A method and device for screening user samples

Abstract

The embodiments of this specification provide a method and device for screening user samples. The first training set contains multiple user samples, and each user sample has multiple sets of risk labels. For any first user sample, the loss values ​​corresponding to the multiple sets of risk labels of the first user sample are obtained, and the multiple loss values ​​are obtained from the training process of the first neural network. Then, the first value is deducted from the first fusion loss corresponding to the multiple loss values ​​of the first user sample to obtain a corrected loss. Among them, the first value is determined using the distribution characteristics of multiple loss values, which is used to reflect the increased loss value caused by the error contained in the first neural network. When the corrected losses of several user samples are obtained, the user samples and their risk labels are screened from the first training set based on the corrected losses, and added to the third training set. The user samples contain privacy data, and privacy protection needs to be performed during the data processing process.

Description

Technical Field

[0001] One or more embodiments of the present specification relate to the field of computer technology, and more particularly, to a method and device for screening user samples. Background Art

[0002] With the development of science and technology, service platforms provide more and more service capabilities to users. Service platforms can provide users with a variety of services with the help of various network models trained by themselves. At the same time, service platforms also need to use risk control models to identify high-risk users, thereby ensuring the security of services. The quality of user samples and labels will directly affect the accuracy of risk control models. User samples contain private data, which needs to be protected during data processing. In order to improve the accuracy of risk control models, a large number of high-quality user samples and labels are required. User samples can be labeled by machine or manually. In practice, multiple rules may be used to label user samples, or multiple people may be hired to label the same set of data sets, so that multiple sets of labels for user samples can be obtained. Multiple sets of labels usually contain noise. Due to the existence of noise, the performance of risk control models trained directly using multiple sets of labels is usually difficult to meet the requirements.

[0003] Therefore, it is hoped that there will be an improved solution that can more effectively utilize multiple sets of labels to improve the accuracy of identifying high-risk users. Summary of the invention

[0004] One or more embodiments of this specification describe a method and device for screening user samples to more effectively utilize multiple sets of labels and improve the accuracy of identifying high-risk users. The specific technical solution is as follows.

[0005] In a first aspect, an embodiment provides a user sample screening method, comprising:

[0006] For any first user sample in the first training set, obtaining loss values ​​corresponding to multiple sets of risk labels of the first user sample respectively; the multiple loss values ​​are obtained from the training process of the first neural network;

[0007] Deducting a first value from a first fusion loss corresponding to a plurality of loss values ​​of the first user sample to obtain a modified loss; the first value is determined using distribution characteristics of the plurality of loss values ​​to reflect an increased loss value caused by an error contained in the first neural network;

[0008] By using the correction losses corresponding to a number of user samples, user samples and their risk labels are selected from the first training set and added into the third training set.

[0009] In one embodiment, the first fusion loss is determined in the following manner:

[0010] Smoothing the multiple loss values ​​of the first user sample respectively to obtain multiple processed loss values;

[0011] The first fusion loss is determined using the multiple loss values ​​after smoothing.

[0012] In one implementation, the first value is determined in the following manner:

[0013] Based on the distribution characteristics of the multiple loss values ​​of the first user sample, an upper limit value of the difference between the first fusion loss and an ideal second fusion loss is determined; the second fusion loss is determined using the ideal multiple loss values, and the ideal multiple loss values ​​are obtained in the process of training a first neural network without errors using the first user sample;

[0014] The first value is determined based on the upper bound value.

[0015] In one embodiment, the step of determining an upper limit of the difference between the first fusion loss and the ideal second fusion loss comprises:

[0016] Obtaining multiple loss values ​​of the first user sample in multiple training cycles;

[0017] Based on the distribution characteristics of the multiple loss values, an upper limit value of the difference is determined.

[0018] In one embodiment, the step of screening user samples and their risk labels from the first training set includes:

[0019] Selecting a set of risk labels from the multiple sets of risk labels of the first user sample;

[0020] A preset proportion of user samples with a smaller correction loss are selected from the plurality of user samples, and the selected user samples and a set of risk labels selected therefrom are added to the third training set.

[0021] In one embodiment, the step of selecting a set of risk labels from the multiple sets of risk labels of the first user sample includes:

[0022] The risk label corresponding to the minimum loss value among the multiple loss values ​​of the first user sample is used as a set of selected risk labels.

