Risk prediction method and device based on dynamic information learning
Through the risk prediction method based on dynamic information learning, the feature filtering and retention modules are used to screen risk characteristics, which solves the problem that traditional risk management methods are difficult to identify complex risk characteristics, and achieves efficient risk prediction and decision support.
Patent Information
- Application Number
- CN202411665819.X
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2024-11-20
- Publication Date
- 2025-02-25
AI Technical Summary
Traditional risk management methods are difficult to effectively identify and deal with complex risk characteristics, and in a massive data environment, manual analysis and empirical judgments are difficult to meet the needs of risk prediction.
The risk prediction method based on dynamic information learning is adopted, and the feature vectors are linearly transformed, feature filtered and retained feature information through the intermediate variable calculation module, the information contribution metric calculation module and the feature retention module, and finally, the risk prediction results are generated using the neural network classifier.
This method can effectively screen key characteristics and generate accurate risk prediction results, improving decision-making efficiency and risk response speed and adaptability.
Smart Images
Figure CN119513698A_ABST
Abstract
Description
Technical Field
[0001] The present invention relates to a risk prediction method, and in particular to a risk prediction method and device based on dynamic information learning. Background Art
[0002] Nowadays, the diversity and complexity of risk factors are increasing. Due to their limitations, traditional risk management methods have been unable to effectively identify and respond to key risk characteristics and cannot meet the needs of risk prediction. In addition, due to the massive influx of data in the field of risk prediction and the rapid update speed of information, relying solely on manual analysis and empirical judgment can no longer meet actual needs. Therefore, it is necessary to conduct in-depth mining and analysis of key risk characteristics in order to more accurately predict and prevent potential risk events. This not only helps to improve the efficiency and quality of decision-making, but also enhances the reaction speed and adaptability of risk response. With the development of big data and artificial intelligence technology, the use of advanced algorithms and models to identify and analyze risk factors has become an important trend in the development of the industry.
[0003] Related research shows that datasets containing irrelevant features can significantly weaken the performance of the model. These irrelevant features not only waste valuable computing resources, but may also introduce noise during model training, reducing the accuracy and stability of the model. Typically, a feature set consists of features that are closely related to the target label and features that are irrelevant to the prediction task. Each subset contributes differently to the classification results, so it is particularly important to screen and optimize the features.
[0004] Existing statistical methods, correlation analysis, principal component analysis and other methods can screen out features that are closely related to the target variable and eliminate redundant or irrelevant features. These methods are based on the same important features for all samples, ignoring the risk feature learning based on sample differences, so it is difficult to make further breakthroughs in risk prediction performance. Summary of the invention
[0005] The present invention is made to solve the above-mentioned problems, and its purpose is to provide a risk prediction method and device based on dynamic information learning.
[0006] The present invention provides a risk prediction method based on dynamic information learning, which is used to obtain risk prediction results according to multiple features of a target, and has the following characteristics, including the following steps: step S1, preprocessing all features respectively to obtain corresponding preprocessing features; step S2, constructing a feature vector according to all preprocessing features, and inputting the feature vector into a prediction model to obtain a risk prediction result, wherein the prediction model includes: an intermediate variable calculation module, which is used to perform linear transformation on the feature vector to calculate the intermediate variable; an information contribution metric calculation module, which is used to perform feature filtering on the intermediate variable to obtain the information contribution metric; a feature retention module, which is used to calculate the retained feature information according to the information contribution metric and the feature vector; a classification module, which includes a neural network classifier, which is used to calculate the corresponding risk prediction label as the risk prediction result according to the retained feature information.
[0007] The risk prediction method based on dynamic information learning provided by the present invention may also have the following characteristics: wherein, in the intermediate variable calculation module, the calculation expression of the intermediate variable obtained according to the feature vector is: In the formula is the intermediate variable, is the feature vector, is the corresponding weight, is the corresponding bias.
[0008] The risk prediction method based on dynamic information learning provided by the present invention may also have the following characteristics: wherein, in the information contribution metric calculation module, the calculation expression of the information contribution metric obtained according to the intermediate variable is: Where γ is the information contribution measure, is the intermediate variable, is the corresponding weight, is the corresponding bias, and σ() is the Sigmoid activation function.
