategy 2, the Mid-Frequency Active Calibration Loop, is an adaptive calibration optimization strategy that combines mid-frequency signal processing, active sample generation, and closed-loop feedback mechanisms. It aims to achieve high-precision calibration in the mid-frequency range by periodically and actively acquiring sample data and dynamically adjusting parameters.

[0039] Therefore, in this embodiment, it is used to periodically calibrate the financial risk assessment model to prevent model drift caused by continuous online learning. Specifically, in practical applications, it can be executed periodically, for example, every 1 to 4 hours, to update the financial risk assessment model using accumulated high-quality manually labeled samples.

[0040] High-quality manually labeled samples can be obtained in the following way: Firstly, it is understandable that as financial transaction data is continuously generated, each financial transaction will be input into the financial risk assessment model to predict the probability of risk. However, for some financial transaction data that the financial risk assessment model is most "unsure" about, it can be used as abnormal transaction data to retrain the model and force the model to "regress" to the correct knowledge.

[0041] To address this, for each financial transaction, its uncertainty value can be calculated to determine whether it constitutes an abnormal transaction. Specifically, please refer to the following evaluation function: ; Among these, This indicates the uncertainty value of the financial transaction data. This represents the risk assessment value of the financial risk assessment model for this financial transaction data, i.e., the predicted risk probability, and the parameter values ​​output by the financial risk assessment model, after normalization, belong to the range (0, 1).

[0042] Therefore, by calculating the uncertainty value of each financial transaction and comparing it with a preset threshold, abnormal transaction data can be identified. For example, financial transaction data with uncertainty values ​​within the range of [0.4, 0.6] are considered abnormal transaction data, thus filtering out all abnormal transaction data. Secondly, the selected abnormal transaction data is pushed to human experts for labeling. Based on the human-labeled abnormal transaction data, i.e. high-quality human-labeled samples, the financial risk assessment model is retrained with a high learning weight, forcing the model to "regress" to the correct knowledge.

[0043] In one specific embodiment, the financial risk assessment model constructed by the Wide & Deep model is still used as an example, while the mid-frequency active sample calibration loop focuses more on fine-tuning the Wide sub-model and the Deep sub-model.

[0044] Specifically, the training samples are a mixture of high-quality manually labeled samples and ordinary samples, i.e., real-time financial transaction data. A small amount of historical samples can also be mixed in, forming a replay buffer to prevent the model from "forgetting historical patterns." Correspondingly, the mixed sample data is sequentially input into the financial risk assessment model, and the results output by the financial risk assessment model are used to calculate a class-weighted cross-entropy loss function based on the true labels of the corresponding sample data. Based on this loss function, the Wide sub-model can be fully updated, i.e., the weight parameters of each layer of the Wide sub-model are optimized. For the Deep sub-model, the layer is frozen to the second-to-last layer, with only the last layer or the last two layers being fine-tuned.

[0045] The parameters used to train the model using sample data are shown below: The batch size for each parameter update can be set between 512 and 2048. The number of epochs the model traverses the entire dataset can be set between 1 and 3. The optimizer for the parameter update method and step size can be Adam with a learning rate of 1e-4. The model is trained based on an early stopping strategy to calculate the corresponding loss function. The loss function is not limited to cross-entropy loss with class weights; Focal Loss (γ=1–2) can also be used.

[0046] Therefore, it can be seen that efficient training can be achieved by using large batches and small epochs, while Adam (1e-4) can ensure stable convergence and balance the speed and accuracy of the model. Using weighted cross-entropy / focal loss can solve class imbalance and thus focus on key samples. Furthermore, using an early stopping strategy can prevent model overfitting and ensure the model's generalization ability.

[0047] Strategy 3, Low-Frequency Offline Adversarial Enhancement Loop, is a closed-loop control strategy that periodically generates adversarial examples offline and optimizes model robustness. Through low-frequency updates (such as daily, weekly, monthly, etc.) and offline batch processing, it systematically improves the model's defense capability against adversarial attacks in non-real-time environments.

[0048] Therefore, in this embodiment, it is used to simulate and generate novel virus samples, improving the model's generalization ability and robustness to unknown risks, which is equivalent to "vaccinating" the model. Specifically, in practical applications, it can also be executed periodically, such as once every 24 hours, using real historical transaction data or the model's risk assessment data to simulate and generate virus transaction data, thereby retraining the financial risk assessment model.