[0023] In one embodiment, the first neural network includes a representation layer and a plurality of prediction layers corresponding to the plurality of sets of risk labels respectively; the plurality of loss values ​​are determined based on the differences between the plurality of sets of risk labels and the prediction values ​​of the plurality of prediction layers respectively.

[0024] In one embodiment, the third training set is used to train a second neural network; the second neural network includes a representation layer and a prediction layer; the representation layer of the second neural network shares model parameters with the representation layer of the first neural network.

[0025] In a second aspect, the embodiment provides a user sample screening device, comprising:

[0026] A loss acquisition module is configured to acquire, for any first user sample in the first training set, loss values ​​corresponding to a plurality of risk labels of the first user sample; the plurality of loss values ​​are obtained from a training process of the first neural network;

[0027] a loss correction module, configured to deduct a first value from a first fusion loss corresponding to a plurality of loss values ​​of the first user sample to obtain a corrected loss; the first value is determined using a distribution characteristic of the plurality of loss values ​​to reflect an increased loss value caused by an error contained in the first neural network;

[0028] The sample screening module is configured to use the correction losses corresponding to a plurality of user samples to screen user samples and their risk labels from the first training set and add them to the third training set.

[0029] In a third aspect, an embodiment provides a computer-readable storage medium having a computer program stored thereon, which, when executed in a computer, causes the computer to execute any one of the methods described in the first aspect.

[0030] In a fourth aspect, an embodiment provides a computing device, comprising a memory and a processor, wherein the memory stores executable code, and when the processor executes the executable code, the method described in any one of the first aspects is implemented.

[0031] In the method and device provided in the embodiments of this specification, by deducting the first value from the first fusion loss of the first user sample, the increased loss value caused by the error contained in the first neural network can be removed from the first fusion loss, so that the corrected loss can more accurately represent the predicted loss of the first user sample, and screening multiple user samples based on the corrected loss can improve the effectiveness of the screened user samples, thereby more effectively utilizing multiple sets of labels, improving the effectiveness of the risk control model, and further improving the accuracy of identifying high-risk users. BRIEF DESCRIPTION OF THE DRAWINGS

[0032] In order to more clearly illustrate the technical solutions of the embodiments of the present invention, the following briefly introduces the drawings required for use in the description of the embodiments. Obviously, the drawings described below are only some embodiments of the present invention, and for ordinary technicians in this field, other drawings can be obtained based on these drawings without creative work.

[0033] Figure 1 A schematic diagram of an implementation scenario of an embodiment disclosed in this specification;

[0034] Figure 2 A flowchart of a user sample screening method provided in an embodiment;

[0035] Figure 3 is a schematic diagram of an overall architecture of this embodiment;

[0036] Figure 4 A schematic block diagram of a user sample screening device provided in an embodiment. DETAILED DESCRIPTION

[0037] The solution provided in this specification is described below in conjunction with the accompanying drawings.

[0038] Figure 1 This is a schematic diagram of an implementation scenario of an embodiment disclosed in this specification. It includes a computing device and a risk control platform. The training set contains multiple user samples, and each user sample has multiple sets of risk control labels. The computing device is used to screen the user samples in the training set, obtain multiple user samples and a set of risk control labels, and send the screened samples to the risk control platform. The risk control platform uses the screened user samples and a set of labels for each user sample to train the risk control model, and uses the trained risk control model to identify risky users. The risk control platform can provide risk control for the service platform. Figure 1 This is just an example of an implementation scenario. In actual applications, the risk control platform can also perform sample screening tasks on computing devices.

[0039] User samples contain user features, which may include user attributes, user behavior features, and user relationship features. Any user sample has multiple sets of risk labels, which are used to indicate the risk of the user sample in a specific task. Specific tasks may be, for example, loan issuance, financial business processing, etc. The risk label may include classification labels and regression labels. For ease of understanding, the following takes classification labels and binary classification as examples for explanation. Each risk label includes two categories: high risk and low risk, and 0 represents low risk and 1 represents high risk. There are erroneous labels, i.e., noise labels, in the multiple risk labels of multiple user samples, but the number of noise labels is relatively small. Inaccurate labels refer to the errors in labels caused by insufficient data information or lack of knowledge of labelers during the process of labeling by multiple labelers. Multiple annotators refer to multiple individuals or sources that label user samples, such as multiple rules.

