Artificial intelligence-based enterprise financing risk management system

Abstract

The present application relates to the technical field of risk management, and more particularly to an enterprise financing risk management system based on artificial intelligence, comprising a data acquisition module, a data processing module, a feature processing module, a data training module, a data analysis module and an execution module.The system filters, divides, trains, compares and calculates the received data; the management state of the system is determined by the risk identification accuracy and the preset threshold; the learning decay rate is adjusted based on the determination result, and the random discard probability and the shortest time data acquisition window are further determined according to the adjustment effect; the present application realizes the unified processing and optimization of the analysis performance of the financing risk management system, effectively optimizes the risk identification accuracy of the financing risk management execution performance of the present application, and improves the resource utilization rate.

Description

Technical Field

[0001] This invention relates to the field of intelligent risk assessment technology, and in particular to an artificial intelligence-based enterprise financing risk management system. Background Technology

[0002] The core of corporate financing risk management lies in making accurate and forward-looking judgments about a company's willingness and ability to repay loans. This leads to severe challenges for traditional management modules as data grows, with systems unable to keep up with feedback and optimize, making it difficult to guarantee the reliability and stability of processing results. To avoid potential risks, quantifying the identification accuracy of different processes and optimizing technical routes place higher demands on the scientific rigor and comparability of analytical results. Many newer and improved measures have been implemented, enabling timely detection of anomalies and significantly improving overall management efficiency, becoming a necessary and effective means. Therefore, traditional manual management operations remain one of the important means of financing risk management today.

[0003] However, existing technologies mostly adopt unified fixed parameters or local adjustment mechanisms, lacking overall optimization of data acquisition modules, data processing modules, feature processing modules, data analysis modules, and execution modules. Existing technologies mostly rely on human experience judgment, making it difficult to achieve global optimization management and efficient collaborative management. Therefore, the management efficiency of risk prediction is low, and it cannot systematically capture the dynamic correlation between various data.

[0004] Meanwhile, existing systems are not adaptable enough to dynamic changes in multi-source data and interference from screening features. The key data involved in the system often rely on experience-set or fixed standard values, lacking a mechanism for dynamic adjustment based on specific real-time monitoring and analysis data, which reduces the reliability of the management and operation efficiency of financing risks.

[0005] Chinese Patent Publication No. CN120047236A discloses a method and system for dynamic financing risk management, including installing equipment in a monitoring warehouse, converting multi-source data into homogeneous data, obtaining value assessment results based on the constructed data model, and simultaneously issuing financing risk warnings based on the risk assessment results.

[0006] Therefore, although the proposed solution can automatically generate risk warnings, it still has the following problems: 1. This solution only uses a single real-time monitoring method for model evaluation. When the system has operational anomalies, it cannot intelligently optimize risk prediction based on the model's fit, thus failing to effectively judge the degree of data training and reducing the efficiency of managing financing risks.

[0007] 2. This scheme uses fixed adjustment instructions, which may lead to risk prediction bias when the training data is not representative enough. This will affect the working status of the subsequent module based on random dropout probability analysis and further reduce the efficiency of managing financing risks. Summary of the Invention

[0008] To address this issue, the present invention provides an artificial intelligence-based enterprise financing risk management system to overcome the problem of low accuracy in risk identification caused by data processing bias resulting from the inability to generate corresponding updated decisions based on changes in model fitting in existing technologies.

[0009] To achieve the above objectives, the present invention provides an artificial intelligence-based enterprise financing risk management system, comprising: The data acquisition module is used to acquire multi-source heterogeneous data from enterprises; A data processing module, which is connected to the data acquisition module, is used to preprocess structured and unstructured data to construct an enterprise feature dataset; A feature processing module, connected to the data processing module, is used to divide the enterprise feature dataset and input the divided dataset into a data training module, wherein the divided dataset includes a training subset and a validation subset. The data training module, which is connected to the feature processing module, is used to build a model based on the received training subset within the period, and to obtain a predicted risk dataset for the enterprise based on the corresponding data in the validation subset. The data training module is also used to determine the risk identification accuracy for the enterprise based on the predicted risk dataset and the risk dataset in the validation subset. A data analysis module, connected to the data processing module and the feature processing module, is used to determine whether to adjust the corresponding module based on the risk identification accuracy output by the data processing module, and to determine whether to adjust the random dropout probability based on the training loss change rate of minority class samples based on the risk identification accuracy obtained after adjustment. The adjustment methods include determining whether to adjust the learning decay rate based on the risk identification accuracy output by the feature processing module, and determining whether to adjust the random dropout probability based on the training loss change rate of minority class samples. An execution module, which is connected to the data analysis module, is used to adjust the running data of the corresponding module to the corresponding value based on the instructions output by the data analysis module.

