Training method of data analysis model, data analysis method, equipment and medium
By alternating updates of generative and discriminative networks, realistic data is generated and subtle differences are identified, which solves the shortcomings of traditional data analysis models in data feature processing and improves the accuracy of data analysis.
Patent Information
- Application Number
- CN202510724414.7
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2025-05-30
- Publication Date
- 2025-09-19
AI Technical Summary
Existing traditional data analysis models are relatively rough when processing data features, resulting in poor accuracy in data analysis, and are particularly difficult to capture small data changes or key information in the structure.
By combining generative networks and discriminative networks, a variety of generated data are generated through simulated generation processing of sample data and noise data, and the performance of the data analysis model is improved by alternately updating network parameters.
The data generated by the generative network is more realistic, and the discriminative network can more keenly identify subtle differences in the data, improving the accuracy and overall performance of data analysis.
Smart Images
Figure CN120671035A_ABST
Abstract
Description
Technical Field
[0001] The present application relates to the field of data analysis technology, and in particular to a data analysis model training method, data analysis method, equipment and medium, and specifically to a data analysis model training method, data analysis method, data processing device, computer equipment and storage medium. Background Art
[0002] In the era of big data, analyzing data, understanding its inherent patterns and potential value, and leveraging it has become a key driver of development in areas such as business decision-making, scientific research, and social management. For example, in internet finance, consumer behavior analysis is used to predict financial product sales trends, while in healthcare, patient data analysis is used to uncover underlying patterns in disease patterns.
[0003] In practical applications, key information often does not appear in obvious, large-scale data patterns, but is hidden in certain subtle data changes or structures. The traditional data analysis models commonly used, such as decision trees and naive Bayes, are relatively rough in processing data features, resulting in poor accuracy in data analysis. Summary of the Invention
[0004] The main purpose of this application is to provide a data analysis model training method, data analysis method, device and medium, through which the performance of the data analysis model can be improved, thereby improving the accuracy of data analysis.
[0005] In a first aspect, the present application provides a method for training a data analysis model, wherein the data analysis model includes a generative network and a discriminative network, and the method includes:
[0006] Acquire sample data and noise data corresponding to the sample data;
[0007] Inputting the sample data and the noise data into the generative network for data simulation generation processing to obtain first generated data corresponding to the sample data, second generated data corresponding to the noise data, and third generated data corresponding to both the sample data and the noise data;
[0008] Inputting the sample data, the first generated data, the second generated data, and the third generated data into the discriminant network for processing to obtain similarity data corresponding to the sample data and the first generated data, the second generated data, and the third generated data respectively;
[0009] Based on the sample data, the first generated data, the second generated data, and the plurality of similarity data, parameters of the generative network and the discriminative network are alternately updated.
[0010] In a second aspect, the present application also provides a data analysis method, comprising:
[0011] Inputting the data to be analyzed into a data analysis model to obtain a probability value that the data to be analyzed is abnormal data; wherein the data analysis model is obtained based on the data analysis model training method described above;
[0012] If the probability value is less than a preset probability threshold, obtaining a plurality of third generated data corresponding to the data to be analyzed, and determining target generated data for representing normal data from the plurality of third generated data;
[0013] The abnormality degree of the data to be analyzed is determined according to the deviation data between the data to be analyzed and the target generated data.
[0014] In a third aspect, the present application further provides a data processing device, comprising:
[0015] A data acquisition module, configured to acquire sample data and noise data corresponding to the sample data;
[0016] a data generation module, configured to input the sample data and the noise data into the generative network for data simulation generation processing, thereby obtaining first generated data corresponding to the sample data, second generated data corresponding to the noise data, and third generated data corresponding to both the sample data and the noise data;
[0017] a data determination module, configured to input the sample data, the first generated data, the second generated data, and the third generated data into the discriminant network for processing, so as to obtain similarity data corresponding to the sample data and the first generated data, the second generated data, and the third generated data, respectively;
[0018] A parameter updating module is used to alternately update the parameters of the generative network and the discriminative network based on the sample data, the first generated data, the second generated data, and the plurality of similarity data.
[0019] In a fourth aspect, the present application also provides a computer device, comprising a processor, a memory, and a computer program stored on the memory and executable by the processor, wherein when the computer program is executed by the processor, it implements the steps of the training method of the data analysis model as described above, or implements the data analysis method as described above.
[0020] In a fourth aspect, the present application also provides a computer-readable storage medium, on which a computer program is stored, wherein when the computer program is executed by a processor, the steps of the training method of the data analysis model as described above are implemented, or the data analysis method as described above are implemented.
[0021] The present application provides a training method, a data analysis method, an apparatus and a medium for a data analysis model. The training method of the data analysis model in the present application obtains sample data and noise data corresponding to the sample data; inputs the sample data and the noise data into a generative network for data simulation generation processing to obtain first generated data corresponding to the sample data, second generated data corresponding to the noise data, and third generated data corresponding to both the sample data and the noise data; inputs the sample data, the first generated data, the second generated data and the third generated data into a discriminant network for processing to obtain similarity data corresponding to the sample data and the first generated data, the second generated data and the third generated data respectively; based on the sample data, the first generated data and the second generated data and a plurality of similarity data, the parameters of the generative network and the discriminant network are alternately updated, so that the data generated by the generative network gradually becomes realistic and can better simulate the details and features of the sample data, while the discriminant network can more keenly identify the subtle differences between the sample data and other generated data. Ultimately, the overall performance of the data analysis model is effectively improved, thereby greatly improving the accuracy of data analysis. BRIEF DESCRIPTION OF THE DRAWINGS
[0022] In order to more clearly illustrate the technical solutions of the embodiments of the present application, the following is a brief introduction to the drawings required for use in the description of the embodiments. Obviously, the drawings described below are some embodiments of the present application. For ordinary technicians in this field, other drawings can be obtained based on these drawings without any creative work.
[0023] Figure 1 A schematic diagram of the steps of a data analysis model training method provided in an embodiment of the present application;
[0024] Figure 2 for Figure 1 A schematic diagram of a sub-step flow chart of a method for training a data analysis model in FIG.
[0025] Figure 3 for Figure 1 A flowchart of another sub-step of the data analysis model training method in FIG.
[0026] Figure 4 A schematic diagram of the steps of a data analysis method provided in an embodiment of the present application;
[0027] Figure 5 A schematic block diagram of a data processing device provided in an embodiment of the present application;
[0028] Figure 6 for Figure 5 A schematic block diagram of a submodule of a data processing device in FIG.