[0040] When using user samples with multiple sets of risk labels, you can train the corresponding neural network for each set of risk labels, learn multiple risk labels at the same time, and aggregate the scores of multiple neural networks. Although this method uses all labels, it does not refine the user samples and does not discard the noisy parts. Alternatively, you can also get a set of risk labels from multiple risk labels by aggregating labels, and use this set of risk labels to train the neural network. This method also lacks sufficient refinement of the data, and when some risk labels are inaccurate, they may aggregate to the wrong labels.

[0041] In order to more effectively utilize multiple sets of labels and improve the accuracy of identifying high-risk users, an embodiment of this specification provides a method for screening user samples. Sample Selection refers to screening multiple user samples in a training set by formulating some screening criteria, screening out user samples with correct risk labels, and eliminating user samples with incorrect risk labels. This method deducts the increased loss value caused by the error contained in the first neural network from the first fusion loss, which can correct the first fusion loss, so that the corrected loss obtained more accurately reflects the correctness of multiple sets of risk labels of the first user sample, so the corrected loss can be used as a screening criterion to screen multiple user samples. In this way, user samples that would have been discarded can have the opportunity to be learned, thereby improving the utilization effect of multiple sets of labels.

[0042] Combine the following Figure 2 This embodiment is described in detail.

[0043] Figure 2A flowchart of a user sample screening method provided in an embodiment. The method can be executed by a computing device. The computing device can be implemented by any device, equipment, platform, device cluster, etc. with computing and processing capabilities. The method includes the following steps.

[0044] Step S210: for any first user sample x in the first training set D i , get the first user sample x i The R sets of risk labels correspond to loss values ​​respectively. The R loss values ​​are obtained from the training process of the first neural network NN1, and R is an integer greater than 1.

[0045] The first training set D contains N user samples, each of which has multiple risk labels. The multiple risk labels are labels given by multiple people who meet the standards, or labels obtained based on multiple rules. The first training set D is the original training set, from which user samples with more accurate labels are selected.

[0046] The first training set D is represented by the following formula:

[0047] D = {x i ,y i} i＝1 N (1)

[0048] Among them, x i is the sample feature of the i-th user sample, and the i-th user sample is taken as any first user sample. i is the first user sample x i The risk label vector, which contains R sets of binary risk labels, can be expanded as:

[0049] y i ={y i r} r＝1 R (2)

[0050] The first neural network NN1 has model parameters Θ, which can be expressed as a neural network f Θ Neural network is a data network model that simulates the human brain's neural thinking mode. There are many types of neural networks, such as feedforward neural networks, convolutional neural networks, recurrent neural networks, and multi-layer perceptrons (MLP).

[0051] The first neural network NN1 learns the first user sample x i and its corresponding risk label (x i ,y ir ), the loss value of the risk label can be calculated using, but is not limited to, the cross entropy loss function:

[0052] L(f Θ (x i ), y i r )= y i r log(f Θ (x i ))+(1-y i r )log(1- f Θ (x i )) (3)

[0053] Among them, f Θ (x i ) is the first user sample x i The predicted value or score obtained after the sample features of the input into the first neural network NN1, y i r It is the rth set of risk labels.

[0054] In one implementation, the first neural network NN1 can be designed using a Multi Tower-based structure. That is, the first neural network NN1 includes a representation layer and R prediction layers corresponding to R sets of risk labels. Multiple prediction layers can share one representation layer. The representation layer is used to extract deep user representations based on user feature outputs, and input the deep user representations into the prediction layer. The prediction layer is used to perform classification based on the deep user representations to obtain prediction values.

[0055] R loss values ​​may be determined based on the differences between the R sets of risk labels and the prediction values ​​of the R prediction layers, respectively.

