Business risk control strategy intelligent prediction and response system

Abstract

The present invention relates to the field of intelligent prediction of business risks, and specifically refers to an intelligent prediction and response system for business risk control strategies, including a data collection and integration module, a risk assessment module, an intelligent prediction and decision module, a real-time monitoring and response module, and a user interface module; the scheme adopts a graph neural network to establish a neighborhood noise purifier, a core node intensifier, and a relationship summarizer to weaken the adverse effects of residual noise; and uses a convolutional neural network and a long short-term memory network to solve the problem of feature weakening in the graph neural network; an optimized Stacking model is used to predict and decide future business risks. The model combines the features extracted by the convolutional neural network, and performs comprehensive feature selection through three feature selection methods to enhance the generalization ability and robustness of the model, and improve the existing loss function to make it more suitable for the needs of risk prediction and decision-making.

Description

Technical Field

[0001] The present invention relates to the field of intelligent prediction of business risks, and specifically to an intelligent prediction and response system for business risk control strategies. Background Art

[0002] In the financial industry, the assessment and prediction of business risk is a crucial task. Existing business risk assessments mostly use big data analysis, machine learning algorithms, and artificial intelligence technologies to build efficient risk prediction models. Selecting the correct variables as input for existing business risk control prediction models is crucial to improving model performance. However, existing business risk control prediction models are unable to accurately extract truly useful features from massive amounts of data, lacking in-depth domain knowledge and data analysis skills. Existing traditional models often have limitations in generalization and robustness when predicting and responding to business risks, which significantly affect their prediction accuracy. Traditional feature selection methods ignore some implicit information that is crucial for prediction. At the same time, existing loss functions may not accurately reflect the model's predictive performance when assessing business risks. Summary of the Invention

[0003] In response to the above situation, in order to overcome the shortcomings of the existing technology, the present invention provides an intelligent prediction and response system for business risk control strategies. To address the problem that existing business risk control prediction models cannot accurately extract truly useful features from massive data and lack deep domain knowledge and data analysis skills, this solution uses a graph neural network to establish a neighborhood noise purifier, a core node strengthener, and a relationship summarizer. The neighborhood noise purifier uses a multi-layer perceptron to construct a similarity evaluation mechanism, and uses a core node strengthening mechanism to implement a weighted strategy for core nodes when aggregating neighborhood information to weaken the adverse effects of residual noise; and uses convolutional neural networks and long short-term memory networks to solve the problem of feature weakening in graph neural networks; to address the problem that existing traditional models often have limitations in generalization and robustness when predicting and responding to business risks, an optimized Stacking model is used to predict and decide future business risks; the model combines the features extracted by convolutional neural networks and performs comprehensive feature selection through three feature selection methods to enhance the generalization and robustness of the model, ensuring that implicit information that is critical to prediction is not missed; and improves the existing loss function to make it more suitable for the needs of risk prediction and decision-making.

[0004] The business risk control strategy intelligent prediction and response system provided by the present invention includes a data collection and integration module, a risk assessment module, an intelligent prediction and decision module, a real-time monitoring and response module and a user interface module;

[0005] The data collection and integration module collects multi-source data related to business risks in the financial industry and integrates the data to obtain primary multi-source data;

[0006] The risk assessment module acquires historical data and industry knowledge to establish a business risk assessment model, analyzes and processes primary multi-source data, evaluates the existing risk levels in different types of business activities, and obtains the current risk level and score of the business;

[0007] The intelligent prediction and decision-making module uses an optimized Stacking model to accurately predict possible future business risks based on current risk levels and scoring data, and outputs the prediction results. Based on the prediction results, it provides management with a decision-making management plan and is directly applied to the automated process for business risk control.

[0008] The real-time monitoring and response module continuously tracks changes in key indicators of business risks. Once abnormal activities are detected, it immediately triggers an early warning notification and initiates the corresponding emergency plan. It also has a certain degree of self-learning ability and can continuously optimize its own judgment criteria as the external environment changes;

[0009] The user interface module provides users with a friendly and easy-to-use operation interface. Non-technical personnel can view the data of the risk assessment module, intelligent prediction decision module and real-time monitoring and response module and set the parameters of each module based on simple operations.

