Automatic risk control management system for financial credit business

Abstract

The invention provides an automatic risk control management system for financial credit business. The automatic risk control management system comprises a data acquisition module used for acquiring user credit behavior data and third-party credit investigation data in real time; the feature processing module is used for performing cleaning, dimension reduction and feature enhancement on the data; the risk assessment module outputs a user risk score based on an improved logistic regression model; the decision execution module generates a loan decision according to the risk score and the dynamic threshold; and the model optimization module performs adaptive adjustment on model parameters through feedback data. According to the invention, efficient and accurate risk assessment and decision can be realized, the risk management capability of credit business is improved, the default risk is reduced, and the business efficiency and the user experience are improved at the same time.

Description

Technical Field

[0001] The present invention relates to the field of financial technology, and more specifically, to an automated risk control management system for financial credit business. Background Art

[0002] In the field of financial credit, risk control is the core link to ensure the sound operation of financial institutions. Traditional credit risk control mainly relies on manual review and simple statistical models. This method is inefficient, easily affected by human factors, and difficult to handle massive and complex user data. With the rapid development of financial technology, the amount of data has exploded, and user behavior patterns have become more diversified. Traditional risk control methods can no longer meet the needs of modern financial credit business. In recent years, although machine learning and big data technologies have been applied to a certain extent in the field of risk control, the existing systems still have deficiencies in feature processing, model optimization, and dynamic decision-making. For example, feature selection is not accurate enough to effectively extract valuable information; model updates lag behind and are difficult to adapt to changes in data distribution; the decision-making process lacks flexibility and cannot dynamically adjust strategies based on real-time risks.

[0003] In the process of implementing the embodiments of the present invention, the inventors found that there are at least the following problems or defects in the prior art: the feature processing module has limited capabilities in data cleaning and feature enhancement, and cannot make full use of multi-source data; the model accuracy and adaptability of the risk assessment module are insufficient, making it difficult to accurately assess user risks; the dynamic adjustment mechanism of the decision execution module is not flexible enough and cannot respond to risk changes in a timely manner; the adaptive adjustment capability of the model optimization module is insufficient, making it difficult to continuously improve model performance. Summary of the invention

[0004] The present invention provides an automated risk control management system for financial credit business, comprising: Data acquisition module, feature processing module, risk assessment module, decision execution module, model optimization module; The data collection module is used to obtain user credit behavior data and third-party credit data in real time; The feature processing module is used to clean, reduce dimension and enhance features of data; The risk assessment module outputs a user risk score based on an improved logistic regression model; The decision execution module generates a loan decision based on the risk score and dynamic threshold; The model optimization module adaptively adjusts the model parameters through feedback data.

[0005] Furthermore, the feature processing module performs the following steps: Step 201: Use the improved random forest algorithm to select features. The specific formula is:

[0006] in, is the comprehensive weight of the i-th feature, is the preset entropy weight coefficient, is the information value weight, Features The value of information, is the input feature variable, is the target variable; Step 202: The missing values ​​are filled using an interpolation algorithm based on the user behavior pattern. The interpolation formula is:

[0007] in, is the dynamic adjustment coefficient, is the historical data smoothing coefficient, It is the exponential moving average of historical data; Step 203: Segment the continuous features using an adaptive binning algorithm based on KL divergence. The binning boundary formula is:

[0008] in, is the kth bin boundary, , are the sample distribution within and between boxes, respectively.

[0009] Furthermore, the improved logistic regression model in the risk assessment module is:

[0010] in, is the intercept term, is the feature weight vector, is the L1-L2 mixed penalty coefficient, is the nonlinear activation penalty coefficient, .

[0011] Furthermore, the mixed penalty coefficient The calculation formula is:

[0012] in, is the number of model training iterations, is a periodic adjustment parameter used to dynamically balance the regularization strength.

[0013] Furthermore, the decision execution module includes: Level division unit: divide users into levels according to risk scores. The division formula is:

[0014] in, , , is the historical risk score average, is the standard deviation, , is the dynamic adjustment coefficient; Credit limit freezing unit: triggers the manual review process for Class C users and freezes the credit limit based on composite rules. The freezing formula is:

[0015] in, , The current and historical average risk scores are respectively.

[0016] Furthermore, the dynamic adjustment coefficient , The update logic is:

[0017] in, is the learning rate, is the AUC impact factor, is the area under the ROC curve of the current model.

[0018] Furthermore, the model optimization module includes the following steps: Step 501: Calculate the model error gradient based on user feedback data, the formula is:

[0019] in, , are the weight coefficients of KL divergence and Wasserstein distance respectively, , For historical and current data distribution; Step 502: Update parameters using an improved adaptive momentum optimizer, the iterative formula of which is:

[0020] in, is the gradient direction correction coefficient, which is used to avoid local optimum.

