Data classification method and system based on deep PU learning and category prior estimation

Abstract

The present invention proposes a data classification method and system based on deep PU learning and category prior estimation. The present invention can simultaneously estimate the category prior and use the obtained prior to estimate the learning deep model without knowing the true prior distribution of the data, thus being more suitable for the application of PU learning in practical scenarios. The iterative framework proposed in the present invention includes modeling the prediction score of the network as a GMM to estimate the positive class prior; performing unbiased PU learning based on the estimated value of the positive class prior; and then combining the average teacher, temperature sharpening and other techniques of semi-supervised learning to improve the performance and stability of the algorithm. The framework can be applied to PU problems in various fields including computer vision, recommendation systems, biomedicine, etc., with excellent results, and has both scientific and practical value.

Description

Technical Field

[0001] The present invention relates to the technical field of positive sample and unlabeled (Positive-Unlabeled, PU) learning in machine learning, and in particular to cost-sensitive deep PU learning. Background Art

[0002] In recent years, with the development of the internet and information technology, humanity has entered the era of big data. Deep learning, powered by massive amounts of data, has garnered widespread attention and achieved groundbreaking progress. However, the superior performance of deep learning algorithms relies on vast amounts of data, particularly complete category labeling. However, in many practical applications, data annotation is difficult and costly to obtain, and labels are often only available for a small portion of the data, often with limited human and material resources. Therefore, learning methods that rely less on data annotation have become popular, and PU learning is one such approach. For example, in the classification of rare diseases, confirmed samples can be considered positive, while undiagnosed samples, known as unlabeled samples, still harbor the potential for rare diseases. This means that unlabeled samples contain both positive and negative examples. Similar situations also arise in tasks such as malicious URL detection, false comment detection, and particle extraction in cryo-electron microscopy. This demonstrates the significant value of learning using only positive and unlabeled data, also known as PU learning.

[0003] The training set for PU learning consists of a subset of labeled positive samples and a subset of unlabeled samples. The key and difficulty lies in how to utilize unlabeled data. Therefore, PU learning is divided into different schools based on how it handles unlabeled data: two-stage methods, biased PU learning, and unbiased PU learning. The two-stage PU learning approach is the most intuitive. This approach first extracts reliable negative and positive samples from unlabeled data based on smoothing or clustering assumptions, thereby transforming the PU problem into a general semi-supervised learning problem. Training is then performed using semi-supervised or supervised learning methods. Another more intuitive approach is biased PU learning, which treats unlabeled data as noisy negative samples. Furthermore, unbiased PU learning, within the framework of cost-sensitive learning, estimates the loss for the positive class using the only positive samples. It also indirectly constructs the loss for the negative class based on the loss resulting from treating all unlabeled data as negative and class prior knowledge. This approach achieves unbiased estimation of common classification optimization objectives, achieving state-of-the-art performance. Summary of the Invention

[0004] When the inventors were conducting research on unbiased PU learning algorithms, they found that such algorithms assume that the true category prior can be obtained in advance. In fact, for most practical application scenarios of PU learning, the category prior is often unknown. If a category prior value is arbitrarily set in advance, especially a positive category prior preset value that is smaller than the true value, the performance of this type of algorithm will be seriously damaged. The above defects greatly reduce the practical value of this type of algorithm. Category prior plays a key role in this type of method and has driven related research on category prior estimation. However, most of these methods can only achieve a relatively good level on small-scale data sets and traditional methods, and are still stretched when faced with deep learning algorithms based on massive data. Therefore, how to accurately estimate the positive category prior at the lowest possible computational cost in the absence of category prior knowledge, so as to perform cost-sensitive deep PU learning, becomes the key to the problem.