[0010] Furthermore, the risk assessment module acquires historical data and industry knowledge to establish a business risk assessment model. The business risk assessment model includes a neighborhood noise purifier, a core node enhancer, and a relationship aggregator, and is constructed using a graph neural network, a convolutional neural network, and a long short-term memory network.

[0011] The neighborhood noise purifier processes information from different historical data and industry knowledge as nodes, uses a multi-layer perceptron to predict node features, and uses L1 distance to calculate the similarity between nodes. The formula used is as follows:

[0012] ;

[0013] ;

[0014] Where, and Represents two different node indexes, and Representation node and nodes In the The feature embedding of the layer, Indicates in Layer Time Node and nodes The dissimilarity measure between Represents the first The output of the layer, represents the L1 paradigm, Representation node and nodes In the Similarity score when layer;

[0015] Sort the neighbors between nodes according to the calculated similarity, select the central node, and select the n most similar neighbors as the effective neighborhood of the central node to obtain the neighbor set. The formula used is as follows:

[0016] ;

[0017] ;

[0018] Where, Represents the relationship between nodes, Indicates in Layer Relationship midpoint The number of samples, Indicates in Layer Relationship Set the filtering threshold, Representation node In relationship The total number of neighbors under Indicates in Layers and relationships Next, node The neighbor set after similarity filtering, Represents the operation of selecting elements from a sorted list. For sorting operations;

[0019] The core node enhancer includes a weighted aggregation unit and a reinforcement learning unit. When updating the information of the central node, the weighted aggregation unit aggregates information from neighbors and increases the weight of the central node itself.

[0020] The reinforcement learning unit uses a reinforcement learning mechanism to dynamically adjust the filtering threshold of neighbor nodes and enhance the characteristics of the central node. The reinforcement learning mechanism determines whether to give rewards or penalties by observing the changes in the average distance and optimizes the filtering threshold;

[0021] After enhancing the central node features, the relationship aggregator performs cross-relationship aggregation, combining the embedding features of each node under different relationships with the embedding features of the previous layer to generate a comprehensive embedding feature;

[0022] Furthermore, a graph neural network is used to construct a neighborhood noise purifier, a core node enhancer, and a relationship aggregator to obtain the final output. A convolutional neural network is used to perform a convolution operation on the final output of the graph neural network to extract local features and perform a pooling operation to obtain a fixed-length feature vector.

[0023] The long short-term memory network converts the fixed-length feature vector output by the convolutional neural network into the input of the bidirectional LSTM model, and combines the forward and reverse hidden states to obtain the integrated hidden layer output.

[0024] Furthermore, the intelligent prediction and decision-making module uses an optimized Stacking model to accurately predict possible future business risks based on the current risk level and scoring data. The optimized Stacking model specifically includes the following steps:

[0025] Step S1: Primary learning layer, which integrates the current risk level and score data into dataset D, randomly divides dataset D into K subsets for cross-validation. The primary learning layer contains T1 learning algorithms, uses K-1 subsets as training sets, and the remaining subset as test set, repeating K times. The data of each subset is predicted by all T1 learning algorithms to construct a new dataset D′;

[0026] Step S2: Meta-learning layer, using SVM as a meta-learner, training with the new dataset D′ as input to learn a more accurate decision boundary;

[0027] Step S3: Based on the CNN-based risk decision model, obtain user data, use CNN to extract features from the user data, and use the CFA method to generate an additional minority class sample balanced data set, and use SVM for classification;

[0028] Step S4: Feature combination and selection: The data in the new dataset D′ is combined with the features extracted by CNN. Three feature selection methods, namely Pearson correlation filtering, unsupervised variance inflation factor method, and recursive feature elimination method, are used for feature selection. The parameter threshold is adjusted, and the performance of different feature subsets is evaluated using logistic regression.

[0029] Step S5: Loss function optimization, define a new loss function , the formula used is as follows:

[0030] ;

[0031] Where, is the original Stacking model loss function, and is the scaling factor, is the prediction result of the CNN-based risk decision model, The prediction results of the optimized Stacking model;

[0032] Step S6: Adjust coefficient and , calculate the dynamic and The value is calculated using the following formula:

[0033] ;

[0034] ;

[0035] Where, is the sum of the distances between the data points in the CNN-based risk decision model and the optimized Stacking model, and are the maximum and minimum values ​​of the sum of all distances, respectively. is the error parameter.