[0021] Furthermore, the data acquisition module 100 includes: Alignment unit: Capture user social network behavior data in real time through a distributed API interface, and use a time alignment algorithm for multi-source data. The alignment formula is:

[0022] in, is the clock deviation, , is the current and maximum allowed clock drift; Attention mechanism unit: A vectorized model using a multi-head attention mechanism for unstructured text data, with the output being:

[0023] in, is the number of attention heads, , , is the query, key, and value matrix of the j-th attention head.

[0024] Furthermore, the system also includes a risk warning module, the logic of which is: When the composite similarity between the user's historical behavior data and the current application information is lower than the threshold, an early warning is triggered. The similarity calculation formula is:

[0025] in, , is the weight coefficient, , are the historical and current feature vectors.

[0026] Furthermore, the decision execution module continuously monitors user behavior after loan issuance. If an abnormal transaction pattern is detected, the credit recovery procedure is initiated. The abnormality determination formula is:

[0027] in, is the anomaly score output by the isolation forest algorithm, is the local outlier factor score, is the algorithm fusion weight.

[0028] The above-mentioned embodiments of the present invention have at least the following beneficial effects: the automated risk control management system of the present invention can realize efficient and accurate risk assessment and decision execution. The data acquisition module obtains user credit behavior data and third-party credit data in real time, and combined with the improved algorithm of the feature processing module, it can effectively screen and optimize features and improve the input quality of the model. The risk assessment module adopts an improved logistic regression model, combined with a dynamically adjusted mixed penalty coefficient and a nonlinear activation penalty, which can more accurately output user risk scores, thereby providing a reliable basis for the decision execution module. The decision execution module divides user levels according to risk scores and dynamically adjusts thresholds, which can flexibly generate loan decisions, and at the same time trigger a manual review process for high-risk users to ensure the rigor of the decision. In addition, the model optimization module can further improve the adaptability and accuracy of the system by adaptively adjusting model parameters through feedback data.

[0029] The system also has powerful data processing and risk warning capabilities. The data collection module captures multi-source user data in real time through a distributed API interface, and uses a vectorized model with a time alignment algorithm and a multi-head attention mechanism to effectively process unstructured data and improve the efficiency and accuracy of data collection. The risk warning module can monitor user behavior data in real time, and when abnormal transaction patterns are detected, it will promptly start the credit recovery procedure to prevent potential risks. Overall, the system can enhance the risk management capabilities of financial credit business, reduce default risks, and improve business efficiency and user experience. BRIEF DESCRIPTION OF THE DRAWINGS

[0030] The above and other objects, features and advantages of the exemplary embodiments of the present invention will become readily understood by reading the detailed description below with reference to the accompanying drawings. In the accompanying drawings, several embodiments of the present invention are shown in an exemplary and non-limiting manner, in which: Figure 1 A schematic diagram of the structure of an automated risk control management system for financial credit business provided by one embodiment of the present invention. DETAILED DESCRIPTION

[0031] The principles and spirit of the present invention will be described below with reference to several exemplary embodiments. It should be understood that these embodiments are provided only to enable those skilled in the art to better understand and implement the present invention, and are not intended to limit the scope of the present invention in any way. On the contrary, these embodiments are provided to make the present invention more thorough and complete, and to fully convey the scope of the present invention to those skilled in the art.

[0032] Those skilled in the art know that the embodiments of the present invention can be implemented as a system, device, apparatus, method or computer program product. Therefore, the present invention can be specifically implemented in the following forms, namely: complete hardware, complete software (including firmware, resident software, microcode, etc.), or a combination of hardware and software.

[0033] It should be noted that any number of elements in the drawings is for illustration rather than limitation, and any naming is only for distinction and does not have any limiting meaning.

[0034] Reference below Figure 1 , Figure 1 This is a schematic diagram of the structure of an automated risk control management system for financial credit business provided by an embodiment of the present invention. Figure 1 As shown, an automated risk control management system for financial credit business includes: Data collection module 100, feature processing module 200, risk assessment module 300, decision execution module 400, model optimization module 500; The data collection module 100 is used to obtain user credit behavior data and third-party credit data in real time; The feature processing module 200 is used to clean, reduce the dimension and enhance the features of the data; The risk assessment module 300 outputs a user risk score based on an improved logistic regression model; The decision execution module 400 generates a loan decision based on the risk score and the dynamic threshold; The model optimization module 500 adaptively adjusts the model parameters through feedback data.

[0035] It should be noted that the system includes a data acquisition module, a feature processing module, a risk assessment module, a decision execution module and a model optimization module. The data acquisition module is used to obtain user credit behavior data and third-party credit data in real time to ensure the comprehensiveness and timeliness of the data. The feature processing module cleans, reduces the dimension and enhances the features of the collected data to improve the data quality and feature expression capabilities. The risk assessment module outputs the user's risk score based on the improved logistic regression model to provide a basis for subsequent decision-making. The decision execution module generates loan decisions based on risk scores and dynamic thresholds to ensure the rationality and security of decisions. The model optimization module adaptively adjusts the model parameters through feedback data to improve the accuracy and adaptability of the model.