[0005] Specifically, in order to overcome the above technical problems, the present invention proposes a data classification method based on deep PU learning and category prior estimation, which includes:

[0006] Step 1: Obtain a training set containing multiple data samples, in which only some data samples are labeled with category labels. Input the training set into two student models and teacher models with the same network structure but different parameters, and obtain the student prediction score and teacher prediction score corresponding to each data sample output by the student model and teacher model respectively;

[0007] Step 2: Input all teacher prediction scores into the Gaussian mixture model to obtain a positive class prior; construct a temperature sharpening loss based on all student prediction scores; construct a consistency loss based on all student prediction scores and teacher prediction scores; obtain a non-negative PU risk based on the positive class prior and all student prediction scores, combine the consistency loss, the non-negative PU risk and the temperature sharpening loss to obtain a target loss, and based on the target loss, use gradient backpropagation to update the parameters of the student model until the target converges or reaches a preset number of iterations, and save the current student model or teacher model as the data classification model, for example, save the current teacher model and the student model with better performance as the data classification model;

[0008] Step 3: Input the data to be classified into the data classification model to obtain the category of the data to be classified.

[0009] The data classification method based on deep PU learning and category prior estimation, when used for malicious URL detection, the data in the training set are some URLs marked with malicious categories and unlabeled URLs, and both the student model and the teacher model are recurrent neural networks; when used for false comment detection, the data in the training set are comments marked with false categories and unlabeled comments, and both the student model and the teacher model are recurrent neural networks; when used for particle picking in cryo-electron microscopy, the data in the training set are particle areas marked with selected categories and unlabeled particle areas, and both the student model and the teacher model are convolutional neural networks.

[0010] The data classification method based on deep PU learning and category prior estimation, wherein step 2 includes:

[0011] The prediction scores of the student model and the teacher model are:

[0012] S = sigmoid(f(X, Θ) t ))

[0013] S′=sigmoid(f(X,Θ′ t ))

[0014] in, X is the training set, S is the student prediction score output by the student model, S′ is the teacher prediction score output by the teacher model, Θ t is the parameter of the student model at time t, Θ′ t are the parameters of the teacher model at time t.

[0015] The data classification method based on deep PU learning and category prior estimation, wherein the consistency loss of the output of the student model and the teacher model for:

[0016]

[0017] Among them, x i ∈X represents the i-th sample in the training set X, c i is its confidence, N = |X|; is an indicator function, which takes the value of 1 when the condition (·) is met, otherwise it takes the value of 0; τ is the confidence threshold, Θ is the parameter of the student model, and Θ′ is the parameter of the teacher model.

[0018] The data classification method based on deep PU learning and category prior estimation, wherein the temperature sharpening loss for:

[0019]

[0020]

[0021] Where T is the temperature of the category distribution and s is the student prediction score output by the student model.

[0022] The present invention also proposes a data classification system based on deep PU learning and category prior estimation, which includes:

[0023] The initialization module is used to obtain a training set including multiple data samples, in which only some data samples are labeled with category labels. The training set is simultaneously input into two student models and teacher models with the same network structure but different parameters. The student prediction score and teacher prediction score corresponding to each data sample output by the student model and the teacher model are obtained respectively.

[0024] The training module is used to input all teacher prediction scores into the Gaussian mixture model to obtain a positive class prior; construct a temperature sharpening loss based on all student prediction scores; construct a consistency loss based on all student prediction scores and teacher prediction scores; obtain a non-negative PU risk based on the positive class prior and all student prediction scores, combine the consistency loss, the non-negative PU risk and the temperature sharpening loss to obtain a target loss, and update the parameters of the student model based on the target loss using gradient backpropagation until the target converges or a preset number of iterations is reached, and save the current student model or teacher model as the data classification model;

[0025] The classification module is used to input the data to be classified into the data classification model to obtain the category of the data to be classified.

[0026] The data classification system based on deep PU learning and category prior estimation, when used for malicious URL detection, the data in the training set are some URLs marked with malicious categories and unlabeled URLs, and both the student model and the teacher model are recurrent neural networks; when used for false comment detection, the data in the training set are comments marked with false categories and unlabeled comments, and both the student model and the teacher model are recurrent neural networks; when used for particle picking in cryo-electron microscopy, the data in the training set are particle areas marked with selected categories and unlabeled particle areas, and both the student model and the teacher model are convolutional neural networks.