[0029] Figure 7 for Figure 5 A schematic block diagram of another submodule of the data processing device in FIG.
[0030] Figure 8 A schematic block diagram of the structure of a computer device provided in an embodiment of the present application.
[0031] The realization of the objectives, functional features and advantages of this application will be further explained in conjunction with embodiments and with reference to the accompanying drawings. DETAILED DESCRIPTION
[0032] The following will be combined with the drawings in the embodiments of this application to clearly and completely describe the technical solutions in the embodiments of this application. Obviously, the embodiments described are part of the embodiments of this application, not all of them. Based on the embodiments in this application, all other embodiments obtained by ordinary technicians in this field without making creative efforts are within the scope of protection of this application.
[0033] The flowcharts shown in the accompanying drawings are for illustrative purposes only and do not necessarily include all contents and operations / steps, nor must they be executed in the order described. For example, some operations / steps may be decomposed, combined, or partially merged, so the actual execution order may vary depending on the actual situation.
[0034] Embodiments of the present application provide a data analysis model training method, data analysis method, device, and medium. The data analysis model training method can be applied to a terminal device or server. The terminal device can be an electronic device such as a mobile phone, tablet computer, laptop computer, desktop computer, personal digital assistant, or wearable device. The server can be a single server or a server cluster consisting of multiple servers.
[0035] The following describes some embodiments of the present application in detail with reference to the accompanying drawings. In the absence of conflict, the following embodiments and features therein may be combined with each other.
[0036] Please refer to Figure 1 , Figure 1 A schematic flow chart of the steps of a data analysis model training method provided in an embodiment of the present application.
[0037] Among them, the data analysis model includes a generative network and a discriminative network. The generative network is used to simulate and generate data, and the discriminative network is used to identify data and determine whether the data is real data or data simulated and generated by the generative network.
[0038] like Figure 1 As shown, the training method of the data analysis model includes steps S101 to S104.
[0039] Step S101: Obtain sample data and noise data corresponding to the sample data.
[0040] The sample data can be any form of data used as a sample. For example, in the healthcare field, the sample data can be medical imaging data, such as X-rays, CT scans, and MRI images. Noise data can be data sampled from a random distribution (such as a standard normal distribution or a uniform distribution), which is used to guide the subsequent generative network to generate diverse and realistic simulated data.
[0041] In one embodiment, obtaining sample data includes: collecting multiple first sample data of the same data type; performing data cleaning and data preprocessing on the multiple first sample data to obtain multiple second sample data; and determining any one of the multiple second sample data as the sample data. Data cleaning and data preprocessing are used to remove noise and abnormal data, thereby improving data quality, enhancing data consistency and usability, and providing a more accurate and reliable sample data input basis for subsequent model training, effectively avoiding model performance deviations caused by abnormal training data.
[0042] It should be noted that in order to further ensure the privacy and security of the above-mentioned sample data and other related information, the above-mentioned sample data and other related information can also be stored in a node of a blockchain. The technical solution of this application can also be applied to adding other data files stored on the blockchain. The blockchain referred to in this application is a new application model of computer technologies such as distributed data storage, point-to-point transmission, consensus mechanism, and encryption algorithm.
[0043] Step S102: Input the sample data and the noise data into the generative network for data simulation generation processing to obtain first generated data corresponding to the sample data, second generated data corresponding to the noise data, and third generated data corresponding to both the sample data and the noise data.
[0044] The first generated data is similar to the sample data in terms of data distribution, but may have certain randomness or ambiguity in details.
[0045] Noise data can be randomly sampled from a random distribution and represents completely random characteristics. The generative network uses the noise data to generate a second generated data that is unrelated to the sample data but still conforms to the data type corresponding to the sample data.
[0046] The third generated data is generated by fusing the sample data and the noise data. Specifically, the third generated data can be generated by combining the sample data and the noise data in some form (such as splicing) and then simulating them using a generative network. Alternatively, the third generated data can be generated by simulating the sample data and the noise data to obtain the first generated data and the second generated data, respectively. The third generated data, as the generated data corresponding to both the sample data and the noise data, retains the key information in the sample data while introducing more diversity, making it richer in structure and details.
[0047] In one embodiment, the generative network includes a data encoding subnetwork, a data decoding subnetwork, a data generation subnetwork, and a data reassembly subnetwork.
[0048] The data encoding subnetwork can be used to encode data, for example, mapping input data to a latent space or a more compact representation. The data encoding subnetwork can be the decoder in a variational autoencoder (VAE) or other types of network structures, such as convolutional neural networks (CNNs), long short-term memory networks (LSTMs), and multi-layer perceptrons (MLPs).
[0049] The data decoding subnetwork can be used to decode the data, for example, by mapping the encoded representation in the latent space output by the data encoding subnetwork back to the original data space, generating reconstructed data similar to the input data or completely new data. The data decoding subnetwork can be the decoder in a variational autoencoder (VAE) or other network structures such as a deconvolutional neural network (DeCNN).
[0050] The data generation subnetwork is used to transform the data, for example, converting data unrelated to the input data into generated data similar to the sample data. The data generation subnetwork can be the generator in a generative adversarial network (GAN), the decoder in a variational autoencoder, or an autoregressive model.
[0051] The data reorganization subnetwork is used to reorganize the data, such as splicing and weighted fusion of the data. The data reorganization subnetwork can be a fusion network, a conditional generative adversarial network (Conditional GAN), or a multimodal generative network.
[0052] In one embodiment, sample data and noise data are input into a generative network for data simulation generation processing to obtain first generated data corresponding to the sample data, second generated data corresponding to the noise data, and third generated data corresponding to both the sample data and the noise data, including: inputting the sample data into a data encoding subnetwork for encoding processing to obtain sample encoded data; inputting the sample encoded data into a data decoding subnetwork for decoding processing to obtain first generated data; inputting the noise data into a data generation subnetwork for data transformation processing to obtain second generated data; inputting the sample encoded data and the noise data into a data recombination subnetwork for data recombination processing to obtain third generated data.
[0053] The sample encoding data can be a low-dimensional representation extracted from the sample data, representing the abstract features of the sample data. These abstract features retain key information that is more beneficial for subsequent generation tasks. For example, in the medical and health field, when the sample data is a medical image (such as a lung CT scan), it is mapped to a latent space through a generative network, resulting in a low-dimensional vector (i.e., the sample encoding data) that contains key feature information of the image, such as the morphology of lung tissue and the shape of the lesion area.