[0036] Specifically, the data collection module captures user social network behavior data in real time through a distributed API interface, and processes multi-source data using a time alignment algorithm. The feature processing module uses an improved random forest algorithm to screen features, and uses an interpolation algorithm based on user behavior patterns to fill missing values. It processes continuous features in segments and uses an adaptive binning algorithm based on KL divergence. The improved logistic regression model of the risk assessment module combines the L1-L2 mixed penalty coefficient and nonlinear activation penalty to provide a more accurate risk score. The decision execution module divides users into levels according to risk scores, and triggers a manual review process for high-risk users to ensure the rigor of decisions. The model optimization module calculates the model error gradient through feedback data, and uses an improved adaptive momentum optimizer to update parameters.

[0037] Preferably, the data acquisition module can vectorize unstructured text data through a multi-head attention mechanism to improve data processing efficiency. The improved random forest algorithm of the feature processing module can screen features according to their importance to ensure the accuracy of feature selection. The improved logistic regression model of the risk assessment module can dynamically balance the regularization strength in combination with periodic adjustment parameters to improve the adaptability of the model. The decision execution module can dynamically adjust the user level division threshold based on the historical risk score mean and standard deviation to ensure the flexibility of decision-making. The model optimization module can improve the accuracy of model parameter updates by calculating the weight coefficients of the KL divergence and the Wasserstein distance.

[0038] In some embodiments, the feature processing module 200 includes the following steps: Step 201: Use the improved random forest algorithm to select features. The specific formula is:

[0039] in, is the comprehensive weight of the i-th feature, is the preset entropy weight coefficient, is the information value weight, Features The value of information, is the input feature variable, is the target variable; Step 202: The missing values ​​are filled using an interpolation algorithm based on the user behavior pattern. The interpolation formula is:

[0040] in, is the dynamic adjustment coefficient, is the historical data smoothing coefficient, It is the exponential moving average of historical data; Step 203: Segment the continuous features using an adaptive binning algorithm based on KL divergence. The binning boundary formula is:

[0041] in, is the kth bin boundary, , are the sample distribution within and between boxes, respectively.

[0042] It should be noted that the feature processing module is a key part of the system for data preprocessing. Its main purpose is to improve data quality and extract valuable information by cleaning, reducing dimensions and enhancing features. Specifically, the feature processing module includes three steps: first, the improved random forest algorithm is used to screen features. The calculation formula of the comprehensive weight combines parameters such as entropy weight coefficient and information value weight to evaluate the importance of features; second, the missing values ​​are filled with an interpolation algorithm based on user behavior patterns, and the missing data is filled by dynamic adjustment coefficients and historical data smoothing coefficients; finally, the continuous features are segmented and an adaptive binning algorithm based on KL divergence is used to optimize the distribution of features. These steps work together to ensure the accuracy and effectiveness of feature processing.

[0043] Specifically, in the improved random forest algorithm, the entropy weight coefficient is used to measure the uncertainty of the feature, and the information value weight reflects the predictive ability of the feature to the target variable. In the calculation formula of the comprehensive weight, the information value (IV) is an important indicator to measure the ability of feature discrimination, while the entropy weight coefficient ( ) can adjust its weight proportion according to the distribution of features. In the process of missing value filling, the dynamic adjustment coefficient ( ) and the historical data smoothing coefficient ( ) can be adjusted according to the volatility of the data to ensure the rationality of the interpolation results. The exponential moving average (EMA) of historical data is used to smooth the impact of historical data and avoid interpolation errors caused by data mutations. For segmented processing of continuous features, KL divergence is used to measure the difference in data distribution before and after binning, so as to dynamically adjust the binning boundaries to ensure that the data distribution in each bin is statistically significant.

[0044] Preferably, the entropy weight coefficient ( ) can be adjusted dynamically according to the diversity of features. For example, when the feature distribution is relatively uniform, the value to reduce the impact of uncertainty. Information value weight ( ) can be optimized by analyzing the correlation between the feature and the target variable, such as using mutual information or chi-square test to evaluate the importance of the feature. In the process of missing value filling, the dynamic adjustment coefficient ( ) can be adjusted according to the missing rate of the data. For example, when the missing rate is high, appropriately increase In addition, for the binning algorithm, other distribution difference metrics (such as JS divergence) can be introduced as an alternative to further optimize the binning effect.

[0045] In some embodiments, the improved logistic regression model in the risk assessment module 300 is:

[0046] in, is the intercept term, is the feature weight vector, is the L1-L2 mixed penalty coefficient, is the nonlinear activation penalty coefficient, .