[0027] The data classification system based on deep PU learning and category prior estimation, wherein the training module is used to predict the scores of the student model and the teacher model based on the following formula:

[0028] S = sigmoid(f(X, Θ) t ))

[0029] S′=sigmoid(f(X,Θ′ t ))

[0030] in, X is the training set, S is the student prediction score output by the student model, S′ is the teacher prediction score output by the teacher model, Θ t is the parameter of the student model at time t, Θ′ t are the parameters of the teacher model at time t.

[0031] The data classification system based on deep PU learning and category prior estimation, wherein the consistency loss of the output of the student model and the teacher model for:

[0032]

[0033] Among them, x i ∈X represents the i-th sample in the training set X, c i is its confidence, N = |X|; is an indicator function, which takes the value of 1 when the condition (·) is met, otherwise it takes the value of 0; τ is the confidence threshold, Θ is the parameter of the student model, and Θ′ is the parameter of the teacher model.

[0034] The data classification system based on deep PU learning and category prior estimation, wherein the temperature sharpening loss for:

[0035]

[0036]

[0037] Where T is the temperature of the category distribution and s is the student prediction score output by the student model.

[0038] The present invention also proposes a storage medium for storing a program for executing any one of the data classification methods based on deep PU learning and category prior estimation.

[0039] The present invention also proposes a client for use in any of the above-mentioned data classification systems based on deep PU learning and category prior estimation.

[0040] From the above scheme, it can be seen that the advantages of the present invention are:

[0041] The present invention proposes an iterative deep PU learning and category prior estimation framework, which can simultaneously estimate the category prior and use the obtained prior to estimate the learning deep model without knowing the true prior distribution of the data, making it more suitable for the application of PU learning in practical scenarios. The iterative framework proposed in the present invention includes the following core points: (1) Modeling the prediction score of the network as a GMM to estimate the positive class prior; (2) Based on the estimated value of the positive class prior, unbiased PU learning is performed; (3) Then, the average teacher, temperature sharpening and other techniques of semi-supervised learning are combined to improve the performance and stability of the algorithm. This framework can be applied to PU problems in various fields including computer vision, recommendation systems, biomedicine, etc., with excellent results, and has both scientific value and practical value. BRIEF DESCRIPTION OF THE DRAWINGS

[0042] Figure 1 It is a flowchart of the present invention. DETAILED DESCRIPTION

[0043] The goal of the present invention is to solve the problem of how to perform unbiased PU learning in the absence of category prior knowledge. Existing unbiased PU learning methods assume that the category prior is known or easy to estimate, while the category prior in real PU problems is often unknown and difficult to estimate. In addition, the existing category prior estimation algorithms are mainly designed for traditional machine learning classifiers, and do not take advantage of the advantages of deep learning in large-scale data sets. In order to overcome the above problems, the present invention proposes an iterative deep PU learning framework based on an unsupervised mixture model. It utilizes the characteristic that deep neural networks give different distributions of prediction scores for samples of different categories (positive samples and negative samples), and uses a Gaussian mixture model to approximate the mixed distribution of prediction scores. Combined with commonly used semi-supervised learning techniques and optimization goals for PU problems, classification performance comparable to that of PU algorithms based on true positive class priors is achieved.

[0044] The present invention includes the following key technical points:

[0045] Key point 1: Before the deep neural network overfits, the prediction scores of positive samples and negative samples show different distributions. The prediction scores of positive samples are concentrated in the interval with higher scores, while the prediction scores of negative samples are concentrated in the interval with lower scores. The two respectively form two bell-shaped curves with high middle and low ends. Based on the above observations, it is proposed to use Gaussian Mixture Model (GMM) to model the prediction scores in an unsupervised manner. GMM has a small number of parameters and is not constrained by the missing negative class labels of PU learning. The time complexity and space complexity required for its solution are both. Therefore, the proposed method occupies less computing resources and can be widely applied to data sets of various sizes;

[0046] Key point 2: Considering both class prior estimation and PU learning, an iterative solution to the PU problem is proposed. This involves iteratively training the model and estimating the class prior using the GMM. Ideally, as the deep neural network's classification performance improves over time, its prediction scores should also become more reliable. As the prediction scores become more reliable, the GMM-estimated positive class prior will become more accurate, further improving the model's classification performance. The model training and the GMM-estimated class prior provide positive feedback to each other, a feature that can be exploited by the iterative framework.