[0054] For example, let the data encoding subnetwork E be a multi-layer perceptron (MLP), the data decoding subnetwork D_{VAE} be the decoder in VAE, the data generation subnetwork G be the generator in GAN, and the data reconstruction subnetwork R be the fusion network.
[0055] Let the sample data be x, and the noise data be the noise vector z_{G} randomly sampled from the normal distribution N(0, 1). Input x and z_{G} into the generative network. The specific process of the generative network simulating the generation of the two is as follows:
[0056] First, use the data encoding subnetwork E to encode x. E includes n_1 hidden layers. The output data h_{i}^{E} of the i-th hidden layer is calculated as follows:
[0057] h_{i}^{E}=\sigma(W_{i}^{E}h_{i-1}^{E}+b_{i}^{E});
[0058] Among them, W_{i}^{E} is the weight matrix, b_{i}^{E} is the bias vector, \sigma is the activation function (such as ReLU function), and h_{0}^{E}=x.
[0059] Correspondingly, it can be seen that the n_1th hidden layer outputs the potential vector z1 = h_{n_1}^{E}.
[0060] In order to make the potential vector z1 conform to the normal distribution, the reparameterization technique is introduced to reparameterize the potential vector z1:
[0061] z=\mu+\epsilon\odot\sigma;
[0062] where \mu and \sigma are the mean and standard deviation of z1, and \epsilon is the noise vector sampled from the standard normal distribution N(0,I).
[0063] As a result, the data encoding subnetwork E outputs the latent vector z, which is the sample encoding data corresponding to the sample data x.
[0064] Next, the latent vector z is input into the data decoding subnetwork D_{VAE}. By decoding the latent vector z, we obtain the reconstructed data \hat{x}_{VAE}. Here, \hat{x}_{VAE} = D_{VAE}(z). The specific calculation process can be referred to the data encoding subnetwork E above. The calculation process is similar, but in opposite directions, so it will not be repeated here. \hat{x}_{VAE} is the first generated data corresponding to the sample data x.
[0065] Again, the noise vector z_{G} is input into the data generation subnetwork G. By performing multiple linear transformations and nonlinear activation function calculations on z_{G}, \hat{x}_{G} is obtained, where \hat{x}_{G}=G(z_{G}), and \hat{x}_{G} is the second generated data corresponding to the noise vector z_{G}.
[0066] Finally, the data reorganization subnetwork R is used to concatenate the latent vector z and the noise vector z_{G} to fuse the two data and obtain the fused data \hat{x}. \hat{x} is the third generated data corresponding to the sample data x and the noise data noise vector z_{G}.
[0067] Finally, the generative network outputs the first generated data \hat{x}_{VAE}, the second generated data \hat{x}_{G}, and the third generated data \hat{x}.
[0068] It should be noted that compared with the existing technology, the generative network in this application can generate data that is closer to the real data distribution, more diverse and more accurate in important dimensions.
[0069] Step S103: input the sample data, the first generated data, the second generated data, and the third generated data into the discriminant network for processing to obtain similarity data corresponding to the sample data and the first generated data, the second generated data, and the third generated data, respectively.
[0070] Discriminative networks can be used to compare and measure the similarity between data. Specifically, they can be models or network structures such as the discriminator in a generative adversarial network, a metric learning network, or a convolutional neural network. Similarity data is used to characterize the degree of similarity between data, such as cosine similarity, Euclidean distance, or Manhattan distance. The specific similarity depends on the data type of the sample data and the task requirements.
[0071] In one embodiment, the discriminant network includes a feature extraction subnetwork, a feature fusion subnetwork, and a data comparison subnetwork.
[0072] The feature extraction subnetwork is used to extract features from the data to obtain important feature representations corresponding to the data, thereby capturing the key information in the data. Depending on the type of input data, the feature extraction subnetwork can adopt different models or network structures. For example, when the input data is image data, the feature extraction subnetwork can be a convolutional neural network (CNN); when the input data is sequence data, the feature extraction subnetwork can be a long short-term memory network (LSTM); when the input data is text data, the feature extraction subnetwork can be a pre-trained language model such as BERT or GPT, and so on.
[0073] The feature fusion subnetwork is used to fuse features from different sources or from different processing stages. This fusion can be achieved by element-by-element addition of features, weighted combination of features, or pooling to aggregate and compress important information from each feature into a compact global feature. The feature fusion subnetwork effectively integrates key information from data or features into a comprehensive representation, effectively improving the accuracy of subsequent data processing.
[0074] The data comparison subnetwork is used to compare data. It can be a multi-layer perceptron or a fully connected layer. The data comparison subnetwork captures subtle features and differences between the data to evaluate the similarity between multiple data.
[0075] In one embodiment, if Figure 2 As shown, step S103 includes: sub-steps S1031 to S1033.
[0076] Sub-step S1031: input the sample data, the first generated data, the second generated data and the third generated data into multiple convolutional layers of the feature extraction subnetwork for feature extraction, and obtain multiple target sample features corresponding to the sample data, multiple first target features corresponding to the first generated data, multiple second target features corresponding to the second generated data and multiple third target features corresponding to the third generated data.
[0077] Among them, the feature extraction subnetwork includes multiple convolutional layers, and multiple specific convolutional layers are determined from these multiple convolutional layers, such as the l_1, l_2 and l_s layers. These specific convolutional layers are used to extract the features corresponding to the input data (any one of the sample data, the first generated data, the second generated data and the third generated data), and the features F_{l_1}, F_{l_2} and F_{l_s} corresponding to the input data are obtained respectively.
[0078] Sub-step S1032: Input multiple target sample features, multiple first target features, second target features and multiple third target features into the feature fusion sub-network for feature fusion, and obtain sample feature fusion data corresponding to the sample data, first fusion feature data corresponding to the first generated data, second fusion feature data corresponding to the second generated data and third fusion feature data corresponding to the third generated data.
[0079] Among them, multiple features corresponding to the same data may differ in scale, spatial position or dimension. In this case, these features can be processed first, such as upsampling or downsampling operations, 1×1 convolution and splicing, to ensure that they can be aligned during fusion.
[0080] Taking the sample feature fusion data corresponding to the sample data x as an example, let the features extracted by the feature extraction subnetwork be feature maps, and let the target sample feature corresponding to the sample data x be F x _{l_1}、F x _{l_2} and F x _{l_s}. First, these target sample features are upsampled or downsampled to make them have the same size. Then, these target feature samples of the same size are fused by element-by-element addition to obtain the fused feature map F x _{fusion},F x _{fusion}=F x _{l_1}+F x _{l_2}+F x _{l_s}.
