CN115083442B - Data processing method, device, electronic device, and computer-readable storage medium - Google Patents

Abstract

The application relates to the technical field of data processing, and discloses a data processing method, a data processing device, electronic equipment and a computer readable storage medium. The method comprises the following steps: acquiring N pieces of data to be inspected, wherein N is a positive integer and is greater than or equal to two; inputting N pieces of data to be inspected to the inspection model for category prediction to obtain a prediction result corresponding to each piece of data to be inspected; determining prediction category distribution corresponding to N pieces of data to be inspected based on a prediction result corresponding to each piece of data to be inspected; if the prediction category distribution does not meet the prior category distribution, determining M pieces of data to be inspected from N pieces of data to be inspected based on the prediction result corresponding to each piece of data to be inspected, and correcting the prediction result of the M pieces of data to be inspected; the prior class distribution is determined based on class labels corresponding to each sample data in the sample data set, M is a positive integer, and M is smaller than or equal to N. By the method, accuracy of the prediction result of the data to be tested can be improved.

Description

Data processing method, device, electronic device, and computer-readable storage medium

technical field

The present application relates to the technical field of data processing, and in particular to a data processing method, device, electronic equipment, and computer-readable storage medium.

Background technique

More and more data needs to be used in the next stage after quality inspection. Therefore, the accuracy of data quality inspection is particularly important. For example, a large amount of voice data is generated in the daily customer service system. Intelligent quality inspection of these voice data can detect irregular content in customer service calls, which can improve the quality of customer service and user satisfaction, and reduce manual work. At the same time, customer service personnel can also be evaluated to improve the evaluation system for customer service personnel. Based on this, how to ensure high-accuracy data quality inspection is one of the key research issues in the field of data quality inspection.

Contents of the invention

The present application provides a data processing method, device, electronic equipment, and computer-readable storage medium, which can improve the accuracy of the prediction result of the data to be inspected.

On the one hand, a data processing method adopted by the present application, the method includes:

Obtain N quality inspection data, N is a positive integer and greater than or equal to two; input the N quality inspection data into the quality inspection model for category prediction, and obtain the prediction result corresponding to each quality inspection data; based on each The prediction results corresponding to the quality inspection data determine the predicted category distribution corresponding to the N quality inspection data; Determine M data to be quality inspected in the data, and correct the prediction results of M data to be quality inspected; wherein, the prior category distribution is determined based on the category label statistics corresponding to each sample data in the sample data set, and M is a positive integer , M is less than or equal to N.

In one aspect, the present application provides a data processing device, the data processing device comprising:

An acquisition unit, configured to acquire N pieces of data to be inspected, where N is a positive integer greater than or equal to two;

A prediction unit, configured to input the N data to be inspected into the quality inspection model for category prediction, and obtain a prediction result corresponding to each data to be inspected;

A determination unit, configured to determine the predicted category distributions corresponding to the N quality inspection data based on the prediction results corresponding to each quality inspection data;

The determining unit is also used to determine M data to be quality inspected from the N data to be quality inspected based on the prediction result corresponding to each data to be quality inspected if the predicted category distribution does not satisfy the prior category distribution; the correction unit uses It is used to correct the prediction results of M data to be inspected; wherein, the prior class distribution is determined based on the class label statistics corresponding to each sample data in the sample data set, M is a positive integer, and M is less than or equal to N.

In one aspect, the present application provides an electronic device, the electronic device includes a processor and a computer storage medium coupled to the processor, a computer program is stored in the computer storage medium, and the processor is used to execute the computer program to implement the above data processing method .

In one aspect, the present application provides a computer-readable storage medium, where the computer-readable storage medium stores a computer program, and when the computer program is executed by a processor, it implements the data processing method provided by the above technical solution.

The beneficial effects of the embodiments of the present application are: different from the prior art, the data processing method, device, electronic equipment and computer-readable storage medium provided by the present application, the method uses the statistical determination of the category labels corresponding to each sample data in the sample data set The prior category distribution of the quality inspection model is used as the judgment basis to determine the data to be inspected that the prediction results of the category prediction by the quality inspection model do not meet the prior category distribution, and the prediction results of the data to be quality inspected that do not meet the prior category distribution are analyzed. The correction makes the corrected prediction results of the data to be inspected satisfy the prior category distribution, which can improve the accuracy of the predicted results of the data to be inspected.

Description of drawings

In order to more clearly illustrate the technical solutions in the embodiments of the present application, the drawings that need to be used in the description of the embodiments will be briefly introduced below. Obviously, the drawings in the following description are only some embodiments of the present application. For those skilled in the art, other drawings can also be obtained based on these drawings without creative effort. in:

FIG. 1 is a schematic diagram of an application scenario of a data processing method provided by the present application;

Fig. 2 is a schematic flow chart of the first embodiment of the data processing method provided by the present application;

FIG. 3 is a schematic diagram of another application scenario of the data processing method provided by the present application;

FIG. 4 is a schematic flow diagram of the second embodiment of the data processing method provided by the present application;

FIG. 5 is a schematic diagram of another application scenario of the data processing method provided by the present application;

FIG. 6 is a schematic structural diagram of an embodiment of a data processing device provided by the present application;

Fig. 7 is a schematic flowchart of an embodiment of an electronic device provided by the present application.

Detailed ways

The technical solutions in the embodiments of the present application will be clearly and completely described below in conjunction with the drawings in the embodiments of the present application. It should be understood that the specific embodiments described here are only used to explain the present application, but not to limit the present application. In addition, it should be noted that, for the convenience of description, only some structures related to the present application are shown in the drawings but not all structures. Based on the embodiments in this application, all other embodiments obtained by persons of ordinary skill in the art without making creative efforts belong to the scope of protection of this application.

Reference herein to an "embodiment" means that a particular feature, structure, or characteristic described in connection with the embodiment can be included in at least one embodiment of the present application. The occurrences of this phrase in various places in the specification are not necessarily all referring to the same embodiment, nor are separate or alternative embodiments mutually exclusive of other embodiments. It is understood explicitly and implicitly by those skilled in the art that the embodiments described herein can be combined with other embodiments.

With the deepening of the application of Internet information technology in the financial field and the continuous strengthening of innovation by various enterprises, the time competition is also becoming more and more fierce. In this fierce time competition, user services have become more and more It has become an important measure to reflect competitive differences, enhance the company's image, and increase user satisfaction. Therefore, the management and control of the service quality of the customer service system has become an important daily work for business managers, and intelligent voice quality inspection is the main component of it. . A large amount of voice data is generated in the daily customer service system. If you can make good use of these data, carry out intelligent quality inspection work according to the specification requirements, and detect irregular points in customer service calls, you can improve the quality of customer service. As well as user satisfaction, reducing manual work, and at the same time evaluating customer service personnel, improving the job evaluation system for customer service personnel.

