A business risk identification method, system, device and medium

Abstract

The present application relates to the technical field of electric power business, and particularly relates to a business risk identification method, system, device and medium, which comprises obtaining a multi-view mask generator and a noise prediction network that have completed joint optimization; inputting a feature vector of a to-be-identified sample into the multi-view mask generator that has completed optimization to generate multiple perturbation views; performing reverse diffusion denoising on each perturbation view to obtain a reconstruction result of each view and aggregate the reconstruction result to generate a reconstruction sample, wherein the reverse diffusion denoising refers to predicting a current noise component through the optimized noise prediction network; and determining a risk according to a difference between the to-be-identified sample and the reconstruction sample. The present application has the beneficial effect that, without the need for labeled samples, the multi-view mask generator and the feature attention diffusion network are jointly optimized, the modeling capability for complex dependency relationships between mixed features is significantly enhanced, and thus the accuracy and robustness of business complaint risk identification are effectively improved.

Description

Technical Field

[0001] This invention relates to the field of power business technology, and in particular to a business risk identification method, system, equipment and medium. Background Technology

[0002] In business scenarios such as smart grid services, the work order data generated after users fail to get answers through self-service consultation often lacks labels and contains multiple types of features, including numerical values ​​and categories. Existing risk identification methods based on fixed rules or supervised learning are unable to effectively cope with the scarcity of labels and the complex dependencies between features, resulting in a high rate of missed and false alarms in complaint risks. Therefore, how to adaptively learn the potential associations in mixed features and identify abnormal samples based on them without relying on a large number of labeled samples has become an urgent problem to be solved. Summary of the Invention

[0003] To solve the above-mentioned technical problems, the present invention provides the following technical solution: In a first aspect, the present invention provides a business risk identification method, including obtaining a multi-view mask generator and a noise prediction network that have been jointly optimized; Input the feature vector of the sample to be identified into the optimized multi-view mask generator to generate multiple perturbation views; Backdiffusion denoising is performed on each perturbation view to obtain the reconstruction results of each view and aggregate them to generate reconstruction samples. Backdiffusion denoising refers to predicting the current noise component through an optimized noise prediction network. Risk is determined based on the differences between the sample to be identified and the reconstructed sample.

[0004] As a preferred embodiment of the business risk identification method of the present invention, the joint optimization of the multi-view mask generator and the noise prediction network includes: The target feature vector is input into a multi-view mask generator containing K parameter-independent sub-networks to obtain K perturbation views, wherein the target feature vector is obtained from the original business data after preprocessing; For each perturbation view, perform forward diffusion noise addition to obtain a noisy view; The noisy view and time step information are input into the noise prediction network to predict the added noise, and the first constraint is calculated based on the difference between the predicted noise and the real noise. The second constraint is calculated based on the differences between the K perturbation views; The multi-view mask generator and noise prediction network are jointly optimized based on the first and second constraints to obtain the optimized multi-view mask generator and noise prediction network.

[0005] As a preferred embodiment of the business risk identification method of the present invention, the construction of the first constraint includes: Obtain the actual noise added during the forward diffusion process; Obtain the predicted noise output by the noise prediction network for the same time step; The mean square error between the actual noise and the predicted noise is calculated and used as the first constraint.

[0006] As a preferred embodiment of the business risk identification method of the present invention, the construction of the second constraint includes: Obtain all perturbation views output by the multi-view mask generator; Calculate the feature distance between any two views in the perturbation view; The average distance is calculated based on all pairs of distances, and a second constraint is constructed with the reverse target that reduces this average distance.

[0007] As a preferred embodiment of the business risk identification method of the present invention, generating a reconstruction sample includes: Each perturbation view is obtained as the initial input for backdiffusion, and the total number of diffusion time steps is determined. For each perturbation view, starting from the maximum time step, the view of the current time step is sequentially input into the optimized noise prediction network to obtain the predicted noise. The view of the previous time step is calculated based on the predicted noise. This process is repeated until the time step is zero, and the reconstruction result of the perturbation view is obtained. After aligning the reconstruction results of all perturbation views according to the feature dimensions, the arithmetic mean of each dimension is calculated to form the final reconstruction sample.