[0036] The beneficial effects achieved by the present invention using the above scheme are as follows:

[0037] (1) In response to the problem that the existing business risk control prediction model cannot accurately extract truly useful features from massive data and lacks deep domain knowledge and data analysis skills, this solution uses graph neural networks to establish neighborhood noise purifiers, core node enhancers, and relationship summarizers. The neighborhood noise purifier uses a multi-layer perceptron (MLP) to build a similarity evaluation mechanism to distinguish and reduce the noise in neighborhood information based on the similarity between nodes. The core node enhancement mechanism is used to implement a weighted strategy for core nodes when aggregating neighborhood information, weakening the adverse effects of residual noise, ensuring the prominence of core node features during the fusion process, and avoiding information overload; and convolutional neural networks and long short-term memory networks are used to solve the problem of feature weakening in graph neural networks;

[0038] (2) To address the limitations of existing traditional models in generalization and robustness when predicting and responding to business risks, we use an optimized Stacking model to predict and make decisions about future business risks. This model combines features extracted by convolutional neural networks and performs comprehensive feature selection using three feature selection methods to enhance the model's generalization and robustness, ensuring that implicit information that is critical to prediction is not missed. In addition, we have improved the existing loss function to make it more suitable for the needs of risk prediction and decision-making. BRIEF DESCRIPTION OF THE DRAWINGS

[0039] Figure 1A schematic diagram of the business risk control strategy intelligent prediction and response system provided by the present invention;

[0040] Figure 2 It is a structural diagram of the risk assessment module;

[0041] The accompanying drawings are used to provide further understanding of the present invention and constitute a part of the specification. They are used to explain the present invention together with the embodiments of the present invention and do not constitute a limitation of the present invention. DETAILED DESCRIPTION

[0042] The technical solutions in the embodiments of the present invention will be clearly and completely described below in conjunction with the drawings in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, rather than all the embodiments; based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative work are within the scope of protection of the present invention.

[0043] Example 1, see Figure 1 , the business risk control strategy intelligent prediction and response system provided by the present invention includes a data collection and integration module, a risk assessment module, an intelligent prediction and decision module, a real-time monitoring and response module and a user interface module;

[0044] The data collection and integration module collects multi-source data related to business risks in the financial industry and integrates the data to obtain primary multi-source data. The multi-source data includes but is not limited to internal enterprise databases, social media, news reports, customer behavior data, market trend information, and competitor dynamics data.

[0045] The risk assessment module acquires historical data and industry knowledge to establish a business risk assessment model, analyzes and processes primary multi-source data, evaluates the existing risk levels in different types of business activities, and obtains the current risk level and score of the business;

[0046] The intelligent prediction and decision-making module uses an optimized Stacking model to accurately predict possible future business risks based on current risk levels and scoring data, and outputs prediction results. The prediction results are provided to management for decision-making and management solutions, and are directly applied to automated processes for business risk control, including automatic adjustment of credit limit limits during the credit approval process.

[0047] The real-time monitoring and response module continuously tracks changes in key indicators of business risks. Once abnormal activities are detected, it immediately triggers an early warning notification and initiates the corresponding emergency plan. It also has a certain degree of self-learning ability and can continuously optimize its own judgment criteria as the external environment changes;

[0048] The user interface module provides users with a friendly and easy-to-use operation interface. Non-technical personnel can view the data of the risk assessment module, intelligent prediction decision module and real-time monitoring and response module and set the parameters of each module based on simple operations.

[0049] Example 2: This example is based on the above example. The risk levels include low risk, medium risk, and high risk. A specific scoring threshold is set for each risk level. A score of 0-20 is low risk, 21-60 is medium risk, and 61-100 is high risk.

[0050] The key indicators include financial indicators, operational risk indicators, market risk indicators, credit risk indicators, compliance risk indicators, supply chain risk indicators, strategic risk indicators and information security risk indicators;

[0051] Financial indicators include current ratio, debt-to-asset ratio, days’ sales outstanding and net profit margin;

[0052] Operational risk indicators include error rates, downtime (the length of time a system or production process is interrupted), employee turnover, and the number of safety incidents;

[0053] Market risk indicators include volatility, VaR (the maximum possible loss at a certain confidence level), market share changes, and customer satisfaction index;

[0054] Credit risk indicators include default rate, overdue loan rate, credit score, and bad debt provision coverage ratio.