In one embodiment, the present application finds that artificial intelligence is widely used in speech data processing when performing quality inspection on the data to be inspected, especially speech data, so the present application can be designed based on artificial intelligence technology. A data processing scheme, mainly through training the quality inspection model, calling the designed quality inspection model to perform quality inspection on the data to be inspected. The general process can be summarized into the following stages: online data pull, data preprocessing, data labeling, data analysis, model training, model online, online badcase collection, and model iteration. It can be seen that from the launch of the model to the next iteration of the model version, it is necessary to rely on the mobile phone of the online badcase dataset and then retrain the model before launching it. Although this data quality inspection method can achieve high quality inspection accuracy to a certain extent, it will take a long time and rely on the collection and processing of badcase data after going online.

In another embodiment, this application proposes another data processing scheme, which adopts a post-processing technique of the quality inspection model to improve the accuracy of quality inspection, and almost no additional cost increase is required. There is a certain effect improvement. The idea of this data processing scheme is very simple, which is to use the prior knowledge of data distribution to optimize the prediction results of the quality inspection model. Suppose there is a binary classification problem, the two categories can be represented by 1 and 0 respectively, the prediction result given by the quality inspection model for the input a is p(a)=[0.01, 0.99], assuming that the prediction category corresponding to the prediction result is 1; Next, for the input b, the prediction result given by the quality inspection model is p(b)=[0.5, 0.5], which is the most uncertain state at this time, and the quality inspection model is not sure which category to output. However, let the quality inspection model know two prior knowledge in advance: 1. The category must be one of 0 or 1; 2. The probability of occurrence of the two categories is 0.5. Under these two points of prior knowledge, since the prediction result of the previous input a sample is 1, then based on the simple uniform idea, it is more inclined to predict the category of the next input b sample as 0, so as to obtain a satisfying the second point Prior prediction results.

Further extension, assuming that the category ratio of the data is already known (that is, prior knowledge), the category ratio of the prediction result obtained by the quality inspection model when predicting the data should also be very close to the category ratio of the prior knowledge. If the difference is large, it can indirectly indicate that the prediction effect of the quality inspection model is not good, and it will make mistakes in the prediction of some data categories. Then, based on the prior knowledge, these incorrectly predicted data can be found out and corrected.

By comparing the above two data processing schemes proposed in this application, it can be seen that the latter data processing scheme does not need to repeatedly perform model iteration and model online compared with the former data processing scheme, which can save time. Therefore, the latter data processing solution is mainly introduced in the following embodiments of the present application.

The latter data processing scheme is specifically summarized as follows:

Obtain N quality inspection data, N is a positive integer and greater than or equal to two; input the N quality inspection data into the quality inspection model for category prediction, and obtain the prediction result corresponding to each quality inspection data; based on each The prediction results corresponding to the quality inspection data determine the predicted category distribution corresponding to the N quality inspection data; Determine M data to be quality inspected in the data, and correct the prediction results of M data to be quality inspected; wherein, the prior category distribution is determined based on the category label statistics corresponding to each sample data in the sample data set, and M is a positive integer , M is less than or equal to N.

The latter data processing scheme uses the prior category distribution determined based on the category label statistics corresponding to each sample data in the sample data set as the basis for judgment, and determines that the prediction result of the category prediction through the quality inspection model does not meet the quality of the prior category distribution. The data to be inspected, and the prediction results of the data to be inspected that do not satisfy the prior category distribution are corrected, so that the predicted results of the revised data to be inspected meet the prior category distribution, which can improve the accuracy of the prediction results of the data to be inspected . Furthermore, the latter data processing scheme does not need to repeatedly perform model iteration and model online, which can save time.

The data processing scheme of the present application can be executed by an electronic device, which can be a terminal device, such as a smart phone, a tablet computer, a notebook computer, a desktop computer, an intelligent voice interaction device, a smart home appliance, a smart watch, a vehicle terminal, an aircraft, etc.; or , the electronic device may also include a server, such as an independent physical server, or a server cluster or distributed system composed of multiple physical servers, or a cloud server that provides cloud computing services.

The above data processing solution can be mainly applied to voice quality inspection of the conversation data between the customer service and the user in the customer service system. Referring to FIG. 1 , it is a diagram of an application scenario provided by the embodiment of the present application. In FIG. 1 , the dialog box contains dialogue data between the customer service and the user. Wherein, the dialogue data may be voice data or text data. Then periodically pull the dialogue data; then use the data processing scheme of this application to call the quality inspection model to perform quality inspection on the dialogue data, and obtain the predicted quality inspection results corresponding to these dialogue data; determine the prediction corresponding to these dialogue data based on these predicted quality inspection results Category distribution; if the predicted category distribution does not satisfy the prior category distribution, based on the predicted quality inspection results corresponding to these dialogue data, several dialogue data are determined from them, and the predicted quality inspection results of several dialogue data are corrected to obtain the final Quality inspection results. Then output the quality inspection result instructions to the customer service personnel; if it is a good result, don’t worry about it, if it is a bad result, the customer service needs to adjust the speech skills.

Based on the above data processing solution, the embodiment of the present application provides a data processing method. Referring to FIG. 2 , it is a schematic flowchart of a data processing method provided in the present application. The data processing method described in FIG. 2 can be executed by an electronic device, specifically, it can be executed by a processor of the electronic device. The data processing method shown in FIG. 2 includes the following steps:

Step 21: Obtain N pieces of data to be inspected, where N is a positive integer greater than or equal to two.

In some embodiments, the data to be checked may be voice data, image data, text data, and the like.

Among them, the data to be inspected can be obtained based on different scenarios. For example, image data may come from monitoring equipment in a monitoring scene. The voice data can come from the dialogue between the customer service and the user. The text data can come from text communication between customer service and users.

Step 22: Input the N data to be inspected into the quality inspection model for category prediction, and obtain the prediction result corresponding to each data to be inspected.

Among them, the quality inspection model can be based on convolutional neural network, deconvolutional neural network, deep convolutional reverse graph network, generative confrontation network and recurrent neural network.

As the name implies, the quality inspection model is for quality inspection of the data to be inspected. Performing quality inspection on the data to be inspected may refer to setting several categories for the data to be inspected in advance, and then calling the quality inspection model to predict the category to which the data to be inspected belongs. For example, if the data to be inspected is the dialogue data between the customer service and the customer from the customer service system, then the pre-set categories for this type of data to be inspected are qualified dialogue data, unqualified dialogue data, or; include Keyword dialogue data and dialogue data that do not include keywords, or, dialogue data that includes positive situations, dialogue data that includes negative emotions; for another example, if the data to be inspected is image data, then pre-set the The categories of the three types of data to be inspected may be: image data containing the target object, and image data not containing the target object.