[0010] Furthermore, the data analysis module is used to determine whether the execution effectiveness of risk management meets the standard based on the risk identification accuracy, and, if it is determined that the execution effectiveness of risk management does not meet the standard, to determine the reason based on the risk prediction loss ratio; wherein, the risk identification accuracy is the weighted sum of the timestamps of a single sample in the validation subset and a preset time decay coefficient; the risk prediction loss ratio is the ratio of the loss function value of the model in the validation subset to the loss function value in the training subset.

[0011] Furthermore, the data analysis module is used to determine the reason for non-compliance based on the risk prediction loss ratio when it is determined that the execution efficiency for the risk management does not meet the standard, and to generate a corresponding processing method based on the determined reason, including adjusting the learning decay rate when the reason is determined to be model overfitting, or adjusting the random discard probability when the reason is determined to be insufficient representativeness of the training data.

[0012] Furthermore, when the data analysis module determines that the cause is model overfitting, it increases the learning rate decay rate based on the confidence distribution change rate, and the increase in the learning rate decay rate is negatively correlated with the confidence distribution change rate; wherein, the confidence distribution change rate is the difference in the average information entropy of the model's predicted probability distribution for the same training subset in two consecutive evaluation periods; and the learning rate decay rate is the sum of the product of the initial confidence distribution change rate and a preset coefficient and an exponential decay factor.

[0013] Furthermore, the data analysis module is also used to adjust the learning decay rate after completing the adjustment of the learning decay rate, and to increase the adjusted learning decay rate according to the average change rate of the model weights in a single period, wherein the adjusted learning decay rate is positively correlated with the average change rate of the model weights; wherein the average change rate of the model weights is the average of the absolute values ​​of the numerical changes of the trainable parameters before and after one iteration update.

[0014] Furthermore, the data analysis module is also used to respond to the first duplicate detection condition, determine that the reason for not meeting the standard is that the training data is not representative enough, and adjust the random discard probability according to the training loss change rate of the minority class samples. Wherein, the first repeated detection condition is that the data analysis module re-determines that the execution efficiency of the risk management does not meet the standard after adjusting the learning decay rate; the training loss change rate of the minority class samples is the average change rate of the loss value of the minority class samples in two adjacent periods; the random discard probability is the probability that each sample is randomly discarded when the majority class samples are randomly sampled.

[0015] Furthermore, the data analysis module is also used to reduce the random drop probability based on the rate of change of the training loss of the minority class samples, and the reduction in the random drop probability is positively correlated with the rate of change of the training loss of the minority class samples.

[0016] Furthermore, the data analysis module is also used to increase the adjusted random discard probability based on the feature selection percentile, and the adjusted random discard probability is positively correlated with the feature selection percentile; wherein, the feature selection percentile is the percentage of features ranked first based on feature importance scores before the feature processing model is trained.

[0017] Furthermore, the data analysis module is also used to respond to the second repeated detection condition, determine that the reason for not meeting the standard is that the quality of the unstructured data acquired by the data acquisition module is biased, and adjust the shortest data acquisition time window according to the minimum number of verification passes. Wherein, the second repeated detection condition is that the data analysis module re-determines that the execution efficiency of the risk management does not meet the standard after adjusting the random discard probability; the minimum number of verification passes is the minimum number of independent data sources contained in the data verification process; the shortest data collection time window is the time length for the system to collect, generate and update data.

[0018] Furthermore, the data analysis module is also used to increase the shortest data collection time window based on the minimum number of successful verifications, and the reduction in the shortest data collection time window is negatively correlated with the minimum number of successful verifications.