[0081] Sub-step S1033: input the sample feature fusion data, the first fusion feature data, the second fusion feature data and the third fusion feature data into the data comparison sub-network to respectively calculate the similarity between the sample feature fusion data and the first fusion feature data, the second fusion feature data and the third fusion feature data to obtain multiple similarity data.
[0082] Exemplarily, the cosine similarity function is used to calculate the similarities between the sample feature fusion and the first fused feature data, the second fused feature data, and the third fused feature data, and the obtained similarities are 0.71, 0.68, and 0.76, respectively.
[0083] Step S104: Based on the sample data, the first generated data, the second generated data, and the plurality of similarity data, the parameters of the generative network and the discriminative network are alternately updated.
[0084] The sample data, the first generated data, and the second generated data can be used to calculate the generation loss of the generative network, which is used to measure the difference between the generated data and the sample data. By adjusting the parameters of the generative network to reduce the difference between the generated data and the sample data (i.e., the real data), the simulation generation capability of the generative network is improved. Multiple similarity data reflect the degree of match between the generated data and the real data, and the discriminative network better distinguishes between the generated data and the real data by adjusting its parameters.
[0085] It should be noted that in each alternating update cycle, the generative network and the discriminative network can be updated once or multiple times. The update times of the generative network and the discriminative network can be the same or different, and can be flexibly set according to task requirements or model training status. In practical applications, in order to improve the performance of the discriminative network, the discriminative network can usually be updated first during the first update.
[0086] For example, after updating the parameters of the discriminative network multiple times, the generative network is updated once or multiple times to ensure that the discriminative network provides more stable and reliable feedback for subsequent updates of the generative network, thereby improving the overall training efficiency of the model and accelerating the convergence of the model.
[0087] Through this alternating updating method, the parameters of the generative network and the discriminative network are continuously optimized so that the two can promote each other, and ultimately the generative network can generate generated data that is closer to real data, more diverse, and more accurate in important feature dimensions, while the discriminative network can distinguish between real data and generated data more effectively and accurately.
[0088] It is understandable that the stopping conditions for alternating parameter updates of the generative network and the discriminative network can be determined based on actual needs. For example, if the purpose of training a data analysis model is to achieve "data augmentation and filling," then the stopping condition could be to stop updating the parameters of the generative network and the discriminative network after the samples output by the generative network reach a certain quality, that is, when the generation error of the generative network is less than a certain threshold. For another example, if the purpose of training a data analysis model is "anomaly detection," in order to achieve accurate anomaly detection of the discrimination results of the input data, then training can be stopped when the discrimination loss of the discriminative network is less than a certain threshold. The stopping condition can also be reaching the training round, etc.
[0089] In one embodiment, if Figure 3 As shown, step S104 includes: sub-step S1041 to sub-step S1042.
[0090] Sub-step S1041: Calculate the discriminant loss value of the discriminant network based on the multiple similarity data, and update the parameters of the discriminant network based on the discriminant loss value.
[0091] The calculation of the discriminant loss value can be based on all similarity data, such as the weighted average or weighted sum of all similarity data; the calculation of the discriminant loss value can also be based on one of all similarity data, such as taking the average or median of all similarity data as the basis for the discriminant loss; the calculation of the discriminant loss value can also be based on some similarity data among all similarity data, such as selecting one or more similarity data that can represent that the generated data is more similar to the sample data (real data), and one or more similarity data that can represent that the generated data is more different from the sample data (real data). The specific calculation formula of the discriminant loss value can adopt the contrast loss function, the cross entropy loss function, etc. It should be noted that with the goal of minimizing the discriminant loss value, the parameters of the discriminant network are updated, thereby continuously enhancing the discriminant network's ability to distinguish subtle differences between real data and generated data, more keenly capturing the subtle features and differences of the data, and thus improving the discriminant performance of the discriminant network.
[0092] In one embodiment, based on multiple similarity data, a discriminant loss value of a discriminant network is calculated, including: based on the numerical size of each similarity data, determining first similarity data with a high value and second similarity data with a low value from the multiple similarity data; using a preset contrast loss function, calculating the first similarity data and the second similarity data to obtain the discriminant loss value of the discriminant network.
[0093] For example, a triple (x, \hat{x}_{VAE}, \hat{x}_{G}) is constructed based on the sample data x, the first generated data \hat{x}_{VAE}, and the second generated data \hat{x}_{G}. The similarity between the data is calculated using the cosine similarity function, and the discriminant loss value L_{contrast} of the discriminant network is calculated using the contrast loss function. The specific process is as follows:
[0094] L_{contrast}=-\log\frac{\exp(sim(x,\hat{x}_{real}) / \tau)}{\exp(sim(x,\hat{x}_{real}) / \tau)+\exp(sim(x,\hat{x}_{fake}) / \tau)}
[0095] Where sim is the cosine similarity function that measures data similarity; \tau is the temperature hyperparameter, which is used to adjust the scale of similarity between data to make the differences between similarities more obvious; \hat{x}_{real} is generated data that is more similar to real data, which can be obtained through screening or adjustment, and \hat{x}_{fake} is generated data that is significantly different from real data.
[0096] It should be noted that by comparing and analyzing the first generated data, the second generated data, and the third generated data generated by different generation methods, and updating the parameters of the discriminant network based on the obtained discriminant loss value, the discriminant network can better distinguish the detailed differences between different generated data during the parameter update process, and further improve the sensitivity and robustness of the discriminant network to small differences between real data and generated data.
[0097] Sub-step S1042: Calculate a generation loss value of the generative network based on the sample data, the first generation data, and the second generation data, and update the parameters of the generative network based on the generation loss value.
[0098] The generation loss value of the generative network can be based on one or more of the sample data and the first generated data or the second generated data. The specific process and reasoning can be referred to the selection process of multiple similarity data in the discriminative network described above and will not be repeated here. It can be understood that by using the generation loss value as the optimization target and updating the parameters of the generative network, the generative network is pushed to generate more realistic data and reduce the difference between the generated data and the real data.
[0099] In one embodiment, a generation loss value of a generative network is calculated based on sample data, first generation data, and second generation data, including: encoding the sample data to obtain sample encoded data; calculating the sample encoded data based on a relative entropy loss function to obtain a first generation loss value for characterizing the difference between the sample encoded data and a standard normal distribution; calculating the first generation data based on a mean square error loss function to obtain a second generation loss value for characterizing the difference between the first generation data and the sample data; determining the importance weights of the second generation data and the sample data in different feature dimensions based on an attention mechanism, and calculating the second generation data based on a binary cross entropy function and the importance weights in different feature dimensions to obtain a third generation loss value; and calculating the sum of the first generation loss value, the second generation loss value, and the third generation loss value as the generation loss value of the generative network.