Optionally, if W categories are set for the data to be quality checked in advance, then the prediction result of each data to be quality checked through the quality control model predicts that each data to be quality checked belongs to each of the W categories probability value. For example, assuming that the data to be quality-checked is dialogue data, the preset categories are dialogue data including positive emotions and dialogue data including negative emotions, assuming that the category of dialogue data including positive emotions is represented by 1, The category of dialogue data including negative emotions is represented by 0. The data to be inspected is input into the quality inspection model, and the output preset result is a probability value of 1 for the data to be inspected and a probability value of 0 for the data to be inspected.

Wherein, the quality inspection model is trained based on each sample data with category labels in the sample data set. For example, in the field of image recognition, sample data is image data with class labels. In the field of voice quality inspection, the sample data can be voice dialogue data with category labels. In the field of text quality inspection, sample data can be text data with category labels.

In some embodiments, multiple data to be quality-tested and sample data may be obtained from the same server at different times in the same time period. The inventor found through research that the data generated in the same time period will obey the corresponding law. If the normal distribution is met, the ratio of positive data and negative data is basically the same. Therefore, the two data are comparable in the future.

Step 23: Determine the predicted category distribution corresponding to the N quality inspection data based on the prediction result corresponding to each quality inspection data.

After the quality inspection model prediction in step 22, the prediction result corresponding to each data to be quality inspection is obtained, and the prediction result of each data to be quality inspection includes the probable value that each data to be quality inspection belongs to each category in W categories , based on the prediction results corresponding to each quality inspection data to determine the prediction category distribution corresponding to the N quality inspection data. First, it is necessary to determine the prediction category corresponding to each quality inspection data based on the prediction results corresponding to each quality inspection data. Among them, W is a positive integer and greater than or equal to two

In the specific implementation, the prediction category corresponding to each quality inspection data is determined based on the prediction result corresponding to each quality inspection data, including: determining the category with the largest probability value among the prediction results corresponding to each quality inspection data as each The prediction category corresponding to the data to be checked. For example, W is equal to 2, that is, the data to be inspected corresponds to two categories, and these two categories are defined as the first category and the second category. The probability value of the first class corresponding to the data to be inspected a is 90%, and the probability value of the corresponding second class is 10%, so the first class is determined as the predicted class corresponding to the data a to be inspected.

For another example, W is equal to 3, that is, the data to be inspected corresponds to 3 categories, and these two categories are defined as the first category, the second category, and the third category. The probability value of the first category corresponding to the quality inspection data b is 70%, the probability value of the corresponding second category is 10%, and the corresponding probability value of the third category is 20%, then the first category is determined to be quality inspection The prediction category corresponding to the detection data b.

After the predicted category corresponding to each data to be quality checked is obtained, the predicted category distributions of the N pieces of data to be quality checked can be calculated based on the predicted category corresponding to each data to be quality checked. In some embodiments, predicted class distributions may be represented by class proportions.

Specifically, count the number of data to be inspected belonging to each of the W categories, and then count the distribution of predicted categories of the data to be inspected based on the number of data to be inspected under each category. For example, assuming that W is equal to 2, that is, the data to be inspected corresponds to two categories, and these two categories are defined as the first category and the second category. Count the number of data to be inspected corresponding to each category. That is, the first category corresponds to the first quantity, and the second category corresponds to the second quantity. Wherein, the sum of the first quantity and the second quantity is equal to N. Thus, the quantity proportion corresponding to each category of data to be quality inspected can be determined.

When W is equal to 3, that is, the data to be inspected corresponds to three categories, and these three categories are defined as the first category, the second category, and the third category. Assume that the quantity of data to be quality inspected under the first category is the first quantity, the quantity of data to be quality inspected under the second category is the second quantity, and the quantity of data to be quality inspected under the third category is the third quantity. Wherein, the sum of the first quantity, the second quantity and the third quantity is equal to N. Thus, the proportion of the quantity of data to be inspected corresponding to each category can be determined.

Step 24: If the predicted class distribution does not satisfy the prior class distribution, then determine M data to be inspected from the N data to be inspected based on the prediction result corresponding to each data to be inspected.

Wherein, the prior class distribution is determined based on the class label statistics corresponding to each sample data in the sample data set, M is a positive integer, and M is less than or equal to N. Optionally, T sample subsets can be obtained by performing T data sampling processing on the sample data set; counting the number of sample data under each category in each sample subset; based on the samples under each category in each sample subset According to the amount of data, determine the category proportion corresponding to each data sampling process; determine the prior category distribution according to the category proportion corresponding to each data sampling data process.

Optionally, each sample subset includes a training subset and a testing subset. Perform T data sampling processing on the sample data set to obtain T sample subsets, including: determining the ratio of sample data between the training subset and the test subset during each data sampling processing. Based on the number ratio, the sample data set is sampled to obtain T training subsets and T testing subsets respectively. Usually the number of sample data in the training subset is larger than the number of sample data in the testing subset. For example, the ratio of sample data between the training subset and the testing subset is 9:1, 8:1 or 7:1.

Thus, according to the quantity ratio determined each time, data sampling is performed on the sample data set to obtain corresponding training subsets and test subsets. After sampling T times, T sample subsets will be obtained.

Further, the quality inspection model can be trained by using the sample data in the training subset and the sample data in the test subset. Specifically, the quality inspection model is trained by using the sample data in the training subset, and then the quality inspection model is tested by using the sample data in the test subset, so as to determine the accuracy of the quality inspection model.

In some embodiments, the categories of sample data include a first category C1 and a second category C2. Then the category ratio of the first category C1 may be c1/(c1+c2), and the category ratio of the second category C2 may be c2/(c1+c2). Wherein, c1 indicates the quantity of the first category C1, and c2 indicates the quantity of the second category C2.

In some embodiments, the categories of sample data include a first category C1, a second category C2 and a third category C3. The category ratio of the first category C1 can be c1/(c1+c2+c3), the category ratio of the second category C2 can be c2/(c1+c2+c3), and the category ratio of the third category C3 can be c3/( c1+c2+c3). Wherein, c3 represents the quantity of the third category C3.

That is, each category corresponds to a category proportion, and the sum of these category proportions is 1.