[0047] It should be noted that the risk assessment module is the core part of the system for assessing user credit risk. It uses an improved logistic regression model to quantify the user's risk. Based on the traditional logistic regression, the model introduces L1-L2 mixed penalty coefficients and nonlinear activation penalty coefficients to enhance the model's generalization ability and ability to capture complex data relationships. Specifically, in the output formula of the model, the intercept term ( ) and the feature weight vector ( ) is used for linear combination features, while the L1-L2 mixed penalty coefficient ( ) and the nonlinear activation penalty coefficient (μ) are used to adjust the regularization strength and nonlinear feature expression ability of the model. This improved logistic regression model can better adapt to the complexity and dynamics of financial credit data, thereby providing more accurate risk scores.

[0048] Specifically, in the improved logistic regression model, the L1-L2 mixed penalty coefficient ( ) is used to balance the effects of L1 regularization (Lasso) and L2 regularization (Ridge), where L1 regularization helps to sparse feature weights and reduce the number of features, while L2 regularization helps to smooth weights and avoid overfitting. Nonlinear activation penalty coefficient ( ) By introducing the nonlinear characteristics of the ReLU function, the model's ability to capture complex data relationships is enhanced. The ReLU function (Rectified Linear Unit) is a commonly used activation function, which is defined as , which can effectively alleviate the gradient vanishing problem. In the model, the feature weight vector ( ) and the intercept term ( ) is learned through training data, and the mixed penalty coefficient ( ) and the nonlinear activation penalty coefficient ( ) needs to be dynamically adjusted according to the performance of the model during training. For example, the mixed penalty coefficient ( ) can be dynamically adjusted by calculating the variance inflation factor (VIF) of the model to ensure the stability and generalization ability of the model.

[0049] Preferably, the mixed penalty coefficient ( ) can be further refined, for example by introducing a periodic adjustment parameter ( ), which is used to dynamically balance the regularization strength. The periodic adjustment parameter can be adjusted according to the number of iterations of model training ( ) to make adjustments, for example using the formula: , where VIF (variance inflation factor) is used to measure the degree of multicollinearity between features. Used to control the periodic change of regularization strength. In addition, the nonlinear activation penalty coefficient ( ) can be adjusted according to the training error of the model. For example, if the training error is large, As an alternative, other nonlinear activation functions (such as LeakyReLU or ELU) can be introduced to replace the ReLU function to further improve the flexibility and adaptability of the model.

[0050] In some embodiments, the hybrid penalty coefficient The calculation formula is:

[0051] in, is the number of model training iterations, is a periodic adjustment parameter used to dynamically balance the regularization strength.

[0052] It should be noted that the calculation formula of the hybrid penalty coefficient is the key mechanism for dynamically adjusting the regularization strength in the system. By introducing the number of model training iterations, periodic adjustment parameters, and variance inflation factor (VIF), the formula can adaptively adjust the regularization strength according to dynamic changes in the model training process. This dynamic adjustment method helps to balance the bias and variance during the model training process, avoid overfitting or underfitting problems, and thus improve the generalization ability and stability of the model.

[0053] Specifically, in the calculation formula of the mixed penalty coefficient, the number of model training iterations ( ) reflects the progress of model training. As the number of iterations increases, the model gradually becomes stable. Periodic adjustment parameters ( ) is used to control the periodic change of regularization strength by adjusting The value of can change the frequency of fluctuations in the regularization strength. Variance inflation factor (VIF) is an indicator of the degree of multicollinearity between features. The higher the VIF value, the stronger the correlation between features. In this case, stronger regularization is needed to suppress overfitting. The calculation formula of the mixed penalty coefficient is: ,in Represents the feature weight vector. In this way, the mixed penalty coefficient can be dynamically adjusted according to the actual situation of model training, ensuring that the model can maintain a good regularization effect at different stages.

[0054] Preferably, the periodic adjustment parameter ( ) can be adjusted according to the specific needs of model training. For example, when the data set has many features and there is a strong correlation, it can be appropriately increased The value of can be adjusted to enhance the periodic adjustment effect of regularization. In addition, in order to further improve the flexibility of the model, other dynamic adjustment mechanisms can be introduced, such as regularization strength adjustment methods based on learning rate decay. For example, as the number of training iterations increases, the learning rate gradually decreases, and the regularization strength can also be adjusted accordingly. As an alternative, the introduction of Bayesian regularization methods can be considered to dynamically adjust the regularization strength through probability distribution, thereby providing a more flexible regularization strategy for the model.

[0055] In some embodiments, the generation logic of the decision execution module 400 includes: Level division unit: divide users into levels according to risk scores. The division formula is:

[0056] in, , , is the historical risk score average, is the standard deviation, , is the dynamic adjustment coefficient; Credit limit freezing unit: triggers the manual review process for Class C users and freezes the credit limit based on composite rules. The freezing formula is:

[0057] in, , The current and historical average risk scores are respectively.

[0058] It should be noted that the generation logic of the decision execution module is the key link in the system for generating loan decisions based on risk scores. This module ensures the flexibility and rigor of decision-making by dividing users into different levels and triggering the manual review process. Specifically, user level division is based on risk scores and dynamic adjustment coefficients, and users are divided into different risk levels so that corresponding decision measures can be taken according to different levels. For high-risk users, the system will trigger a manual review process and freeze the credit limit based on composite rules to further assess the risk and take necessary risk control measures.