[0008] As a preferred embodiment of the business risk identification method of the present invention, the risk is determined based on the difference between the sample to be identified and the reconstructed sample, including: Calculate the differences between the sample to be identified and the reconstructed sample in each feature dimension to form a residual vector; The anomalous score is obtained by squaring each component in the residual vector and taking the arithmetic mean. The abnormal score is compared with a preset threshold. When the abnormal score is greater than or equal to the threshold, it is determined that there is a risk.

[0009] In a preferred embodiment of the business risk identification method of the present invention, the determination of the preset threshold includes: Obtain a training sample set consisting of normal business data; Perform backdiffusion denoising and aggregation on each sample in the training sample set to obtain the corresponding reconstructed sample; Calculate the anomaly score between each training sample and its reconstructed sample to form a score set; The 95th percentile of the score set is used as the judgment threshold.

[0010] In a second aspect, the present invention provides a business risk identification system, comprising: an acquisition module for acquiring a multi-view mask generator and a noise prediction network that have been jointly optimized; The first generation module is used to input the feature vector of the sample to be identified into the optimized multi-view mask generator to generate multiple perturbation views. The second generation module is used to perform backdiffusion denoising on each perturbation view, obtain the reconstruction results of each view and aggregate them to generate reconstruction samples. Backdiffusion denoising refers to predicting the current noise component through an optimized noise prediction network. The judgment module is used to determine the risk based on the differences between the sample to be identified and the reconstructed sample.

[0011] Thirdly, the present invention provides a computer device, including a memory and a processor, wherein the memory stores a computer program, and the processor executes the computer program to implement the steps of the method described above.

[0012] Fourthly, the present invention provides a computer-readable storage medium having a computer program stored thereon, wherein the computer program, when executed by a processor, implements the steps of the method described above.

[0013] Compared with existing technologies, the beneficial effects of this invention are as follows: By introducing a trainable multi-view mask generator and combining it with the noise-adding and denoising prediction process of the diffusion model, the model can capture implicit dependencies between features in the training task of noise reconstruction, effectively identify potential anomalies, enhance the ability to identify and adapt to distribution changes and unknown samples, and improve the overall robustness and generalization performance of the model. Furthermore, this method proposes a mechanism for jointly optimizing the mask diversity loss function with the noise prediction loss of the diffusion model. This mechanism imposes explicit diversity constraints on the mask generation process, which can prevent mask pattern collapse, ensure that the generator can produce multiple differentiated mask strategies, and expand the coverage of training data in the mask space. This allows the model to maintain stable and effective representation learning under diverse data patterns, improving the adaptability and accuracy of risk identification. Attached Figure Description

[0014] To more clearly illustrate the technical solutions of the embodiments of the present invention, the drawings used in the following description of the embodiments will be briefly introduced. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0015] Figure 1 This is a flowchart illustrating a business risk identification method.

[0016] Figure 2This is a schematic diagram of the internal workflow of a multi-view mask generator.

[0017] Figure 3 The flowchart shows the structure for noise prediction. Detailed Implementation

[0018] To make the above-mentioned objects, features, and advantages of the present invention more apparent and understandable, specific embodiments of the present invention will be described in detail below with reference to the accompanying drawings. Obviously, the described embodiments are only a part of the embodiments of the present invention, and not all of them. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort should fall within the protection scope of the present invention.

[0019] Example 1, referring to Figure 1 This is the first embodiment of the present invention, which provides a method for improving short-term high-frequency energy storage efficiency, comprising: To address the current problem of how to adaptively learn the potential correlations in mixed features and thereby identify anomalous samples without relying on a large number of labeled samples, this application provides a method that can effectively solve the aforementioned problem. The following embodiment will elaborate on how to implement this business risk identification method. Figure 1 A flowchart of a business risk identification method is shown, including: S100: Obtain the jointly optimized multi-view mask generator and noise prediction network.

