Model-based training method and device for screening out noise samples, equipment and medium

Abstract

The application discloses a noise sample screening method and device based on model training, equipment and medium, belonging to the field of data processing. The method comprises the following steps: performing multiple model training by using a training sample set and a verification sample set to obtain multiple target models; after each model training, determining noise samples according to actual labels and prediction values of samples in a target sample set, and screening the noise samples from the target sample set until a noise sample screening stop condition is met; each model training comprises the following steps: training a first model by using a newly updated training sample set; based on the newly updated target sample set and a comparison result of the first model and a second model, iteratively updating model parameters of the first model and model parameters of the second model until the comparison result of the first model and the second model meets an iteration stop condition. According to the embodiment of the application, the purity of the sample set used for model training can be improved.

Description

TECHNICAL FIELD

[0001] The application belongs to the field of data processing, and particularly relates to a noise sample screening method and device based on model training, equipment and medium. BACKGROUND

[0002] With the continuous development of electronic computing technology, data modeling is applied in more and more fields. In the process of data modeling, samples are needed to train the model, so as to train a model that can accurately process data. However, a part of the samples with mislabeled labels in the samples used for training the model are noise samples. The existence of noise samples makes the purity of the sample set used for model training lower. Using such a sample set to participate in model training will make the model learn incorrectly, resulting in a poor accuracy of the trained model and being difficult to achieve the expectation. SUMMARY

[0003] The embodiments of the application provide a noise sample screening method and device based on model training, equipment and medium, which can improve the purity of the sample set used for model training.

[0004] In a first aspect, the embodiments of the application provide a noise sample screening method based on model training, comprising: performing multiple model training by using a training sample set and a verification sample set to obtain multiple target models, the training sample set comprising multiple samples, and the verification sample set comprising multiple samples; after each model training, determining noise samples in at least part of the samples in the target sample set according to actual labels and predicted values of the at least part of the samples, and screening the noise samples from the target sample set until a noise sample screening stop condition is met, the target sample set comprising the training sample set and / or the verification sample set, and the predicted value being a value predicted by a target model obtained by model training according to an input sample;

[0005] In each model training, the first model is trained by using the most recently updated training sample set; the model parameters of the first model and the model parameters of the second model are iteratively updated based on the most recently updated target sample set and the comparison result of the first model and the second model until the comparison result of the first model and the second model meets an iteration stop condition, and the iteratively updated second model is determined as the target model, the second model being the first model before training.

[0006] In a second aspect, an embodiment of the present application provides a noise sample screening device based on model training, comprising: a training module configured to perform multiple model training by using a training sample set and a verification sample set, to obtain multiple target models, the training sample set comprising multiple samples, and the verification sample set comprising multiple samples; and a noise screening module configured to, after each model training, determine noise samples in at least part of samples in a target sample set according to actual labels and predicted values of the at least part of samples, and screen the noise samples from the target sample set until a noise sample screening stop condition is met, the target sample set comprising the training sample set and / or the verification sample set, and the predicted value being a value predicted by a target model obtained through model training according to an input sample.

[0007] The training module comprises: a training unit configured to train a first model by using a newly updated training sample set; and an updating unit configured to iteratively update model parameters of the first model and model parameters of a second model based on a newly updated target sample set and a comparison result of the first model and the second model until the comparison result of the first model and the second model meets an iteration stop condition, and determine the second model after the iteration as the target model, the second model being the first model before the training.

[0008] In a third aspect, an embodiment of the present application provides a noise sample screening device based on model training, comprising: a processor and a memory storing computer program instructions; and the processor is configured to implement the noise sample screening method based on model training of the first aspect when executing the computer program instructions.

[0009] In a fourth aspect, an embodiment of the present application provides a computer readable storage medium, the computer readable storage medium storing computer program instructions, and the computer program instructions are configured to implement the noise sample screening method based on model training of the first aspect when executed by a processor.

[0010] The embodiment of the present application provides a noise sample screening method, device, equipment and medium based on model training, which can utilize a training sample set and a verification sample set to participate in multiple model training. In each model training, the training sample set is utilized to train a first model, and according to a comparison result of the first model and a second model, model parameters of the first model and model parameters of the second model are iteratively updated. The second model is the first model before training. The second model can supervise and guide the training of the first model, avoid error learning and overfitting phenomenon, and ensure the accuracy of the obtained target model. After each model training, part of samples can be input into the obtained target model to obtain prediction values corresponding to the samples. According to the real labels and the prediction values of the samples, whether the samples are noise samples can be judged. In the case that the samples are noise samples, the samples are screened from the training sample set or the verification sample set. After each model training, part of noise samples can be screened, so that the training sample set and the verification sample set with higher accuracy can be obtained, that is, the purity of the sample set used for model training is improved, and then the accuracy of the model trained by using the training sample set and the verification sample set after screening noise samples can be improved. BRIEF DESCRIPTION OF DRAWINGS

[0011] In order to more clearly illustrate the technical solutions of the embodiments of the present application, the drawings needed to be used in the embodiments of the present application will be briefly introduced. Those skilled in the art can obtain other drawings according to these drawings without creating any creative labor.

[0012] Figure 1 The flowchart of the noise sample screening method based on model training provided by an embodiment of the present application is provided.

[0013] Figure 2 The flowchart of the model training process in an embodiment of the present application is provided.

[0014] Figure 3 The flowchart of the model training process in another embodiment of the present application is provided.

[0015] Figure 4 The flowchart of the noise sample screening method based on model training provided by another embodiment of the present application is provided.

[0016] Figure 5 The logic diagram of an example of the noise sample screening method based on model training provided by the embodiment of the present application is provided.

[0017] Figure 6 The flowchart of an example of the noise sample screening process based on model training provided by the embodiment of the present application is provided.

[0018] Figure 7A structural schematic diagram of a noise sample screening device based on model training provided by an embodiment of the present application is provided.

[0019] Figure 8 A structural schematic diagram of a noise sample screening device based on model training provided by an embodiment of the present application is provided. DETAILED DESCRIPTION

[0020] The features and exemplary embodiments of various aspects of the present application will be described in detail below with reference to the accompanying drawings and specific embodiments. It should be understood that the specific embodiments described herein are intended to explain the present application, but not to limit the present application. The present application can be implemented without some of the specific details by those skilled in the art. The following description of the embodiments is merely intended to provide a better understanding of the present application by showing examples of the present application.