[0009] The risk prediction method based on dynamic information learning provided by the present invention may also have the following characteristics: wherein, in the feature retention module, the calculation expression for retaining feature information obtained according to the information contribution metric and the feature vector is: In the formula To retain feature information, γ is the information contribution measure, is the eigenvector, and ⊙ is the element-by-element multiplication.
[0010] In the risk prediction method based on dynamic information learning provided by the present invention, it can also have the following characteristics: wherein, the risk prediction label is 0 or 1, when the risk prediction label is 1, the risk prediction result is that the target has risk, and when the risk prediction label is 0, the risk prediction result is that the target does not have risk.
[0011] In the risk prediction method based on dynamic information learning provided by the present invention, it can also have the following characteristics: wherein, step S1 includes the following sub-steps: step S1-1, label encoding the features of the character category to generate a numerical value corresponding to the feature; step S1-2, normalizing each feature to obtain the corresponding preprocessing feature.
[0012] The risk prediction method based on dynamic information learning provided by the present invention may also have the following characteristics: wherein, a training data set is constructed by using existing data, and the prediction model is trained in combination with a loss function, the training data set includes multiple samples, the samples include multiple features and corresponding true labels, and the calculation expression of the loss function is: loss = loss β +loss v +λ·‖l‖1, Where loss is the loss function, N is the number of samples, and y is k is the true label corresponding to the kth sample, is the risk prediction probability generated by the prediction model for the kth sample, y l is the true label corresponding to the lth sample, is the risk prediction probability generated by the prediction model for the lth sample, λ is the regularization strength, γ l is the information contribution measure corresponding to the lth sample.
[0013] The present invention also provides a risk prediction device based on dynamic information learning, which is used to obtain risk prediction results based on multiple features of a target, and has the following characteristics: a preprocessing unit, which is used to preprocess all features separately to obtain corresponding preprocessed features; a prediction unit, which includes a prediction model, which is used to construct a feature vector based on all preprocessed features, and input the feature vector into the prediction model to obtain a risk prediction result, wherein the prediction model includes: an intermediate variable calculation module, which is used to perform linear transformation on the feature vector to calculate the intermediate variable; an information contribution metric calculation module, which is used to perform feature filtering on the intermediate variable to obtain the information contribution metric; a feature retention module, which is used to calculate the retained feature information based on the information contribution metric and the feature vector; a classification module, which includes a neural network classifier, which is used to calculate the corresponding risk prediction label as the risk prediction result based on the retained feature information.
[0014] Functions and Effects of the Invention
[0015] According to the risk prediction method and device based on dynamic information learning involved in the present invention, because the feature vector is linearly transformed and feature filtered through the intermediate variable calculation module and the information contribution measurement calculation module, the corresponding information contribution amount is obtained, and then the feature vector to be retained is screened according to the information contribution amount through the feature retention model as the retained feature information, and finally the risk prediction result is obtained using the retained feature information. Therefore, the risk prediction method and device based on dynamic information learning of the present invention can screen key features and generate accurate risk prediction results. BRIEF DESCRIPTION OF THE DRAWINGS
[0016] Figure 1 It is a risk prediction device based on dynamic information learning in an embodiment of the present invention;
[0017] Figure 2 is a block diagram of a prediction model in an embodiment of the present invention;
[0018] Figure 3 It is a flow chart of a risk prediction method based on dynamic information learning in an embodiment of the present invention. DETAILED DESCRIPTION
[0019] In order to make the technical means, creative features, objectives and effects achieved by the present invention easy to understand, the following embodiments and the accompanying drawings specifically illustrate the risk prediction method based on dynamic information learning of the present invention.
[0020] In this embodiment, a risk prediction device based on dynamic information learning is provided, which is used to obtain risk prediction results based on multiple characteristics of the target. In this embodiment, the risk prediction method based on dynamic information learning is applied to the financial field, the target is a financial user, the characteristics are the relevant data of the user, such as occupation, age, etc., and the risk prediction result is whether the user is a high-risk user. In other embodiments, the risk prediction method based on dynamic information learning can be applied to other fields that require risk prediction, such as the field of food safety, then the target is a specific food, the characteristics are the relevant data of the food, such as the manufacturer, date, ingredients, etc., and the risk prediction result is whether the food has a food safety risk.