[0049] In one specific embodiment, such as Figure 3 As shown, abnormal transaction data is obtained from real historical transaction data as the initial sample set. Random noise conforming to a certain distribution (Gaussian, uniform, etc.) is added to each sample through random noise perturbation to generate initial adversarial samples, i.e., simulated abnormal transaction data. Based on this, the adversarial samples can be continuously optimized through multiple rounds of small perturbations, combined with the feedback of the discriminator, to form more "deceptive" high-quality abnormal transaction data that is difficult for the model to simulate, i.e., virus transaction data.

[0050] It should be noted that multi-round adversarial game refers to the generator and discriminator updating alternately in the same training round. In each round, the generator adjusts its perturbation strategy based on the feedback from the discriminator in the previous round, thereby generating new adversarial samples until the final virus transaction data is obtained (usually set to 5-10 rounds).

[0051] Specifically, in practical applications, the Generator in the GAN framework can be used to add learnable small perturbations to the initial samples, the Discriminator scores the real and fake samples, and the parameters of the Generator are updated by feedback loss.

[0052] It's also worth noting that the Discriminator takes two types of samples as input: real samples (real abnormal transaction data) and adversarial samples (simulated abnormal transaction data) output by the Generator. It outputs a probability score representing the "real" probability. If the score is higher than a preset threshold, it's judged as "real"; otherwise, it's judged as "adversarial." Simultaneously, a corresponding feedback loss is calculated based on the Discriminator's judgment results. For example, cross-entropy loss can be used to maximize the probability of judging real samples as real and adversarial samples as fake, defining the corresponding loss function. Finally, the parameters in the Generator are optimized using the aforementioned loss function to make it "harder to fool."

[0053] It is understood that the number of generator / discriminator layers, hidden units, activation functions, etc. used in the above adversarial enhancement loop are only examples for reference. In practical applications, corresponding adjustments can be made according to actual needs. Modifications and refinements made by those skilled in the art to the embodiments of the present invention without departing from the spirit of the present invention still fall within the scope of the invention application patent of the present invention.

[0054] In one specific embodiment, taking the financial risk assessment model built by the Wide&Deep model as an example, the low-frequency offline adversarial enhancement loop focuses more on supplementing training on real + adversarial datasets, gradually unfreezing more levels of Deep sub-models.

[0055] Specifically, normal financial transaction data, abnormal transaction data, and virus transaction data can be selected and mixed in a 1:1:1 ratio to form training samples. Further, the training samples are sequentially input into the financial risk assessment model, and a first loss function is calculated based on the model's output and the corresponding sample data's true labels. The last layer of the Wide sub-model and the last layer of the Deep sub-model are then optimized and updated according to the first loss function. The second and third to last layers of the Deep sub-model are then gradually unfrozen, and the updated financial risk assessment model is repeatedly trained using these training samples to calculate a second loss function. The second and third to last layers of the Deep sub-model are then optimized and updated based on this second loss function.

[0056] The specific training parameters are as follows: The number of times the model traverses the full sample data (Epochs) can be set between 2 and 3, and the learning rate can be set between 5e-5-1e-4. Cosine or Plateau can be used to schedule the learning rate.

[0057] It should also be noted that the Wide & Deep model is only a preferred option for financial risk assessment. The Wide sub-model can be linearly modeled after hashing / bugging to avoid an excessively large weight matrix in the linear model and accelerate parameter updates. The Deep sub-model, through the "semantic mapping" of the embedding layer and the "non-linear interaction" of MLP (such as a 3-4 layer structure (256→128→64) + ReLU + Dropout), mines the implicit correlations between sparse features, compensating for the Wide model's insufficient capture of long-tail patterns and new correlations. In addition, a tree model + linear model architecture can also be used, such as offline GBDT feature transformation + online LR-FTRL prediction as a financial risk assessment model. This is not a limitation. Modifications and refinements made by those skilled in the art to the embodiments of this invention without departing from the spirit of this invention still fall within the scope of the invention application.

[0058] Based on the three strategies mentioned above, the financial risk assessment model can be automatically updated by automatically generating scheduling instructions, thereby suppressing the performance degradation of the model and improving the identification rate of new risks.