[0020] In one specific implementation, the joint optimization of the view mask generator and the noise prediction network includes the following steps: A1: Input the target feature vector into a multi-view mask generator containing K parameter-independent sub-networks to obtain K perturbation views, where the target feature vector is obtained from the original business data after preprocessing.

[0021] Understandably, preprocessing includes imputation of missing fields, label encoding for categorical fields with discrete values, and Z-score standardization for continuous numerical fields. Imputation employs a multivariate imputation method based on K-nearest neighbors (K=5), which estimates missing values ​​using the similarity of samples across other feature dimensions. Label encoding maps non-numerical symbolic variables to integer values. Furthermore, Z-score standardization eliminates dimensional differences between features by normalizing the mean and standard deviation.

[0022] Furthermore, the Z-score is calculated as follows: ; In the formula: For the first in the original data The values ​​of each sample; This is the mean of the feature; The standard deviation of this feature; This is the standardized result.

[0023] Specifically, such as Figure 2 The internal workflow of the multi-view mask generator shown is as follows: The preprocessed target feature vector x is input into the multi-view mask generator, which contains K parameter-independent sub-networks. Each subnetwork receives The output is the raw mask with the same dimension as the input, which is then normalized to the nearest integer using the sigmoid function. The interval forms a soft mask, and then... Operations and raw inputs Multiplying element by element, we get the first... A mask view It should be emphasized that the dashed line on the left side of the diagram represents each... All operations use the same raw input. .because The parameters of each sub-network are independent of each other, so they will learn different feature attention patterns, ultimately... The views are stacked to form the output. This provides multiple sets of differentiated feature representations for downstream tasks to simulate different scenarios of feature loss or abnormal perturbation.

[0024] A2: Perform forward diffusion noise addition on each perturbation view to obtain a noisy view.

[0025] Specifically, the operation of adding noise to individual views, each view According to the diffusion model, Gaussian noise is added at random time steps, defined as: ; In the formula: For the first Noisy samples of the step; For the original first One view; For standard normally distributed noise, This represents the cumulative retention rate during the diffusion process. in: The retention factor for each step; It is a linear or cosine noise scheduling parameter.

[0026] A3: Input the noisy view and time step information into the noise prediction network to predict the added noise, and calculate the first constraint based on the difference between the predicted noise and the real noise.

[0027] Specifically, such as Figure 3 As shown: The input consists of numerical features and time step features. It should be noted that the numerical features refer to the parts of the original values ​​in the noisy view, while the time step features are obtained by high-dimensional embedding of the scalar time step t using sine-cosine position encoding. Specifically: the target embedding dimension is set. First take half the dimension Generate frequency sequence ,in The maximum period parameter controls the decay rate from high to low frequency. Use the frequency index. Then calculate the angle parameters. This is the product of the time step and each frequency. Low-frequency components capture large-scale changes in the number of steps, while high-frequency components distinguish between adjacent steps. Finally, the cosine and sine components are concatenated to obtain the embedding vector. ,like If the number is odd, trailing zeros are used for alignment. Preferably, this embedding allows different time steps to have distinguishable and smooth representations in the embedding space. Furthermore, the time steps are embedded... Input features By directly splicing along the feature dimension, a sequence is formed. The tokens are then fed into the feature embedding module and converted into a unified token sequence (CLS removed).

[0028] Furthermore, the token sequence is input into a backbone network composed of multiple Transformer encoding blocks. Each encoding block sequentially contains two substructures: "layer normalization—multi-head self-attention—residual summation" and "layer normalization—feedforward neural network—residual summation." The self-attention mechanism models the relationships between different features, thereby achieving the interaction and fusion of multi-dimensional feature information. It can be understood that after multiple layers of Transformer encoding, a token feature representation containing rich contextual information can be obtained.