[0047] Key point 3: The framework incorporates common semi-supervised learning techniques, including the Mean Teacher and Temperature Sharpening. Without these techniques, the class prior estimate is unstable, fluctuating with the number of training epochs and exhibiting large variance. With their introduction, the Mean Teacher ensures stable prediction scores by averaging historical parameters; while Temperature Sharpening encourages prediction scores to approach 0 or 1, making the bell-shaped curves corresponding to different classes more distinguishable and thus enhancing the GMM fit. These two techniques work together to stabilize the class prior estimate and effectively improve the algorithm's classification performance.

[0048] In order to make the above features and effects of the present invention more clearly understood, embodiments are given below and described in detail with reference to the accompanying drawings.

[0049] This paper proposes a new iterative deep PU learning and category prior estimation framework, the process architecture of which is shown in the attached Figure 1 As shown. First, obtain the training set. When used for malicious URL detection, the data obtained is usually a small number of malicious URLs and a large number of unlabeled URLs. When detecting false comments, the identified false comments often have significant features and can be regarded as positive samples, while the authenticity of unidentified comments is difficult to determine and should be regarded as unlabeled samples. When picking particles with cryo-electron microscopy, the selected particles are positive samples. However, due to the low signal-to-noise ratio of cryo-electron microscopy images and the diversity of negative distribution, the unselected areas may contain both positive and negative examples. The particles in such areas are unlabeled samples. Input the PU training set X into two student models f(., Θ) and teacher models f(·, Θ′) with the same network structure and different parameters (Θ′ can be obtained by performing an exponential moving average operation on Θ); S = f(X, Θ), S′ = f(·, Θ′) are the prediction scores of the student model and the teacher model respectively. The prediction score S′ of the teacher model is solved using Gaussian mixture modeling GMM to estimate the positive prior. Thus, we can use Calculating non-negative PU risk Combined with the average teacher's consistency loss and temperature sharpening losses Then calculate the optimization target Finally, the gradient back propagation algorithm is used to update the parameters Θ of the student model. The above steps are repeated until the optimization target is reached. Convergence. Next, the calculation process of each step is explained in detail.

[0050] For text-based tasks, such as malicious URL detection and fake review detection, recurrent neural networks can be used as student and teacher models. For image-based tasks, such as particle picking in cryo-electron microscopy, convolutional neural networks can be used. The model, confidence threshold τ, temperature T, and hyperparameters λ1 and λ2 can be freely configured for different applications. The prediction score can be viewed as the probability of the positive class given by the model. It is obtained by passing the sample through a deep neural network and a sigmoid transform.

[0051] (1) Calculate the prediction scores S, S′;

[0052] Given a triplet (x, y, z), x is the input feature, y is its class label, and z∈{1, 0} indicates the presence or absence of the class label. In PU learning, since the true y is unknown, the PU training set usually consists of several binary pairs (x, z). In addition, all labeled samples are positive, that is, Pr(y=1|z=1)=1. The sample set X=X l ∪X u , Xl is a positive sample subset with labels, X u is a subset of unlabeled samples. The parameters of the student model and the teacher model at time t are Θ t ,Θ′ t Indicates that the corresponding prediction functions are defined as f(·, Θ t ), f(·, Θ′ t ): Then, the prediction scores of the student model and the teacher model are:

[0053] S = sigmoid(f(X, Θ) t )),

[0054] S′=sigmoid(f(X,Θ′ t )).

