Noise label identification method and system based on prior confidence and heterogeneous consensus strategy

Through the two-stage noise label recognition method of prior confidence and heterogeneous consensus strategy, the noise label recognition problem under the influence of category imbalance in the existing technology is solved, accurate identification and data set quality optimization are achieved, and the classification model is improved .

CN119026032AActive Publication Date: 2024-11-26SHANDONG UNIV OF FINANCE & ECONOMICS

Patent Information

Application Number
CN202411523294.6
Authority / Receiving Office
CN · China
Patent Type
Applications(China)
Current Assignee / Owner
Filing Date
2024-10-30
Publication Date
2024-11-26
Estimated Expiration
2044-10-30

AI Technical Summary

Technical Problem

The existing noise label recognition method fails to effectively consider the category imbalance of the data set, resulting in inaccurate identification of noise labels, and filtering may aggravate the category imbalance of the data set, affecting the generalization performance of the classification model.

Method used

Using a two-stage noise label recognition method based on a priori confidence and heterogeneous consensus strategy, the gradient contribution of the sample is calculated as a confidence indicator through cross-validation, and combining a priori confidence strategy and a heterogeneous consensus strategy, the noise label is identified and filtered to ensure that the noise label is Accurate identification of noise labels under category imbalance conditions does not aggravate the imbalance of the dataset.

Benefits of technology

Accurate noise label recognition under category imbalance conditions is realized, potential minority class pseudo-samples are mined, the generalization performance of the classification model is improved, the category imbalance aggravation caused by existing methods is avoided, and the data set quality is optimized.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN119026032A_ABST
    Figure CN119026032A_ABST
Patent Text Reader

Abstract

The invention discloses a noise label identification method and system based on a prior confidence and heterogeneous consensus strategy, and relates to the technical field of artificial intelligence and unbalanced classification, and the method comprises the steps: obtaining a plurality of sample data labeled with labels, and constructing an original data set; obtaining gradient contribution of each sample data in the original data set by using a dichotomy model; by taking class imbalance as a priori condition and gradient contribution as a confidence degree measurement index of label labeling, identifying noise label samples in the original data set by adopting a priori confidence strategy, and constructing a noise data set; removing a noise data set in the original data set, and training a heterogeneous classifier by using a clean data set; and performing pseudo label prediction on each sample data in the noise data set by using a heterogeneous classifier, and screening potential pseudo minority class sample data in the noise data set by using a heterogeneous consensus strategy based on the pseudo labels and a preset probability consensus threshold. According to the invention, accurate noise label identification under the condition of class imbalance can be realized.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] The present invention relates to the field of artificial intelligence and imbalanced classification technology, and in particular to a noise label recognition method and system based on prior confidence and heterogeneous consensus strategy. Background Art

[0002] At present, classification technology based on artificial intelligence has become one of the research hotspots. This classification technology can be divided into supervised classification tasks and unsupervised classification tasks. Among them, the data set of supervised classification tasks usually has noise labels due to crowdsourcing, human errors and other reasons. The noise labels referred to here are broad. It may be a label that is incorrectly labeled, or it may be a label that is correctly labeled but reduces the normal prediction ability of the classification model. The noise label is usually defined as a label that has a counterproductive effect on the fitting of the machine learning model. The classification model is forced to fit these noise labels, which leads to deviations in the prediction ability of normal samples, seriously affecting the generalization performance of the machine learning classifier. Therefore, effectively identifying and processing these noise labels can improve the classification accuracy at the level of optimizing data quality without designing a very complex classification model.