[0021] With the continuous development of electronic computing technology, data modeling is applied in more and more fields. In the process of data modeling, samples are needed to train the model, so as to train a model that can accurately process data. However, some of the samples used to train the model contain mislabeled labels, which are noise samples. The existence of noise samples makes the purity of the sample set used for model training lower. Using such a sample set to participate in model training will make the model learn incorrectly, resulting in a model with poor accuracy and difficult to achieve the expected result.

[0022] The present application provides a noise sample screening method, device, equipment and medium based on model training, which can use a training sample set and a verification sample set to participate in multiple model training. After each model training, part of the samples can be input into the target model obtained by training to obtain the predicted value corresponding to the sample. Using the true label of the sample and the predicted value, it can be judged whether the sample is a noise sample. In the case of a noise sample, the sample is screened out from the training sample set or the verification sample set. A part of the noise samples can be screened out after each model training, so as to obtain a training sample set and a verification sample set with higher accuracy. The model trained by using the training sample set and the verification sample set after screening out the noise samples has higher accuracy.

[0023] The noise sample screening method, device, equipment and medium based on model training provided by the present application are described below.

[0024] The first aspect of the present application provides a noise sample screening method based on model training, which can be applied to a model training scene and executed by a noise sample screening device, equipment, etc., which is not limited herein. Figure 1A flowchart of a method for screening noise samples based on model training provided by an embodiment of the present application is shown in FIG. 1. The method for screening noise samples based on model training can include steps S101 and S102. Figure 1

[0025] In step S101, multiple model training is performed using the training sample set and the validation sample set to obtain multiple target models.

[0026] The training sample set includes multiple samples. The validation sample set includes multiple samples. Model training can include both training and validation. The samples in the training sample set are used to train the model. The samples in the validation sample set are used to validate the model.

[0027] In an embodiment of the present application, the number of model training is more than twice. Each model training can obtain a target model. Correspondingly, more than twice model training using the training sample set and the validation sample set can obtain more than two target models. The target model is the final model obtained by this model training.

[0028] In step S102, after each model training, the actual label and the predicted value of at least part of the samples in the target sample set are determined to determine the noise samples in at least part of the samples, and the noise samples are screened out from the target sample set until the noise sample screening stop condition is met.

[0029] The target sample set includes the training sample set and / or the validation sample set, which can be selected according to the requirement. For example, if it is required to screen out the noise samples in the training sample set, the training sample set can be selected as the target sample set; if it is required to screen out the noise samples in the validation sample set, the validation sample set can be selected as the target sample set; if it is required to screen out the noise samples in the training sample set and the validation sample set, the training sample set and the validation sample set can be selected as the target sample set.

[0030] ​In some examples, all samples in the target sample set can be used to filter out noise samples after each model training. In other examples, part of the samples in the target sample set can be selected to filter out noise samples after each model training. It should be noted that the training sample set and the validation sample set used in the first model training are the original training sample set and the original validation sample set, and the training sample set and the validation sample set used in the model training after the first model training are the training sample set and the validation sample set obtained after the noise samples are filtered out in the last model training. Similarly, the target sample set used in the filtering out of the noise samples after the first model training includes the original training sample set and the original validation sample set, and the target sample set used in the filtering out of the noise samples after the model training after the first model training includes the training sample set and the validation sample set obtained after the noise samples are filtered out in the last model training.

[0031] The predicted value is a value predicted by the target model obtained through the model training according to the input sample. The predicted label is a label output by the target model obtained through the model training according to the input sample. Whether a sample is a noise sample can be determined based on a comparison result of the direct label and the predicted label of the sample. If the sample is a noise sample, the sample is filtered out from the target sample set to improve the accuracy of the training sample set and / or the validation sample set. The training sample set and the validation sample set obtained after the noise samples are filtered out in the last model training can be used in the current model training, that is, the current model training is performed by using the training sample set and the validation sample set updated most recently.

[0032] The noise sample filtering out stop condition is a condition for stopping the filtering out of the noise samples, which can be set according to scenarios, requirements, experience, etc., and is not limited herein. The noise sample filtering out stop condition can be related to one or more of the number of times of model training, the number of noise samples determined, the validation effect of the target model, etc. When the noise sample filtering out stop condition is met, the training sample set and the validation sample set obtained after the current noise filtering out can be used as the training sample set and the validation sample set suitable for model training, or the training sample set and the validation sample set obtained after the last noise filtering out can be used as the training sample set and the validation sample set suitable for model training.

[0033] In some examples, the noise sample filtering out stop condition includes one or more of the following:

[0034] The number of times of model training reaches a preset number threshold;

[0035] The number of noise samples filtered out from the training sample set this time satisfies a first condition or a second condition, the first condition being that the number of noise samples filtered out from the training sample set this time is less than a first preset number threshold, and the second condition being that the proportion of the number of noise samples filtered out from the training sample set this time in the number of samples in the training sample set is less than a first preset ratio threshold.

[0036] The number of noise samples filtered out from the verification sample set this time satisfies a third condition or a fourth condition, the third condition being that the number of noise samples filtered out from the verification sample set this time is less than a second preset number threshold, and the fourth condition being that the proportion of the number of noise samples filtered out from the verification sample set this time in the number of samples in the verification sample set is less than a second preset ratio threshold.

[0037] The number of noise samples filtered out from the target sample set this time satisfies a fifth condition or a sixth condition, the fifth condition being that the number of noise samples filtered out from the target sample set this time is less than a third preset number threshold, and the sixth condition being that the proportion of the number of noise samples filtered out from the target sample set this time in the number of samples in the target sample set is less than a third preset ratio threshold.

[0038] According to the actual label and the prediction value of at least part of the samples in the target sample set, no noise sample is found;

[0039] The first model effect parameter is less than the second model effect parameter, the first model effect parameter being a model effect parameter obtained by verifying the target model trained this time by using the verification sample set, and the second model effect parameter being a model effect parameter obtained by verifying the target model trained last time by using the verification sample set.