[0100] Exemplarily, the KL divergence loss function (i.e., relative entropy loss function) is used to measure the difference between the distribution of the sample encoded data z and the standard normal distribution N(0, 1) to obtain the first generation loss value L_{KL}. The specific calculation formula is:
[0101] L_{KL}=D_{KL}(q_{\phi}(z|x)||p(z))=\mathbb{E}_{q_{\phi}(z|x)}[\log\frac{q_{\phi}(z|x)}{p(z)}];
[0102] Among them, q_{\phi}(z|x) is the distribution of the sample encoded data z output by the first generation sub-network, and p(z) is the standard normal distribution.
[0103] The mean square error (MSE) is used to measure the difference between the first generated data and the sample data x, and the second generated loss value L_{recon} is obtained. The specific calculation formula is:
[0104] L_{recon}=\mathbb{E}_{x\sim p_{data}(x)}[(x-\hat{x}_{VAE})^2]
[0105] In addition, the calculation process of the sample encoded data and the first generated data is similar but in opposite directions. Therefore, a hyperparameter is introduced to balance the first generation loss value and the second generation loss value. Specifically, the loss corresponding to the sample encoded data and the first generated data is L_{VAE}, and the specific calculation formula is:
[0106] L_{VAE}=L_{recon}+\lambda L_{KL};
[0107] Among them, \lambda is a hyperparameter that balances the first generation loss value and the second generation loss value.
[0108] The attention mechanism is used to calculate the importance weights \alpha of the second generated data and the sample data in different feature dimensions, and the importance weights \alpha and the binary cross entropy function are used to measure the difference between the second generated data and the sample data x to obtain the third generated loss value L_{G_{new}}. The specific calculation formula is:
[0109] L_{G_{new}}=-\mathbb{E}_{z\simp(z)}[\sum_{i=1}^{d}\alpha_i\log(D(G(z))_i)].
[0110] The total generation loss value L_{generator} of the final generative network is:
[0111] L_{generator}=L_{VAE}+L_{G_{new}}.
[0112] By using the generated loss value to update the parameters of the generative network, the generative network is encouraged to generate data that is closer to the real data distribution, more diverse, and more accurate in important feature dimensions.
[0113] The data analysis model training method provided in the above embodiment, by continuously and alternately updating the parameters of the generative network and the discriminative network, makes the data generated by the generative network increasingly realistic and better simulates the details and features of the sample data, while the discriminative network can more keenly identify subtle differences between the sample data and other generated data. Ultimately, this effectively improves the overall performance of the data analysis model, thereby greatly enhancing the accuracy of data analysis.
[0114] like Figure 4 As shown, the embodiment of the present application also provides a data analysis method, which includes:
[0115] Step S201: input the data to be analyzed into the data analysis model to obtain a probability value of the data to be analyzed being abnormal data; wherein the data analysis model is obtained based on the data analysis model training method in any of the above embodiments.
[0116] Step S202: If the probability value is less than the preset probability threshold, a plurality of third generated data corresponding to the data to be analyzed is obtained, and target generated data for representing normal data is determined from the plurality of third generated data.
[0117] Step S203: Determine the degree of abnormality of the data to be analyzed based on the deviation data between the data to be analyzed and the target generated data.
[0118] In some cases, the data to be analyzed can be determined to be abnormal data directly based on the probability value corresponding to the data to be analyzed. In other cases, one or more of the multiple generated data corresponding to the data to be analyzed can be compared, such as the distribution of the data.
[0119] For example, the data to be analyzed Q is input into the data analysis model, and after being processed by the generative network and the discriminative network, the encoded data Q corresponding to the Q output by the generative network is obtained. Z , the first generated data Q1, the second generated data Q2 and a plurality of third generated data Q3, and the probability value P of the discriminant network output for Q Q ; Among them, the coded data Q Z It is obtained by encoding the data Q to be analyzed.
[0120] The coded data Q Z The distribution of the normal data corresponding to the data to be analyzed Q is compared, and the probability value P corresponding to the data to be analyzed Q is Q Compare with the preset probability threshold P (P = 0.5). If P Q Less than P, and the encoded data Q Z The distribution of the data deviates greatly from the distribution of the normal data, then the coded data Q is determined Z is abnormal data. Combined with the distribution centers of multiple third generated data Q3, calculate the Mahalanobis distance d(x_{anomaly},\mu) from the data to be analyzed Q to the distribution centers of multiple third generated data Q3. The specific calculation formula is as follows:
[0121] d(x_{anomaly},\mu)=\sqrt{(x_{anomaly}-\mu)^T\Sigma^{-1}(x_{anomaly}-\mu)}.
[0122] Where x_{anomaly} is the data to be analyzed Q, \mu is the mean vector of the distribution centers of multiple third-generated data Q3, and \Sigma is the covariance matrix.
[0123] The larger the Mahalanobis distance d(x_{anomaly},\mu), the greater the deviation between the data to be analyzed Q and the normal data distribution, that is, the more serious the anomaly of the data to be analyzed Q is.
[0124] Please refer to Figure 5 , Figure 5 A schematic block diagram of a data processing device provided in an embodiment of the present application.
[0125] like Figure 5 As shown, the data processing device 300 includes:
[0126] The data acquisition module 301 is used to acquire sample data and noise data corresponding to the sample data;
[0127] A data generation module 302 is configured to input sample data and noise data into a generative network for data simulation generation processing to obtain first generated data corresponding to the sample data, second generated data corresponding to the noise data, and third generated data corresponding to both the sample data and the noise data;
[0128] The data determination module 303 is used to input the sample data, the first generated data, the second generated data, and the third generated data into the discriminant network for processing to obtain similarity data corresponding to the sample data and the first generated data, the second generated data, and the third generated data respectively;
[0129] The parameter updating module 304 is configured to alternately update the parameters of the generative network and the discriminative network based on the sample data, the first generated data, the second generated data, and the plurality of similarity data.
[0130] In one embodiment, the generative network includes a data encoding subnetwork, a data decoding subnetwork, a data generation subnetwork, and a data reassembly subnetwork. The data generation module 302 is further configured to:
[0131] Input the sample data into the data encoding sub-network for encoding processing to obtain sample encoded data;
[0132] Inputting the sample encoded data into the data decoding subnetwork for decoding processing to obtain first generated data;
[0133] Inputting the noise data into the data generation subnetwork for data transformation processing to obtain the second generated data;
[0134] The sample coded data and the noise data are input into the data recombination subnetwork for data recombination processing to obtain the third generated data.