[0059] Specifically, the level division unit of the decision execution module divides the user level according to the risk score, where the dynamic adjustment coefficient ( ) is used to calculate the risk score based on the historical risk score average ( ) and standard deviation ( ) Dynamically adjust the threshold of the classification. This dynamic adjustment mechanism can flexibly adjust the risk classification criteria according to changes in historical data to ensure the adaptability of decision-making. The credit limit freezing unit triggers the manual review process for C-level users and freezes the credit limit based on composite rules. In the formula for freezing the credit limit, the current risk score ( ) and the historical average risk score ( ) is used to evaluate the changing trend of user risk, and the dynamic adjustment coefficient ( ) is used to adjust the proportion of frozen quotas according to changes in risk. In this way, the system can further reduce risks based on automated decision-making and manual review.

[0060] Preferably, the dynamic adjustment coefficient ( ) can be regularly updated based on historical data accumulated during system operation to ensure its adaptability. For example, the distribution changes of historical risk scores can be analyzed to dynamically adjust and value to make it closer to the current risk distribution characteristics.

[0061] Furthermore, for the manual review process, an auxiliary decision-making mechanism of the expert system can be introduced to conduct a more detailed risk assessment of high-risk users through expert rules. As an alternative, the calculation method of freezing the credit limit can introduce more dimensional risk indicators, such as user credit history, industry risk, etc., to more comprehensively assess user risks and formulate corresponding credit limit adjustment strategies.

[0062] In some embodiments, the dynamic adjustment coefficient , The update logic is:

[0063] in, is the learning rate, is the AUC impact factor, is the area under the ROC curve of the current model.

[0064] It should be noted that the update logic of the dynamic adjustment coefficient is the key mechanism for optimizing the decision execution module in the system. ) is used as the update formula. The system can dynamically adjust the threshold of user classification based on the performance indicators of the current model (such as the area under the ROC curve), thereby ensuring the flexibility and adaptability of decision-making. This update mechanism combines the learning rate ( )、AUC impact factor( ) and other parameters, so that the system can dynamically optimize the decision threshold according to the actual performance of the model to better adapt to changes in data distribution.

[0065] Specifically, in the update formula of the dynamic adjustment coefficient, the learning rate ( ) is used to control the update step size and determines the magnitude of the coefficient change at each update. AUC influence factor ( ) reflects the influence of model performance on the dynamic adjustment coefficient, which is calculated by calculating the difference between the area under the ROC curve (AUC) of the current model and the target AUC ( ), the system can dynamically adjust and The update formula is:

[0066] in, is a smoothing parameter used to avoid the denominator being zero; is the historical adjustment range, In this way, the dynamic adjustment coefficient can be dynamically optimized according to the performance of the model to ensure that the threshold of user level classification is always in the optimal state.

[0067] Preferably, the learning rate ( ) can be adaptively adjusted according to the convergence speed during model training. For example, in the initial training stage of the model, a higher learning rate can be set to quickly adjust the coefficients; as the training progresses, the learning rate can be gradually reduced to fine-tune the adjustment. In addition, the AUC influencing factor ( ) can be adjusted according to business needs. For example, in scenarios with higher risk control requirements, As an alternative, other performance indicators (such as the area under the precision-recall curve, PR-AUC) can be introduced to replace ROC-AUC to better meet the needs of different business scenarios.

[0068] In some embodiments, the model optimization module 500 includes the following steps: Step 501: Calculate the model error gradient based on user feedback data, the formula is:

[0069] in, , are the weight coefficients of KL divergence and Wasserstein distance respectively, , For historical and current data distribution; Step 502: Update parameters using an improved adaptive momentum optimizer, the iterative formula of which is:

[0070] in, is the gradient direction correction coefficient, which is used to avoid local optimum.

[0071] It should be noted that the model optimization module is a key link in the system for improving model performance and adaptability. This module calculates the model error gradient based on user feedback data and uses an improved adaptive momentum optimizer to update model parameters, thereby achieving dynamic optimization of the model. This method combines statistical measurement tools such as KL divergence and Wasserstein distance, which can effectively capture changes in data distribution and avoid local optimal problems through gradient direction correction coefficients, ensuring that the model maintains good performance in a dynamic environment.

[0072] Specifically, in the model optimization module, KL divergence (Kullback-Leibler Divergence) is used to measure the difference between the model's predicted distribution and the true distribution, while Wasserstein distance (also known as Earth Mover's Distance) is used to measure the distance between two probability distributions, which can handle the difference between distributions more smoothly. When calculating the model error gradient, the weight coefficients of KL divergence and Wasserstein distance ( and ) can be dynamically adjusted according to the changes in data distribution to balance the impact of the two metrics. The improved adaptive momentum optimizer introduces a gradient direction correction coefficient ( ), correct the gradient direction to prevent the model from falling into the local optimum. The parameter update formula is:

[0073] in, is the learning rate, is the smoothing parameter, is the momentum term, is the loss function, P and Q are the historical and current data distributions respectively.