[0040] For example, the noise sample filtering-out cutoff condition includes that the number of model training reaches a preset number threshold, which can be set according to the scene, demand, experience, etc., and is not limited herein. If the preset number threshold is 3, after the third time of model training, according to the actual label and the prediction value of at least part of the samples in the target sample set, the noise sample is determined and filtered out from the target sample set. The training sample set and the verification sample set obtained after the noise sample is filtered out this time are the final training sample set and the verification sample set suitable for model training, and after that, no noise sample is found by using model training.

[0041] For example, the noise sample screening cut-off condition includes that the number of noise samples screened out from the training sample set this time meets a first condition. The first preset number threshold can be set according to the scene, demand, experience, etc., and is not limited herein. Assuming that the first preset number threshold is set to 4, after the model is trained this time, if the number of noise samples determined in the training sample set according to the actual label and the predicted value of at least part of the samples in the target sample set is 3, the 3 noise samples are screened out from the training sample set, and the training sample set and the verification sample set obtained after the noise sample screening this time are the final training sample set and the verification sample set suitable for model training, after which, noise samples are no longer searched for by using the model.

[0042] For example, the noise sample screening cut-off condition includes that the number of noise samples screened out from the training sample set this time meets a second condition. The first preset ratio threshold can be set according to the scene, demand, experience, etc., and is not limited herein. Assuming that the first preset ratio threshold is set to 5%, the most recently updated training sample set used for training the model this time includes 100 samples, if the number of noise samples determined in the training sample set according to the actual label and the predicted value of at least part of the samples in the target sample set is 3, the 3 noise samples are screened out from the training sample set, and the training sample set and the verification sample set obtained after the noise sample screening this time are the final training sample set and the verification sample set suitable for model training, after which, noise samples are no longer searched for by using the model.

[0043] The specific content of the noise sample screening cut-off condition that the number of noise samples screened out from the verification sample set this time meets a third condition or a fourth condition can be referred to the above related description of the number of noise samples screened out from the training sample set this time meeting a first condition or a second condition, with the difference being that the noise samples are screened out from the verification sample set in the present example. The second preset number threshold can be set according to the scene, demand, experience, etc., and is not limited herein. The second preset ratio threshold can be set according to the scene, demand, experience, etc., and is not limited herein.

[0044] The specific content of the noise sample screening cut-off condition that the number of noise samples screened out from the target sample set this time meets a fifth condition or a sixth condition can be referred to the above related description of the number of noise samples screened out from the training sample set this time meeting a first condition or a second condition, with the difference being that the noise samples are screened out from the target sample set in the present example. The third preset number threshold can be set according to the scene, demand, experience, etc., and is not limited herein. The third preset ratio threshold can be set according to the scene, demand, experience, etc., and is not limited herein.

[0045] For example, the noise sample screening cut-off condition comprises that no noise sample is found according to the actual label and the prediction value of at least part of the samples in the target sample set. After the current model training, if no noise sample is found according to the actual label and the prediction value of at least part of the samples in the target sample set, the training sample set and the verification sample set obtained after the last noise sample screening are taken as the final training sample set and the final verification sample set suitable for model training, and after that, no noise sample is found by using model training.

[0046] For example, the noise sample screening cut-off condition comprises that the first model effect parameter is less than the second model effect parameter. After obtaining the target model after each model training, the target model can be verified by using the verification sample set to obtain the model effect parameter of the target model. The model effect parameter can represent the effect of the target model. The larger the model effect parameter is, the better the effect of the target model is. For example, the model effect parameter can be the overall classification accuracy (i.e. Accuracy), the precision (i.e. Precision), the recall (i.e. recall), the F value, the ROC (i.e. Receiver Operating Characteristic), and the like, which are not limited herein. If the first model effect parameter is greater than or equal to the second model effect parameter, the next model training needs to be performed, and the noise sample is continuously found and screened after the next model training. If the first model effect parameter is less than the second model effect parameter, the training sample set and the verification sample set obtained after the last noise sample screening are taken as the final training sample set and the final verification sample set suitable for model training, and after that, no noise sample is found by using model training.

[0047] In order to facilitate understanding, the process of any model training is described below. Figure 2 The flowchart of the model training process in an embodiment of the present application is shown in FIG. 2. Figure 2 As shown in FIG. 2, the model training process can comprise steps S201 to S202.

[0048] In step S201, the first model is trained by using the most recently updated training sample set.

[0049] For the first model training, the most recently updated training sample set is the original training sample set. For the i-th model training, the most recently updated training sample set is the training sample set obtained after the noise sample screening after the (i-1)-th model training, and i is an integer greater than 1.

[0050] In step S202, based on the latest updated target sample set and the comparison result of the first model and the second model, the model parameters of the first model and the model parameters of the second model are iteratively updated until the comparison result of the first model and the second model meets the iteration stopping condition, and the second model updated iteratively is determined as the target model.

[0051] The second model is the first model before training. The training of the first model can be supervised by using the second model to avoid the phenomenon of error learning or overfitting in model training. In order to facilitate understanding, an example is described below. The initial first model is M train0 , the first model after training by using the latest updated training sample set is M train1 , the second model at this time is M guide1 , M guide1 is the same as M train0 ; the first model M train1 and the second model M guide1 are compared by using the updated target sample set, and the model parameters of the first model and the model parameters of the second model are updated according to the comparison result, that is, the updated first model M train2 and the updated second model M guide2 are obtained. In this way, the first model and the second model are iteratively updated until the comparison result meets the iteration stopping condition, and the second model updated iteratively is determined as the target model.

[0052] The second model, that is, the first model before training and the first model after training, are compared and iteratively updated for multiple times, which can prevent the phenomenon of error learning or overfitting in each model training process, and continuously optimize the second model to improve the accuracy of model training.

[0053] The iteration stopping condition is a stopping condition for iteratively updating the model parameters of the first model and the model parameters of the second model, which can be set according to the scene, demand, experience, etc., and is not limited herein. In some examples, the iteration stopping condition can include that the consistency loss function value of the first model and the second model reaches convergence. The consistency loss function value of the first model and the second model reaching convergence indicates that the first model and the second model reach a relatively optimal consistent state, and no error learning or overfitting phenomenon occurs. After the iteration stopping condition is met, the current model training is ended, and the second model updated iteratively is output as the target model.

[0054] In some examples, the initial model used in different model training is different. Specifically, the first model in the i-th model training is different from the first model in the j-th model training, and the second model in the i-th model training is different from the second model in the j-th model training, i and j are positive integers, and i≠j.