[0056] Figure 3 Schematic diagram of the overall architecture of this embodiment. Taking R=3 as an example, each user sample x has three risk labels y1, y2 and y3. The values ​​of y1, y2 and y3 are 0 or 1, 1 represents high risk and 0 represents low risk. Figure 3 The three sets of risk labels for the three user samples are listed as examples. The first neural network NN1 has a representation layer and three prediction layers tower1, tower2 and tower3. The first neural network NN1 is trained using the first training set D. For example, the first user sample x i The sample features are input into the first neural network NN1, and the prediction values ​​​​score output by the three prediction layers can be obtained 1 、score 2 and score 3Based on the predicted value score 1 The difference between y1 and y1 can determine the loss value L1, based on the predicted value score 2 The difference between y2 and y2 can determine the loss value L2, based on the predicted value score 3 The difference between y and y3 can determine the loss value L3. In this way, the first user sample x is obtained. i 3 loss values.

[0057] In the risk control scenario, the number of high-risk users is significantly smaller than that of low-risk users. In order to avoid the estimation bias caused by the imbalance of the positive and negative sample ratios and improve the performance of the trained model, the conventional cross entropy loss function can be weighted by the sample ratio. The weighted cross entropy loss function is

[0058]

[0059] Among them, L BCE is the aggregated loss function. r is the ratio of the number of negative samples to positive samples in the rth set of risk labels, which is used to balance positive and negative samples. For the rth set of risk labels, assuming that the number of negative samples in N user samples is N0 r , the number of positive samples is N1 r , then w r =N0 r / N1 r The first user sample x i The R loss values ​​can also be obtained based on formula (4), that is, the loss values ​​obtained by weighting the sample proportion.

[0060] Step S220: from the first user sample x i The first fusion loss μ corresponding to the R loss values i Deduct the first value d from the equation to get the corrected loss C i .

[0061] The first value d is determined using the distribution characteristics of the R loss values, and is used to reflect the increased loss value caused by the error contained in the first neural network N1. Since the first neural network N1 is trained using the first training set D with R sets of risk labels, and the first training set D contains label noise, the first neural network N1 contains errors (that is, inaccuracies) caused by these label noises. Such errors will cause the loss value to increase, and the increased loss value can be reflected by the first value.

[0062] The following describes the methods for determining the first fusion loss and the first value respectively.

[0063] Here, a method for determining the first fusion loss is described. In one implementation, for any first user sample xi , we can directly take the average of its R loss values ​​to get the first fusion loss μ i , expressed as

[0064]

[0065] Alternatively, the first fusion loss μ i It can also be obtained by taking a weighted average of R loss values.

[0066] In one implementation, in order to avoid normal samples being discarded as much as possible, the first user sample x i The multiple loss values ​​of are smoothed respectively to obtain multiple loss values ​​after processing, and the first fusion loss is determined by using the multiple loss values ​​after smoothing. i φ To express.

[0067] The purpose of smoothing multiple loss values ​​is to reduce those relatively large values. Generally, user samples with small loss values ​​are more likely to be considered as samples with accurate labels. User samples with small loss values ​​are more likely to be selected, and samples with large loss values ​​are more likely to be eliminated. Therefore, smoothing the loss values ​​can avoid normal samples from being discarded to a certain extent.

[0068] When smoothing the loss value, a non-decreasing function such as a Taylor polynomial or a truncation function can be used. The following uses the Taylor expansion as an example to illustrate the smoothing process. Specifically, the following smoothing function obtained from the Taylor polynomial can be used to smooth the R loss values:

[0069] φ(L)=log(1+L+L 2 / 2) (6)

[0070] The smoothed loss values ​​are averaged using the following formula to obtain the following first fusion loss μ i φ :

[0071]

[0072] The method for determining the first value is described below. The first value can be determined by following steps 1 and 2.

[0073] Step 1: Based on the first user sample x i The distribution characteristics of the R loss values ​​determine the first fusion loss μ i With the ideal second fusion loss ρ i The distribution characteristics may include statistical characteristics such as variance and mean square error.

[0074] When determining the above lower limit value, it can be determined based on the above distribution characteristics and the number R of risk labels.

[0075] The second fusion loss ρ i The ideal R loss values ​​are determined using the first user sample x. i The second fusion loss ρ is obtained by training the first neural network NN1 without errors. That is, in an ideal case, the first neural network NN1 is trained with user samples without noise, which does not contain errors. The ideal R loss values ​​are used to theoretically determine the second fusion loss ρ i The process of using R loss values ​​to determine the first fusion loss μ i The process is the same and will not be repeated here.