In some embodiments, the prior class distribution may be represented by a probability interval. Optionally, according to the category proportion corresponding to each data sampling process, determine the prior category distribution, including: according to the category proportion corresponding to each data sampling process, count the mean value and variance corresponding to the T times of data sampling process; The weighted difference operation is performed on the variance, and the operation result after the weighted difference is taken as the minimum value of the probability interval, and the weighted sum operation is performed on the mean and variance, and the operation result after the weighted sum is taken as the maximum value of the probability interval.

Among them, the mean value can be determined by the following formula:

Among them, the variance can be determined by the following formula:

Among them, T indicates the number of category proportions corresponding to each category, _Xi indicates the i-th category proportion, μ indicates the mean, and σ indicates the variance.

When the data to be inspected corresponds to the first category and the second category, after counting T times of data sampling, category proportions corresponding to T first categories and category proportions corresponding to T second categories will be obtained. Thus, the first category will correspond to a mean and variance, and the second category will also correspond to a mean and variance.

When the data to be inspected corresponds to the first category, the second category, and the third category, after counting T times of data sampling, the category proportions corresponding to T first categories, the category proportions corresponding to T second categories, and The proportion of categories corresponding to T third categories. Thus, the first category will correspond to a mean and variance, the second category will also correspond to a mean and variance, and the third category will also correspond to a mean and variance.

Then perform a weighted difference operation on the mean and variance, and use the weighted difference operation result as the minimum value of the probability interval, and perform a weighted sum operation on the mean and variance, and use the weighted sum operation result as the probability interval the maximum value.

In some embodiments, the weights of the mean and variance can be set according to actual needs, for example, the weight of the mean is set to 1, and the weight of the variance is set to 3. The weighted difference operation is performed on the mean value and variance, and the obtained operation result is μ-3σ, and the weighted sum operation is performed on the mean value and variance, and the obtained operation result is μ+3σ. That is, the prior class distribution can be expressed as [μ-3σ,μ+3σ].

If the predicted category distribution is within the range of the probability interval, it means that the predicted category distribution satisfies the prior category distribution.

If the predicted category distribution is not within the range of the probability interval, it means that the predicted category distribution does not satisfy the prior category distribution. That is, it can be determined that there are errors in the prediction results of the N pieces of data to be quality inspected, and it is necessary to find out the data to be inspected that have wrong prediction results.

In some embodiments, the difference value can be determined by using the predicted category distribution and the prior category distribution; by using the difference value and the number N of data to be quality checked, the number M of data to be quality checked with abnormal prediction results can be determined; M pieces of data to be quality checked are determined from the N pieces of data to be quality checked.

Step 25: Correct the prediction results of the M data to be inspected.

When W is equal to 2, that is, the data to be inspected corresponds to two categories, and these two categories are defined as the first category and the second category. If it is determined that the prediction results of the M data to be quality inspection belong to the first category of data to be quality inspection, then modifying the prediction results of the M data to be quality inspection can be to make the prediction result of the M data to be quality inspection belong to the first category. modifying the prediction result of one category of data to be quality inspected to the second category; and modifying the prediction result to the first category of the second category of quality inspection data.

When W is equal to 3, that is, the data to be inspected corresponds to 3 categories, and these two categories are defined as the first category, the second category and the second category. Determine the data to be inspected whose prediction results belong to the first category among the M data to be inspected, determine the data to be inspected whose prediction result belongs to the second category among the data to be inspected, and determine the data to be inspected The predicted results belong to the third category of data to be inspected.

Among them, in the prediction process, there is a probability value corresponding to each category of the data to be inspected, and the category corresponding to the maximum probability value will be used as the prediction result. That is, there is a probability value corresponding to the first category, the second category, and the second category in the data to be inspected. After determining the data to be inspected that the prediction results belong to the first category, the data to be inspected to belong to the second category, and the data to be inspected to belong to the third category among the M data to be quality inspected, determine each data to be inspected The category corresponding to the second largest probability value. Modify the prediction results of the M data to be quality inspected to the category with the second largest probability value.

In this embodiment, the prior category distribution determined based on the category label statistics corresponding to each sample data in the sample data set is used as the judgment basis, and it is determined that the prediction result of the category prediction through the quality inspection model does not meet the quality of the prior category distribution. inspection data, and correct the prediction results of the quality inspection data that do not satisfy the prior category distribution. On the one hand, it can improve the abnormal prediction of the quality inspection model due to training accuracy problems, and does not need to retrain the quality inspection model, reducing the need for The training cost of the quality inspection model, on the other hand, correcting the prediction results of the data to be quality inspection that do not satisfy the prior category distribution can improve the accuracy of the prediction results of the data to be quality inspection.

In an application scenario, it will be described with reference to Figure 3:

Step 301: Online data fetching.

In step 301, data is randomly pulled from the line to ensure that the pulled data is consistent with the real data distribution on the line. In the voice quality inspection scenario, the voice conversation data between customer service and users can be pulled. The online here refers to the system that generates the data. For example, voice data is generated by a customer service system.

Step 302: Data labeling.

The extracted data can be marked manually to determine the category corresponding to each data. For example, mark the data obtained online according to the business logic and the set mark format.

Step 303: Divide training subsets and test subsets.

Randomly scramble the marked data, and then divide the training subset and test subset according to a certain ratio.

Step 304: Statistical category proportions.

Count the number of data of each category in the training subset and test subset respectively, and calculate the corresponding category proportion.

Step 305: Calculate the mean and variance corresponding to each category.

Step 303 and step 304 are repeated T times, and each category corresponds to T category proportions. The mean and variance corresponding to each category can be calculated based on the proportions of T categories.

Step 306: Quality inspection model training.

Use the labeled data to iteratively train the quality inspection model. Among them, the quality inspection model can be iteratively trained by using the divided training subset and test subset.

Step 307: Model testing.

Use the trained QC model to make predictions on the data in the test subset.

Step 308: Calculation of category proportions of prediction results.

The class proportion is determined according to the prediction result in step 307 .

Step 309: Compare.

After the mean value and variance corresponding to each category are calculated in step 305, the prior category distribution of each category is determined based on the mean value and variance.

If the category proportion of each category in the prediction result does not satisfy its corresponding prior category distribution, step 310 is executed. If the category proportion of each category in the prediction result satisfies its corresponding prior category distribution, it is determined that the prediction is normal.

Step 310: Determine the data with abnormal prediction results.

Wherein, the entropy corresponding to the data can be determined according to the probability value corresponding to each category in the prediction result, and the data with abnormal prediction results can be determined according to the entropy. For details, reference may be made to any of the foregoing embodiments, and details are not described here.