[0021] Figure 1 It is a risk prediction device based on dynamic information learning in an embodiment of the present invention.
[0022] like Figure 1 As shown, the risk prediction device 100 based on dynamic information learning includes a preprocessing unit 11 and a prediction unit 12 .
[0023] The preprocessing unit 11 is used to preprocess all features respectively to obtain corresponding preprocessed features.
[0024] The preprocessing section 11 includes a character processing subsection 111 and a normalization processing subsection 112 .
[0025] The character processing sub-unit 111 is used to perform label encoding on the features of the character category and generate a numerical value corresponding to the feature. For example, for the feature of "occupation", the original manufacturer name of the character category can be replaced with the corresponding numerical value.
[0026] The normalization processing sub-unit 112 is used to perform normalization processing on each feature to obtain corresponding pre-processed features.
[0027] The expression for normalization processing in this embodiment is:
[0028]
[0029] In the formula is the normalized feature, i.e., the preprocessed feature, x l is the feature before normalization, min(x l ) is the feature x l The minimum value in the feature column, max(x l ) is the feature x l The maximum value in the feature column.
[0030] In this embodiment, the features include character categories and numerical categories. The preprocessing unit 11 first performs label encoding on the features of the character categories in all the features through the character processing sub-unit 111, converts them into features of the numerical categories, and then converts all the features of the numerical categories into corresponding preprocessing features through the normalization processing sub-unit 112.
[0031] The prediction unit 12 includes a prediction model, which is used to construct a feature vector based on all preprocessed features, and input the feature vector into the prediction model to obtain a risk prediction result.
[0032] Figure 2 is a block diagram of a prediction model in an embodiment of the present invention.
[0033] like Figure 2 As shown, the prediction model 200 includes an intermediate variable calculation module 21, an information contribution metric calculation module 22, a feature retention module 23 and a classification module 24.
[0034] The intermediate variable calculation module 21 is used to perform linear transformation on the feature vector and calculate the intermediate variable.
[0035] Among them, in the intermediate variable calculation module 21, the calculation expression of the intermediate variable obtained according to the characteristic vector is:
[0036]
[0037] In the formula is the intermediate variable, is the feature vector, is the corresponding weight, is the corresponding bias.
[0038] The information contribution metric calculation module 22 is used to perform feature filtering on the intermediate variables to obtain the information contribution metric.
[0039] Among them, in the information contribution metric calculation module 22, the calculation expression of the information contribution metric is obtained according to the intermediate variable:
[0040]
[0041] Where γ is the information contribution measure, is the intermediate variable, is the corresponding weight, is the corresponding bias, and σ() is the Sigmoid activation function.
[0042] In this embodiment, the value of the information contribution measure γ is 0 or 1. When γ=1, it means that the corresponding feature vector The information of is completely retained, and when γ = 0, it means the corresponding eigenvector information is completely discarded.
[0043] The feature retention module 23 is used to calculate and obtain the retained feature information according to the information contribution metric and the feature vector.
[0044] Among them, in the feature retention module 23, the calculation expression of the retained feature information is obtained according to the information contribution metric and the feature vector:
[0045]
[0046] In the formula To retain feature information, γ is the information contribution measure, is the eigenvector, and ⊙ is the element-by-element multiplication.
[0047] In this embodiment, the feature vector data is screened by the feature retention module 23, thereby reducing the amount of data storage required for subsequent calculations by the device for loading the prediction model 200 and improving the calculation efficiency of the subsequent classification module 24.
[0048] The classification module 24 includes a neural network classifier, which is used to calculate the corresponding risk prediction label as the risk prediction result based on the retained feature information.
[0049] Among them, the risk prediction label is 0 or 1. When the risk prediction label is 1, the risk prediction result is that the target has risk. When the risk prediction label is 0, the risk prediction result is that the target does not have risk. In other embodiments, the risk prediction label can be replaced with other values or characters according to actual needs, so as to distinguish different risk prediction results by using different values or characters.