[0055] Example 3, see Figure 1 and Figure 2 This embodiment is based on the above embodiment. The risk assessment module acquires historical data and industry knowledge to establish a business risk assessment model. The business risk assessment model includes a neighborhood noise purifier, a core node enhancer, and a relationship aggregator, and is constructed using a graph neural network, a convolutional neural network, and a long short-term memory network.

[0056] The neighborhood noise purifier processes information from different historical data and industry knowledge as nodes, uses a multi-layer perceptron to predict node features, and uses L1 distance to calculate the similarity between nodes. The formula used is as follows:

[0057] ;

[0058] ;

[0059] Where, and Represents two different node indexes, and Representation node and nodes In the The feature embedding of the layer, Indicates in Layer Time Node and nodes The dissimilarity measure between Represents the first The output of the layer, represents the L1 paradigm, Representation node and nodes In the Similarity score when layer;

[0060] Sort the neighbors between nodes according to the calculated similarity, select the central node, and select the n most similar neighbors as the effective neighborhood of the central node to obtain the neighbor set. The formula used is as follows:

[0061] ;

[0062] ;

[0063] Where, Represents the relationship between nodes, Indicates in Layer Relationship midpoint The number of samples, Indicates in Layer Relationship Set the filtering threshold, Representation node In relationship The total number of neighbors under Indicates in Layers and relationships Next, node The neighbor set after similarity filtering, Represents the operation of selecting elements from a sorted list. For sorting operations;

[0064] The core node enhancer includes a weighted aggregation unit and a reinforcement learning unit. When updating the information of the central node, the weighted aggregation unit aggregates information from neighbors and increases the weight of the central node itself.

[0065] The reinforcement learning unit uses a reinforcement learning mechanism to dynamically adjust the filtering threshold of neighbor nodes and enhance the characteristics of the central node. The reinforcement learning mechanism determines whether to give rewards or penalties by observing the changes in the average distance and optimizes the filtering threshold;

[0066] After enhancing the central node features, the relationship aggregator performs cross-relationship aggregation, combines the embedding features of each node under different relationships with the embedding features of the previous layer to generate a comprehensive embedding feature.

[0067] Example 4, see Figure 1 and Figure 2 Based on the above embodiment, this embodiment uses a graph neural network to construct a neighborhood noise purifier, a core node enhancer, and a relationship aggregator to obtain the final output. A convolutional neural network is used to perform a convolution operation on the final output obtained by the graph neural network to extract local features and perform a pooling operation to obtain a feature vector of fixed length.

[0068] The long short-term memory network converts the fixed-length feature vector output by the convolutional neural network into the input of the bidirectional LSTM model, and combines the forward and reverse hidden states to obtain the integrated hidden layer output.

[0069] Example 5: This example is based on the above example. The experimental environment is configured to run Python 3.8.5 on Ubuntu 18.04.1, and is equipped with a GeForce RTX 2080 Ti GPU, 16GB of memory, and an Intel Xeon Silver 4214 processor (2.20GHz);

[0070] Given the large number of nodes in the dataset, we adopted a mini-batch training strategy to efficiently train the risk assessment model. We used the Adam optimizer for parameter optimization, with an initial learning rate of 0.01, hyperparameters of E=50, L=1, λ=2, and a reinforcement learning step size of τ=0.02. The batch size of the dataset was set to 1024, and the learning rate was 0.01.

[0071] By performing the above operations, in order to address the problem that the existing business risk control prediction model cannot accurately extract truly useful features from massive data and lacks deep domain knowledge and data analysis skills, this solution uses graph neural networks to establish a neighborhood noise purifier, core node strengthener and relationship summarizer. The neighborhood noise purifier uses a multi-layer perceptron (MLP) to build a similarity evaluation mechanism to distinguish and reduce the noise in neighborhood information based on the similarity between nodes. The core node strengthening mechanism is used to implement a weighted strategy for core nodes when aggregating neighborhood information, weakening the adverse effects of residual noise, ensuring the prominence of core node features during the fusion process, and avoiding information overload; and uses convolutional neural networks and long short-term memory networks to solve the problem of feature weakening in graph neural networks.