[0003] At present, the scale of data is getting larger and larger, and the amount of noise data in it is also increasing. Therefore, the identification and processing of noise labels is an indispensable task for optimizing the quality of data sets. At present, there have been many studies in this area at home and abroad. For the identification of noise data in data processing, a confidence learning method is currently proposed. This method is a method proposed by MIT and Google for identifying wrong labels in labeled data sets, which has been proven to be effective in many real-world scenario tasks. Confidence learning is a noise label filtering method based on predicted probability difference. It identifies and filters wrongly labeled samples by estimating the joint distribution of noise labels and true labels. The true label refers to a pseudo-label that is not originally labeled and is considered correct by confidence learning. In addition, the current recognition of noisy labels also includes the label confidence reasoning method (LISR) based on statistics and hypothesis testing and the Benjamini-Hochberg network method (BHN), the method based on relative probability density (RPD), the fast transition matrix estimation method for learning with noisy labels (FASTEN), the three-way decision-based noise correction method (TDNC), and the label denoising method for anomaly detection (LDAAD).

[0004] However, in reality, many datasets for supervised classification tasks are class-imbalanced, that is, the number of samples in one category is much larger than the number of samples in other categories. For example, default prediction and fraudulent transaction detection in the field of credit risk control, abnormal intrusion traffic identification in the field of network security, equipment fault detection in the field of smart industry, early disease screening in the field of medicine, etc. The datasets for these task scenarios are usually severely class-imbalanced, that is, the number of normal samples is much larger than the number of abnormal samples, and the accurate identification of abnormal samples is more important for scenario tasks.

[0005] Although the above existing methods can achieve a certain noise label recognition effect, thereby optimizing the quality of the data set and improving the accuracy of the classification model, these methods do not consider the impact of the class imbalance of the data set on the noise label recognition method. For example, the confidence learning method relies on the prediction probability of the sample, but due to the class imbalance factor, the prediction probability of samples of different categories is biased, which will make the imbalance rate of the data set greater after filtering; the ProMix method will assume that there are no noise labels in the minority class samples in the data set with severe class imbalance. In other words, the above methods do not consider the class imbalance problem in the data set, resulting in inaccurate noise labels finally identified. Summary of the invention

[0006] In order to address the deficiencies of the above-mentioned prior art, the present invention provides a noise label recognition method and system based on prior confidence and heterogeneous consensus strategy, which takes category imbalance as a priori condition, adopts prior confidence strategy and heterogeneous consensus strategy, and performs two-stage noise label recognition, thereby achieving accurate noise label recognition under category imbalance conditions; when processing category imbalance data sets, this method will not aggravate the degree of category imbalance of the data set, and can effectively improve the generalization performance of the classification model.

[0007] In a first aspect, the present invention provides a noise label recognition method based on prior confidence and heterogeneous consensus strategy.

[0008] A noise label recognition method based on prior confidence and heterogeneous consensus strategy, including: Obtain a number of labeled sample data to construct an original data set; wherein the sample data in the original data set is divided into majority class samples and minority class samples according to the number of labels; Using a binary classification model based on cross-validation, calculate the gradient contribution of each sample data in the original data set; Taking category imbalance as a priori condition, taking the gradient contribution of sample data as the confidence metric of the annotation label, adopting the prior confidence strategy, identifying the noise label samples in the original data set, and constructing the noise data set; Obtain a clean data set by removing the noise data set from the original data set, and use the clean data set to train a heterogeneous classifier; Heterogeneous classifiers are used to predict pseudo labels for each sample data in the noise dataset. Based on the pseudo labels and the preset probability consensus threshold, a heterogeneous consensus strategy is adopted to screen out potential pseudo minority sample data in the noise dataset.

[0009] In a second aspect, the present invention provides a noise label recognition system based on prior confidence and heterogeneous consensus strategy.

[0010] A noise label recognition system based on prior confidence and heterogeneous consensus strategy, including: The label sample acquisition module is used to obtain a number of labeled sample data to construct an original data set; wherein the sample data in the original data set is divided into majority class samples and minority class samples according to the number of labels; The noise label sample identification module is used to calculate the gradient contribution of each sample data in the original data set using a binary classification model based on cross-validation; taking category imbalance as a priori condition, the gradient contribution of sample data is used as a confidence metric for the annotation label, and a priori confidence strategy is adopted to identify the noise label samples in the original data set and construct a noise data set; The pseudo minority sample identification module is used to obtain a clean data set from the original data set by removing the noise data set, and use the clean data set to train a heterogeneous classifier; use the heterogeneous classifier to predict the pseudo labels of each sample data in the noise data set, and use the heterogeneous consensus strategy to screen out potential pseudo minority sample data in the noise data set based on the pseudo labels and the preset probability consensus threshold.