In some embodiments, the data with abnormal prediction results are determined according to the proportion of categories. If M data with abnormal prediction results are determined in the data corresponding to the first category, there will also be M data with abnormal prediction results in the data corresponding to the second category. Then it is necessary to determine the data with abnormal prediction results from the second category. That is, when the prediction result of the data in one of the multiple categories is abnormal, the data with abnormal prediction results will also appear in the other categories.

Step 311: Correct.

The prediction result of the data whose prediction result is abnormal is corrected.

In an application scenario, the categories that data may correspond to are the first category and the second category. The data whose prediction result is abnormal corresponds to the first category, and the prediction result of the data is changed to the second category.

Similarly, if data with an abnormal prediction result is determined from the first category, data with an abnormal prediction result also exists in the data corresponding to the second category. These data also require revisions to forecasts.

In another application scenario, the categories that data may correspond to are the first category, the second category, and the third category. The data whose prediction result is abnormal corresponds to the first category, and then it is determined that the data corresponds to the probability value of the second category and the probability value of the third category. A larger one between the probability value of the second category and the probability value of the third category is determined. Modify the first category corresponding to the data with abnormal prediction results to the category corresponding to the larger one. For example, if the probability value of the second category is relatively large, the prediction result of the data whose prediction result is abnormal is modified to the second category.

Similarly, if the data with abnormal prediction results is determined from the first category, there will also be data with abnormal prediction results in the corresponding data of the second category and the data of the third category. These data also require revisions to forecasts.

In other embodiments, manual correction may be used. For details, please refer to the description in any other embodiment, and details are not repeated here.

Referring to FIG. 4 , FIG. 4 is a schematic flowchart of a second embodiment of the data processing method provided by the present application. The method includes:

Step 41: Obtain N pieces of data to be inspected, where N is a positive integer greater than or equal to two.

Step 42: Input the N data to be inspected into the quality inspection model for category prediction, and obtain a prediction result corresponding to each data to be inspected.

Steps 41 to 42 have the same or similar technical solutions as those in any of the above-mentioned embodiments, which will not be repeated here.

Step 43: Determine the category with the highest probability value among the prediction results corresponding to each data to be inspected as the predicted category corresponding to each data to be inspected.

In this embodiment, the prediction result corresponding to each data subject to quality inspection includes a probability value that each data subject to quality inspection belongs to each of the W categories.

Thus, the category with the highest probability value among the prediction results corresponding to each data to be quality checked can be determined as the predicted category corresponding to each data to be quality checked.

Step 44: According to the predicted category corresponding to each quality-inspected data, count the quantity of the quality-inspected data belonging to each category, and obtain the predicted category distribution corresponding to the N quality-inspected data.

In some embodiments, when W is equal to 2, that is, the data to be inspected corresponds to two categories, and these two categories are defined as the first category and the second category.

After the prediction of the quality inspection model and the determination of the predicted category in steps 42-43, each data subject to quality inspection corresponds to a category. In this way, the quantity of data to be inspected corresponding to each category is counted. That is, the first category corresponds to the first quantity, and the second category corresponds to the second quantity. Wherein, the sum of the first quantity and the second quantity is equal to N. Thus, the quantity proportion corresponding to each category of data to be quality inspected can be determined.

When W is equal to 3, that is, the data to be inspected corresponds to 3 categories, and these two categories are defined as the first category, the second category and the third category.

After the prediction of the quality inspection model in steps 42-43 and the determination of the predicted category, each data to be quality-checked corresponds to a category. In this way, the quantity of data to be inspected corresponding to each category is counted. That is, the first category corresponds to the first quantity, the second category corresponds to the second quantity, and the third category corresponds to the third quantity. Wherein, the sum of the first quantity, the second quantity and the third quantity is equal to N. Thus, the proportion of the quantity of data to be inspected corresponding to each category can be determined.

Step 45: If the predicted category distribution does not satisfy the prior category distribution, then use the predicted category distribution and the prior category distribution to determine the difference value.

Wherein, the prior class distribution is determined based on the class label statistics corresponding to each sample data in the sample data set.

In some embodiments, the predicted category distribution can be expressed by the proportion of each category, the prior category distribution can be expressed by the probability interval of each category proportion, and then can be determined by using each category proportion and the probability interval of each category proportion out the difference value.

If the proportion of each category is greater than the maximum value in the probability interval of each category proportion, the corresponding difference value is obtained by subtracting the maximum value in the probability interval of each category proportion from each category proportion.

If the proportion of each category is less than the minimum value in the probability interval of each category proportion, the minimum value in the probability interval of each category proportion is used to subtract each category proportion to obtain the corresponding difference value.

Specifically, the following formula can be used to express:

Among them, Δ indicates the difference, Y indicates the proportion of each category, μ-3σ indicates the maximum value in the probability interval of each category proportion, and (μ-3σ) indicates the minimum value in the probability interval of each category proportion.

Step 46: Using the difference value and the number N of data to be inspected, determine the number M of data to be inspected with abnormal prediction results.

Wherein, M is a positive integer, and M is less than or equal to N.

Then, according to the difference value, the quantity M of data to be inspected with abnormal prediction results is determined. For example, the difference value is multiplied by the number N of data to be quality checked to obtain the number M of data to be quality checked with abnormal prediction results.

It can be understood that because the values in the probability interval of each category proportion and each category proportion are less than 1, the difference value is also a decimal, and the number M of data to be inspected with abnormal prediction results will be less than the data to be inspected The number of N.

Step 47: Determine M data to be inspected from the N data to be inspected.

In some embodiments, step 47 includes: using the probability value corresponding to each category in the prediction result corresponding to each data to be quality checked to determine the entropy corresponding to each data to be quality checked. According to the entropy corresponding to each data to be quality inspection, M pieces of data to be quality checked are determined from the N pieces of data to be quality checked.

For example, in the quality inspection model prediction process, the probability value of each category corresponding to each data to be quality inspection is predicted, and then the corresponding label category is finally determined according to the magnitude of the probability value. Specifically, the following formula can be used to determine the entropy corresponding to each data to be quality checked:

Among them, H(X) represents entropy, p( _xi ) represents the probability value of x _i , here refers to the probability value of each category predicted by the quality inspection model, m here represents the number of categories, and x _i represents the i-th one to be inspected data.

After determining the entropy corresponding to the N pieces of data to be quality inspected, sort these entropies, such as from large to small or from small to large.

When these entropies are sorted from large to small, M entropies are obtained from front to back, and then the data to be inspected corresponding to these entropies are determined. The data to be inspected is the data to be inspected with abnormal prediction results.

When these entropies are sorted from small to large, M entropies are obtained from back to front, and then the data to be inspected corresponding to these entropies is determined. These samples to be labeled are the data to be quality inspected with abnormal prediction results.

Step 48: Correct the prediction results of the M data to be inspected.