[0072] Example 6: Based on the above example, the intelligent prediction and decision module uses an optimized Stacking model to accurately predict possible future business risks based on the current risk level and scoring data. The optimized Stacking model specifically includes the following steps:

[0073] Step S1: Primary learning layer, which integrates the current risk level and score data into dataset D, randomly divides dataset D into K subsets for cross-validation. The primary learning layer contains T1 learning algorithms, uses K-1 subsets as training sets, and the remaining subset as test set, repeating K times. The data of each subset is predicted by all T1 learning algorithms to construct a new dataset D′;

[0074] Step S2: Meta-learning layer, using SVM as a meta-learner, training with the new dataset D′ as input to learn a more accurate decision boundary;

[0075] Step S3: Based on the CNN-based risk decision model, obtain user data, use CNN to extract features from the user data, and use the CFA method to generate an additional minority class sample balanced data set, and use SVM for classification;

[0076] Step S4: Feature combination and selection: The data in the new dataset D′ is combined with the features extracted by CNN. Three feature selection methods, namely Pearson correlation filtering, unsupervised variance inflation factor method, and recursive feature elimination method, are used for feature selection. The parameter threshold is adjusted, and the performance of different feature subsets is evaluated using logistic regression.

[0077] Step S5: Loss function optimization, define a new loss function , the formula used is as follows:

[0078] ;

[0079] Where, is the original Stacking model loss function, and is the scaling factor, is the prediction result of the CNN-based risk decision model, The prediction results of the optimized Stacking model;

[0080] Step S6: Adjust coefficient and , calculate the dynamic and The value is calculated using the following formula:

[0081] ;

[0082] ;

[0083] Where, is the sum of the distances between the data points in the CNN-based risk decision model and the optimized Stacking model, and are the maximum and minimum values ​​of the sum of all distances, respectively. is the error parameter.

[0084] Embodiment 7: This embodiment is based on the above embodiment. In step S4, the Pearson correlation filtering is used to select features. Specifically, the Pearson correlation coefficient between each feature and the target variable is calculated, and features are selected based on the correlation.

[0085] The unsupervised variance inflation factor method is used to detect high correlation between features. If the variance inflation factor of a feature exceeds a threshold, it indicates that the feature is highly correlated with other features and the feature is discarded. Otherwise, the feature is retained.

[0086] The recursive feature elimination method is usually combined with a specific learning algorithm to gradually eliminate the least important features until a predetermined number is reached;

[0087] For each feature, each feature selection method will give a decision, which is recorded using 1 and 0, where 1 means retention and 0 means discard. The decisions of the three methods for the same feature are added together and then the average is taken. If the average is greater than 0.5, the feature is retained; otherwise, the feature is discarded.

[0088] By performing the above operations, we address the generalization and robustness limitations of existing traditional models for predicting and responding to business risks. We then employ an optimized Stacking model to predict and make decisions about future business risks. This model combines features extracted by convolutional neural networks with comprehensive feature selection using three methods to enhance the model's generalization and robustness, ensuring that implicit information critical to prediction is not missed. Furthermore, we have improved the existing loss function to make it more suitable for risk prediction and decision-making.

[0089] It should be noted that, in this document, relational terms such as first and second, etc., are used only to distinguish one entity or operation from another entity or operation, and do not necessarily require or imply any actual relationship or order between these entities or operations. Moreover, the terms "comprises," "comprising," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that includes a list of elements includes not only those elements but also other elements not explicitly listed, or elements inherent to such process, method, article, or apparatus.

[0090] While embodiments of the present invention have been shown and described, it will be appreciated by those skilled in the art that various changes, modifications, substitutions, and variations may be made to these embodiments without departing from the principles and spirit of the invention, and that the scope of the invention is defined by the appended claims and their equivalents.

[0091] The present invention and its embodiments are described above. This description is not restrictive. The drawings show only one embodiment of the present invention, and the actual structure is not limited thereto. In short, if a person skilled in the art is inspired by this and, without departing from the purpose of the present invention, designs structures and embodiments similar to this technical solution without inventiveness, they shall fall within the scope of protection of the present invention.