[0050] In this embodiment, a training data set is constructed by using existing data, and the prediction model 200 is trained in combination with a loss function, so as to obtain a trained prediction model 200 for risk prediction.
[0051] The training data set includes multiple samples, and the samples include multiple features and corresponding true labels. The samples in the training data set of this embodiment are unbalanced samples, so the calculation expression of the corresponding loss function is:
[0052] loss=loss β +loss v +λ·‖l‖1,
[0053]
[0054]
[0055] Where loss is the loss function, N is the number of samples, and y is k is the true label corresponding to the kth sample, is the risk prediction probability generated by the prediction model for the kth sample, y l is the true label corresponding to the lth sample, is the risk prediction probability generated by the prediction model for the lth sample, λ is the regularization strength, γ l is the information contribution measure corresponding to the lth sample.
[0056] In this embodiment, loss β The prediction results of the binary classification problem are trained by optimizing the prediction model 200. The loss function guides the adjustment of the parameters of the prediction model 200 by quantifying the difference between the probability distribution predicted by the prediction model 200 and the true label, thereby improving the accuracy and robustness of the prediction.
[0057] In this embodiment, loss v Measuring the overlap between the predicted results and the true labels emphasizes the focus on the minority categories, thereby more effectively adjusting the parameters of the prediction model 200 so that it can more accurately identify the minority categories or targets on the imbalanced dataset.
[0058] In this embodiment, ‖l‖1 regularization is used to approximate the sparsity effect of the ‖l‖0 norm while maintaining the feasibility and stability of the optimization, thereby further increasing the sparsity of key features, so that the prediction model 200 is more focused on the most influential features. Therefore, in the ‖l‖1 regularization of the training process, |γ l The value of | approaches zero, thus achieving dynamic feature information filtering.
[0059] In this embodiment, an ablation experiment is performed on the prediction model 200 on multiple existing data sets.
[0060] In the ablation experiment, the prediction model 200 without the intermediate variable calculation module 21, the information contribution measurement calculation module 22 and the feature retention module 23 is used as the ablation model, and the overall performance evaluation index KS and the category prediction performance index Recall are used. class And the comprehensive performance index of the classification model f1 class As a performance evaluation indicator, the specific results are shown in the following table:
[0061]
[0062] The first column in the table above is the existing data sets, the second column is the different models used on the existing data sets, and the third to sixth columns are the overall performance evaluation index KS of each model on different existing data sets, and the comprehensive performance index f1 of the classification model of samples with label 0. class , Recall, a performance indicator for class prediction with label 1 class And the comprehensive performance index f1 of the classification model with samples with label 1 class For example, the cell in the second row and third column indicates that the value of the overall performance evaluation index KS of the prediction model 200 in the existing data set 1 is 0.396. It can be seen that the dynamic feature learning strategy composed of the intermediate variable calculation module 21, the information contribution measurement calculation module 22 and the feature retention module 23 can effectively improve the prediction accuracy.
[0063] The following describes the process of using the risk prediction device 100 based on dynamic information learning to perform a risk prediction method based on dynamic information learning in conjunction with the accompanying drawings.
[0064] Figure 3 It is a flow chart of a risk prediction method based on dynamic information learning in an embodiment of the present invention.
[0065] like Figure 3 As shown, the risk prediction method based on dynamic information learning includes the following steps:
[0066] Step S1 - 1 , using the character processing sub-unit 111 to perform label encoding on the features of the character category, and generate a numerical value corresponding to the feature.
[0067] Step S1-2, using the normalization processing sub-unit 112 to normalize each feature to obtain the corresponding pre-processed features.
[0068] Step S2, using the prediction unit 12 to construct a feature vector based on all preprocessed features, and input the feature vector into the prediction model 200 to obtain a risk prediction result.
[0069] Functions and Effects of the Embodiments
[0070] According to the risk prediction method and device based on dynamic information learning involved in this embodiment, the feature vector is linearly transformed and feature filtered through the intermediate variable calculation module and the information contribution measurement calculation module to obtain the corresponding information contribution amount, and then the feature vector to be retained is obtained as the retained feature information according to the information contribution amount through the feature retention model, and finally the risk prediction result is obtained using the retained feature information. In short, this method can screen key features and generate accurate risk prediction results.