In some embodiments, step 48 may correct the prediction results of the M data to be quality-checked by manual modification. For example, a modification interface is displayed, wherein, the modification interface displays M data to be quality inspected and prediction results of M data to be quality inspected; receive correction information for any data to be quality inspected in the M data to be quality inspected, and use the correction information to The prediction results of any pending quality inspection data are corrected.

Combined with Figure 5 for illustration:

As shown in FIG. 5 , the modification interface displays data to be inspected a, data to be inspected b, and data to be inspected c, as well as corresponding prediction results. Each data to be inspected corresponds to a modification button.

After selecting the modify button, an input field or a selection field may pop up. If the input field pops up, the correction information input by the user is received, and the prediction result of any data to be inspected is corrected by using the correction information. For example, modify the first category of the data to be quality checked a to the second category.

If a selection bar pops up, each category available for selection can be displayed in the selection bar. Receives the category selected by the user. Use the selected category to correct the prediction results of any data to be inspected. For example, modify the second category of the data b to be quality checked to the first category.

In this embodiment, the prior category distribution determined based on the category label statistics corresponding to each sample data in the sample data set is used as the judgment basis, and it is determined that the prediction result of the category prediction through the quality inspection model does not meet the quality of the prior category distribution. inspection data, and correct the prediction results of the quality inspection data that do not satisfy the prior category distribution. On the one hand, it can improve the abnormal prediction of the quality inspection model due to training accuracy problems, and does not need to retrain the quality inspection model, reducing the need for The training cost of the quality inspection model, on the other hand, correcting the prediction results of the data to be quality inspection that do not satisfy the prior category distribution can improve the accuracy of the prediction results of the data to be quality inspection.

Further, by counting the number of data to be inspected under each category, the predicted category distribution corresponding to N data to be inspected can be obtained, and the number of data to be inspected under each category can be determined, and then N The predicted category distribution corresponding to each quality-inspection data is convenient for determining the subsequent quality-inspection data whose predicted category distribution does not satisfy the prior category distribution.

In addition, the method of any of the above-mentioned embodiments of the present application can not only be used for data quality inspection, but also can be used in other fields, such as data labeling. In the field of data labeling, the data to be quality checked is the data to be labeled. After the prediction category of the data to be labeled is predicted by the model, the abnormally predicted data to be labeled is determined, and the prediction result is corrected, and then the corrected prediction category is used as the label of the data to be labeled. And the uncorrected predicted category is used as the label of the uncorrected data to be labeled. In this way, on the one hand, using the prediction results of the network model as labels can improve labeling efficiency; Anomaly prediction, thereby improving the accuracy of labeling.

In an application scenario, 10,000 pieces of data are randomly obtained from the line and marked, assuming that the positive and negative ratio is 3:1, and this is repeated n times independently, and the positive and negative sample ratio of each group is calculated [x1, x2, ..., xn ], this data obeys the normal distribution, and then the mean value is assumed to be 3.01, and the variance is 0.05. It can be seen that the proportion of categories is mainly distributed in the range of [2.86, 3.16], and then the trained model is used to predict the samples to be labeled, then Theoretically, the category ratio of the predicted result should be within the range of [2.86, 3.16]. If it exceeds this range, it means that there is an error in the model prediction, and it needs to be corrected. Assuming that the predicted category ratio is 3.2, it can be seen that there are 1000* in the predicted result (3.2-3.16) = 40 pieces of data have errors, and then use the method of calculating entropy to filter out the top40 data based on the probability value of the predicted result, and then correct it.

Referring to FIG. 6 , FIG. 6 is a schematic structural diagram of an embodiment of a data processing device provided in the present application. The data processing device 60 includes: an acquisition unit 61 , a prediction unit 62 , a determination unit 63 and a correction unit 64 .

Wherein, the acquiring unit 61 is used to acquire N pieces of data to be inspected, where N is a positive integer greater than or equal to two.

The predicting unit 62 is configured to input the N data to be inspected into the quality inspection model to perform category prediction, and obtain a prediction result corresponding to each data to be inspected.

The determining unit 63 is configured to determine the predicted category distribution corresponding to the N quality-checked data based on the prediction result corresponding to each quality-checked data; and if the predicted category distribution does not satisfy the prior category distribution, then based on the corresponding As a result of the prediction, M data to be inspected are determined from N data to be inspected.

The correction unit 64 is used to correct the prediction results of the M data to be quality inspected; wherein, the prior category distribution is determined based on the category label statistics corresponding to each sample data in the sample data set, M is a positive integer, and M is less than or equal to N .

In one embodiment, the prediction result corresponding to each data subject to quality inspection includes a probability value that each data subject to quality inspection belongs to each category in W categories; the determination unit 63 is based on each When the prediction result corresponding to the data to be quality checked determines the predicted category distribution corresponding to the N pieces of data to be quality checked, the following steps are performed:

Determining the category with the largest probability value among the prediction results corresponding to each data to be quality checked as the predicted category corresponding to each data to be quality checked;

According to the prediction category corresponding to each of the data to be quality inspection, the number of data to be quality inspection belonging to each category is counted, and the predicted category distribution corresponding to the N data to be quality inspection is obtained.

In one embodiment, the determining unit 63 executes the following steps when determining M data to be inspected from the N data to be inspected based on the prediction result corresponding to each data to be inspected:

determining a difference value using the predicted class distribution and the prior class distribution;

Using the difference value and the number N of data to be quality checked, determine the number M of data to be quality checked with abnormal prediction results;

M pieces of data to be quality checked are determined from the N pieces of data to be quality checked.

In one embodiment, the determining unit 63 performs the following steps when determining M numbers of data to be quality checked from the N pieces of data to be quality checked:

Using the probability value corresponding to each category in the prediction result corresponding to each data to be quality inspection, determine the entropy corresponding to each data to be quality inspection;

According to the entropy corresponding to each of the data to be quality checked, M pieces of data to be quality checked are determined from the N pieces of data to be quality checked.

In one embodiment, the data processing device further includes a processing unit 65, the processing unit 65 is used for:

performing T times of data sampling processing on the sample data set to obtain T sample subsets;

Count the number of sample data under each category in each sample subset;

Determine the proportion of categories corresponding to each data sampling process based on the number of sample data under each category in each sample subset;

According to the category proportion corresponding to each data sampling data processing, the prior category distribution is determined.

In one embodiment, each sample subset includes a training subset and a test subset, and the processing unit 65 performs the following steps when performing T data sampling on the sample data set to obtain T sample subsets:

Determine the proportion of sample data between the training subset and the test subset for each data sampling process;

Based on the number ratio, data sampling is performed on the sample data set to obtain T training subsets and T testing subsets respectively.