[0029] Furthermore, tokens corresponding to the original numerical features are selected from the output token feature representation, and their features are transformed through a linear mapping layer to achieve the prediction output of feature noise in each dimension.

[0030] Preferably, this method constructs a noise prediction network with a Transformer based on a feature attention mechanism as its core, and trains the model with noise restoration as its goal. Compared with the U-Net structure for images in traditional diffusion models, this method can effectively capture complex interactions between features.

[0031] It should be noted that the construction of the first constraint includes obtaining the actual noise added during the forward diffusion process and the predicted noise output by the noise prediction network at the same time step, and calculating the mean square error between the actual noise and the predicted noise as the first constraint. Specifically, it manifests as follows: ; in: : Parameters are Noise prediction network; To predict noise; Real noise; Noisy samples; : Number of diffusion steps.

[0032] A4: Calculate the second constraint based on the differences between the K perturbation views.

[0033] It should be noted that the construction of the second constraint includes first obtaining all perturbation views output by the multi-view mask generator; calculating the feature distance between any two views in the perturbation view; and finally calculating the average value based on all pairwise distances, and constructing the second constraint with the reverse target of reducing the average distance.

[0034] Specifically, the second constraint Represented as: ; In the formula: : The k-th view sample; The total number of views generated.

[0035] Understandable It is the result of the original input after being perturbed by a mask; the second constraint The smaller the value, the greater the difference between the generated perturbation views, and the better the model's robustness to feature loss.

[0036] A5: Based on the first and second constraints, the multi-view mask generator and noise prediction network are jointly optimized to obtain the optimized multi-view mask generator and noise prediction network.

[0037] Specifically, the training loss function is expressed as: ; In the formula: : Final training loss. : Hyperparameter used to control the weight of multi-view difference error in the total loss.

[0038] In one alternative implementation, the multi-view mask generator and the noise prediction network can be optimized in stages. First, the parameters of the noise prediction network are fixed, and only the multi-view mask generator is optimized, using the inter-view difference loss as the objective function, to train the generator to output diverse perturbation views. Then, the parameters of the mask generator are fixed, and the noise prediction network is optimized separately, using the noise prediction loss in forward diffusion as the objective function, to train the network to accurately predict the added noise. It should be noted that these two stages are performed sequentially and do not share the loss signal with each other.

[0039] In an alternative implementation, optimizing the multi-view mask generator and noise prediction network can also involve indirectly constraining the mask generator using only the noise prediction loss. This means directly feeding the perturbed view output by the multi-view mask generator into the noise prediction network, calculating the noise prediction loss, and then updating the parameters of both the mask generator and the noise prediction network simultaneously based solely on this loss through backpropagation, without introducing explicit inter-view difference losses. In this approach, the mask generator can only indirectly adjust its generated mask pattern by minimizing the noise prediction error.

[0040] S200: Input the feature vector of the sample to be identified into the optimized multi-view mask generator to generate multiple perturbation views.

[0041] In one specific implementation, step S200 further includes the following steps: B1: Input the feature vectors of the samples to be identified into the K parameter-independent subnetworks of the multi-view mask generator in parallel.

[0042] It should be noted that in this step, the sample to be identified also needs to be preprocessed before the input operation is performed.

[0043] B2: Each sub-network independently outputs a soft mask with values ​​in the range [0,1]. This soft mask is multiplied element by element with the feature vector of the sample to be identified to obtain a corresponding perturbation view.

[0044] B3: Collect all perturbation views of the outputs of K sub-networks to form multiple perturbation view sets.

[0045] S300: Perform backdiffusion denoising on each perturbation view to obtain the reconstruction results of each view and aggregate them to generate reconstruction samples. Backdiffusion denoising refers to predicting the current noise component through an optimized noise prediction network.

[0046] In one specific implementation, step S300 further includes the following steps: C1: Use each perturbation view as the initial input for backdiffusion and determine the total number of diffusion time steps.