[0074] Preferably, the weight coefficient and It can be adjusted according to the complexity of data distribution. For example, when the data distribution changes greatly, you can increase The value of to enhance the sensitivity of Wasserstein distance to distribution changes. Gradient direction correction coefficient It can be adjusted dynamically according to the convergence speed of the model optimization process. For example, when the model is close to convergence, the to reduce the correction amplitude in the gradient direction.

[0075] Furthermore, as an alternative, other optimization algorithms (such as AdamW or RMSprop) can be introduced to replace the improved adaptive momentum optimizer, or combined with Bayesian optimization methods to further improve the flexibility and adaptability of model optimization.

[0076] In some embodiments, the data acquisition module 100 includes: Alignment unit: Capture user social network behavior data in real time through a distributed API interface, and use a time alignment algorithm for multi-source data. The alignment formula is:

[0077] in, is the clock deviation, , is the current and maximum allowed clock drift; Attention mechanism unit: A vectorized model using a multi-head attention mechanism for unstructured text data, with the output being:

[0078] in, is the number of attention heads, , , is the query, key, and value matrix of the j-th attention head.

[0079] It should be noted that the data collection module is a key part of the system for obtaining user credit behavior data and third-party credit data. The module captures user social network behavior data in real time through a distributed API interface, and uses a time alignment algorithm to process multi-source data to ensure the time consistency of the data. At the same time, for unstructured text data, the module uses a vectorization model with a multi-head attention mechanism to process it, thereby converting text data into a numerical form that can be used by machine learning models. This data collection and processing method can effectively improve the integrity and availability of data, and provide a solid foundation for subsequent risk assessment and decision-making.

[0080] Specifically, the distributed API interface is used to connect multiple data sources, including user credit behavior data, third-party credit data, and social network behavior data. The time alignment algorithm calculates the clock deviation ( ) and the maximum allowed clock drift ( ) to adjust the timestamp of the data to ensure the temporal consistency of data from different sources. The formula is:

[0081] in, is the current data timestamp, and are the maximum and minimum allowed timestamps, respectively. For unstructured text data, the vectorization model of the multi-head attention mechanism encodes the text through multiple attention heads (H), and the query (Q), key (K), and value (V) matrices of each attention head are used to capture the semantic relationship in the text. The output formula is:

[0082] in, is a weight matrix used to integrate the outputs of different attention heads.

[0083] Preferably, the clock bias in the time alignment algorithm ( ) can be adjusted dynamically according to the stability of the data source. For example, for a data source with higher stability, the adjustment range of the clock deviation can be appropriately reduced. For the multi-head attention mechanism, the number of attention heads (H) can be optimized according to the complexity of the text data. For example, when processing long texts or texts with complex semantic relationships, the number of attention heads can be increased to improve the expressiveness of the model.

[0084] Furthermore, as an alternative, the position encoding mechanism in the Transformer architecture can be introduced to further enhance the model's perception of text order, thereby improving the processing effect of unstructured text data.

[0085] In some embodiments, the system further comprises a risk warning module (600), the logic of which is: When the composite similarity between the user's historical behavior data and the current application information is lower than the threshold, an early warning is triggered. The similarity calculation formula is:

[0086] in, , is the weight coefficient, , are the historical and current feature vectors.

[0087] It should be noted that the risk warning module is a key part of the system for real-time monitoring of user behavior data and triggering warnings. The module calculates the composite similarity between the user's historical behavior data and the current application information, and triggers a warning when the similarity is lower than the set threshold, thereby timely discovering potential risks. This method combines the intersection of feature vectors and cosine similarity, which can evaluate the consistency of user behavior from different angles and ensure the accuracy and timeliness of warnings.

[0088] Specifically, in the similarity calculation formula of the risk warning module, the weight coefficient ( and ) is used to balance the influence of intersection-over-union and cosine similarity. Jaccard Similarity measures similarity by calculating the ratio of the intersection and union of the historical feature vector (H) and the current feature vector (C). The formula is:

[0089] Cosine Similarity measures the directional similarity of two vectors by calculating the cosine value of the angle between them. The formula is:

[0090] In the risk warning module, and It can be adjusted according to actual business needs and data characteristics. For example, when the dimension of historical data is high, you can increase to emphasize the role of cosine similarity; conversely, when data sparsity is high, you can increase In this way, the risk warning module can flexibly adapt to the risk monitoring needs in different scenarios.

[0091] Preferably, the weight coefficient ( and ) can be adjusted dynamically through machine learning methods. For example, use historical warning data to train a classifier and dynamically adjust the classifier based on its performance. and value to optimize the warning effect.