In one embodiment, the prior class distribution is represented by a probability interval, and the determining unit 63 performs the following steps when determining the prior class distribution according to the class ratio corresponding to each data sampling data processing:

According to the category ratio corresponding to each data sampling process, count the mean value and variance corresponding to T times of data sampling process;

Perform weighted difference calculation on the mean value and variance, and use the operation result after weighted difference calculation as the minimum value of the probability interval, and perform weighted sum operation on the mean value and variance, and use the weighted sum operation result as the maximum value of the probability interval.

In one embodiment, the modifying unit 64 performs the following steps when correcting the prediction results of the M data to be quality inspected:

Displaying a modification interface, wherein the modification interface displays the M data to be inspected and prediction results of the M data to be inspected;

Receive correction information for any data to be inspected among the M data to be inspected, and use the correction information to correct a prediction result of any data to be inspected.

The data processing device provided in this application is based on the prior category distribution determined by the statistics of the category labels corresponding to each sample data in the sample data set as the judgment basis, and determines that the prediction result of the category prediction through the quality inspection model does not meet the quality of the prior category distribution. The data to be inspected, and the prediction results of the data to be inspected that do not satisfy the prior category distribution are corrected, so that the predicted results of the revised data to be inspected meet the prior category distribution, which can improve the accuracy of the prediction results of the data to be inspected .

Referring to FIG. 7 , FIG. 7 is a schematic structural diagram of an embodiment of an electronic device provided in the present application. The electronic device 70 includes a processor 71 , an input interface 72 , an output interface 73 and a computer storage medium 74 . Wherein, the processor 71 is respectively coupled to the input interface 72 , the output interface 73 and the computer storage medium 74 . The input interface 72 can be connected with an external device for receiving data input by the external device. The output interface 73 can be connected with an external device, and is used for outputting data to the external device.

Wherein, a computer program is stored in the computer storage medium 74, and the processor 71 is used to execute the computer program to realize the following method:

Obtain N quality inspection data, N is a positive integer and greater than or equal to two; input N quality inspection data into the quality inspection model for category prediction, and obtain the prediction result corresponding to each quality inspection data; based on each quality inspection data The prediction results corresponding to the data to be inspected determine the predicted category distribution corresponding to the N quality inspection data; if the predicted category distribution does not satisfy the prior category distribution, based on the prediction results corresponding to each quality inspection data, from the N quality inspection data Determine the M data to be quality-inspected, and correct the prediction results of the M data to be quality-inspected; wherein, the prior category distribution is determined based on the category label statistics corresponding to each sample data in the sample data set, M is a positive integer, M is less than or equal to N.

The present application also provides a computer-readable storage medium, wherein a computer program is stored in the computer-readable storage medium, and when the computer program is executed by the processor 71, the following methods are implemented:

Obtain N quality inspection data, N is a positive integer and greater than or equal to two; input N quality inspection data into the quality inspection model for category prediction, and obtain the prediction result corresponding to each quality inspection data; based on each quality inspection data The prediction results corresponding to the data to be inspected determine the predicted category distribution corresponding to the N quality inspection data; if the predicted category distribution does not satisfy the prior category distribution, based on the prediction results corresponding to each quality inspection data, from the N quality inspection data Determine the M data to be quality-inspected, and correct the prediction results of the M data to be quality-inspected; wherein, the prior category distribution is determined based on the category label statistics corresponding to each sample data in the sample data set, M is a positive integer, M is less than or equal to N.

In one embodiment, the prediction result corresponding to each data subject to quality inspection includes a probability value that each data subject to quality inspection belongs to each category in W categories; the processor 71 is based on each When the prediction result corresponding to the data to be quality checked determines the predicted category distribution corresponding to the N pieces of data to be quality checked, the following steps are performed:

Determining the category with the largest probability value among the prediction results corresponding to each data to be quality checked as the predicted category corresponding to each data to be quality checked;

According to the prediction category corresponding to each of the data to be quality inspection, the number of data to be quality inspection belonging to each category is counted, and the predicted category distribution corresponding to the N data to be quality inspection is obtained.

In one embodiment, the processor 71 performs the following steps when determining M data to be inspected from the N data to be inspected based on the prediction result corresponding to each data to be inspected:

determining a difference value using the predicted class distribution and the prior class distribution;

Using the difference value and the number N of data to be quality checked, determine the number M of data to be quality checked with abnormal prediction results;

M pieces of data to be quality checked are determined from the N pieces of data to be quality checked.

In one embodiment, the processor 71 executes the following steps when determining M numbers of data to be quality-checked from the N pieces of data to be quality-checked:

Using the probability value corresponding to each category in the prediction result corresponding to each data to be quality inspection, determine the entropy corresponding to each data to be quality inspection;

According to the entropy corresponding to each of the data to be quality checked, M pieces of data to be quality checked are determined from the N pieces of data to be quality checked.

In one embodiment, the processor 71 is also used for:

performing T times of data sampling processing on the sample data set to obtain T sample subsets;

Count the number of sample data under each category in each sample subset;

Determine the proportion of categories corresponding to each data sampling process based on the number of sample data under each category in each sample subset;

According to the category proportion corresponding to each data sampling data processing, the prior category distribution is determined.

In one embodiment, each sample subset includes a training subset and a test subset, and the processor 71 performs the following steps when performing T data sampling on the sample data set to obtain T sample subsets:

Determine the proportion of sample data between the training subset and the test subset for each data sampling process;

Based on the number ratio, data sampling is performed on the sample data set to obtain T training subsets and T testing subsets respectively.

In one embodiment, the prior class distribution is represented by a probability interval, and the processor 71 performs the following steps when determining the prior class distribution according to the class ratio corresponding to each data sampling data processing:

According to the category ratio corresponding to each data sampling process, count the mean value and variance corresponding to T times of data sampling process;

Perform weighted difference calculation on the mean value and variance, and use the operation result after weighted difference calculation as the minimum value of the probability interval, and perform weighted sum operation on the mean value and variance, and use the weighted sum operation result as the maximum value of the probability interval.

In one embodiment, the processor 71 performs the following steps when correcting the prediction results of the M data to be quality checked:

Displaying a modification interface, wherein the modification interface displays the M data to be inspected and prediction results of the M data to be inspected;

Receive correction information for any data to be inspected among the M data to be inspected, and use the correction information to correct a prediction result of any data to be inspected.