[0076] Ideal R loss values ​​and ideal second fusion loss ρ i It is an ideal value, not a value actually calculated, but a value that exists in theoretical derivation due to the need for theoretical analysis.

[0077] Theoretically, the second fusion loss ρ i is the mean of multiple ideal loss values ​​that cannot be observed, and the first fusion loss μ i is the observed value. Based on the observed value and the distribution characteristics of the R loss values, the first fusion loss μ can be derived i and the second fusion loss ρ i The range of the difference between the two, which includes the upper and lower bounds. The specific derivation process can refer to the existing technology and will not be described in detail here. Here we give the first fusion loss μ i φ and the second fusion loss ρ i The upper limit of the difference between

[0078]

[0079] The right side of the above inequality is the upper bound of the difference. is the variance of the R loss values, μ i φ is the first fusion loss determined by the R loss values ​​after smoothing. Similarly, the first fusion loss μ can be obtained i and the second fusion loss ρ i The upper limit of the difference between them is relatively more complicated and will not be described in detail here.

[0080] Step 2: Determine the first value based on the upper limit value.

[0081] Specifically, the upper limit value may be directly used as the first value, or the upper limit value may be used as the first value after a preset transformation, both of which are feasible.

[0082] Transform the above formula (8) to obtain

[0083]

[0084] Formula (9) is the second fusion loss ρ i The lower bound value is also from the first fusion loss μ i φ The corrected loss is the ideal second fusion loss ρ i The minimum value that can be achieved with high probability is used as the standard for screening user samples C i ,Right now:

[0085]

[0086] In practical applications, the upper limit value can be made smaller to obtain the first value. For example, the upper limit value is subtracted from a preset value or reduced by a certain proportion to obtain the first value. In other words, the correction loss C i It can be based on the ideal second fusion loss ρ i For example, we can take the ideal second fusion loss ρ i The lower bound of the second fusion loss ρ can also be taken as i The lower bound of is greater than a certain value. Correction loss C i can be the ideal first fusion loss ρ i The lower bound of is close to the value.

[0087] Using this corrected loss C i As a screening criterion, it can give as many samples as possible that are discarded due to problems in tower predictions in some prediction layers a chance to be learned.

[0088] In one embodiment, to improve accuracy, step 1 determines the first fusion loss μ i φ With the ideal second fusion loss ρ i When the lower bound of the difference between , multiple training cycles (epochs) can be used. Specifically, it may include obtaining the first user sample x in T training cycles. i A plurality of loss values ​​are obtained, and a lower bound value of the difference is determined based on the distribution characteristics of the plurality of loss values.

[0089] In each training cycle, all user samples in the first training set D are used to train the first neural network NN1. In each training cycle, the first user sample x iCorresponding R loss values. In T training cycles, n = RT loss values ​​can be obtained. The lower bound of the difference can be determined by using the distribution characteristics of n loss values. For example, R on the right side of equation (8) can be replaced by n. In this way, the corrected loss C i It can be expressed as

[0090]

[0091] Where n = RT, T is the number of training cycles, R is the number of risk labels, is the variance obtained using RT loss values.

[0092] Step S230: using the correction loss C corresponding to a plurality of user samples i , filter user samples and their risk labels from the first training set D and add them to the third training set D*.

[0093] Generally speaking, the ideal loss value of a user sample can represent the accuracy of the label of the user sample. The smaller the ideal loss value, the higher the label accuracy. i To a certain extent, it is close to the ideal loss value, which can reflect the possibility of wrong labels in multiple sets of risk labels of corresponding user samples. Corrected loss C i The smaller it is, the less likely it is that there are wrong labels in the multiple risk labels of the user sample. Therefore, the correction loss C can be screened out from the first training set D. i The smaller user samples are the user samples without wrong labels in multiple sets of risk labels, and these user samples are added to the third training set D*.

[0094] At the same time, it is also necessary to select a set of risk labels with higher credibility from multiple sets of risk labels for each user sample and add them to the third training set D*.

[0095] In one implementation, step S230 of screening user samples and their risk labels from the first training set D may be performed using steps 3 and 4 below.

[0096] Step 3: From the first user sample x i Select a set of risk labels from multiple sets of risk labels.

[0097] Step 4: Select the corrected loss C from N user samples i A smaller preset proportion p% of user samples, the selected user samples and their selected set of risk labels are added to the third training set D*.