[0135] In one embodiment, the discriminant network includes a feature extraction subnetwork, a feature fusion subnetwork, and a data comparison subnetwork. Figure 6 As shown, the data determination module 303 includes:
[0136] The feature extraction submodule 3031 is used to input the sample data, the first generated data, the second generated data and the third generated data into multiple convolutional layers of the feature extraction subnetwork for feature extraction, thereby obtaining multiple target sample features corresponding to the sample data, multiple first target features corresponding to the first generated data, multiple second target features corresponding to the second generated data and multiple third target features corresponding to the third generated data.
[0137] The feature fusion submodule 3032 is used to input multiple target sample features, multiple first target features, second target features and multiple third target features into the feature fusion subnetwork for feature fusion, and obtain sample feature fusion data corresponding to the sample data, first fusion feature data corresponding to the first generated data, second fusion feature data corresponding to the second generated data and third fusion feature data corresponding to the third generated data.
[0138] The data comparison submodule 3033 is used to input the sample feature fusion data, the first fusion feature data, the second fusion feature data and the third fusion feature data into the data comparison subnetwork to respectively calculate the similarity between the sample feature fusion data and the first fusion feature data, the second fusion feature data and the third fusion feature data to obtain multiple similarity data.
[0139] In one embodiment, Figure 7 As shown, the data determination module 304 includes:
[0140] The first parameter updating submodule 3041 is used to calculate the discriminant loss value of the discriminant network based on the multiple similarity data, and update the parameters of the discriminant network based on the discriminant loss value.
[0141] The second parameter updating submodule 3042 is used to calculate the generation loss value of the generative network based on the sample data, the first generation data and the second generation data, and update the parameters of the generative network based on the generation loss value.
[0142] In one embodiment, the data determination module 304 is further configured to:
[0143] Based on the numerical values of each similarity data, determining first similarity data with a high numerical value and second similarity data with a low numerical value from the plurality of similarity data;
[0144] The first similarity data and the second similarity data are calculated using a preset contrast loss function to obtain a discriminant loss value of the discriminant network.
[0145] In one embodiment, the data determination module 304 is further configured to:
[0146] Performing encoding processing on the sample data to obtain sample encoded data;
[0147] Calculating the sample coded data based on the relative entropy loss function to obtain a first generation loss value for characterizing the difference between the sample coded data and a standard normal distribution;
[0148] Calculating the first generated data based on a mean square error loss function to obtain a second generated loss value for characterizing the difference between the first generated data and the sample data;
[0149] Determine the importance weights of the second generated data and the sample data in different feature dimensions based on the attention mechanism, and calculate the second generated data based on the binary cross entropy function and the importance weights in different feature dimensions to obtain a third generated loss value;
[0150] The sum of the first generation loss value, the second generation loss value, and the third generation loss value is calculated as the generation loss value of the generative network.
[0151] It should be noted that those skilled in the art can clearly understand that, for the convenience and brevity of description, the specific working processes of the above-described devices and modules and units can refer to the corresponding processes in the aforementioned data analysis model training method embodiment, and will not be repeated here.
[0152] The apparatus provided in the above embodiment can be implemented in the form of a computer program. The computer program can be used in Figure 8 Runs on the computer device shown.
[0153] See also Figure 8 , Figure 8 A schematic block diagram of the structure of a computer device provided in an embodiment of the present application.
[0154] like Figure 8 As shown, the computer device includes a processor, a memory and a network interface connected through a system bus, wherein the memory may include a storage medium and an internal memory, and the storage medium may be non-volatile or volatile.
[0155] The storage medium may store an operating system and a computer program. The computer program includes program instructions that, when executed, enable the processor to perform any data analysis model training method.
[0156] The processor is used to provide computing and control capabilities and support the operation of the entire computer equipment.
[0157] The internal memory provides an environment for the operation of the computer program in the storage medium. When the computer program is executed by the processor, the processor can execute any training method of the data analysis model.
[0158] The network interface is used for network communication, such as sending assigned tasks, etc. Those skilled in the art will understand that Figure 8 The structure shown in the figure is only a block diagram of a part of the structure related to the solution of the present application, and does not constitute a limitation on the computer device to which the solution of the present application is applied. The specific computer device may include more or fewer components than shown in the figure, or combine certain components, or have a different component arrangement.
[0159] It should be understood that the processor may be a central processing unit (CPU), or other general-purpose processors, digital signal processors (DSP), application-specific integrated circuits (ASIC), field-programmable gate arrays (FPGA), or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, etc. The general-purpose processor may be a microprocessor or any conventional processor, etc.
[0160] In one embodiment, the processor is configured to execute a computer program stored in the memory to implement the following steps:
[0161] Obtain sample data and noise data corresponding to the sample data;
[0162] Inputting the sample data and the noise data into the generative network for data simulation generation processing to obtain first generated data corresponding to the sample data, second generated data corresponding to the noise data, and third generated data corresponding to both the sample data and the noise data;
[0163] Inputting the sample data, the first generated data, the second generated data, and the third generated data into the discriminant network for processing to obtain similarity data corresponding to the sample data and the first generated data, the second generated data, and the third generated data respectively;
[0164] Based on the sample data, the first generated data, the second generated data, and the plurality of similarity data, parameters of the generative network and the discriminative network are alternately updated.
[0165] In one embodiment, the generative network includes a data encoding subnetwork, a data decoding subnetwork, a data generation subnetwork, and a data recombination subnetwork. When the processor inputs sample data and noise data into the generative network for data simulation generation processing to obtain first generated data corresponding to the sample data, second generated data corresponding to the noise data, and third generated data corresponding to both the sample data and the noise data, it is configured to implement:
[0166] Input the sample data into the data encoding sub-network for encoding processing to obtain sample encoded data;
[0167] Inputting the sample encoded data into the data decoding subnetwork for decoding processing to obtain first generated data;
[0168] Inputting the noise data into the data generation subnetwork for data transformation processing to obtain the second generated data;
[0169] The sample coded data and the noise data are input into the data recombination subnetwork for data recombination processing to obtain the third generated data.