[0071] Those skilled in the art should understand that the present invention is not limited to the above embodiments, and the above embodiments and descriptions are only for explaining the principles of the present invention. Without departing from the spirit and scope of the present invention, the present invention may have various changes and improvements, and these changes and improvements fall within the scope of the present invention to be protected. The scope of protection of the present invention is defined by the attached claims and their equivalents.
Claims
1. A risk prediction method based on dynamic information learning, used to obtain risk prediction results according to multiple characteristics of a target, characterized in that: The following steps are involved: Step S1, preprocessing all the features respectively to obtain corresponding preprocessing features; Step S2, constructing a feature vector based on all the preprocessing features, and inputting the feature vector into a prediction model to obtain the risk prediction result, Wherein, the prediction model includes: An intermediate variable calculation module, used for performing a linear transformation on the feature vector to calculate the intermediate variable; An information contribution metric calculation module, used to perform feature filtering on the intermediate variables to obtain information contribution metrics; A feature retention module, used for calculating and obtaining retained feature information according to the information contribution metric and the feature vector; The classification module includes a neural network classifier, which is used to calculate the corresponding risk prediction label as the risk prediction result based on the retained feature information.
2. The risk prediction method based on dynamic information learning according to claim 1 is characterized in that: in, In the intermediate variable calculation module, the calculation expression of the intermediate variable is obtained according to the feature vector: In the formula is the intermediate variable, is the feature vector, is the corresponding weight, is the corresponding bias.
3. The risk prediction method based on dynamic information learning according to claim 1 is characterized in that: in, In the information contribution metric calculation module, the calculation expression of the information contribution metric is obtained according to the intermediate variable: Where γ is the information contribution measure, is the intermediate variable, is the corresponding weight, is the corresponding bias, and σ( ) is the Sigmoid activation function.
4. The risk prediction method based on dynamic information learning according to claim 1 is characterized in that: in, In the feature retention module, the calculation expression for obtaining the retained feature information according to the information contribution metric and the feature vector is: In the formula is the retained feature information, γ is the information contribution measure, is the eigenvector, and ⊙ is the element-by-element multiplication.
5. The risk prediction method based on dynamic information learning according to claim 1 is characterized in that: in, The risk prediction label is 0 or 1. When the risk prediction label is 1, the risk prediction result is that the target has risk. When the risk prediction label is 0, the risk prediction result is that the target has no risk.
6. The risk prediction method based on dynamic information learning according to claim 1 is characterized in that: in, The step S1 comprises the following sub-steps: Step S1-1, label encoding the feature of the character category to generate a numerical value corresponding to the feature; Step S1-2, normalizing each of the features to obtain the corresponding preprocessing features.
7. The risk prediction method based on dynamic information learning according to claim 1 is characterized in that: in, A training data set is constructed using existing data, and the prediction model is trained using a loss function. The training data set includes multiple samples. The sample includes a plurality of the features and corresponding true labels, The calculation expression of the loss function is: loss=loss β +loss v +λ·||l||1, Where loss is the loss function, N is the number of samples, and y k is the true label corresponding to the kth sample, is the risk prediction probability generated by the prediction model for the kth sample, y l is the true label corresponding to the lth sample, is the risk prediction probability generated by the prediction model for the lth sample, λ is the regularization strength, γ l is the information contribution measure corresponding to the lth sample.
8. A risk prediction device based on dynamic information learning, used to obtain risk prediction results based on multiple characteristics of a target, characterized in that: include: A preprocessing unit, used for preprocessing all the features respectively to obtain corresponding preprocessing features; The prediction unit includes a prediction model, which is used to construct a feature vector based on all the preprocessing features, and input the feature vector into the prediction model to obtain the risk prediction result. Wherein, the prediction model includes: An intermediate variable calculation module, used for performing a linear transformation on the feature vector to calculate the intermediate variable; An information contribution metric calculation module, used to perform feature filtering on the intermediate variables to obtain information contribution metrics; A feature retention module, used for calculating and obtaining retained feature information according to the information contribution metric and the feature vector; The classification module includes a neural network classifier, which is used to calculate the corresponding risk prediction label as the risk prediction result based on the retained feature information.