[0098] In this embodiment, the step of selecting a set of risk labels in step 3 and the step of selecting the modified loss C in step 4 are iThe steps for a smaller preset percentage of user samples are executed in no particular order.

[0099] In step 3, from the first user sample x i When selecting a set of risk labels from the R sets of risk labels, multiple selection methods can be used. For example, voting can be used to select the first user sample x. i The risk labels with more occurrences are selected from the R sets of risk labels as a set of risk labels with greater credibility.

[0100] Alternatively, the first user sample x i The risk label corresponding to the minimum loss value among the R loss values ​​is selected as a set of risk labels with greater credibility.

[0101] For samples with different risk labels, we can obtain a set of risk labels y through step 3. i *, for samples with R sets of risk labels that are completely consistent, we naturally also get a set of risk labels y i *. Through the above method, the original first data set D with R sets of risk labels is merged into only one set of risk labels y i *The second data set D0, where D0=(x i ,y i *).

[0102] When step 4 is executed, user samples with a preset proportion p% having a smaller correction loss Ci can be selected from the N user samples included in the second training set D0, and the selected user samples and their set of risk labels are added to the third training set D*. p% is the screening proportion, which can be set as needed.

[0103] The third training set D* is a data set with a smaller proportion of error labels obtained through screening by the embodiments of this specification, and can be used to train the second neural network NN2. The second neural network NN2 can be used to identify risky users. In one embodiment, the second neural network NN2 includes a representation layer and a prediction layer, and the representation layer of the second neural network NN2 and the representation layer of the first neural network NN1 can share model parameters.

[0104] During the training process, the first neural network NN1 and the second neural network NN2 can be trained jointly, and the model training process can be carried out jointly with the user sample selection process. Figure 3 As shown in Figure 3, R takes 3 as an example, takes a batch of user samples and their 3 sets of risk labels from the first training set D, inputs them into the first neural network NN1 for model training, and obtains 3 prediction values ​​score 1 、score 2 and score 3Using these three predicted values, we can get the first fusion loss. By deducting the first value from the first fusion loss, we can get the corrected loss C. i At the same time, a set of risk labels with higher credibility is selected from the three sets of risk labels, and the corrected loss C of each user sample is used. i , p% of the user samples with smaller correction loss are selected from the batch to train the second neural network NN2. Then the next batch of sample training is performed.

[0105] The above training mode and screening mode are only examples and do not constitute a limitation on the embodiments of this specification. In practical applications, a more appropriate mode can be selected from a variety of implementation modes.

[0106] In the technical solution of this embodiment, the information of multiple sets of risk labels is comprehensively utilized, and the underlying representation of the model is modeled and trained to enhance the underlying representation capability. At the same time, the embodiment uses the balance function to help alleviate the problem of outliers in the loss value, explores and utilizes the lower bound of the ideal loss mean theory in multi-label modeling, and uses this as a standard for sample screening.

[0107] This embodiment screens user samples based on the loss value of the user samples. Generally, the smaller the loss value, the smaller the difference between the predicted value of the neural network model and the label value, and the smaller the probability that the label value is an incorrect label.

[0108] If the R loss values ​​of each user sample are used and the first fusion loss obtained by formula (5) is directly used as the standard to screen reliable samples, the error in the first neural network NN1 may cause the loss value to increase, that is, the first fusion loss contains a large error value. Using the first fusion loss to screen multiple user samples (selecting some user samples with smaller first fusion losses), some normal user samples may be discarded and cannot continue to participate in model training. Similar to the common practice of selecting samples with small loss in a single set of labels, this type of outliers may cause some normal samples to be discarded. By correcting the first fusion loss and removing the increased loss caused by the error of the first neural network, the correctness of the screening criteria can be greatly improved, thereby avoiding discarding normal samples.

[0109] In this specification, the word "first" in terms such as first user sample, first fusion loss, first value and first neural network, as well as the corresponding "second" (if any) in the text, are merely for the convenience of distinction and description and do not have any limiting meaning.