[0055] The initial model used in different model training is different, and the target model obtained by different model training is also different. The more different the initial model used in different model training, the stronger the comprehensiveness of the noise samples determined by the different target models, the larger the range of the noise samples covered, the better the noise samples can be filtered out, and the purity of the training sample set and / or the verification sample set is further improved.

[0056] In some examples, the number of at least part of the samples used to determine the noise samples in each model training can be the same.

[0057] In some examples, the number of at least part of the samples used to determine the noise samples in different model training can be different. For example, the number of at least part of the samples used to determine the noise samples in the i-th model training is greater than the number of at least part of the samples used to determine the noise samples in the (i-1)-th model training, i is a positive integer greater than 1. As the number of model training increases, the purity of the training sample set and the verification sample set also increases, and the accuracy of the target model obtained by the model training also increases. As the number of model training increases, the number of at least part of the samples used to determine the noise samples can be gradually increased, the possibility of normal samples being misjudged as noise samples can be reduced, and the purity of the training sample set and the verification sample set can be further improved.

[0058] In the embodiments of the present application, the training sample set and the verification sample set can be used for multiple model training. In each model training, the first model is trained using the training sample set, and the model parameters of the first model and the model parameters of the second model are iteratively updated according to the comparison result of the first model and the second model. The second model is the first model before training. The second model can be used to supervise and guide the training of the first model, avoid error learning and overfitting, and ensure the accuracy of the target model. After each model training, part of the samples can be input into the target model obtained by training to obtain the predicted value corresponding to the sample. Using the true label and the predicted value of the sample, it can be determined whether the sample is a noise sample. In the case that the sample is a noise sample, the sample is filtered out from the training sample set or the verification sample set. A part of the noise samples can be filtered out after each model training, so that the training sample set and the verification sample set with higher accuracy can be obtained, that is, the purity of the sample set used for model training is improved, and the accuracy of the model trained by using the training sample set and the verification sample set after filtering out the noise samples is further improved.

[0059] In some embodiments, the comparison result in the above embodiments includes a consistency loss function value. The consistency loss function value of the first model and the second model can be used to supervise the iteratively updated first model by the iteratively updated second model, and avoid error learning and overfitting. Figure 3 The flowchart of the model training process in another embodiment of the present application is shown in FIG.Figure 3 Different from Figure 2 , Figure 2 Step S202 in Figure 3 may be specifically refined as steps S2021 to S2024 in

[0060] In step S2021, the samples in the recently updated target sample set are respectively input into the first model and the second model to obtain the predicted values of the samples output by the first model and the predicted values of the samples output by the second model.

[0061] The same samples are input into the first model and the second model respectively, and since the first model and the second model are different, the predicted values of the samples output by the first model and the predicted values of the samples output by the second model are different.

[0062] In step S2022, based on the actual labels of the samples, the predicted values of the samples output by the first model and the predicted values of the samples output by the second model, a consistency loss function value of the first model and the second model is obtained.

[0063] According to the actual labels of the samples and the predicted values of the samples output by the first model, a prediction error parameter of the first model can be calculated, which can represent the error between the prediction of the first model on the samples and the actual situation of the samples.

[0064] According to the actual labels of the samples and the predicted values of the samples output by the second model, a prediction error parameter of the second model can be calculated, which can represent the error between the prediction of the second model on the samples and the actual situation of the samples.

[0065] Based on the prediction error parameter of the first model and the prediction error parameter of the second model, a consistency loss function value of the first model and the second model can be calculated. For example, the consistency loss function value of the first model and the second model can be calculated according to the following formula (1):

[0066] Loss =‖l train -l guide ‖ 2 (1)

[0067] Wherein, Loss is the consistency loss function value of the first model and the second model; l train is the prediction error parameter of the first model; l guide is the prediction error parameter of the second model; ‖l train -l guide ‖ is the L1 norm of the difference between l train and l guide .

[0068] In step S2023, the model parameters of the first model and the model parameters of the second model are updated according to the consistency loss function value, to obtain an updated first model and an updated second model.

[0069] The model parameters of the first model can be updated according to the consistency loss function, and the updated first model is obtained according to the updated model parameters of the first model. The model parameters of the second model can be updated according to the updated model parameters of the first model, and the updated second model is obtained according to the updated model parameters of the second model.

[0070] Specifically, the first model parameters can be calculated based on the consistency loss function value, and the model parameters of the first model are updated to the first model parameters. The first learning rate control coefficient is obtained, and the second model parameters are calculated according to the first learning rate control coefficient, the first model parameters and the model parameters of the second model updated in the last iteration, and the model parameters of the second model are updated to the second model parameters. Based on the first model parameters and the second model parameters, the updated first model and the updated second model are obtained.

[0071] The first model parameters can minimize the consistency loss function value, that is, the first model parameters are the model parameters of the first model that can minimize the consistency loss function value. The first learning rate control parameter can control the learning rate of the model parameter update of the second model, and the first learning rate control parameter is greater than or equal to 0 and less than or equal to 1. Specifically, it can be set according to the scene, demand, experience, etc., which is not limited here. For example, the first model parameters can be calculated according to the following formula (2), and the second model parameters can be calculated according to the following formula (3):

[0072] θ train = argmin(Loss) (2)

[0073] θ guide,t = α1·θ guide,t-1 +(1-α1)·θ train,t (3)

[0074] Wherein, θ train is the first model parameter; θ guide,t is the second model parameter of the second model updated in this iteration (i.e. the tth iteration); Loss is the consistency loss function value; argmin() is the function of the minimum value function; α1 is the first learning rate control coefficient; θ guide,t-1 is the model parameter of the second model updated in the last iteration (i.e. the (t-1)th iteration); θ train,t is the first model parameter of the first model updated in this iteration.

[0075] In step S2024, the samples in the latest updated target sample set are input into the updated first model and the updated second model respectively again to obtain a consistency loss function value of the updated first model and the updated second model, until the consistency loss function value meets an iteration stopping condition, and the second model updated this time is determined as the target model.

[0076] Step S2024 is equivalent to re-executing steps S2021 to S2023 until the consistency loss function value meets the iteration stopping condition, and the second model updated this time is determined as the target model.