[0055] in,

[0056] Use GMM to model S′ to obtain the positive class prior π p Estimated value of In this formula, · is any real number. There is no closed-form solution, so it is necessary to use the EM algorithm for iterative approximation (see the M step and parameter update equation). This step uses GMM to model S′ to estimate the positive class prior, which is one of the inventive points of the present invention. The technological progress it brings is that it is applicable to large-scale data sets in deep learning scenarios (previous methods are not applicable), and the time and space complexity of solving GMM is low. In addition, the unique technology of this application also includes a positive class prior estimation stabilizer (average teacher and temperature sharpening), which brings about a technological progress that the positive class prior estimation process is more accurate and stable. As well as the positive class prior estimation and the deep model are iteratively performed, the technological progress it brings is that the two promote each other, making the positive class prior estimation more accurate and the model classification performance better.

[0057] GMM is an unsupervised modeling method that models the predicted score S′ as follows:

[0058]

[0059]

[0060]

[0061] Among them, the category label y is a latent variable; π p is the mixing coefficient, which also represents the positive class prior; π n =1-π p ; Respectively represent the Gaussian distribution obeyed by the positive (negative) sample prediction score; Gaussian distribution μ, σ represent the mean and variance respectively; because the prediction scores of positive samples are generally greater than those of negative samples, μ n <μ p .

[0062] The Expectation Maximization (EM) algorithm is usually used to solve the GMM. represents the parameter value of GMM at time t, The parameter is Φ (t) Gaussian distribution, the EM algorithm performs (1) expectation (E) steps alternately and iteratively, that is, calculating s by The resulting conditional probability is:

[0063]

[0064] and (2) maximize (M) steps, that is:

[0065]

[0066]

[0067] Until convergence. According to the literature, Φ (t+1) The parameter update equation can be expanded into the following form:

[0068]

[0069]

[0070]

[0071]

[0072] (2) Calculate the optimization target

[0073] Step (2) obtains the estimated value of the positive class prior Thus, with the help of Calculating non-negative PU risk

[0074]

[0075]

[0076]

[0077] The confidence level c of the model output is defined as the maximum category probability of the sample obtained by the student model output, that is:

[0078] c = max(s, 1-s).

[0079] Using the confidence-based masking technique, the confidence threshold is set to τ, then the consistency loss of the output of the student model and the teacher model is for:

[0080]

[0081] Among them, x i ∈X represents the i-th sample in the sample set X, c i is its confidence, N = |X|; Is an indicator function, when the condition (·) is met, the function value is 1, otherwise it is 0. And the consistency loss It can stabilize the category prior estimation process and ultimately improve the classification performance of the deep model.

[0082] Given a predicted score s, a sharpening function is used to reduce its information entropy about the category distribution. Temperature sharpening achieves the above purpose by adjusting the temperature (Temperature, T) of the category distribution. The formula is as follows:

[0083]

[0084] Combined with the confidence-based masking technique mentioned in the average teacher, only the reliable output is temperature sharpened, so the loss for:

[0085]

[0086] In summary, the ultimate optimization goal is

[0087]

[0088] Among them, λ1 and λ2 are hyperparameters.

[0089] (3) Update model parameters Θ, Θ′

[0090] Use the gradient back propagation algorithm to update the parameters Θ of the student model:

[0091]

[0092] Among them, η represents the learning rate, which means Optimization goal Derivatives with respect to the model parameters Θ. Update the parameters Θ′ of the teacher model through EMA operation:

[0093] Θ′ t+1 =αΘ′ t +(1-α)Θ t+1 .

[0094] Where α is the smoothing coefficient, and its value is in [0, 1].

[0095] Iterate steps (1)-(4) until the optimization target is reached. Convergence. When the algorithm converges is the estimate of the final positive prior, and f(·, Θ) and f(·, Θ′) are the resulting classifiers.

[0096] The following is a system embodiment corresponding to the above method embodiment. This embodiment can be implemented in conjunction with the above embodiment. The relevant technical details mentioned in the above embodiment are still valid in this embodiment and will not be repeated here to reduce repetition. Accordingly, the relevant technical details mentioned in this embodiment can also be applied to the above embodiment.

[0097] The present invention also proposes a data classification system based on deep PU learning and category prior estimation, which includes:

[0098] The initialization module is used to obtain a training set including multiple data samples, in which only some data samples are labeled with category labels. The training set is simultaneously input into two student models and teacher models with the same network structure but different parameters. The student prediction score and teacher prediction score corresponding to each data sample output by the student model and the teacher model are obtained respectively.