[0110] The foregoing describes certain embodiments of the present specification, and other embodiments are within the scope of the appended claims. In some cases, the actions or steps described in the claims may be performed in an order different from that in the embodiments, and the desired results may still be achieved. In addition, the processes depicted in the accompanying drawings do not necessarily have to be performed in the specific order or sequential order shown to achieve the desired results. In some embodiments, multitasking and parallel processing are also possible or may be advantageous.

[0111] Figure 4 A schematic block diagram of a user sample screening device provided in an embodiment. Figure 2 The method embodiment shown corresponds to the embodiment. The device 400 is deployed in a computing device. The device 400 includes: a loss acquisition module 410, a loss correction module 420 and a sample screening module 430. Among them, the loss acquisition module 410 is configured to obtain the loss values ​​corresponding to multiple sets of risk labels of the first user sample for any first user sample in the first training set, and the multiple loss values ​​are obtained from the training process of the first neural network. The loss correction module 420 is configured to deduct the first value from the first fusion loss corresponding to the multiple loss values ​​of the first user sample to obtain the corrected loss. The first value is determined using the distribution characteristics of multiple loss values, and is used to reflect the increased loss value caused by the error contained in the first neural network. The sample screening module 430 is configured to use the corrected losses corresponding to several user samples to screen user samples and their risk labels from the first training set, and add them to the third training set.

[0112] In one embodiment, the device 400 also includes a loss fusion module 440, which is configured to determine the first fusion loss in the following manner: smoothing multiple loss values ​​of the first user sample respectively to obtain multiple loss values ​​after processing; and determining the first fusion loss using the multiple loss values ​​after smoothing.

[0113] In one embodiment, the device 400 further includes a first value determination module 450, which is configured to determine the first value, specifically including: a lower limit value determination submodule 451 and a first value determination submodule 452. The lower limit value determination submodule 451 is configured to determine the upper limit value of the difference between the first fusion loss and the ideal second fusion loss based on the distribution characteristics of multiple loss values ​​of the first user sample. The second fusion loss is determined using multiple ideal loss values, and the multiple ideal loss values ​​are obtained in the process of training the first neural network that does not contain errors using the first user sample. The first value determination submodule 452 is configured to determine the first value based on the upper limit value.

[0114] In one implementation, the lower limit value determination submodule 451 is specifically configured to: obtain multiple loss values ​​of the first user sample in multiple training cycles, and determine the upper limit value of the difference based on the distribution characteristics of the multiple loss values.

[0115] In one embodiment, the sample screening module 430 includes: a label screening submodule 431 and a sample screening submodule 432. The label screening submodule 431 is configured to select a set of risk labels from multiple sets of risk labels of the first user sample. The sample screening submodule 432 is configured to select a preset proportion of user samples with a smaller correction loss from a plurality of user samples, and add the selected user samples and the selected set of risk labels to the third training set.

[0116] In one embodiment, the label screening submodule 431 is specifically configured to use the risk label corresponding to the minimum loss value among multiple loss values ​​of the first user sample as a set of selected risk labels.

[0117] In one embodiment, the first neural network includes a representation layer and multiple prediction layers corresponding to multiple sets of risk labels respectively; multiple loss values ​​are determined based on the differences between the multiple sets of risk labels and the predicted values ​​of the multiple prediction layers respectively.

[0118] In one embodiment, the third training set is used to train a second neural network; the second neural network includes a representation layer and a prediction layer; the representation layer of the second neural network shares model parameters with the representation layer of the first neural network.

[0119] The above device embodiments correspond to the method embodiments. For specific descriptions, please refer to the description of the method embodiments, which will not be repeated here. The device embodiments are obtained based on the corresponding method embodiments and have the same technical effects as the corresponding method embodiments. For specific descriptions, please refer to the corresponding method embodiments.

[0120] The present specification also provides a computer-readable storage medium having a computer program stored thereon. When the computer program is executed in a computer, the computer is caused to execute Figures 1 to 3 Any of the methods described above.

[0121] The embodiment of the present specification also provides a computing device, including a memory and a processor, wherein the memory stores an executable code, and when the processor executes the executable code, Figures 1 to 3 Any of the methods described above.

[0122] Each embodiment in this specification is described in a progressive manner, and the same or similar parts between the embodiments can be referred to each other, and each embodiment focuses on the differences from other embodiments. In particular, for the storage medium and computing device embodiments, since they are basically similar to the method embodiments, the description is relatively simple, and the relevant parts can be referred to the partial description of the method embodiments.