[0092] Furthermore, other similarity measurement methods (such as Euclidean distance or Manhattan distance) can be introduced as supplements to further improve the robustness of the risk warning module. As an alternative, the user's behavior pattern change rate (such as transaction frequency, amount change, etc.) can be combined as an auxiliary indicator for early warning, so as to more comprehensively evaluate the abnormality of user behavior.

[0093] In some embodiments, the decision execution module 400 continuously monitors user behavior after loan issuance. If an abnormal transaction pattern is detected, the credit recovery procedure is initiated. The abnormality determination formula is:

[0094] in, is the anomaly score output by the isolation forest algorithm, is the local outlier factor score, is the algorithm fusion weight.

[0095] It should be noted that the function of the decision execution module to continuously monitor user behavior after loan issuance is an important mechanism in the system for preventing potential risks. This module effectively reduces the risk of default by detecting abnormal transaction patterns in user behavior data and initiating the credit recovery procedure when anomalies are detected. This method combines the isolation forest algorithm and the local outlier factor algorithm to evaluate the degree of abnormality of user behavior by calculating the anomaly score and the outlier factor score, ensuring that the system can respond to potential risks in a timely manner.

[0096] Specifically, in the determination formula of abnormal trading patterns, the Isolation Forest algorithm detects outliers in the data by constructing an isolation tree and outputs anomaly scores ( The Local Outlier Factor (LOF) algorithm evaluates the degree of outlier of a data point by calculating the density deviation of the data point in the local neighborhood and outputs the local outlier factor score ( ). The formula is:

[0097] Among them, n is the number of user behavior features, m is the number of features for isolation forest detection, The algorithm fusion weight is used to balance the influence of the isolation forest and local outlier factor algorithms. In practical applications, the number of trees and subsample size of the isolation forest algorithm can be adjusted according to the scale and complexity of the data, and the neighborhood size (k) of the local outlier factor algorithm can be set according to the density distribution of the data. In this way, the system can flexibly adapt to the characteristics of different user behavior data and accurately detect abnormal transaction patterns.

[0098] Preferably, the algorithm fusion weights ( ) can be adjusted dynamically based on the detection performance of the Isolation Forest and Local Outlier Factor algorithms. For example, when the detection accuracy of the Isolation Forest algorithm is high, to enhance its role in abnormality determination.

[0099] Furthermore, other anomaly detection algorithms (such as deep learning-based autoencoders) can be introduced as supplements to further improve the accuracy of anomaly detection. As an alternative, the threshold for anomaly determination can be dynamically adjusted based on the user's historical transaction behavior patterns to more accurately identify potential risks.

[0100] The above-mentioned embodiments of the present invention have the following beneficial effects: The automated risk control management system of the present invention can achieve comprehensive risk control and management by integrating multiple modules. The data acquisition module can obtain user credit behavior data and third-party credit data in real time to ensure the timeliness and accuracy of the data. The feature processing module uses an improved random forest algorithm to screen features, and cleans and optimizes the data through an interpolation algorithm and an adaptive binning algorithm based on user behavior patterns, which can effectively improve data quality and feature expression capabilities. The risk assessment module is based on an improved logistic regression model, combined with the L1-L2 mixed penalty coefficient and nonlinear activation penalty, to provide a more accurate risk score. The decision execution module generates a loan decision based on the risk score and dynamic threshold, and triggers a manual review process for high-risk users, which can ensure the rationality and security of the decision.

[0101] The model optimization module can continuously improve the accuracy and adaptability of the model by adaptively adjusting the model parameters through feedback data. The data acquisition module can effectively process multi-source data and unstructured text data through distributed API interfaces and multi-head attention mechanisms to improve data processing efficiency. The risk warning module can monitor user behavior data in real time. When abnormal transaction patterns are detected, the credit recovery program is initiated in a timely manner to effectively prevent potential risks. Overall, the system can significantly improve the risk management capabilities of financial credit business, reduce default risks, and improve business efficiency and user experience.

[0102] Furthermore, the storage medium of the embodiment of the present application stores program instructions that can implement all the above methods, wherein the program instructions can be stored in the above storage medium in the form of a software product, including several instructions for enabling a computer device (which can be a personal computer, server, or network device, etc.) or a processor to execute all or part of the steps of the methods described in each embodiment of the present application. The aforementioned storage medium includes: various media that can store program codes, such as a USB flash drive, a mobile hard disk, a read-only memory (ROM), a random access memory (RAM), a magnetic disk or an optical disk, or terminal devices such as a computer, a server, a mobile phone, and a tablet.

[0103] The above descriptions are only some preferred embodiments of the present invention and an explanation of the technical principles used. Those skilled in the art should understand that the scope of the invention involved in the embodiments of the present invention is not limited to the technical solutions formed by a specific combination of the above technical features, but should also cover other technical solutions formed by any combination of the above technical features or their equivalent features without departing from the above inventive concept. For example, the above features are replaced with (but not limited to) technical features with similar functions disclosed in the embodiments of the present invention.