Claims

1. Business risk control strategy intelligent prediction and response system, characterized by: It includes data collection and integration module, risk assessment module, intelligent prediction and decision-making module, real-time monitoring and response module and user interface module; The data collection and integration module collects multi-source data related to business risks in the financial industry and integrates the data to obtain primary multi-source data; The risk assessment module acquires historical data and industry knowledge to establish a business risk assessment model, analyzes and processes primary multi-source data, assesses the existing risk level in different types of business activities, and obtains the current risk level and score of the business; The intelligent prediction and decision-making module uses an optimized Stacking model to accurately predict possible future business risks based on current risk levels and scoring data, and outputs prediction results, which are provided to management for decision-making and management solutions, and directly applied to the automated process for business risk control; The real-time monitoring and response module continuously tracks changes in key indicators of business risks, immediately triggers early warning notifications once abnormal activities are detected, and initiates corresponding emergency plans. It also has a certain self-learning ability and can continuously optimize its own judgment criteria as the external environment changes; The user interface module provides users with a friendly and easy-to-use operation interface, and non-technical personnel can view the data of the risk assessment module, the intelligent prediction decision module and the real-time monitoring and response module and set the parameters of each module according to simple operations; The risk assessment module acquires historical data and industry knowledge to establish a business risk assessment model, which includes a neighborhood noise purifier, a core node enhancer, and a relationship aggregator, and is constructed using a graph neural network, a convolutional neural network, and a long short-term memory network; The neighborhood noise purifier processes information in different historical data and industry knowledge as nodes, uses a multi-layer perceptron to predict node features, and uses L1 distance to calculate the similarity between nodes. The formula used is as follows: ； ； In the formula, and Represents two different node indexes, and Representation Node and nodes In the The feature embedding of the layer, Indicated in Layer Time Node and nodes The dissimilarity measure between Represents the multilayer perceptron The output of the layer, represents the L1 paradigm, Representation Node and nodes In the Similarity score when stratified; Sort the neighbors between nodes according to the calculated similarity scores, select the central node, and select the n most similar neighbors as the effective neighborhood of the central node to obtain the neighbor set; The core node enhancer includes a weighted aggregation unit and a reinforcement learning unit. When updating the information of the central node, the weighted aggregation unit aggregates information from neighbors and increases the weight of the central node itself; The reinforcement learning unit uses a reinforcement learning mechanism to dynamically adjust the filtering threshold of neighbor nodes and enhance the characteristics of the central node. The reinforcement learning mechanism determines whether to give rewards or penalties by observing the changes in the average distance and optimizes the filtering threshold; After enhancing the central node features, the relationship aggregator performs cross-relation aggregation, combining the embedding features of each node under different relationships with the embedding features of the previous layer to generate a comprehensive embedding feature.

2. The business risk control strategy intelligent prediction and response system according to claim 1 is characterized by: Use graph neural networks to build neighborhood noise purifiers, core node enhancers, and relationship aggregators to obtain the final output. Use convolutional neural networks to perform convolution operations on the final output obtained by the graph neural network, extract local features, and perform pooling operations to obtain a feature vector of fixed length. The long short-term memory network converts the fixed-length feature vector output by the convolutional neural network into the input of the bidirectional LSTM model, and combines the forward and reverse hidden states to obtain the integrated hidden layer output.

3. The business risk control strategy intelligent prediction and response system according to claim 1 is characterized by: The intelligent prediction decision module uses an optimized Stacking model to accurately predict possible future business risks based on the current risk level and scoring data. The optimized Stacking model specifically includes the following steps: Step S1: Primary learning layer, integrate the current risk level and score data into data set D, randomly divide data set D into K subsets for cross-validation, the primary learning layer contains T1 learning algorithms, use K-1 subsets as training sets, and the remaining subset as test set, repeat K times, the data of each subset is predicted by all T1 learning algorithms to construct a new data set D′; Step S2: Meta-learning layer, using SVM as a meta-learner, training with the new dataset D′ as input to learn a more accurate decision boundary; Step S3: Based on the CNN-based risk decision model, obtain user data, use CNN to extract features from the user data, and use the CFA method to generate an additional minority class sample balanced data set, and use SVM for classification; Step S4: feature combination and selection, combining the data in the new dataset D′ as features with the features extracted by CNN, using three feature selection methods for feature selection, and adjusting the parameter threshold; Step S5: Loss function optimization, define a new loss function , the formula used is as follows: ； In the formula, is the original Stacking model loss function, and is the scaling factor, is the prediction result of the risk decision model based on CNN, This is the prediction result of the optimized Stacking model; Step S6: Adjust coefficient and , calculate the dynamic and Value, distance It is the sum of the distances between data points in the CNN-based risk decision model and the optimized Stacking model.

Images