[0019] Compared with the prior art, the beneficial effects of the present invention are as follows: by setting up a data analysis module, the present invention can effectively determine whether there are execution deviations or preprocessing deviations in the risk management system by analyzing the risk identification accuracy of each data. At the same time, the data analysis module can also output corresponding processing methods based on the determined actual situation, thereby effectively eliminating the deviation caused by abnormal data during the data processing. While effectively improving the risk identification processing method for data, it effectively avoids the impact of model overfitting on the risk identification accuracy, thereby effectively improving the risk identification accuracy of the solution described in the present invention for financing risk management.

[0020] Furthermore, the data analysis module of the present invention is used to determine whether the execution effectiveness of risk management meets the standard based on the risk identification accuracy rate, and, if it is determined that the execution effectiveness of risk management does not meet the standard, to determine the cause based on the risk prediction loss ratio, thereby making the determination result of meeting the standard more scenario-based and intelligent, thereby improving the risk identification accuracy of the solution of the present invention for the risk management.

[0021] Furthermore, the data analysis module of the present invention is used to determine the reasons for non-compliance based on the risk prediction loss ratio when it is determined that the execution efficiency of the risk management does not meet the standard, and to generate corresponding processing methods based on the determined reasons. This makes the judgment results more complete and refined. While effectively avoiding misjudgment and omission of the risk identification accuracy under a single fixed threshold, the data analysis module further improves the risk identification accuracy of the solution of the present invention for the risk management.

[0022] Furthermore, when the cause is determined to be model overfitting, the data analysis module of the present invention increases the learning rate decay rate according to the confidence distribution change rate, which can intuitively determine the relationship between excessive, appropriate and insufficient data processing in the current data processing process. While further improving the agility of the data analysis module, it further avoids resource waste, thereby further improving the risk identification accuracy of the solution of the present invention for risk management.

[0023] Furthermore, the data analysis module of the present invention is also used to increase the adjusted learning decay rate according to the average change rate of model weights within a single period after adjusting the learning decay rate, thereby effectively outputting the corresponding processing decision. While further improving the learning decay rate, it further avoids the influence of the average change rate of weights on model fitting, thereby further improving the risk identification accuracy of the scheme of the present invention for risk management.

[0024] Furthermore, the data analysis module of the present invention is also used to respond to the first repeated detection condition, determine that the reason for not meeting the standard is that the training data is not representative enough, and adjust the random discard probability according to the training loss change rate of minority class samples. While further improving the data processing in the region, it further avoids resource waste, thereby further improving the risk identification accuracy of the solution of the present invention for the risk management.

[0025] Furthermore, the data analysis module of the present invention is also used to reduce the random discard probability based on the training loss change rate of minority class samples, which can effectively improve the accuracy of the system in terms of random discard probability, thereby effectively avoiding the occurrence of deviation in the comparison results, and further improving the risk identification accuracy of the scheme of the present invention for the risk management.

[0026] Furthermore, the data analysis module of the present invention is also used to filter the percentile increase of the adjusted random discard probability based on the feature, which can specifically and accurately adjust the random discard probability, further refine the adjustment range, and thus further improve the risk identification accuracy of the solution of the present invention for the risk management.

[0027] Furthermore, the data analysis module of the present invention is also used to respond to the second repeated detection condition and determine that the reason for non-compliance with the standard is that the quality of the unstructured data acquired by the data acquisition module is biased. The shortest data acquisition time window is adjusted according to the minimum pass count, and multiple preset values ​​of adjustment range are stored in advance. Thus, the data processing module adjusted by the adjustment coefficient obtains a more accurate processing result, thereby further improving the risk identification accuracy of the solution of the present invention for risk management.

[0028] Furthermore, the data analysis module of the present invention is also used to increase the shortest data collection time window based on the minimum number of verification passes, making the adjustment of the shortest data collection time window more logical and precise. At the same time, multiple minimum number of verification passes are pre-stored to ensure that the collected data can be used to obtain the results, thereby effectively improving the accuracy of processing decisions and further improving the risk identification accuracy of the solution of the present invention for risk management. Attached Figure Description

[0029] Figure 1 This is a structural block diagram of the financing risk management system described in this invention; Figure 2 This is a flowchart illustrating how the system described in this invention determines whether the execution effectiveness of risk management meets the standards based on the accuracy of risk identification. Figure 3 This is a flowchart of the optimization process for fitting the system model described in this invention; Figure 4 This is an optimization flowchart representing the training data of the system described in this invention. Detailed Implementation

[0030] To make the objectives and advantages of the present invention clearer, the present invention will be further described below with reference to embodiments; it should be understood that the specific embodiments described herein are merely for explaining the present invention and are not intended to limit the present invention.