[0059] In one specific embodiment, the first, second, and third instructions can be generated continuously at preset frequencies. Specifically, the first instruction is generated continuously at a preset first frequency, the second instruction at a preset second frequency, and the third instruction at a preset third frequency, ensuring that the three strategies—the high-frequency online incremental adaptation loop, the mid-frequency active sample calibration loop, and the low-frequency offline adversarial enhancement loop—operate collaboratively at all times. Here, because the high-frequency online incremental adaptation loop has a higher execution frequency requirement, the low-frequency offline adversarial enhancement loop has a lower execution frequency requirement, and the mid-frequency active sample calibration loop only needs a normal execution frequency, the first frequency is higher than the second frequency, and the second frequency is higher than the third frequency.

[0060] In this regard, it is understandable that the first frequency can be set according to the update frequency of actual financial transaction data, so that the high-frequency online incremental adaptation loop can update and optimize the financial risk assessment model based on one or more financial transaction data updated in real time.

[0061] Therefore, the first time mentioned above can correspond to the time interval of the first frequency, or it can be a custom-set time to ensure that the model is updated using real-time financial transaction data, without any restrictions.

[0062] Similarly, the second time mentioned above can correspond to the time interval of the second frequency, that is, the time interval between the last time the financial risk assessment model was updated using the mid-frequency active sample calibration loop to optimize the financial risk assessment model. The second time can also be customized without any restrictions.

[0063] It is understood that the aforementioned scheduling instructions (first instruction, second instruction, and third instruction) are passively generated to achieve automatic optimization and updating of the financial risk assessment model, ensuring its reliability and accuracy in predicting abnormal risks. Furthermore, scheduling instructions can also be actively input externally. For example, after operations and maintenance personnel discover an anomaly in the model, they can input the corresponding scheduling instructions, and then utilize one or more of the above three strategies to actively optimize and update the model. This is not restricted, as the reliability and stability of the model are ensured through both active and passive methods. Modifications and refinements made by those skilled in the art to the embodiments of this invention without departing from the spirit of this invention still fall within the scope of the invention application.

[0064] The self-learning method of the aforementioned financial risk assessment model also includes: updating the financial risk assessment model based on the cumulative number of abnormal transaction data, so that when the cumulative number of abnormal transaction data exceeds a preset first threshold, the accumulated abnormal transaction data is pushed to human experts for annotation, and the financial risk assessment model is retrained based on the human-annotated abnormal transaction data.

[0065] Understandably, as described above, the financial risk assessment model will be automatically updated using Strategy Two at a preset frequency. However, in practical applications, there may be situations where massive amounts of financial transaction data are generated in a short period of time. To avoid model drift due to continuous online learning, the financial risk assessment model can be proactively optimized once the accumulated abnormal transaction data reaches a certain amount.

[0066] It should be noted that in Strategy 2, the financial risk assessment model is retrained using manually labeled abnormal transaction data. Here, the manually labeled abnormal transaction data can be obtained by manually labeling all abnormal transaction data or by manually labeling only some abnormal transaction data, such as real-time accumulated abnormal transaction data. There are no restrictions on this.

[0067] In one specific embodiment, when Strategy 2 is passively executed at a preset frequency, real-time accumulated abnormal transaction data, i.e., abnormal transaction data accumulated in the second time period, can be used to optimize the financial risk assessment model after being manually labeled. When the accumulated abnormal transaction data reaches a certain amount, i.e., the cumulative number of abnormal transaction data reaches a preset first threshold, all accumulated abnormal transaction data can be used to optimize the financial risk assessment model after being manually labeled.

[0068] Therefore, since all abnormal transaction data is required, there is a risk of duplicate labeling. That is, when Strategy 2 is passively executed at a preset frequency, some abnormal transaction data has already been manually labeled. To address this, once a certain amount of abnormal transaction data has accumulated, for each abnormal transaction, it is determined whether it has been manually labeled: if so, the push is stopped; if not, it is pushed to a human expert, and feedback is awaited. Finally, all manually labeled abnormal transaction data is integrated to retrain the financial transaction risk assessment model, thus avoiding the problem of duplicate labeling of abnormal transaction data.

[0069] Furthermore, the aforementioned proactive optimization of the financial risk assessment model after the cumulative number of abnormal transaction data reaches a preset first threshold can also be set to a situation within a certain period of time. For example, the proactive optimization of the financial risk assessment model after the cumulative number of abnormal transaction data reaches a preset first threshold within a certain period of time is not limited in this regard. Modifications and refinements made by those skilled in the art to the embodiments of the present invention without departing from the spirit of the present invention still fall within the scope of the invention application patent of the present invention.