[0077] In each model training, the first model and the second model are iteratively updated through the consistency loss function, and the model parameters of the second model are optimized according to the weight represented by the first learning rate control coefficient, using the model parameters of the first model, to ensure the consistency of the first model and the second model, and further ensure the accuracy of the target model.

[0078] In some embodiments, in the process of determining the noise sample, the prediction value of the sample in the target model of the last model training and the prediction value of the sample in the target model of the current model training are learned, the prediction value of the sample in the target model of the current model training is updated, and the updated prediction value of the sample is used to determine whether the sample is a noise sample. Figure 4 The flowchart of the noise sample screening method based on model training provided by another embodiment of the present application is as follows, Figure 4 Different from Figure 1 The difference between steps S102 in Figure 1 Steps S1021 to S1023 in Figure 4

[0079] In step S1021, after each model training, a second learning rate control coefficient is obtained, a first prediction value is calculated according to the second learning rate control coefficient, the prediction value of the sample in the target model of the current model training and the prediction value of the sample in the target model of the last model training, and the prediction value of the sample in the target model of the current model training is updated to the first prediction value.

[0080] The second learning rate control parameter can control the learning rate of the prediction value of the sample in the target model of the current model training, and the second learning rate control parameter is greater than or equal to 0 and less than or equal to 1, which can be set according to the scene, demand, experience, etc., and is not limited herein. The sample is input into the target model of the current model training, and the target model can output the prediction value of the current time. Similarly, after the last model training, the sample is input into the target model of the last model training, and the target model can output the prediction value of the last time. In some examples, the first prediction value can be calculated according to the following formula (4) ​

[0081] ep i = a2 ep i-1 + (1 - a2) ep i (4)

[0082] wherein, ep i is the prediction value of the sample in the target model of the present (i.e., the i-th) model training; a2 is the second learning rate control parameter; ep i-1 is the prediction value of the sample in the target model of the last (i.e., the (i-1)-th) model training.

[0083] In step S1022, the prediction label of the sample in the target model of the present model training is obtained according to the first prediction value.

[0084] Both the actual label and the prediction label are identifiers capable of representing the classification result, which can be represented by numbers or other characters, and are not limited herein.

[0085] In some examples, the prediction label can be obtained according to the first prediction value and a preset judgment threshold. For example, if the first prediction value is greater than or equal to the judgment threshold, the prediction label is 1; and if the first prediction value is less than the judgment threshold, the prediction label is 0.

[0086] In other examples, the first prediction value can be substituted into a preset function to obtain the corresponding prediction label. For example, the prediction label can be calculated according to the following formula (5):

[0087] y' = softmax(ep i ) (5)

[0088] wherein, y' is the prediction label of the sample in the target model of the present model training; ep i is the first prediction value; and softmax() is a normalized exponential function.

[0089] It should be noted that if the present model training is the first model training, the prediction value of the sample in the target model of the present model training in step S1021 does not need to be updated, and the prediction label of the sample in the target model of the present model training can be directly obtained by using the prediction value of the target model of the present model training.

[0090] In step S1023, in the case that the actual label of the sample is different from the prediction label of the sample in the target model of the present model training, the sample is determined as a noise sample, and the noise sample is excluded from the target sample set until a noise sample exclusion cutoff condition is met.

[0091] The actual label of the sample is different from the predicted label of the sample in the target model of the current model training, indicating that the prediction result of the target model on the sample is inconsistent with the actual label of the sample, and the label of the sample may be a wrong label. Therefore, the sample is determined as a noise sample.

[0092] The target model obtained by training is used to assist in determining the noise sample, so that the screening of the noise sample is closely combined with the model training process, and the noise sample can be identified while the model is trained, thereby improving the efficiency of model training and noise sample screening.

[0093] For ease of understanding, the noise sample screening cutoff condition includes an example in which the first model effect parameter is less than the second model effect parameter, and the noise sample screening method based on model training provided by the embodiments of the present application is described.

[0094] Figure 5 A logic diagram of an example of the noise sample screening method based on model training provided by the embodiments of the present application is shown in FIG. 1. Figure 5 As shown in FIG. 1, in the iterative training process of the noise sample screening, a training sample set and a validation sample set are involved. The training sample set includes noise samples and non-noise samples. The validation sample set also includes noise samples and non-noise samples. The training sample set can participate in the model ensemble learning process. The model ensemble learning process includes a first model and a second model. The second model is a supervised model of the first model, and the two models are iteratively updated. The consistency loss function value between the first model and the second model can be used to determine when to stop the iterative update. When the iterative update is stopped, the second model obtained by the iterative update is output as a target model. The validation sample set can be used to verify the effect of the target model. If the model effect of the target model obtained this time is better than or equal to the model effect of the target model obtained last time, the next model training is performed; if the model effect of the target model obtained this time is worse than the model effect of the target model obtained last time, the next model training is not performed. The target model can be used to predict the samples in the training sample set and the samples in the validation sample set, so as to obtain the noise samples according to the predicted values of the samples and the actual labels of the samples. The noise sample screening function is used to screen the noise samples from the training sample set and the validation sample set. After the noise samples are screened, the training sample set can continue to participate in the next model training, and the validation sample set can participate in the effect verification of the target model obtained by the next model training.

[0095] Figure 6 A flowchart of an example of the noise sample screening process based on model training provided by the embodiments of the present application is shown in FIG. 2. Figure 6 As shown in FIG. 2, the noise sample screening process based on model training can include steps S301 to S310.

[0096] In step S301, the first model is trained by using the training sample set.

[0097] In step S302, at least part of the samples in the target sample set are input into the first model and the second model, to obtain the predicted value of the sample by the first model and the predicted value of the sample by the second model.

[0098] In step S303, based on the actual value of the sample, the predicted value of the sample by the first model and the predicted value of the sample by the second model, the consistency loss function value of the first model and the second model is obtained.

[0099] In step S304, the model parameters of the first model and the model parameters of the second model are updated by using the consistency loss function value, to obtain the updated first model and second model, and steps S302 to S304 are repeated until the consistency loss function value of the first model and the second model reaches convergence.

[0100] In step S305, when the consistency loss function value of the first model and the second model reaches convergence, the most recently updated second model is output as the target model.

[0101] In step S306, at least part of the samples in the target sample set are input into the target model, to obtain the predicted value of the sample output by the target model, and the predicted label of the sample is obtained according to the predicted value of the sample.