[0031] Preferred embodiments of the present invention will now be described with reference to the accompanying drawings. Those skilled in the art should understand that these embodiments are merely illustrative of the technical principles of the present invention and are not intended to limit the scope of protection of the present invention.

[0032] It should be noted that in the description of this invention, the terms "upper", "lower", "left", "right", "inner", "outer", etc., which indicate directions or positional relationships, are based on the directions or positional relationships shown in the accompanying drawings. This is only for the convenience of description and is not intended to indicate or imply that the device or element must have a specific orientation, or be constructed and operated in a specific orientation. Therefore, it should not be construed as a limitation of this invention.

[0033] Furthermore, it should be noted that, in the description of this invention, unless otherwise explicitly specified and limited, the terms "installation," "connection," and "linking" should be interpreted broadly. For example, they can refer to a fixed connection, a detachable connection, or an integral connection; they can refer to a mechanical connection or an electrical connection; they can refer to a direct connection or an indirect connection through an intermediate medium; and they can refer to the internal connection of two components. Those skilled in the art can understand the specific meaning of the above terms in this invention according to the specific circumstances.

[0034] It should be noted that the data in this embodiment are all derived from a comprehensive analysis and evaluation of historical test data and corresponding historical test results from the system described in this invention over the three months prior to this test. Before this test, the system described in this invention comprehensively determines the preset values ​​stored in the database based on the analysis results of 25,863 cases cumulatively tested in the previous three months and the processing results after handling 19,584 cases under specific circumstances. Those skilled in the art will understand that the system described in this invention can determine the above-mentioned parameters for a single item by selecting the value with the highest proportion based on the data distribution as the preset standard parameter, using weighted summation to obtain the value as the preset standard parameter, substituting each historical data into a specific formula and using the value obtained by the formula as the preset standard parameter, or other selection methods, as long as the system described in this invention can clearly define different specific situations in the single-item judgment process through the obtained values.

[0035] Please see Figure 1 The diagram shown is a structural block diagram of the financing risk management system of the present invention. The financing risk management system of the present invention includes a data acquisition module, a data processing module, a feature processing module, a data training module, a data analysis module, and an execution module. The data acquisition module is used to acquire multi-source heterogeneous data from the enterprise; The data processing module, which is connected to the data acquisition module, is used to preprocess structured and unstructured data to construct an enterprise feature dataset. The feature processing module is connected to the data processing module and is used to divide the enterprise feature dataset and input the divided dataset into a data training module, wherein the divided dataset includes a training subset and a validation subset. The data training module, which is connected to the feature processing module, is used to build a model based on the received training subset within the period, and to obtain a predicted risk dataset for the enterprise based on the corresponding data in the validation subset. The data training module is also used to determine the risk identification accuracy for the enterprise based on the predicted risk dataset and the risk dataset in the validation subset. The data analysis module, which is connected to the data processing module and the feature processing module, is used to determine whether to adjust the corresponding module according to the risk identification accuracy output by the data processing module, and to determine whether to adjust the random dropout probability based on the training loss change rate of minority class samples based on the risk identification accuracy obtained after adjustment. The adjustment method includes determining whether to adjust the learning decay rate according to the risk identification accuracy output by the feature processing module, and determining whether to adjust the random dropout probability based on the training loss change rate of minority class samples. The execution module is connected to the data analysis module and is used to adjust the running data of the corresponding module to the corresponding value based on the instructions output by the data analysis module.

[0036] Specifically, during the operation of the system described in this invention, the data processing module preprocesses the heterogeneous data output by the data acquisition module, the feature processing module divides the preprocessed data and inputs it into the data training module, the data analysis module combines the trained data to start the corresponding control mode, and compares the trained data with preset parameters to determine whether the execution efficiency for the current risk management meets the standard, and generates the corresponding adjustment instruction based on the determined reason if it does not meet the standard.

[0037] Specifically, the system described in this invention stores and manages the processing results of received multi-source data and constructs a database that can be continuously updated.