[0070] The self-learning method of the aforementioned financial risk assessment model further includes: updating the financial risk assessment model according to the performance evaluation index of the financial risk assessment model, so that when the performance evaluation index of the financial risk assessment model is lower than a preset second threshold, simulating and generating virus transaction data, and retraining the financial risk assessment model based on it.

[0071] Similarly, as described above, the financial risk assessment model will be automatically updated using Strategy 3 at a preset frequency. However, since the frequency of passive execution of Strategy 3 is usually set low, there may be a sudden drop in model performance, and the inability to optimize and update in a timely manner may affect its reliability and accuracy.

[0072] Therefore, by monitoring the performance evaluation indicators of the financial risk assessment model in real time, the model can be proactively optimized once it declines to a certain level.

[0073] In summary, the scheduling methods for the three strategies mentioned above are not limited to the situations described above. The main purpose of this application is to improve the reliability and accuracy of financial risk assessment models by utilizing the synergistic operation of these three strategies.

[0074] In one specific embodiment, please refer to Table 1 below: Group A indicates that only Strategy 1 was used to optimize and update the financial risk assessment model, and the model's prediction accuracy is [missing data]. Group B indicates that both Strategy 1 and Strategy 3 were used to optimize and update the financial risk assessment model, and the model's prediction accuracy is [missing data]. Group C indicates that both Strategy 1 and Strategy 2 were used to optimize and update the financial risk assessment model, and the model's prediction accuracy is [missing data]. Group D indicates that a combination of Strategy 1, Strategy 2, and Strategy 3 were used to optimize and update the financial risk assessment model, and the model's prediction accuracy is [missing data].

[0075] Table 1

[0076] Therefore, compared with groups A and B, group D has the flattest AUC curve and the smallest performance degradation over 30 days, fully demonstrating the crucial role of strategy two. Furthermore, the coordinated operation of the three strategies consistently outperforms any single strategy or any combination of strategies, further demonstrating an unexpected technical synergy among the three strategies, achieving the optimal balance between robustness, adaptability, and accuracy.

[0077] It should be noted that the steps of the various methods described above are only for clarity. In practice, they can be combined into one step or some steps can be split into multiple steps. As long as they contain the same logical relationship, they are all within the scope of protection of this patent. Adding insignificant modifications or introducing insignificant designs to the algorithm or process, but without changing the core design of the algorithm and process, are also within the scope of protection of this patent.

[0078] Example 2 Please see Figure 4 As shown, this application also provides a self-learning system for a financial risk assessment model, comprising: The central scheduling module 10 is used to generate scheduling instructions for optimizing the financial risk assessment model and to update the financial risk assessment model according to the scheduling instructions. The model update module 20 is used to adjust the parameters of the financial risk assessment model based on a first instruction and one or more financial transaction data updated within a preset first time period; based on a second instruction, push the abnormal transaction data accumulated within a preset second time period to human experts for annotation, and retrain the financial risk assessment model based on the annotated abnormal transaction data; based on a third instruction, simulate and generate virus transaction data, and retrain the financial risk assessment model based on it.

[0079] It should be noted that the self-learning system for the financial risk assessment model provided in the above embodiments and the self-learning method for the financial risk assessment model provided in Embodiment 1 belong to the same concept. The specific methods by which each module and unit performs operations have been described in detail in the method embodiments and will not be repeated here. In practical applications, the self-learning method for the financial risk assessment model provided in Embodiment 1 can be used to allocate the above functions to different functional modules as needed, that is, to divide the internal structure of the device into different functional modules to complete all or part of the functions described above. This is not a limitation here.

[0080] Example 3 Please see Figure 5 As shown, embodiments of this application also provide an electronic device, including a memory 2, a processor 1, and a program stored in the memory and executable on the processor, wherein the processor executes the steps of any of the methods described above.

[0081] The memory includes at least one type of readable storage medium, such as flash memory, portable hard drive, multimedia card, card-type memory (e.g., SD or DX memory), magnetic memory, magnetic disk, optical disk, etc. In some embodiments, the memory can be an internal storage unit of an electronic device, such as a portable hard drive. In other embodiments, the memory can be an external storage device of the electronic device, such as a plug-in portable hard drive, Smart Media Card (SMC), Secure Digital (SD) card, Flash Card, etc. Furthermore, the memory can include both internal and external storage units of the electronic device. The memory can be used not only to store application software and various types of data installed on the electronic device, but also to temporarily store data that has been output or will be output.