[0102] In step S307, in the case where the predicted label of the sample is different from the actual label of the sample, the sample is determined as a noise sample, and the noise sample is excluded from the target sample set.

[0103] In step S308, the effect of the target model is verified by using the verification sample set, to obtain the model effect parameter of this time.

[0104] In step S309, it is judged whether the model effect parameter of this time is greater than or equal to the model effect parameter of the last time, if the model effect parameter of this time is greater than or equal to the model effect parameter of the last time, step S301 is returned to be executed, but it should be noted that the first model in step 301 is a new first model, which is different from the first model used this time, if the model effect parameter of this time is less than the model effect parameter of the last time, step S310 is executed.

[0105] In step S310, the noise sample exclusion process is ended.

[0106] The specific contents of steps S301 to S310 can be referred to the related description in the above embodiments, which will not be repeated here.

[0107] The second aspect of the present application provides a noise sample exclusion device based on model training.Figure 7 A structural schematic diagram of a noise sample screening device based on model training is provided for an embodiment of the present application. As shown in the figure, the noise sample screening device 400 based on model training can include a training module 401 and a noise screening module 402, wherein the training module 401 can include a training unit 4011 and an updating unit 4012. Figure 7

[0108] The training module 401 can be used to perform multiple model training by using a training sample set and a verification sample set, and obtain multiple target models.

[0109] The training sample set includes multiple samples. The verification sample set includes multiple samples.

[0110] The noise screening module 402 can be used to determine noise samples in at least part of the samples in the target sample set according to the actual labels and predicted values of the at least part of the samples in the target sample set after each model training, and screen out the noise samples from the target sample set until the noise sample screening stop condition is met.

[0111] The target sample set includes the training sample set and / or the verification sample set. The predicted value is a value predicted by the target model according to the input sample.

[0112] The training unit 4021 can be used to train the first model by using the recently updated training sample set.

[0113] The updating unit 4022 can be used to iteratively update the model parameters of the first model and the model parameters of the second model based on the recently updated target sample set and the comparison result of the first model and the second model until the comparison result of the first model and the second model meets the iteration stop condition, and determine the iteratively updated second model as the target model.

[0114] The second model is the first model before training.

[0115] In some examples, the number of at least part of the samples used to determine the noise samples in the i-th model training is greater than the number of at least part of the samples used to determine the noise samples in the i-1-th model training, i is a positive integer greater than 1.

[0116] In some examples, the first model in the i-th model training is different from the first model in the j-th model training, i and j are positive integers, and i≠j.

[0117] ​In the embodiments of the present application, the training sample set and the verification sample set can be used to participate in multiple model training. In each model training, the first model is trained by using the training sample set, and the model parameters of the first model and the model parameters of the second model are iteratively updated according to the comparison result of the first model and the second model. The second model is the first model before training. The training of the first model can be supervised and guided by using the second model to avoid the phenomenon of error learning and overfitting, and to ensure the accuracy of the obtained target model. After each model training, part of the samples can be input into the obtained target model to obtain the prediction value corresponding to the sample. By using the real label and the prediction value of the sample, it can be judged whether the sample is a noise sample. In the case that the sample is a noise sample, the sample is excluded from the training sample set or the verification sample set. A part of noise samples can be excluded after each model training, so that a training sample set and a verification sample set with higher accuracy can be obtained, that is, the purity of the sample set used for model training is improved, and the accuracy of the model trained by using the training sample set and the verification sample set after excluding noise samples is improved.

[0118] In some embodiments, the comparison result includes a consistency loss function value.

[0119] The updating unit 4022 can be configured to: input the samples in the recently updated target sample set into the first model and the second model respectively to obtain the prediction values of the samples output by the first model and the prediction values of the samples output by the second model; obtain a consistency loss function value of the first model and the second model based on the actual labels of the samples, the prediction values of the samples output by the first model and the prediction values of the samples output by the second model; update the model parameters of the first model and the model parameters of the second model according to the consistency loss function value to obtain an updated first model and an updated second model; input the samples in the recently updated target sample set into the updated first model and the updated second model again to obtain a consistency loss function value of the updated first model and the updated second model, until the consistency loss function value meets an iteration stopping condition, and determine the iteratively updated second model as the target model.

[0120] In some embodiments, the updating unit 4022 can be configured to: obtain the first model parameters based on the consistency loss function value, and update the model parameters of the first model to the first model parameters, the first model parameters making the consistency loss function value minimum; obtain a first learning rate control coefficient, and obtain the second model parameters based on the first learning rate control coefficient, the first model parameters and the model parameters of the second model obtained in the last iteration update, and update the model parameters of the second model to the second model parameters; and obtain the updated first model and the updated second model based on the first model parameters and the second model parameters.

[0121] In some embodiments, the iteration stopping condition comprises that the consistency loss function value reaches convergence.

[0122] In some embodiments, the noise screening module 402 can be configured to obtain a second learning rate control coefficient, calculate a first prediction value according to the second learning rate control coefficient, the prediction value of the sample in the target model in the current model training and the prediction value of the sample in the target model in the last model training, and update the prediction value of the sample in the target model in the current model training to the first prediction value; obtain a prediction label of the sample in the target model in the current model training according to the first prediction value; and determine the sample as a noise sample in a case where the actual label of the sample is different from the prediction label of the sample in the target model in the current model training.

[0123] In some embodiments, the noise sample screening stopping condition comprises one or more of the following:

[0124] the number of times of model training reaches a preset number threshold;

[0125] the number of noise samples screened out from the training sample set in the current time satisfies a first condition or a second condition, the first condition being that the number of noise samples screened out from the training sample set in the current time is less than a first preset number threshold, and the second condition being that the proportion of the number of noise samples screened out from the training sample set in the current time in the number of samples in the training sample set is less than a first preset ratio threshold;

[0126] the number of noise samples screened out from the verification sample set in the current time satisfies a third condition or a fourth condition, the third condition being that the number of noise samples screened out from the verification sample set in the current time is less than a second preset number threshold, and the fourth condition being that the proportion of the number of noise samples screened out from the verification sample set in the current time in the number of samples in the verification sample set is less than a second preset ratio threshold;

[0127] the number of noise samples screened out from the target sample set in the current time satisfies a fifth condition or a sixth condition, the fifth condition being that the number of noise samples screened out from the target sample set in the current time is less than a third preset number threshold, and the sixth condition being that the proportion of the number of noise samples screened out from the target sample set in the current time in the number of samples in the target sample set is less than a third preset ratio threshold;

[0128] no noise sample is found according to the actual label and the prediction value of at least part of the samples in the target sample set;

[0129] the first model effect parameter is less than the second model effect parameter, the first model effect parameter being a model effect parameter obtained by verifying the target model after the current model training by using the verification sample set, and the second model effect parameter being a model effect parameter obtained by verifying the target model after the last model training by using the verification sample set.