[0047] C2: For each perturbation view, starting from the maximum time step, the view of the current time step is sequentially input into the optimized noise prediction network to obtain the predicted noise. The view of the previous time step is calculated based on the predicted noise. This process is repeated until the time step is zero, and the reconstruction result of the perturbation view is obtained.

[0048] It should be noted that during the backdiffusion process, the sample at the current time step... As input, the noise prediction network infers the corresponding noise components by combining sample and time step information. Then, based on the noise prediction results, the mean and variance of the conditional distribution are calculated, and samples are taken from the conditional distribution to obtain the samples of the previous time step. The process iterates step by step, gradually restoring the data from a high-noise state to a clear, noise-free sample, thus reconstructing the original data structure.

[0049] Specifically, at each step In the reverse diffusion, the sampling formula is as follows: ; In the formula: Time step The sample is the current sample. The noise-reduced version; : Random noise vector.

[0050] It should be noted that if This indicates that the final step is to reconstruct the sampling, at which point no noise term is added. .

[0051] Wherein: the mean term in the denoising process The calculation is as follows: ; Standard deviation of noise term for: ; In the formula: : Denoising model at time Predicted noise levels; : The cumulative retention rate at the previous moment.

[0052] It is important to emphasize that all perturbation views share the same pre-trained noise prediction network, and each view performs this iterative process independently.

[0053] C3: After aligning the reconstruction results of all perturbation views according to the feature dimensions, calculate the arithmetic mean of each dimension to form the final reconstruction sample.

[0054] Specifically, the reconstruction result of each view obtained through the backdiffusion process is represented as follows: ; In the formula: The k-th view is restored back to time during the diffusion process. The final sample obtained at that time; : No. The original denoised version of each view.

[0055] Furthermore, the reconstruction results of all views are aggregated, and an averaging operation is used to obtain the final unified reconstruction sample, as shown below: ; in: The final reconstructed sample after integrating all view information; Total number of views.

[0056] Preferably, this method enhances the diffusion model through masking to model the risk data of smart grid services under unsupervised conditions, and obtains the internal patterns of the data through the reconstruction process to achieve stable learning of the sample distribution. This not only effectively reduces the dependence on manually labeled samples and avoids the situation where model training is limited due to insufficient samples, but also ensures that high recognition ability and stability can still be maintained under the conditions of limited labels or insufficient label quality.

[0057] S400: Determine risk based on the difference between the sample to be identified and the reconstructed sample.

[0058] In one specific implementation, step S400 further includes the following steps: D1: Calculate the difference between the sample to be identified and the reconstructed sample in each feature dimension to form a residual vector.

[0059] Specifically, by obtaining the feature vector of the sample to be identified. and the corresponding reconstructed sample vector , where D represents the total number of feature dimensions.

[0060] Furthermore, for each feature dimension d (d=1,2,...,D), the difference between the original value and the reconstructed value in that dimension is calculated. .

[0061] Finally, by arranging the differences across all dimensions in their original order, we obtain the residual vector. .

[0062] D2: The arithmetic mean of the squares of each component in the residual vector is used to obtain the anomaly score.

[0063] D3: Compare the abnormal score with a preset threshold. If the abnormal score is greater than or equal to the threshold, it is determined that there is a risk.

[0064] Specifically, when the anomaly score is greater than or equal to the threshold, the sample is judged as abnormal and the label "1" is output, indicating that there is a potential complaint risk; when the anomaly score is less than the threshold, the label "0" is output, indicating that the sample is in a normal state and no warning is required.

[0065] In one alternative implementation, determining the preset threshold includes the following steps: D31: Obtain a training sample set consisting of normal business data.

[0066] D32: Perform backdiffusion denoising and aggregation on each sample in the training sample set to obtain the corresponding reconstructed sample.

[0067] D33: Calculate the outlier scores between each training sample and its reconstructed sample to form a score set.

[0068] D34: Take the 95th percentile of the score set as the judgment threshold.