[0082] In some embodiments, a processor may be composed of integrated circuits, such as a single packaged integrated circuit or multiple integrated circuits packaged with the same or different functions. This includes combinations of one or more central processing units (CPUs), microprocessors, digital processing chips, graphics processors, and various control chips. The processor is the control unit of the electronic device, connecting various components of the device via various interfaces and lines. It executes programs or modules stored in the memory and calls data stored in the memory to perform various functions and process data within the electronic device.

[0083] The processor executes the operating system of the electronic device and various installed applications. The processor executes the applications to implement the steps in the above method embodiments.

[0084] For example, the program may be divided into one or more modules, which are stored in the memory and executed by the processor to complete the present invention. The one or more modules may be a series of program instruction segments capable of performing a specific function, which describe the execution process of the program in the electronic device.

[0085] The integrated unit implemented as a software functional module described above can be stored in a computer-readable storage medium. This software functional module, stored in a storage medium, includes several instructions to cause a computer device (which may be a personal computer, computer equipment, or network device, etc.) or processor to execute some of the functions of the various embodiments of the present invention.

[0086] In summary, the beneficial effects of this invention are: by employing three model optimization strategies—cooperative online learning, sample calibration, and adversarial enhancement—the accuracy and reliability of the financial risk assessment model are improved, thereby enhancing the predictive ability for novel and abnormal risks.

[0087] This invention is applicable to various scenarios such as regression testing for map data version upgrades, scenario library construction for autonomous driving simulation testing, and traffic control impact assessment (generating detour path test cases by modifying link attributes (such as closed state)).

[0088] The above embodiments are merely illustrative of the principles and effects of the present invention and are not intended to limit the invention. Any person skilled in the art can modify or alter the above embodiments without departing from the spirit and scope of the present invention. Therefore, all equivalent modifications or alterations made by those skilled in the art without departing from the spirit and technical concept disclosed in the present invention should still be covered by the claims of the present invention.

Claims

1. A self-learning method for a financial risk assessment model, characterized in that, The financial risk assessment model is used to assess the risk probability of real-time input financial transaction data; the method includes: Generate scheduling instructions for optimizing the financial risk assessment model, and update the financial risk assessment model according to the scheduling instructions; Based on the first instruction, the parameters of the financial risk assessment model are adjusted according to one or more financial transaction data updated within a preset first time period. Based on the second instruction, the abnormal transaction data accumulated within a preset second time period is pushed to human experts for annotation, and the financial risk assessment model is retrained based on the annotated abnormal transaction data. Based on the third instruction, virus transaction data is simulated and generated, and the financial risk assessment model is retrained accordingly.

2. The self-learning method for the financial risk assessment model according to claim 1, characterized in that, The steps for generating scheduling instructions to optimize the financial risk assessment model include: The first instruction is generated continuously at a preset first frequency; The second instruction is generated continuously at a preset second frequency; The third instruction is continuously generated according to a preset third frequency; Wherein, the first frequency is higher than the second frequency, and the second frequency is higher than the third frequency.

3. The self-learning method for the financial risk assessment model according to claim 1, characterized in that, The steps for adjusting the parameters of the financial risk assessment model include: The financial risk assessment model is incrementally updated using the FTRL algorithm.

4. The self-learning method for the financial risk assessment model according to claim 3, characterized in that, The financial risk assessment model is built based on the Wide & Deep model, and the steps for incrementally updating the financial risk assessment model using the FTRL algorithm include: One or more financial transaction data updated within a preset first time period are sequentially input into the financial risk assessment model, and a weighted cross-entropy loss function is calculated based on the output of the financial risk assessment model and the true label of the corresponding financial transaction data. Based on the weighted cross-entropy loss function, gradient calculations are performed on the weight parameters of the Wide sub-model layer in the financial risk assessment model and the output layer parameters of the financial risk assessment model. The historical gradients and corresponding regularization constraints are accumulated using the FTRL algorithm to update the parameters of the Wide sub-model and the output layer. The parameters of the last layer of the Deep sub-model in the financial risk assessment model are updated based on the weighted cross-entropy loss function.

5. The self-learning method for the financial risk assessment model according to claim 1, characterized in that, Also includes: The financial risk assessment model is updated based on the cumulative number of abnormal transaction data. When the cumulative number of abnormal transaction data exceeds a preset first threshold, the accumulated abnormal transaction data is pushed to human experts for annotation, and the financial risk assessment model is retrained based on the annotated abnormal transaction data.