[0123] Those skilled in the art should be aware that in one or more of the above examples, the functions described in the embodiments of the present invention may be implemented using hardware, software, firmware, or any combination thereof. When implemented using software, these functions may be stored in a computer-readable medium or transmitted as one or more instructions or codes on a computer-readable medium.

[0124] The specific implementation methods described above further describe the purpose, technical solutions and beneficial effects of the embodiments of the present invention in detail. It should be understood that the above description is only a specific implementation method of the embodiments of the present invention and is not intended to limit the scope of protection of the present invention. Any modification, equivalent replacement, improvement, etc. made on the basis of the technical solution of the present invention shall be included in the scope of protection of the present invention.

Claims

1. A user sample screening method, comprising: For any first user sample in the first training set, obtaining loss values ​​corresponding to multiple sets of risk labels of the first user sample; A plurality of loss values ​​are obtained from the training process of the first neural network; Deducting a first value from a first fusion loss corresponding to a plurality of loss values ​​of the first user sample to obtain a modified loss; the first value is determined using distribution characteristics of the plurality of loss values ​​to reflect an increased loss value caused by an error contained in the first neural network; By using the correction losses corresponding to a number of user samples, user samples and their risk labels are selected from the first training set and added into the third training set.

2. According to the method of claim 1, the first fusion loss is determined in the following manner: Smoothing the multiple loss values ​​of the first user sample respectively to obtain multiple processed loss values; The first fusion loss is determined using the multiple loss values ​​after smoothing.

3. The method according to claim 1, wherein the first value is determined in the following manner: Based on the distribution characteristics of the multiple loss values ​​of the first user sample, an upper limit value of the difference between the first fusion loss and an ideal second fusion loss is determined; the second fusion loss is determined using the ideal multiple loss values, and the ideal multiple loss values ​​are obtained in the process of training a first neural network without errors using the first user sample; The first value is determined based on the upper bound value.

4. The method according to claim 3, wherein the step of determining an upper limit of the difference between the first fusion loss and the ideal second fusion loss comprises: Obtaining multiple loss values ​​of the first user sample in multiple training cycles; Based on the distribution characteristics of the multiple loss values, an upper limit value of the difference is determined.

5. According to the method of claim 1, the step of screening user samples and their risk labels from the first training set comprises: Selecting a set of risk labels from the multiple sets of risk labels of the first user sample; A preset proportion of user samples with a smaller correction loss are selected from the plurality of user samples, and the selected user samples and a set of risk labels selected therefrom are added to the third training set.

6. The method according to claim 5, wherein the step of selecting a set of risk labels from the plurality of risk labels of the first user sample comprises: The risk label corresponding to the minimum loss value among the multiple loss values ​​of the first user sample is used as a set of selected risk labels.

7. The method according to claim 1, wherein the first neural network comprises a representation layer and a plurality of prediction layers corresponding to the plurality of risk labels respectively; The plurality of loss values ​​are determined based on differences between the plurality of sets of risk labels and the predicted values ​​of the plurality of prediction layers, respectively.

8. According to the method of claim 7, the third training set is used to train a second neural network; the second neural network includes a representation layer and a prediction layer; the representation layer of the second neural network shares model parameters with the representation layer of the first neural network.

9. A user sample screening device, comprising: A loss acquisition module, configured to acquire, for any first user sample in the first training set, loss values ​​corresponding to multiple sets of risk labels of the first user sample; A plurality of loss values ​​are obtained from the training process of the first neural network; a loss correction module, configured to deduct a first value from a first fusion loss corresponding to a plurality of loss values ​​of the first user sample to obtain a corrected loss; the first value is determined using a distribution characteristic of the plurality of loss values ​​to reflect an increased loss value caused by an error contained in the first neural network; The sample screening module is configured to use the correction losses corresponding to a plurality of user samples to screen user samples and their risk labels from the first training set and add them to the third training set.

10. A computer-readable storage medium having a computer program stored thereon, which, when executed in a computer, causes the computer to execute the method according to any one of claims 1 to 8.

11. A computing device, comprising a memory and a processor, wherein the memory stores executable code, and when the processor executes the executable code, the method according to any one of claims 1 to 8 is implemented.

Images