[0069] In one alternative implementation, risk assessment based on the difference between the sample to be identified and the reconstructed sample can also be performed using a single-class support vector machine (SVM). This involves inputting the feature vectors of all normal samples in the training set into the SVM. The model learns the distribution boundaries of normal samples in a high-dimensional feature space, constructing a hypersphere or hyperplane that encloses most normal samples. During the assessment phase, the feature vector of the sample to be identified is input into the trained SVM. If the sample falls within the normal boundary, it is classified as normal; otherwise, it is classified as risky.

[0070] In another alternative implementation, risk assessment based on the difference between the sample to be identified and the reconstructed sample can also be achieved using an autoencoder reconstruction error-based approach. This involves training an autoencoder network using normal samples. The encoder compresses the input features into low-dimensional latent variables, which are then reconstructed by the decoder. After training, the sample to be identified is input into the autoencoder to obtain the reconstructed sample. The mean squared error between the original sample and the reconstructed sample is calculated as an anomaly score. When this score exceeds a preset threshold, a risk is considered to exist.

[0071] In summary, the beneficial effects of the business risk identification method of this invention, by introducing a trainable multi-view mask generator and combining it with the noise-adding and denoising prediction process of the diffusion model, enable the model to capture implicit dependencies between features in the training task of noise reconstruction, effectively identify potential anomalies, enhance the ability to identify and adapt to distribution changes and unknown samples, and improve the overall robustness and generalization performance of the model. Furthermore, this method proposes a mechanism for jointly optimizing the mask diversity loss function with the noise prediction loss of the diffusion model. This mechanism imposes explicit diversity constraints on the mask generation process, which can prevent mask pattern collapse, ensure that the generator can produce multiple differentiated mask strategies, and expand the coverage of training data in the mask space. This allows the model to maintain stable and effective representation learning under diverse data patterns, improving the adaptability and accuracy of risk identification.

[0072] Example 2: This example also provides a business risk identification method system, including: The acquisition module is used to acquire the multi-view mask generator and noise prediction network that have completed joint optimization. The first generation module is used to input the feature vector of the sample to be identified into the optimized multi-view mask generator to generate multiple perturbation views. The second generation module is used to perform backdiffusion denoising on each perturbation view, obtain the reconstruction results of each view and aggregate them to generate reconstruction samples. Backdiffusion denoising refers to predicting the current noise component through an optimized noise prediction network. The judgment module is used to determine the risk based on the differences between the sample to be identified and the reconstructed sample.

[0073] This embodiment also provides an electronic device applicable to the business risk identification method, comprising: a memory and a processor; the memory is used to store computer-executable instructions, and the processor is used to execute the computer-executable instructions to implement the business risk identification method proposed in the above embodiment.

[0074] This embodiment also provides a storage medium on which a computer program is stored. When the program is executed by a processor, it implements the business risk identification method proposed in the above embodiments.

[0075] The storage medium proposed in this embodiment and the method for implementing business risk identification proposed in the above embodiments belong to the same inventive concept. Technical details not described in detail in this embodiment can be found in the above embodiments, and this embodiment has the same beneficial effects as the above embodiments.

[0076] Based on the above description of the implementation methods, those skilled in the art can clearly understand that the present invention can be implemented using software and necessary general-purpose hardware, and of course, it can also be implemented using hardware. Based on this understanding, the technical solution of the present invention, or the part that contributes to the prior art, can be embodied in the form of a software product. This computer software product can be stored in a computer-readable storage medium, such as a computer floppy disk, read-only memory (ROM), random access memory (RAM), flash memory, hard disk, or optical disk, etc., including several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute the methods of the various embodiments of the present invention.

[0077] It should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and not to limit it. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can be made to the technical solutions of the present invention without departing from the spirit and scope of the technical solutions of the present invention, and all such modifications or substitutions should be covered within the scope of the claims of the present invention.