6. The self-learning method for the financial risk assessment model according to claim 1 or 5, characterized in that, For each piece of financial transaction data, the steps to identify whether it is abnormal transaction data include: The uncertainty value corresponding to the financial transaction data is determined according to the evaluation function, and financial transaction data with uncertainty values ​​exceeding a preset threshold are classified as abnormal transaction data; wherein, the formula of the evaluation function is as follows: ； This indicates the uncertainty value of the financial transaction data. This represents the risk assessment value of the financial transaction data by the financial risk assessment model, and the normalized risk assessment value is (0, 1).

7. The self-learning method for the financial risk assessment model according to claim 1 or 5, characterized in that, The steps of pushing accumulated abnormal transaction data to human experts for annotation, and retraining the financial risk assessment model based on the annotated abnormal transaction data include: For each abnormal transaction, identify whether it has been manually labeled: If so, then stop pushing notifications; If not, the request will be forwarded to a human expert, and feedback will be awaited. Integrate all manually labeled abnormal transaction data and retrain the financial risk assessment model based on it.

8. The self-learning method for the financial risk assessment model according to claim 1, characterized in that, The financial risk assessment model is built based on the Wide & Deep model, and the steps for retraining the financial risk assessment model based on manually labeled abnormal transaction data include: The manually labeled abnormal transaction data are sequentially input into the financial risk assessment model, and the weighted cross-entropy loss function is calculated based on the output of the financial risk assessment model and the true label of the corresponding abnormal transaction data. The weight parameters of each level of the Wide submodel in the financial risk assessment model are updated according to the cross-entropy loss function with class weights. Based on the cross-entropy loss function with class weights, the parameters of the last layer or the last two layers of the Deep sub-model in the financial risk assessment model are updated.

9. The self-learning method for the financial risk assessment model according to claim 1, characterized in that, Also includes: The financial risk assessment model is updated based on its performance evaluation index. When the performance evaluation index of the financial risk assessment model is lower than a preset second threshold, virus transaction data is simulated and generated, and the financial risk assessment model is retrained based on the data.

10. The self-learning method for the financial risk assessment model according to claim 1 or 9, characterized in that, The steps to simulate and generate virus transaction data include: Initial virus transaction data was generated by simulating random noise based on abnormal transaction data. The final virus transaction data is obtained by conducting multiple rounds of adversarial games on the initial virus transaction data.

11. The self-learning method for the financial risk assessment model according to claim 1, characterized in that, The financial risk assessment model is built based on the Wide & Deep model, and the steps for retraining the financial risk assessment model based on virus transaction data include: The simulated virus transaction data is sequentially input into the financial risk assessment model, and the first loss function is calculated based on the output of the financial risk assessment model and the real label of the corresponding virus transaction data. Based on the first loss function, the parameters of the last layer of the Wide sub-model in the financial risk assessment model are updated, and the parameters of the last layer of the Deep sub-model in the financial risk assessment model are also updated. The virus transaction data is re-input into the updated financial risk assessment model, and a second loss function is calculated based on the output of the financial risk assessment model and the true label of the virus transaction data. Based on the second loss function, the parameters of the penultimate and penultimate layers of the Deep sub-model in the financial risk assessment model are updated.

12. A self-learning system for a financial risk assessment model, characterized in that, The financial risk assessment model is used to assess the risk probability of real-time input financial transaction data; the system includes: The central scheduling module is used to generate scheduling instructions for optimizing the financial risk assessment model and to update the financial risk assessment model according to the scheduling instructions. The model update module is used to adjust the parameters of the financial risk assessment model based on a first instruction and one or more financial transaction data updated within a preset first time period; based on a second instruction, push the abnormal transaction data accumulated within a preset second time period to human experts for annotation, and retrain the financial risk assessment model based on the annotated abnormal transaction data; based on a third instruction, simulate and generate virus transaction data, and retrain the financial risk assessment model based on it.

13. An electronic device, characterized in that, The method includes a processor coupled to a memory storing program instructions, which, when executed by the processor, implement the method of any one of claims 1 to 11.

14. A computer-readable storage medium, characterized in that, Includes a program that, when run on a computer, causes the computer to perform the method as described in any one of claims 1 to 11.

Images