[0099] The training module is used to input all teacher prediction scores into the Gaussian mixture model to obtain a positive class prior; construct a temperature sharpening loss based on all student prediction scores; construct a consistency loss based on all student prediction scores and teacher prediction scores; obtain a non-negative PU risk based on the positive class prior and all student prediction scores, combine the consistency loss, the non-negative PU risk and the temperature sharpening loss to obtain a target loss, and update the parameters of the student model based on the target loss using gradient backpropagation until the target converges or a preset number of iterations is reached, and save the current student model or teacher model as the data classification model;

[0100] The classification module is used to input the data to be classified into the data classification model to obtain the category of the data to be classified.

[0101] The data classification system based on deep PU learning and category prior estimation, when used for malicious URL detection, the data in the training set are some URLs marked with malicious categories and unlabeled URLs, and both the student model and the teacher model are recurrent neural networks; when used for false comment detection, the data in the training set are comments marked with false categories and unlabeled comments, and both the student model and the teacher model are recurrent neural networks; when used for particle picking in cryo-electron microscopy, the data in the training set are particle areas marked with selected categories and unlabeled particle areas, and both the student model and the teacher model are convolutional neural networks.

[0102] The data classification system based on deep PU learning and category prior estimation, wherein the training module is used to predict the scores of the student model and the teacher model based on the following formula:

[0103] S = sigmoid(f(X, Θ) t ))

[0104] S′=sigmoid(f(X,Θ′ t ))

[0105] in, X is the training set, S is the student prediction score output by the student model, S′ is the teacher prediction score output by the teacher model, Θ t is the parameter of the student model at time t, Θ′ t are the parameters of the teacher model at time t.

[0106] The data classification system based on deep PU learning and category prior estimation, wherein the consistency loss of the output of the student model and the teacher model for:

[0107]

[0108] Among them, x i ∈X represents the i-th sample in the training set X, c i is its confidence, N = |X|; is an indicator function, which takes the value of 1 when the condition (·) is met, otherwise it takes the value of 0; τ is the confidence threshold, Θ is the parameter of the student model, and Θ′ is the parameter of the teacher model.

[0109] The data classification system based on deep PU learning and category prior estimation, wherein the temperature sharpening loss for:

[0110]

[0111]

[0112] Where T is the temperature of the category distribution and s is the student prediction score output by the student model.

[0113] The present invention also proposes a storage medium for storing a program for executing any one of the data classification methods based on deep PU learning and category prior estimation.

[0114] The present invention also proposes a client for use in any of the above-mentioned data classification systems based on deep PU learning and category prior estimation.

Claims

1. A data classification method based on deep PU learning and category prior estimation, characterized in that: include: Step 1: Obtain a training set containing multiple data samples, in which only some data samples are labeled with category labels. Input the training set into two student models and teacher models with the same network structure but different parameters, and obtain the student prediction score and teacher prediction score corresponding to each data sample output by the student model and teacher model respectively; Step 2: Input all teacher prediction scores into the Gaussian mixture model to obtain a positive class prior; construct a temperature sharpening loss based on all student prediction scores; construct a consistency loss based on all student prediction scores and teacher prediction scores; obtain a non-negative PU risk based on the positive class prior and all student prediction scores, combine the consistency loss, the non-negative PU risk and the temperature sharpening loss to obtain a target loss, and based on the target loss, use gradient backpropagation to update the parameters of the student model until the target converges or reaches a preset number of iterations, and save the current student model or teacher model as the data classification model; Step 3: input the data to be classified into the data classification model to obtain the category of the data to be classified; Among them, when used for malicious URL detection, the data in the training set are some URLs marked with malicious categories and unlabeled URLs, and both the student model and the teacher model are recurrent neural networks; when used for false comment detection, the data in the training set are comments marked with false categories and unlabeled comments, and both the student model and the teacher model are recurrent neural networks; when used for particle picking in cryo-electron microscopy, the data in the training set are particle areas marked with selected categories and unlabeled particle areas, and both the student model and the teacher model are convolutional neural networks.