[0170] In one embodiment, when the processor inputs the sample data, the first generated data, the second generated data, and the third generated data into the discriminant network for processing to obtain similarity data corresponding to the sample data and the first generated data, the second generated data, and the third generated data, respectively, the processor is configured to implement:
[0171] The sample data, first generated data, second generated data and third generated data are respectively input into multiple convolutional layers of the feature extraction subnetwork for feature extraction to obtain multiple target sample features corresponding to the sample data, multiple first target features corresponding to the first generated data, multiple second target features corresponding to the second generated data and multiple third target features corresponding to the third generated data.
[0172] Multiple target sample features, multiple first target features, second target features and multiple third target features are respectively input into the feature fusion subnetwork for feature fusion to obtain sample feature fusion data corresponding to the sample data, first fusion feature data corresponding to the first generated data, second fusion feature data corresponding to the second generated data and third fusion feature data corresponding to the third generated data.
[0173] The sample feature fusion data, the first fusion feature data, the second fusion feature data and the third fusion feature data are input into the data comparison subnetwork to respectively calculate the similarity between the sample feature fusion data and the first fusion feature data, the second fusion feature data and the third fusion feature data to obtain multiple similarity data.
[0174] In one embodiment, when the processor implements the alternate updating of parameters of the generative network and the discriminative network based on the sample data, the first generated data, the second generated data, and the plurality of similarity data, the processor is configured to implement:
[0175] Based on multiple similarity data, the discriminant loss value of the discriminant network is calculated, and the parameters of the discriminant network are updated based on the discriminant loss value;
[0176] Based on the sample data, the first generated data, and the second generated data, a generation loss value of the generative network is calculated, and parameters of the generative network are updated based on the generation loss value.
[0177] In one embodiment, when calculating the discriminant loss value of the discriminant network based on multiple similarity data, the processor is configured to implement:
[0178] Based on the numerical values of each similarity data, determining first similarity data with a high numerical value and second similarity data with a low numerical value from the plurality of similarity data;
[0179] The first similarity data and the second similarity data are calculated using a preset contrast loss function to obtain a discriminant loss value of the discriminant network.
[0180] In one embodiment, when calculating the generation loss value of the generative network based on the sample data, the first generation data, and the second generation data, the processor is configured to implement:
[0181] Performing encoding processing on the sample data to obtain sample encoded data;
[0182] Calculating the sample coded data based on the relative entropy loss function to obtain a first generation loss value for characterizing the difference between the sample coded data and a standard normal distribution;
[0183] Calculating the first generated data based on a mean square error loss function to obtain a second generated loss value for characterizing the difference between the first generated data and the sample data;
[0184] Determine the importance weights of the second generated data and the sample data in different feature dimensions based on the attention mechanism, and calculate the second generated data based on the binary cross entropy function and the importance weights in different feature dimensions to obtain a third generated loss value;
[0185] The sum of the first generation loss value, the second generation loss value, and the third generation loss value is calculated as the generation loss value of the generative network.
[0186] In one embodiment, the processor is configured to execute a computer program stored in the memory to implement the following steps:
[0187] Inputting the data to be analyzed into a data analysis model to obtain a probability value of the data to be analyzed being abnormal data; wherein the data analysis model is obtained based on the training method of the data analysis model in any of the above embodiments;
[0188] If the probability value is less than a preset probability threshold, obtaining a plurality of third generated data corresponding to the data to be analyzed, and determining target generated data for representing normal data from the plurality of third generated data;
[0189] The degree of abnormality of the data to be analyzed is determined based on the deviation data between the data to be analyzed and the target generated data.
[0190] It should be noted that those skilled in the art can clearly understand that, for the sake of convenience and brevity of description, the specific working process of the computer device described above can refer to the corresponding process in the embodiment of the training method of the aforementioned data analysis model, and will not be repeated here.
[0191] The present application can be used in many general or special computer system environments or configurations. For example: personal computers, server computers, handheld or portable devices, tablet devices, multiprocessor systems, microprocessor-based systems, set-top boxes, programmable consumer electronics, network PCs, minicomputers, mainframe computers, distributed computing environments including any of the above systems or devices, and the like. The present application can be described in the general context of computer-executable instructions executed by a computer, such as program modules. Generally, program modules include routines, programs, objects, components, data structures, and the like that perform specific tasks or implement specific abstract data types. The present application can also be practiced in distributed computing environments in which tasks are performed by remote processing devices connected via a communication network. In a distributed computing environment, program modules can be located in local and remote computer storage media, including storage devices.
[0192] An embodiment of the present application also provides a computer-readable storage medium, on which a computer program is stored. The computer program includes program instructions. The method implemented when the program instructions are executed can refer to the various embodiments of the training method of the data analysis model of the present application.
[0193] The computer-readable storage medium may be an internal storage unit of the computer device described in the aforementioned embodiment, such as a hard disk or memory of the computer device. The computer-readable storage medium may also be an external storage device of the computer device, such as a plug-in hard disk, a SmartMedia Card (SMC), a Secure Digital (SD) card, a flash memory card, etc., equipped on the computer device.
[0194] Furthermore, the computer-usable storage medium may mainly include a program storage area and a data storage area, wherein the program storage area may store an operating system, an application required for at least one function, etc.; the data storage area may store data created based on the use of blockchain nodes, etc. The blockchain referred to in this application is a new application model of computer technologies such as distributed data storage, point-to-point transmission, consensus mechanism, and encryption algorithm. Blockchain is essentially a decentralized database, a string of data blocks generated by cryptographic methods. Each data block contains information about a batch of network transactions, which is used to verify the validity of its information (anti-counterfeiting) and generate the next block. The blockchain may include the blockchain underlying platform, the platform product service layer, and the application service layer.
[0195] It should be understood that the terms used in this specification are only for the purpose of describing specific embodiments and are not intended to limit the present application. As used in this specification and the appended claims, the singular forms "a", "an", and "the" are intended to include the plural forms unless the context clearly indicates otherwise.
[0196] It should also be understood that the term "and / or" used in this specification and the appended claims refers to any combination of one or more of the associated listed items and all possible combinations, including these combinations. It should be noted that, in this article, the terms "include", "comprise" or any other variations thereof are intended to cover non-exclusive inclusion, so that a process, method, article or system that includes a series of elements includes not only those elements, but also other elements that are not explicitly listed, or also includes elements that are inherent to such process, method, article or system. In the absence of further restrictions, an element defined by the sentence "including a..." does not exclude the presence of other identical elements in the process, method, article or system that includes the element.