[0130] The third aspect of the present application further provides a noise sample screening device based on model training. Figure 8 A structural schematic diagram of a noise sample screening device based on model training is provided for an embodiment of the present application. As shown in Figure 8 The noise sample screening device based on model training 500 includes a memory 501, a processor 502, and a computer program stored in the memory 501 and executable on the processor 502.

[0131] In one example, the processor 502 can include a central processing unit (CPU), or an application specific integrated circuit (ASIC), or one or more integrated circuits configured to implement one or more embodiments of the present application.

[0132] The memory 501 can include a read-only memory (ROM), a random access memory (RAM), a magnetic disk storage medium device, an optical storage medium device, a flash memory device, an electrical, optical, or other physical / tangible memory storage device. Therefore, generally, the memory includes one or more tangible (non-transitory) computer-readable storage media (e.g., memory devices) encoded with software including computer-executable instructions and when the software is executed (e.g., by one or more processors), it is operable to perform operations described with reference to the noise sample screening method based on model training according to embodiments of the present application.

[0133] The processor 502 runs a computer program corresponding to the executable program code stored in the memory 501 by reading the executable program code, for implementing the noise sample screening method based on model training in the above-described embodiments.

[0134] In one example, the noise sample screening device based on model training 500 can further include a communication interface 503 and a bus 504. As shown in Figure 8 The memory 501, the processor 502, and the communication interface 503 are connected through the bus 504 and complete communication therebetween.

[0135] The communication interface 503 is mainly used to realize communication between various modules, devices, units, and / or equipment in embodiments of the present application. Input devices and / or output devices can also be accessed through the communication interface 503.

[0136] Bus 504 includes hardware, software, or both, to couple components of model training based noise sample culling device 500 to each other and to couple model training based noise sample culling device 500 to one or more other devices. While FIG. 4 shows bus 504 as a single bus, bus 504 can be composed of multiple buses or separate communication links within model training based noise sample culling device 500. Also, while FIG. 4 shows bus 504 as a single bus that interconnects all of the components of model training based noise sample culling device 500, multiple buses can be used, and different buses can interconnect different combinations of the components of model training based noise sample culling device 500. For example, bus 504 can include an Accelerated Graphics Port (AGP) or other graphics bus, an Enhanced Industry Standard Architecture (EISA) bus, a Front Side Bus (FSB), a Hyper Transport (HT) Interconnect, an Industry Standard Architecture (ISA) bus, an Infmiband interconnect, a Low Pin Count (LPC) bus, a memory bus, a Micro Channel Architecture (MCA) bus, a Peripheral Component Interconnect (PCI) bus, a PCI-Express (PCI-E) bus, a Serial Advanced Technology Attachment (SATA) bus, a Video Electronics Standards Association Local Bus (VLB) bus, or another suitable bus or interconnect in addition to or instead of one or more of the buses described above. As will be apparent from the following description, bus 504, or components thereof, can be implemented using any combination of known bus structures.

[0137] The fourth aspect of the present application further provides a computer readable storage medium, which stores computer program instructions, and the computer program instructions are executed by a processor to implement the model training based noise sample culling method in the above embodiments and achieve the same technical effects. To avoid repetition, details are not described herein. The computer readable storage medium can include a non-transitory computer readable storage medium, such as a Read-Only Memory (ROM), a Random Access Memory (RAM), a magnetic disk or an optical disk, and the like, which are not limited herein.

[0138] The embodiments of the present application provide a computer program product, and instructions in the computer program product are executed by a processor of an electronic device to enable the electronic device to perform the model training based noise sample culling method in the above embodiments and achieve the same technical effects. To avoid repetition, details are not described herein.

[0139] It should be clarified that the various embodiments in this specification are described in a progressive manner, and the same or similar parts between the various embodiments can be referred to mutually. Each embodiment focuses on describing the differences from other embodiments. For device embodiments, equipment embodiments, computer-readable storage medium embodiments, and computer program products, relevant parts can be referred to the description section of the method embodiments. This application is not limited to the specific steps and structures described above and shown in the figures. Those skilled in the art can make various changes, modifications, and additions, or change the order of steps, after understanding the spirit of this application. Furthermore, for the sake of brevity, detailed descriptions of known methods and techniques are omitted here.

[0140] The aspects of this application have been described above with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of this application. It should be understood that each block in the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, a special-purpose computer, or other programmable data processing apparatus to produce a machine such that these instructions, executable via the processor of the computer or other programmable data processing apparatus, enable the implementation of the functions / actions specified in one or more blocks of the flowchart illustrations and / or block diagrams. Such a processor can be, but is not limited to, a general-purpose processor, a special-purpose processor, a special application processor, or a field-programmable logic circuit. It is also understood that each block in the block diagrams and / or flowcharts, and combinations of blocks in the block diagrams and / or flowcharts, can also be implemented by dedicated hardware performing the specified functions or actions, or can be implemented by a combination of dedicated hardware and computer instructions.

[0141] Those skilled in the art will understand that the above embodiments are exemplary and not restrictive. Different technical features appearing in different embodiments can be combined to achieve beneficial effects. Based on a study of the drawings, specification, and claims, those skilled in the art should be able to understand and implement other variations of the disclosed embodiments. In the claims, the term "comprising" does not exclude other means or steps; the quantifier "a" does not exclude a plurality; the terms "first" and "second" are used to identify names and not to indicate any particular order. No reference numerals in the claims should be construed as limiting the scope of protection. The functionality of multiple parts appearing in the claims can be implemented by a single hardware or software module. The appearance of certain technical features in different dependent claims does not mean that these technical features cannot be combined to achieve beneficial effects.