2. The data classification method based on deep PU learning and category prior estimation according to claim 1, characterized in that: This step 2 includes: The prediction scores of the student model and the teacher model are: S=sigmoid(f(X,Θ t )) S′=sigmoid(f(X,Θ′ t )) in, X is the training set, S is the student prediction score output by the student model, S' is the teacher prediction score output by the teacher model, Θ t is the parameter of the student model at time t, Θ' t are the parameters of the teacher model at time t.

3. The data classification method based on deep PU learning and category prior estimation according to claim 1, characterized in that: The consistency loss of the output of the student model and the teacher model for: Among them, x i ∈X represents the i-th sample in the training set X, c i is its confidence, N = |X|; is an indicator function, which takes the value of 1 when the condition (·) is met, otherwise it takes the value of 0; τ is the confidence threshold, Θ is the parameter of the student model, and Θ' is the parameter of the teacher model.

4. The data classification method based on deep PU learning and category prior estimation according to claim 3, characterized in that: The temperature sharpening loss for: Where T is the temperature of the category distribution and s is the student prediction score output by the student model.

5. A data classification system based on deep PU learning and category prior estimation, characterized in that: include: The initialization module is used to obtain a training set including multiple data samples, in which only some data samples are labeled with category labels. The training set is simultaneously input into two student models and teacher models with the same network structure but different parameters. The student prediction score and teacher prediction score corresponding to each data sample output by the student model and the teacher model are obtained respectively. The training module is used to input all teacher prediction scores into the Gaussian mixture model to obtain a positive class prior; construct a temperature sharpening loss based on all student prediction scores; construct a consistency loss based on all student prediction scores and teacher prediction scores; obtain a non-negative PU risk based on the positive class prior and all student prediction scores, combine the consistency loss, the non-negative PU risk and the temperature sharpening loss to obtain a target loss, and update the parameters of the student model based on the target loss using gradient backpropagation until the target converges or a preset number of iterations is reached, and save the current student model or teacher model as the data classification model; The classification module is used to input the data to be classified into the data classification model to obtain the category of the data to be classified Among them, when used for malicious URL detection, the data in the training set are some URLs marked with malicious categories and unlabeled URLs, and both the student model and the teacher model are recurrent neural networks; when used for false comment detection, the data in the training set are comments marked with false categories and unlabeled comments, and both the student model and the teacher model are recurrent neural networks; when used for particle picking in cryo-electron microscopy, the data in the training set are particle areas marked with selected categories and unlabeled particle areas, and both the student model and the teacher model are convolutional neural networks.

6. The data classification system based on deep PU learning and category prior estimation according to claim 5, characterized in that: The training module is used to predict the scores of the student model and the teacher model based on the following formula: S=sigmoid(f(X,Θ t )) S′=sigmoid(f(X,Θ′ t )) in, X is the training set, S is the student prediction score output by the student model, S' is the teacher prediction score output by the teacher model, Θ t is the parameter of the student model at time t, Θ' t are the parameters of the teacher model at time t.

7. The data classification system based on deep PU learning and category prior estimation according to claim 5, characterized in that The consistency loss of the output of the student model and the teacher model for: Among them, x i ∈X represents the i-th sample in the training set X, c i is its confidence, N = |X|; is an indicator function, which takes the value of 1 when the condition (·) is met, otherwise it takes the value of 0; τ is the confidence threshold, Θ is the parameter of the student model, and Θ' is the parameter of the teacher model.

8. The data classification system based on deep PU learning and category prior estimation according to claim 7, characterized in that: The temperature sharpening loss for: Where T is the temperature of the category distribution and s is the student prediction score output by the student model.

9. A storage medium for storing a program for executing any one of the data classification methods based on deep PU learning and category prior estimation as claimed in claims 1 to 4.

10. A client, used in any one of the data classification systems based on deep PU learning and category prior estimation according to claims 5 to 8.

Images