[0197] The serial numbers of the embodiments of the present application are for description only and do not represent the advantages or disadvantages of the embodiments. The above description is only a specific embodiment of the present application, but the scope of protection of the present application is not limited thereto. Any technician familiar with the technical field can easily think of various equivalent modifications or replacements within the technical scope disclosed in this application, and these modifications or replacements should be included in the scope of protection of this application. Therefore, the scope of protection of this application should be based on the scope of protection of the claims.
Claims
1. A method for training a data analysis model, characterized in that: The data analysis model includes a generative network and a discriminative network, and the method includes: Acquire sample data and noise data corresponding to the sample data; Inputting the sample data and the noise data into the generative network for data simulation generation processing to obtain first generated data corresponding to the sample data, second generated data corresponding to the noise data, and third generated data corresponding to both the sample data and the noise data; Inputting the sample data, the first generated data, the second generated data, and the third generated data into the discriminant network for processing to obtain similarity data corresponding to the sample data and the first generated data, the second generated data, and the third generated data respectively; Based on the sample data, the first generated data, the second generated data, and the plurality of similarity data, parameters of the generative network and the discriminative network are alternately updated.
2. The data analysis model training method according to claim 1, wherein: The step of alternately updating the parameters of the generative network and the discriminative network based on the sample data, the first generated data, the second generated data, and the plurality of similarity data includes: Calculating a discrimination loss value of the discriminant network based on the plurality of similarity data, and updating parameters of the discriminant network based on the discrimination loss value; and Based on the sample data, the first generated data, and the second generated data, a generation loss value of the generative network is calculated, and parameters of the generative network are updated based on the generation loss value.
3. The data analysis model training method according to claim 2, characterized in that: The calculating the discriminant loss value of the discriminant network based on the plurality of similarity data includes: Based on the numerical values of the respective similarity data, determining first similarity data with a high numerical value and second similarity data with a low numerical value from the plurality of similarity data; The first similarity data and the second similarity data are calculated using a preset contrast loss function to obtain a discriminant loss value of the discriminant network.
4. The data analysis model training method according to claim 2, wherein: The calculating the generation loss value of the generative network based on the sample data, the first generation data, and the second generation data includes: performing encoding processing on the sample data to obtain sample encoded data; Calculating the sample coded data based on a relative entropy loss function to obtain a first generation loss value for characterizing the difference between the sample coded data and a standard normal distribution; Calculating the first generated data based on a mean square error loss function to obtain a second generated loss value used to characterize the difference between the first generated data and the sample data; determining, based on an attention mechanism, importance weights of the second generated data and the sample data in different feature dimensions, and calculating the second generated data based on a binary cross entropy function and the importance weights in different feature dimensions to obtain a third generated loss value; A sum of the first generation loss value, the second generation loss value, and the third generation loss value is calculated as a generation loss value of the generative network.
5. The data analysis model training method according to claim 1, wherein: The discriminant network includes a feature extraction subnetwork, a feature fusion subnetwork and a data comparison subnetwork; The step of inputting the sample data, the first generated data, the second generated data, and the third generated data into the discriminant network for processing to obtain similarity data corresponding to the sample data and the first generated data, the second generated data, and the third generated data, respectively, includes: Inputting the sample data, the first generated data, the second generated data, and the third generated data into the multiple convolutional layers of the feature extraction subnetwork for feature extraction, respectively, to obtain multiple target sample features corresponding to the sample data, multiple first target features corresponding to the first generated data, multiple second target features corresponding to the second generated data, and multiple third target features corresponding to the third generated data; Inputting the multiple target sample features, the multiple first target features, the second target features, and the multiple third target features into a feature fusion subnetwork for feature fusion, respectively, to obtain sample feature fusion data corresponding to the sample data, first fusion feature data corresponding to the first generated data, second fusion feature data corresponding to the second generated data, and third fusion feature data corresponding to the third generated data; The sample feature fusion data, the first fusion feature data, the second fusion feature data and the third fusion feature data are input into the data comparison subnetwork to respectively calculate the similarity between the sample feature fusion data and the first fusion feature data, the second fusion feature data and the third fusion feature data to obtain a plurality of similarity data.
6. The data analysis model training method according to claim 1, wherein: The generative network includes a data encoding subnetwork, a data decoding subnetwork, a data generation subnetwork and a data reorganization subnetwork; The step of inputting the sample data and the noise data into the generative network for data simulation generation processing to obtain first generated data corresponding to the sample data, second generated data corresponding to the noise data, and third generated data corresponding to both the sample data and the noise data includes: Inputting the sample data into the data encoding subnetwork for encoding processing to obtain sample encoded data; Inputting the sample encoded data into the data decoding subnetwork for decoding processing to obtain the first generated data; Inputting the noise data into the data generation subnetwork for data transformation processing to obtain the second generated data; The sample coded data and the noise data are input into the data recombination subnetwork for data recombination processing to obtain the third generated data.
7. A data analysis method, characterized in that: include: Inputting the data to be analyzed into a data analysis model to obtain a probability value of the data to be analyzed being abnormal data; wherein the data analysis model is obtained based on the data analysis model training method according to any one of claims 1 to 6; If the probability value is less than a preset probability threshold, obtaining a plurality of third generated data corresponding to the data to be analyzed, and determining target generated data for representing normal data from the plurality of third generated data; The abnormality degree of the data to be analyzed is determined according to the deviation data between the data to be analyzed and the target generated data.
8. A data processing device, characterized in that: The data processing device includes: A data acquisition module, configured to acquire sample data and noise data corresponding to the sample data; a data generation module, configured to input the sample data and the noise data into the generative network for data simulation generation processing, thereby obtaining first generated data corresponding to the sample data, second generated data corresponding to the noise data, and third generated data corresponding to both the sample data and the noise data; a data determination module, configured to input the sample data, the first generated data, the second generated data, and the third generated data into the discriminant network for processing, so as to obtain similarity data corresponding to the sample data and the first generated data, the second generated data, and the third generated data, respectively; A parameter updating module is used to alternately update the parameters of the generative network and the discriminative network based on the sample data, the first generated data, the second generated data, and the plurality of similarity data.
9. A computer device, characterized in that: The computer device includes a processor, a memory, and a computer program stored in the memory and executable by the processor, wherein when the computer program is executed by the processor, it implements the training method of the data analysis model as described in any one of claims 1 to 6, or implements the data analysis method as described in claim 7.
10. A computer-readable storage medium, characterized in that The computer-readable storage medium stores a computer program, wherein when the computer program is executed by a processor, it implements the training method of the data analysis model as described in any one of claims 1 to 6, or implements the data analysis method as described in claim 7.