[0038] Please see Figure 2 As shown, this is a flowchart illustrating how the system of the present invention determines whether the execution effectiveness of risk management meets the standards based on the accuracy of risk identification. The process includes: Specifically, the data analysis module in this embodiment of the invention is used to determine whether the execution effectiveness of risk management meets the standard based on the risk identification accuracy rate, and, if it is determined that the execution effectiveness of risk management does not meet the standard, to determine the reason based on the risk prediction loss ratio; wherein, the risk identification accuracy rate is the weighted sum of the timestamps of a single sample in the validation subset and a preset time decay coefficient; the risk prediction loss ratio is the ratio of the loss function value of the model in the validation subset to the loss function value in the training subset; Specifically, the feature processing module further includes: The risk identification and determination submodule calculates the risk identification accuracy by validating the training parameters of the samples, thereby determining the effectiveness of risk management. This value reflects the judgment ability and reliability of the risk identification model itself. A higher risk identification accuracy indicates a stronger ability to capture high-risk parameters, fewer missed or false positives, higher reliability of subsequent adjustment decisions, and more targeted adjustments. Conversely, a lower risk identification accuracy indicates a less accurate model, more missed or false positives, and a risk of overfitting and performance degradation during model training.

[0039] It is understood that the database pre-stores the maximum allowable risk identification accuracy for different types of data. Therefore, the present invention does not impose specific restrictions on the value of the preset risk identification accuracy, as long as the comparison result between the obtained risk identification accuracy and the preset risk identification accuracy can directly characterize the analysis status of the data analysis module.

[0040] Specifically, the data analysis module described in this embodiment of the invention is used to determine the reason for non-compliance with the standard when it is determined that the execution efficiency for the risk management does not meet the standard, based on the comparison result of the risk prediction loss ratio and the risk prediction loss ratio of the snake, and to generate a corresponding processing method based on the determined reason, including adjusting the learning decay rate when the reason is determined to be model overfitting, or adjusting the random discard probability when the reason is determined to be insufficient representativeness of the training data.

[0041] Specifically, the feature processing module described in this embodiment of the invention includes: The risk prediction determination submodule calculates the risk prediction loss ratio by comparing the current risk prediction loss value with the preset risk prediction loss value. This ratio reflects the instantaneous or fluctuating deviation of the model's risk prediction performance relative to the benchmark level. A larger risk prediction loss ratio indicates less stable model performance, decreased prediction reliability, and potential overfitting, requiring adjustment and intervention. Conversely, a smaller risk prediction loss ratio indicates that the model's risk prediction performance exceeds the benchmark and the prediction results are more reliable. If the current risk identification accuracy still does not meet the standard, it indicates insufficient representativeness of the training data, meaning the risk prediction task and the benchmark task have different objectives, requiring correlation adjustment.

[0042] Please see Figure 3 As shown, it is an optimization flowchart for the system model fitting described in this invention, and the process includes: Specifically, in this embodiment of the invention, when the cause is determined to be model overfitting, the data analysis module increases the learning rate decay rate based on the comparison between the confidence distribution change rate and the preset confidence distribution change rate, and the increase in the learning rate decay rate is negatively correlated with the confidence distribution change rate; wherein, the confidence distribution change rate is the difference in the average information entropy of the model's predicted probability distribution for the same training subset in two consecutive evaluation periods; and the learning rate decay rate is the sum of the product of the initial confidence distribution change rate and the preset coefficient and the exponential decay factor.

[0043] Specifically, the data analysis module described in this embodiment of the invention is further configured to, after adjusting the learning decay rate, increase the adjusted learning decay rate based on the comparison result between the average change rate of model weights within a single period and the preset average change rate of weights, wherein the adjusted learning decay rate is positively correlated with the average change rate of model weights; wherein the average change rate of model weights is the average of the absolute values ​​of the numerical changes of trainable parameters before and after one iteration update.