[0011] In a third aspect, the present invention further provides an electronic device comprising a memory and a processor, and computer instructions stored in the memory and executed on the processor, wherein when the computer instructions are executed by the processor, the steps of the method described in the first aspect are completed.

[0012] In a fourth aspect, the present invention further provides a computer-readable storage medium for storing computer instructions, which, when executed by a processor, complete the steps of the method described in the first aspect.

[0013] One or more of the above technical solutions have the following beneficial effects: 1. The present invention provides a noise label recognition method and system based on prior confidence and heterogeneous consensus strategy. It takes category imbalance as a priori condition, and adopts prior confidence strategy and heterogeneous consensus strategy to perform two-stage noise label recognition. It can not only realize accurate noise label recognition under category imbalance conditions, but also mine potential minority class pseudo samples in the noise label data set. Compared with the existing methods, the present invention can better solve the problem of aggravated category imbalance caused by noise label filtering, and the optimized data set can further improve the generalization performance of the classifier.

[0014] 2. In the present invention, the cross entropy gradient contribution of sample data is used as the label confidence measurement index of the sample. Compared with the prediction probability and cross entropy loss, the measurement index of the present invention can not only reflect whether the sample is correctly identified, but also reflect the contribution of the sample to the model fitting. To a certain extent, it can represent the distance of the sample from the decision boundary in space. On this basis, a priori confidence strategy with category imbalance as a priori condition is proposed. Based on the average gradient contribution of samples of different categories, an expansion factor is designed according to the prior condition for the average gradient contribution of minority class samples, so as to increase the threshold for minority class samples to be identified as noise labels, and retain the precious minority class samples to a greater extent. Considering that the above-mentioned priori confidence method retains minority class samples, it does not mean that minority class samples have no noise labels. Therefore, the present invention also proposes a heterogeneous consensus strategy that integrates the sample's own characteristics and spatial feature decision-making capabilities. Through this strategy, potential minority class pseudo samples in the noise samples screened out in the priori confidence stage are mined, further alleviating the category imbalance problem of the data set.

[0015] 3. The noise label identification method based on prior confidence and heterogeneous consensus strategy proposed in the present invention is suitable for binary classification tasks of numerical type category imbalanced data sets, and can be used for default prediction and fraudulent transaction detection in the field of credit risk control, abnormal intrusion traffic identification in the field of network security, equipment failure detection in the field of smart industry, early disease screening in the field of medicine, etc. In the face of category imbalanced data sets in the above-mentioned fields, by solving the problem of increased category imbalance rate of data sets caused by the existing noise label filtering method not considering the condition of category imbalance, the noise labels in the imbalanced data sets are identified and processed to improve the quality of the data sets. Moreover, when processing category imbalanced data sets, the category imbalance degree of the data sets will not be aggravated, which can effectively improve the generalization performance of machine learning classification models on imbalanced data sets. BRIEF DESCRIPTION OF THE DRAWINGS

[0016] The accompanying drawings in the specification, which constitute a part of the present invention, are used to provide a further understanding of the present invention. The exemplary embodiments of the present invention and their descriptions are used to explain the present invention and do not constitute improper limitations on the present invention.

[0017] Figure 1This is an overall flow chart of the noise label identification method based on prior confidence and heterogeneous consensus strategy described in an embodiment of the present invention; Figure 2 Flow chart of the a priori confidence strategy in an embodiment of the present invention. DETAILED DESCRIPTION

[0018] It should be noted that the following detailed descriptions are exemplary only, are intended to describe specific embodiments, are intended to provide further explanation of the present invention, and are not intended to limit exemplary embodiments according to the present invention. Unless otherwise specified, all technical and scientific terms used herein have the same meanings as those of ordinary skill in the art to which the present invention belongs. In addition, it should also be understood that when the terms "comprising" and / or "including" are used in this specification, it indicates the presence of features, steps, operations, devices, components and / or combinations thereof.