To sum up, the data processing method, device, electronic equipment, and computer-readable storage medium provided by this application use the prior category distribution determined based on the category label statistics corresponding to each sample data in the sample data set as the basis for judgment, and determine The prediction results of the category prediction by the quality inspection model do not meet the prior category distribution of the data to be quality inspection, and the prediction results of the data to be quality inspection that do not meet the prior category distribution are corrected, so that the corrected data to be quality inspection The prediction results satisfy the prior category distribution, which can improve the accuracy of the prediction results of the data to be inspected.

In the several implementation manners provided in this application, it should be understood that the disclosed methods and devices may be implemented in other ways. For example, the device implementation described above is only illustrative. For example, the division of the circuit or unit is only a logical function division. In actual implementation, there may be another division method. For example, multiple units or components can be Incorporation may either be integrated into another system, or some features may be omitted, or not implemented.

The units described as separate components may or may not be physically separated, and the components shown as units may or may not be physical units, that is, they may be located in one place, or may be distributed to multiple network units. Part or all of the units can be selected according to actual needs to achieve the purpose of the solution of this embodiment.

In addition, each functional unit in each embodiment of the present application may be integrated into one processing unit, each unit may exist separately physically, or two or more units may be integrated into one unit. The above-mentioned integrated units can be implemented in the form of hardware or in the form of software functional units.

The above is only the implementation of the application, and does not limit the patent scope of the application. All equivalent structures or equivalent process transformations made according to the contents of the specification and drawings of the application, or directly or indirectly used in other related technologies fields, are all included in the scope of patent protection of this application in the same way.

Claims (11)

1. A method of data processing, the method comprising:

acquiring N pieces of data to be inspected, wherein N is a positive integer and is greater than or equal to two;

inputting the N pieces of data to be inspected to the inspection model for category prediction to obtain a prediction result corresponding to each piece of data to be inspected;

determining prediction category distribution corresponding to the N pieces of data to be inspected based on the prediction result corresponding to each piece of data to be inspected;

if the prediction category distribution does not meet the prior category distribution, determining M pieces of data to be inspected from the N pieces of data to be inspected based on the prediction results corresponding to the data to be inspected, and correcting the prediction results of the M pieces of data to be inspected; each piece of data to be inspected corresponds to an entropy, and each entropy of the M pieces of data to be inspected is at least larger than each entropy of the N-M pieces of data to be inspected; the prior class distribution is determined based on class labels corresponding to each sample data in the sample data set, M is a positive integer, and M is smaller than or equal to N.

2. The method of claim 1, wherein the prediction result corresponding to each piece of data to be inspected includes a probability value of each of W categories to which each piece of data to be inspected belongs; the determining the prediction category distribution corresponding to the N pieces of data to be inspected based on the prediction result corresponding to each piece of data to be inspected includes:

determining the category with the maximum probability value in the prediction result corresponding to each piece of data to be inspected as the prediction category corresponding to each piece of data to be inspected;

and counting the number of the data to be inspected under each category according to the prediction category corresponding to each data to be inspected, and obtaining the prediction category distribution corresponding to the N data to be inspected.

3. The method of claim 1, wherein the

Based on the prediction result corresponding to each piece of quality inspection data, determining M pieces of quality inspection data from the N pieces of quality inspection data, including:

determining a difference value by using the predicted category distribution and the prior category distribution;

determining the quantity M of the data to be inspected with abnormal prediction results by using the difference value and the quantity N of the data to be inspected;

and determining M pieces of data to be inspected from the N pieces of data to be inspected.

4. A method according to claim 3, wherein said determining M pieces of quality inspection data from said N pieces of quality inspection data comprises:

determining entropy corresponding to each piece of data to be inspected by utilizing a probability value corresponding to each category in a prediction result corresponding to each piece of data to be inspected;

and determining M pieces of data to be inspected from the N pieces of data to be inspected according to the entropy corresponding to each piece of data to be inspected.

5. The method of claim 1, wherein the method further comprises:

performing T times of data sampling processing on the sample data set to obtain T sample subsets;

counting the number of sample data under each category in each sample subset;

determining a class proportion corresponding to each data sampling process based on the number of sample data in each class in each sample subset;

and determining prior category distribution according to category proportion corresponding to data processing of each data sampling.

6. The method of claim 5, wherein each sample subset comprises a training subset and a test subset, wherein the T data samples are taken from the sample data set to obtain T sample subsets, comprising:

determining the quantity proportion of sample data between the training subset test subsets during each data sampling process;

And based on the quantity proportion, carrying out data sampling on the sample data set to respectively obtain T training subsets and T testing subsets.

7. The method of claim 5, wherein the prior class distribution is represented by a probability interval, the determining the prior class distribution from the class proportions corresponding to each data sample data processing, comprising:

according to the class proportion corresponding to each data sampling process, counting the mean value and variance corresponding to T times of data sampling processes;

and carrying out weighted difference calculation on the mean value and the variance, taking the calculation result after weighted difference calculation as the minimum value of the probability interval, carrying out weighted summation calculation on the mean value and the variance, and taking the calculation result after weighted summation as the maximum value of the probability interval.

8. The method of claim 1, wherein the modifying the predicted outcome of the M pieces of quality control data comprises:

displaying a modification interface, wherein the modification interface displays the M pieces of data to be inspected and prediction results of the M pieces of data to be inspected;

and receiving correction information of any to-be-inspected data in the M to-be-inspected data, and correcting a prediction result of the any to-be-inspected data by using the correction information.

9. A data processing apparatus, characterized in that the data processing apparatus comprises:

the acquisition unit is used for acquiring N pieces of data to be inspected, wherein N is a positive integer and is greater than or equal to two;

the prediction unit is used for inputting the N pieces of data to be inspected to the quality inspection model for category prediction to obtain a prediction result corresponding to each piece of data to be inspected;

the determining unit is used for determining prediction category distribution corresponding to the N pieces of data to be inspected based on the prediction result corresponding to each piece of data to be inspected; if the prediction category distribution does not meet the prior category distribution, determining M pieces of data to be inspected from the N pieces of data to be inspected based on a prediction result corresponding to each piece of data to be inspected;

the correction unit is used for correcting the prediction results of the M pieces of data to be inspected; each piece of data to be inspected corresponds to an entropy, and each entropy of the M pieces of data to be inspected is at least larger than each entropy of the N-M pieces of data to be inspected; the prior class distribution is determined based on class labels corresponding to each sample data in the sample data set, M is a positive integer, and M is smaller than or equal to N.

10. An electronic device comprising a processor and a computer storage medium coupled to the processor, the computer storage medium having a computer program stored therein, the processor configured to execute the computer program to implement the method of any of claims 1-8.

11. A computer readable storage medium, characterized in that the computer readable storage medium stores a computer program which, when executed by a processor, implements the method according to any of claims 1-8.