[0044] Specifically, the data processing module described in this embodiment of the invention includes: The confidence level determination submodule calculates the rate of change of the confidence level distribution using the predicted probability distribution of the same validation sample to determine the stability of the model output, i.e., the degree of fluctuation during training. A larger rate of change in the confidence level distribution indicates more unstable model predictions and more drastic output fluctuations. This is more likely to occur in the early stages of training and when the training samples are too difficult, preventing the model from forming a stable understanding. Conversely, a smaller rate of change in the confidence level distribution indicates more stable model predictions and more consistent outputs. This is more likely to occur during the convergence phase in the later stages of training, leading to the model forming a local optimum.

[0045] The learning decay rate determination submodule calculates the learning decay rate by measuring the rate at which the learning rate decreases over time, thus determining the desired rate at which the system reduces the parameter update step size. A larger learning decay rate results in faster learning decay and a more rapid convergence of the optimization process; a smaller learning decay rate results in slower learning decay, slower optimization speed, and higher computational cost.

[0046] The weight determination submodule calculates the average rate of change of weights by measuring the difference between the weights of key layers in consecutive training cycles, thus determining the activity and magnitude of parameter updates. A larger average rate of change indicates more drastic weight updates, more active training, and a rapid learning phase; a smaller average rate of change indicates slower weight updates, the model is close to convergence, and may fail to learn due to vanishing gradients or an excessively low learning rate.

[0047] It is understood that the database described in this embodiment of the invention is based on the monitored average rate of change of weights, and compares it with the average rate of change of weights obtained through real-time calculation, thereby achieving accurate evaluation and graded response of parameter adjustment. The larger the average rate of change of weights, the greater the required learning decay rate and the greater the required adjustment range. Therefore, a corresponding correction coefficient needs to be selected to ensure that the average rate of change of weights is proportional to the corrected learning decay rate. Thus, the correction coefficient is proportional to the learning decay rate.

[0048] Specifically, the data analysis module described in this embodiment of the invention is also used to respond to the first repeated detection condition, determine that the reason for not meeting the standard is that the training data is not representative enough, and adjust the random discard probability according to the training loss change rate of minority class samples. Wherein, the first repeated detection condition is that the data analysis module re-determines that the execution efficiency of the risk management does not meet the standard after adjusting the learning decay rate; the training loss change rate of the minority class samples is the average change rate of the loss value of the minority class samples in two adjacent periods; the random discard probability is the probability that each sample is randomly discarded when the majority class samples are randomly sampled.

[0049] Please see Figure 4 As shown, it is a flowchart of the optimization process for representative training data of the system described in this invention, and the process includes: Specifically, the data analysis module described in this embodiment of the invention is also used to reduce the random dropout probability based on the rate of change of the training loss of minority class samples, and the reduction in the random dropout probability is positively correlated with the rate of change of the training loss of minority class samples. Specifically, the data training module described in this embodiment of the invention further includes: The training loss determination submodule calculates the rate of change of training loss for minority class samples by measuring the relative change in the loss value of minority class samples between the current and previous periods. This determines the learning progress and stability of the model for the minority class in the imbalanced class. A larger rate of change in training loss indicates a faster increase in loss for minority class samples, resulting in a worse fit of the model to these samples, and the possibility of forgetting or a shift in the classification boundary unfavorable to the minority class. Conversely, a smaller rate of change in training loss indicates a more subtle change in loss, a more stable decrease in loss, higher effective learning intensity, and more significant optimization of the fit to the minority class.

[0050] The random dropout probability determination submodule calculates the random dropout probability through neuron outputs to determine the strength of network regularization and the degree of model integration. This value is between 0 and 1. The higher the random dropout probability, the more neurons are dropped, the sparser the network, and the stronger the regularization effect. The lower the random dropout probability, the denser the network, the stronger the ability, and the higher the learning efficiency.

[0051] Specifically, the data analysis module described in this embodiment of the invention is further used to increase the adjusted random discard probability based on the comparison results between the feature selection percentile and the preset feature selection percentile, and the adjusted random discard probability is positively correlated with the feature selection percentile; wherein, the feature selection percentile is the percentage of features ranked first based on feature importance scores before the feature processing model is trained.

[0052] Specifically, the data training module described in this embodiment of the invention further includes: The feature selection submodule calculates the feature selection percentage using feature scoring gradients and attention weights to determine the selectivity and sparsity of the input information. A higher feature selection percentage retains more features and provides richer information; a lower percentage retains fewer features, resulting in sparser information and a greater risk of information loss.