Claims

1. An automated risk control management system for financial credit business, characterized in that: Includes the following modules: Data acquisition module, feature processing module, risk assessment module, decision execution module, model optimization module; The data collection module is used to obtain user credit behavior data and third-party credit data in real time; The feature processing module is used to clean, reduce dimension and enhance features of data; include: Step 201: Using an improved random forest algorithm to select features; Step 202: Fill missing values ​​using an interpolation algorithm based on user behavior patterns; Step 203: segment the continuous features using an adaptive binning algorithm based on KL divergence; The risk assessment module outputs a user risk score based on an improved logistic regression model; The decision execution module generates a loan decision based on the risk score and dynamic threshold; The model optimization module adaptively adjusts the model parameters through feedback data; The decision execution module includes: Level division unit: divide users into levels according to risk scores. The division formula is: in, , , is the historical risk score average, is the standard deviation, , is the dynamic adjustment coefficient; Credit limit freezing unit: triggers the manual review process for Class C users and freezes the credit limit based on composite rules. The freezing formula is: in, , are the current and historical average risk scores, The total credit limit of the user; The dynamic adjustment coefficient , The update logic is shown in the following formula: in, is the learning rate, is the AUC impact factor, is the area under the ROC curve of the current model, is the value of the mth dynamic adjustment coefficient at the tth iteration, is the mth dynamic adjustment coefficient at the tth The value at the iteration, is the false positive rate at the tth iteration, is the true rate at the tth iteration; The model optimization module includes the following steps: Step 501: Calculate the model error gradient based on user feedback data, the formula is: in, , are the weight coefficients of KL divergence and Wasserstein distance respectively, , is the historical and current data distribution, are model parameters; Step 502: Update parameters using an improved adaptive momentum optimizer, the iterative formula of which is: in, is the gradient direction correction coefficient, used to avoid local optimum, is the model parameter at the tth iteration, For the tth The model parameters at the iteration, is the first-order moment estimate, which represents the mean of the smoothed gradient, is the second-order moment estimate, which represents the variance of the smoothed gradient, is the numerical stability coefficient; The data acquisition module comprises: Alignment unit: Capture user social network behavior data in real time through a distributed API interface, and use a time alignment algorithm for multi-source data. The alignment formula is: in, is the clock deviation, , is the current and maximum allowed clock drift, is the timestamp after alignment, is the original timestamp; Attention mechanism unit: A vectorized model using a multi-head attention mechanism for unstructured text data, with the output being: in, is the number of attention heads, , , is the query, key, and value matrix of the jth attention head, is the i-th position in the text sequence, is the length of the text sequence, is the dimension of the feature vector at the i-th position.

2. The system according to claim 1, characterized in that Said Step 201: Use the improved random forest algorithm to select features. The specific formula is: in, is the comprehensive weight of the i-th feature, is the preset entropy weight coefficient, is the information value weight, Features The value of information, is the input feature variable, is the target variable; is the feature index, indicating the jth feature, Features The information entropy of Features The mutual information between the target variable Y; Step 202: The missing values ​​are filled using an interpolation algorithm based on the user behavior pattern. The interpolation formula is: in, is the dynamic adjustment coefficient, is the historical data smoothing coefficient, is the exponential moving average of historical data, is the filling result of missing values, is the mean of the feature, is the mode of the feature; Step 203: Segment the continuous features using an adaptive binning algorithm based on KL divergence. The binning boundary formula is: in, is the kth bin boundary, , are the sample distribution within and between boxes, respectively. is the bin index, indicating the Boxes, is the boundary among all bin boundaries that minimizes the objective function.

3. The system according to claim 1, characterized in that The improved logistic regression model in the risk assessment module is: in, is the intercept term, is the feature weight vector, is the L1-L2 mixed penalty coefficient, is the nonlinear activation penalty coefficient, .

4. The system according to claim 3, characterized in that The mixed penalty coefficient The calculation formula is: in, is the number of model training iterations, is a periodic adjustment parameter used to dynamically balance the regularization strength, is the feature matrix, is the sample size, is the natural logarithm of the sample size, is a cosine function that represents periodic changes.

5. The system according to claim 1, characterized in that The system also includes a risk warning module, the logic of which is: When the composite similarity between the user's historical behavior data and the current application information is lower than the threshold, an early warning is triggered. The similarity calculation formula is: in, , is the weight coefficient, , are the historical and current feature vectors.

6. The system according to claim 1, characterized in that The decision execution module continuously monitors user behavior after loan issuance. If an abnormal transaction pattern is detected, the credit recovery procedure is initiated. The abnormality determination formula is: in, is the anomaly score output by the isolation forest algorithm, is the local outlier factor score, is the algorithm fusion weight, is the current value of the kth feature, is the historical mean of the kth feature, is the historical standard deviation of the kth feature, x is the feature vector of the current sample, is the current value of the lth feature.

Images