[0019] Embodiment 1 In order to solve the problem that the existing noise label recognition method does not consider the impact of the imbalance of data set categories, which ultimately leads to inaccurate recognition results, this embodiment provides a noise label recognition method based on prior confidence and heterogeneous consensus strategy, such as Figure 1 As shown, the method specifically comprises the following steps: Step S1, obtaining a number of labeled sample data to construct an original data set; wherein the sample data in the original data set is divided into majority class samples and minority class samples according to the number of labels; Step S2: using a binary classification model based on cross-validation, calculate the gradient contribution of each sample data in the original data set; taking class imbalance as a priori condition, using the gradient contribution of the sample data as a confidence measurement indicator of the annotation label, adopting a priori confidence strategy, identifying the noise label samples in the original data set, and constructing a noise data set; Step S3, obtain a clean data set from the original data set by removing the noise data set, and use the clean data set to train a heterogeneous classifier; use the heterogeneous classifier to predict pseudo labels for each sample data in the noise data set, and use a heterogeneous consensus strategy to screen out potential pseudo minority sample data in the noise data set based on the pseudo labels and a preset probability consensus threshold.

[0020] The following content introduces the noise label recognition method based on prior confidence and heterogeneous consensus strategy proposed in this embodiment in more detail.

[0021] In step S1, a number of labeled sample data are obtained to construct an original data set. In this embodiment, data processing in the field of financial risk control is used as an example. First, labeled data in this field is obtained as sample data, and an original data set is constructed based on the obtained sample data. , where the sample data in the original data set is divided into majority class samples and minority class samples according to the number of labels. For example, a number of publicly available credit card consumption record data labeled as normal consumption and fraudulent consumption are obtained as sample data to construct the original credit card consumption record data set, in which the label of normal consumption samples is 0 and the label of fraudulent consumption samples is 1; or, for example, a number of publicly available credit loan data labeled as normal repayment and repayment default are obtained as sample data to construct the original credit loan data set, in which the label of normal repayment samples is 0 and the label of repayment default samples is 1. The above sample data content and its labels can be provided by banks or banking financial institutions. Under normal circumstances, there are few abnormal samples (such as fraudulent consumption samples and repayment default samples) and a lot of normal samples (such as normal consumption samples and normal repayment samples). Therefore, the sample data in the above data set can be divided into minority class samples and majority class samples.

[0022] like Figure 2 As shown, step S2 is the prior confidence filtering stage. Through this stage, the original data set is Divide into clean data sets and noisy dataset , initially from the original data set Remove the noisy label samples.

[0023] Step S2.1: Use a binary classification model based on cross-validation to calculate the gradient contribution of each sample data in the original data set, and convert the gradient contribution ( gradient contribution, gc ) is used as a confidence metric for the label.

[0024] Specifically, first, the prediction probability of each sample is obtained by cross-validation on the original data set. , that is: input each sample data in the original data set into the binary classification model based on cross-validation, and output the probability that each sample data is predicted to be a minority class sample The above binary classification refers to predicting whether the labeled sample data is a majority class sample or a minority class sample. The classification model based on cross-validation is Logistic Regression (LR). Since LR is a classification model with cross entropy as the loss function, and this embodiment derives the label confidence measurement indicator based on cross entropy, when the LR cross-validation result is optimal, the cross entropy loss of each sample can be guaranteed to be the lowest.

[0025] Secondly, according to the probability of predicting the minority class sample , calculate and determine the cross entropy loss for the binary classification. Cross entropy loss for the binary classification for: ,in Indicates the label of the sample. In binary classification, the value is 0 or 1, indicating the majority class or minority class respectively; represents the probability that the model predicts the sample as label 1 (i.e., the minority class), then Represents the probability that the model predicts the sample as label 0 (i.e., the majority class).