[0053] It is understood that the feature importance score is generated by the feature processing module during training and is obtained by normalization calculation based on the class mixing degree of the training samples.

[0054] It is understandable that the higher the percentile of feature selection, the higher the proportion of system noise, the greater the probability of random discarding, and the greater the adjustment range required. Therefore, the correction coefficient is proportional to the probability of random discarding.

[0055] Specifically, the data analysis module described in this embodiment of the invention is also used to respond to the second repeated detection condition, determine that the reason for not meeting the standard is that the quality of the unstructured data acquired by the data acquisition module is biased, and adjust the shortest data acquisition time window according to the minimum number of verification passes. Wherein, the second repeated detection condition is that the data analysis module re-determines that the execution efficiency of the risk management does not meet the standard after adjusting the random discard probability; the minimum number of verification passes is the minimum number of independent data sources contained in the data verification process; the shortest data collection time window is the time length for the system to collect, generate and update data.

[0056] Specifically, the data analysis module described in this embodiment of the invention is also used to increase the shortest data collection time window based on the minimum number of successful verifications, and the reduction in the shortest data collection time window is negatively correlated with the minimum number of successful verifications.

[0057] Specifically, the data analysis module described in this embodiment of the invention further includes: The minimum pass determination submodule calculates the minimum number of validation passes based on risk tolerance to determine the minimum number of times the model performance must reach the standard across multiple consecutive validation cycles. A higher minimum pass number indicates a higher decision threshold, requiring the system to act only after achieving more stable and sustained performance, thus minimizing misjudgments and extending the evaluation period. Conversely, a lower minimum pass number indicates a lower decision threshold, making the system more susceptible to noise interference and more prone to misjudgments based on unstable signals.

[0058] Specifically, the data acquisition module described in this embodiment of the invention includes: The time data acquisition window determination submodule calculates the shortest time data acquisition window based on the data generation frequency and timeliness to determine the shortest time range. A larger shortest time data acquisition window results in a longer data time window and a larger data volume, but also reduces sensitivity to recent sudden changes and increases response latency. Conversely, a smaller shortest time data acquisition window results in a shorter data window, faster updates, and allows the system to more accurately capture the latest trends and respond more quickly.

[0059] It is understandable that the higher the minimum pass count for validation, the more stable and reliable the model, and the less intervention is required in the system. Therefore, an appropriate adjustment coefficient needs to be selected. Hence, the minimum pass count for validation is inversely proportional to the shortest data acquisition window.

[0060] The technical solution of the present invention has been described above with reference to the preferred embodiments shown in the accompanying drawings. However, it will be readily understood by those skilled in the art that the scope of protection of the present invention is obviously not limited to these specific embodiments. Without departing from the principles of the present invention, those skilled in the art can make equivalent changes or substitutions to the relevant technical features, and the technical solutions after these changes or substitutions will all fall within the scope of protection of the present invention.

[0061] The above description is merely a preferred embodiment of the present invention and is not intended to limit the invention. Various modifications and variations can be made to the present invention by those skilled in the art. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.

Claims

1. An artificial intelligence-based enterprise financing risk management system, characterized in that, include: The data acquisition module is used to acquire multi-source heterogeneous data from enterprises; A data processing module, which is connected to the data acquisition module, is used to preprocess structured and unstructured data to construct an enterprise feature dataset; A feature processing module, connected to the data processing module, is used to divide the enterprise feature dataset and input the divided dataset into a data training module, wherein the divided dataset includes a training subset and a validation subset. The data training module, which is connected to the feature processing module, is used to build a model based on the received training subset within the period, and to obtain a predicted risk dataset for the enterprise based on the corresponding data in the validation subset. The data training module is also used to determine the risk identification accuracy for the enterprise based on the predicted risk dataset and the risk dataset in the validation subset. A data analysis module, connected to the data processing module and the feature processing module, is used to determine whether to adjust the corresponding module based on the risk identification accuracy output by the data processing module, and to determine whether to adjust the random dropout probability based on the training loss change rate of minority class samples based on the risk identification accuracy obtained after adjustment. The adjustment methods include determining whether to adjust the learning decay rate based on the risk identification accuracy output by the feature processing module, and determining whether to adjust the random dropout probability based on the training loss change rate of minority class samples. An execution module, which is connected to the data analysis module, is used to adjust the running data of the corresponding module to the corresponding value based on the instructions output by the data analysis module.