Claims

1. A method for identifying business risks, characterized in that: include: Obtain the jointly optimized multi-view mask generator and noise prediction network; Input the feature vector of the sample to be identified into the optimized multi-view mask generator to generate multiple perturbation views; Backdiffusion denoising is performed on each of the perturbation views to obtain the reconstruction results of each view and aggregate them to generate reconstruction samples. Backdiffusion denoising refers to predicting the current noise component through an optimized noise prediction network. Risk is determined based on the differences between the sample to be identified and the reconstructed sample.

2. The business risk identification method as described in claim 1, characterized in that: Joint optimization of the multi-view mask generator and noise prediction network includes: The target feature vector is input into a multi-view mask generator containing K parameter-independent sub-networks to obtain K perturbation views, wherein the target feature vector is obtained by preprocessing the original service data; Forward diffusion noise is applied to each of the perturbation views to obtain a noisy view; The noisy view and time step information are input into the noise prediction network to predict the added noise, and the first constraint is calculated based on the difference between the predicted noise and the real noise. The second constraint is calculated based on the differences between the K perturbation views; The multi-view mask generator and the noise prediction network are jointly optimized based on the first constraint and the second constraint to obtain the optimized multi-view mask generator and noise prediction network.

3. The business risk identification method as described in claim 2, characterized in that: The construction of the first constraint includes: Obtain the actual noise added during the forward diffusion process; Obtain the predicted noise output by the noise prediction network for the same time step; The mean square error between the actual noise and the predicted noise is calculated and used as the first constraint.

4. The business risk identification method as described in claim 2, characterized in that: The construction of the second constraint includes: Obtain all perturbation views output by the multi-view mask generator; Calculate the feature distance between any two views in the perturbation view; The average distance is calculated based on all pairs of distances, and a second constraint is constructed with the reverse target that reduces this average distance.

5. A business risk identification method as described in any one of claims 1-4, characterized in that: The generation of reconstructed samples includes: Each perturbation view is obtained as the initial input for backdiffusion, and the total number of diffusion time steps is determined. For each perturbation view, starting from the maximum time step, the view of the current time step is sequentially input into the optimized noise prediction network to obtain the predicted noise. The view of the previous time step is calculated based on the predicted noise. This process is repeated until the time step is zero, and the reconstruction result of the perturbation view is obtained. After aligning the reconstruction results of all perturbation views according to the feature dimensions, the arithmetic mean of each dimension is calculated to form the final reconstruction sample.

6. The business risk identification method as described in claim 5, characterized in that: The risk determination based on the difference between the sample to be identified and the reconstructed sample includes: Calculate the differences between the sample to be identified and the reconstructed sample in each feature dimension to form a residual vector; The anomalous score is obtained by squaring each component of the residual vector and taking the arithmetic mean. The anomaly score is compared with a preset threshold, and a risk is determined when the anomaly score is greater than or equal to the threshold.

7. The business risk identification method as described in claim 6, characterized in that: The determination of the preset threshold includes: Obtain a training sample set consisting of normal business data; Perform backdiffusion denoising and aggregation on each sample in the training sample set to obtain the corresponding reconstructed sample; Calculate the anomaly score between each training sample and its reconstructed sample to form a score set; The 95th percentile of the score set is taken as the judgment threshold.

8. A business risk identification system, employing the method described in any one of claims 1-7, characterized in that, include: The acquisition module is used to acquire the multi-view mask generator and noise prediction network that have completed joint optimization. The first generation module is used to input the feature vector of the sample to be identified into the optimized multi-view mask generator to generate multiple perturbation views. The second generation module is used to perform backdiffusion denoising on each of the perturbation views, obtain the reconstruction results of each view and aggregate them to generate reconstruction samples. Backdiffusion denoising refers to predicting the current noise component through an optimized noise prediction network. The determination module is used to determine the risk based on the difference between the sample to be identified and the reconstructed sample.

9. A computer device comprising a memory and a processor, wherein the memory stores a computer program, characterized in that, When the processor executes the computer program, it implements the steps of the method according to any one of claims 1 to 7.

10. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by a processor, it implements the steps of the method according to any one of claims 1 to 7.

Images