Claims

1. A method for screening noisy samples based on model training, characterized in that, include: Multiple model training sessions are conducted using a training sample set and a validation sample set to obtain multiple target models. The training sample set includes multiple samples, and the validation sample set includes multiple samples. After each model training, based on the actual labels and predicted values ​​of at least a portion of the samples in the target sample set, noise samples are determined and removed from the target sample set until the noise sample removal cutoff condition is met. The target sample set includes the training sample set and / or the validation sample set, and the predicted value is the value predicted by the target model based on the input samples. Each model training session includes: The first model is trained using the most recently updated training sample set; Based on the most recently updated target sample set and the comparison results between the first model and the second model, the model parameters of the first model and the second model are iteratively updated until the comparison results between the first model and the second model meet the iteration cutoff condition. The iteratively updated second model is then determined as the target model, and the second model is the first model before training.

2. The method according to claim 1, characterized in that, The comparison results include the consistency loss function value. The step of iteratively updating the model parameters of the first model and the second model based on the most recently updated target sample set and the comparison results between the first model and the second model, until the comparison results between the first model and the second model meet the iteration cutoff condition, and then determining the iteratively updated second model as the target model, includes: The samples in the most recently updated target sample set are input into the first model and the second model respectively to obtain the predicted values ​​of the samples output by the first model and the predicted values ​​of the samples output by the second model. Based on the actual labels of the samples, the predicted values ​​of the samples output by the first model, and the predicted values ​​of the samples output by the second model, the consistency loss function values ​​of the first model and the second model are obtained. Based on the consistency loss function value, update the model parameters of the first model and the model parameters of the second model to obtain the updated first model and the updated second model; The samples from the most recently updated target sample set are then input into the updated first model and the updated second model respectively to obtain the consistency loss function values ​​of the updated first model and the updated second model. This process continues until the consistency loss function values ​​meet the iteration cutoff condition, at which point the iteratively updated second model is determined as the target model.

3. The method according to claim 2, characterized in that, The step of updating the model parameters of the first model and the model parameters of the second model based on the consistency loss function value to obtain the updated first model and the updated second model includes: Based on the consistency loss function value, the first model parameters are calculated, and the model parameters of the first model are updated to the first model parameters, wherein the first model parameters minimize the consistency loss function value; Obtain the first learning rate control coefficient, calculate the second model parameters based on the first learning rate control coefficient, the first model parameters, and the model parameters of the second model obtained in the previous iteration, and update the model parameters of the second model to the second model parameters; Based on the first model parameters and the second model parameters, the updated first model and the updated second model are obtained.

4. The method according to claim 2, characterized in that, The iteration cutoff condition includes: the consistency loss function value reaches convergence.

5. The method according to claim 1, characterized in that, The step of determining the noise samples in at least a subset of samples based on their actual labels and predicted values ​​includes: Obtain the second learning rate control coefficient, and calculate the first predicted value based on the second learning rate control coefficient, the predicted value of the sample in the target model during this model training and the predicted value of the sample in the target model during the previous model training. Then update the predicted value of the sample in the target model during this model training to the first predicted value. Based on the first predicted value, the predicted label of the sample in the target model during this model training is obtained; If the actual label of a sample differs from the predicted label of the target model during the current model training, the sample is identified as a noise sample.

6. The method according to claim 1, characterized in that, The number of samples used to determine at least a portion of the noise samples in the i-th model training is greater than the number of samples used to determine at least a portion of the noise samples in the (i-1)-th model training, where i is a positive integer greater than 1.

7. The method according to any one of claims 1 to 6, characterized in that, The noise sample screening cutoff criteria include one or more of the following: The number of training iterations of the model has reached the preset threshold. The number of noise samples screened out from the training sample set in this instance satisfies either a first condition or a second condition. The first condition is that the number of noise samples screened out from the training sample set in this instance is less than a first preset number threshold. The second condition is that the proportion of the number of noise samples screened out from the training sample set in this instance to the number of samples in the training sample set is less than a first preset ratio threshold. The number of noise samples removed from the verification sample set in this instance satisfies either the third condition or the fourth condition. The third condition is that the number of noise samples removed from the verification sample set in this instance is less than the second preset number threshold. The fourth condition is that the proportion of the number of noise samples removed from the verification sample set in this instance to the number of samples in the verification sample set is less than the second preset ratio threshold. The number of noise samples removed from the target sample set in this instance satisfies either the fifth condition or the sixth condition. The fifth condition is that the number of noise samples removed from the target sample set in this instance is less than a third preset number threshold. The sixth condition is that the proportion of the number of noise samples removed from the target sample set in this instance to the number of samples in the target sample set is less than a third preset ratio threshold. Based on the actual labels and predicted values ​​of at least some samples in the target sample set, no noise samples were found. The first model performance parameter is less than the second model performance parameter. The first model performance parameter is the model performance parameter obtained by validating the target model after this model training using the validation sample set. The second model performance parameter is the model performance parameter obtained by validating the target model after the previous model training using the validation sample set.

8. The method according to any one of claims 1 to 6, characterized in that, The first model in the i-th training iteration is different from the first model in the j-th training iteration, where i and j are positive integers and i ≠ j.

9. A noise sample screening device based on model training, characterized in that, include: The training module is used to perform multiple model trainings using a training sample set and a validation sample set to obtain multiple target models. The training sample set includes multiple samples, and the validation sample set includes multiple samples. The noise removal module is used to determine the noise samples in the at least part of the samples after each model training based on the actual labels and predicted values ​​of at least part of the samples in the target sample set, and remove the noise samples from the target sample set until the noise sample removal cutoff condition is met. The target sample set includes the training sample set and / or the validation sample set, and the predicted value is the value predicted by the target model based on the input samples obtained by the model training. The training module includes: The training unit is used to train the first model using the most recently updated set of training samples. An update unit is used to iteratively update the model parameters of the first model and the model parameters of the second model based on the most recently updated target sample set and the comparison results between the first model and the second model, until the comparison results between the first model and the second model meet the iteration cutoff condition, and the iteratively updated second model is determined as the target model, where the second model is the first model before training.

10. A noise sample screening device based on model training, characterized in that, include: Processor and memory storing computer program instructions; When the processor executes the computer program instructions, it implements the noise sample screening method based on model training as described in any one of claims 1 to 8.

11. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores computer program instructions that, when executed by a processor, implement the model-trained noise sample removal method as described in any one of claims 1 to 8.

Images