2. The enterprise financing risk management system based on artificial intelligence according to claim 1, characterized in that, The data analysis module is used to determine whether the execution effectiveness of risk management meets the standard based on the risk identification accuracy, and, if it is determined that the execution effectiveness of risk management does not meet the standard, to determine the reason based on the risk prediction loss ratio; wherein, the risk identification accuracy is the weighted average of the timestamp of a single sample in the validation subset and a preset time decay coefficient; the risk prediction loss ratio is the ratio of the loss function value of the model in the validation subset to the loss function value in the training subset.

3. The enterprise financing risk management system based on artificial intelligence according to claim 2, characterized in that, The data analysis module is used to determine the reasons for non-compliance with the standard based on the risk prediction loss ratio when it is determined that the execution efficiency for the risk management does not meet the standard, and to generate corresponding processing methods based on the determined reasons, including adjusting the learning decay rate when the reason is determined to be model overfitting, or adjusting the random discard probability when the reason is determined to be insufficient representativeness of the training data.

4. The enterprise financing risk management system based on artificial intelligence according to claim 3, characterized in that, When the data analysis module determines that the cause is model overfitting, it increases the learning rate decay rate based on the confidence distribution change rate, and the increase in the learning rate decay rate is negatively correlated with the confidence distribution change rate. The confidence distribution change rate is the difference in the average information entropy of the model's predicted probability distribution for the same training subset over two consecutive evaluation periods. The learning rate decay rate is the sum of the product of the initial confidence distribution change rate and a preset coefficient and an exponential decay factor.

5. The enterprise financing risk management system based on artificial intelligence according to claim 3, characterized in that, The data analysis module is further configured to, after adjusting the learning decay rate, increase the adjusted learning decay rate based on the average change rate of model weights within a single period, wherein the adjusted learning decay rate is positively correlated with the average change rate of model weights; wherein the average change rate of model weights is the average of the absolute values ​​of the numerical changes of trainable parameters before and after one iteration update.

6. The enterprise financing risk management system based on artificial intelligence according to claim 5, characterized in that, The data analysis module is also used to respond to the first duplicate detection condition, determine that the reason for not meeting the standard is that the training data is not representative enough, and adjust the random discard probability according to the training loss change rate of minority class samples. Wherein, the first repeated detection condition is that the data analysis module re-determines that the execution efficiency of the risk management does not meet the standard after adjusting the learning decay rate; the training loss change rate of the minority class samples is the average change rate of the loss value of the minority class samples in two adjacent periods; the random discard probability is the probability that each sample is randomly discarded when the majority class samples are randomly sampled.

7. The enterprise financing risk management system based on artificial intelligence according to claim 6, characterized in that, The data analysis module is also used to reduce the random drop probability based on the rate of change of the training loss of minority class samples, and the reduction in the random drop probability is positively correlated with the rate of change of the training loss of minority class samples.

8. The enterprise financing risk management system based on artificial intelligence according to claim 7, characterized in that, The data analysis module is also used to increase the adjusted random discard probability based on the feature selection percentile, and the adjusted random discard probability is positively correlated with the feature selection percentile; wherein, the feature selection percentile is the percentage of features ranked first based on feature importance scores before the feature processing model is trained.

9. The enterprise financing risk management system based on artificial intelligence according to claim 8, characterized in that, The data analysis module is also used to respond to the second repeated detection condition and determine that the reason for not meeting the standard is that the quality of the unstructured data acquired by the data acquisition module is biased, and adjust the shortest data acquisition time window according to the minimum number of verification passes. Wherein, the second repeated detection condition is that the data analysis module re-determines that the execution efficiency of the risk management does not meet the standard after adjusting the random discard probability; the minimum number of verification passes is the minimum number of independent data sources contained in the data verification process; the shortest data collection time window is the time length for the system to collect, generate and update data.

10. The enterprise financing risk management system based on artificial intelligence according to claim 9, characterized in that, The data analysis module is also used to increase the shortest data collection time window based on the minimum number of successful verifications, and the reduction in the shortest data collection time window is negatively correlated with the minimum number of successful verifications.

Images