[0026] Finally, the cross entropy loss for the two categories The probability of predicting a sample as a minority class The derivative of is the gradient contribution of the sample data. That is, the gradient contribution Defined as the binary cross entropy loss versus probability The derivative of right The derivative of ,when hour, ,when hour, , therefore, the gradient contribution is uniformly expressed as the following formula: ; On the basis of the above, the preset prediction probability threshold is 0.5. , the sample is correctly identified; otherwise, the sample is incorrectly identified. It can directly reflect whether a sample is correctly identified. It can also reflect the difficulty of fitting the sample, that is, the contribution of a sample to the model fitting. The larger it is, the more difficult it is to fit the sample and the closer it is to the decision boundary in space.

[0027] After calculating and obtaining the gradient contribution of each sample data, the influence of class imbalance is considered. For this purpose, this embodiment proposes a priori confidence strategy with class imbalance as a priori condition. Based on the average gradient contribution of samples of different categories, an expansion factor is designed for the average gradient contribution of minority class samples according to the prior condition, so as to increase the threshold for minority class samples to be identified as noise labels, so as to retain precious minority class samples to a greater extent, that is: Step S2.2, taking category imbalance as a priori condition, taking the gradient contribution of sample data as the confidence measurement indicator of the annotation label, adopting the prior confidence strategy, identifying the noise label samples in the original data set, and constructing the noise data set.

[0028] Specifically, first, based on the gradient contribution of each sample data in the original data set, the average gradient contribution of the majority class samples and the minority class samples is calculated respectively. and , for each category, the samples whose gradient contribution is greater than the average gradient contribution and whose identification errors are initially marked as noise label samples, which can be expressed as the following formula: ; Secondly, taking class imbalance as a priori condition, we design an expansion factor for the average gradient contribution of minority class samples according to the priori condition , optimize the average gradient contribution of minority class samples according to the expansion factor, so as to increase the threshold for minority class samples to be identified as noise samples. are the hyperparameters that need to be adjusted.

[0029] Finally, for the majority class samples and minority class samples, screening is performed based on their corresponding average gradient contribution and preset prediction probability threshold, and the noise label samples in the majority class samples and minority class samples are identified. That is, the final screening formula for noise samples is: ; That is to say, for the majority class samples, determine the maximum value of the average gradient contribution of the majority class samples and the preset prediction probability threshold, compare the gradient contribution of each sample data in the majority class samples with the maximum value, if it is greater than the maximum value, it is judged as a noise label sample; for the minority class samples, determine the maximum value of the average gradient contribution of the minority class samples after optimization and the preset prediction probability threshold, compare the gradient contribution of each sample data in the minority class samples with the maximum value, if it is greater than the maximum value, it is judged as a noise label sample.

[0030] According to the above method, identify the noise label samples in the original data set and use them to construct the noise data set Then, step S3 is performed, that is, in the original data set Subtract the noise dataset , initially obtain a clean data set : .

[0031] Furthermore, although step S2 can retain precious minority samples to a greater extent, the filtered There may still be potential minority class samples in the dataset. Therefore, this embodiment proposes step S3 to enter the heterogeneous consensus recall stage. Through this step, the noise dataset filtered out from the first stage (i.e., step S2) Mining potential minority class samples from the majority class samples.

[0032] First, the noisy dataset The sample labels in the clean data set are discarded (or regarded as unlabeled data). In order to obtain the fusion decision-making ability of the sample's own characteristics and the sample space characteristics, Two heterogeneous classifiers are trained on the same machine: multi-layer perceptron MLP and K nearest neighbor model KNN, denoted as and .

[0033] Secondly, the noise data sets predicted by the two heterogeneous classifiers after training are The pseudo label of each sample data in the , obtains the corresponding prediction probability; at the same time, the probability consensus threshold is pre-set , according to the consensus strategy of the following formula, The potential minority class data in the formula is: ; In the above formula, Representative category, represent The samples in the dataset, and They represent the predicted probabilities of the samples by the models (i.e., the multi-layer perceptron MLP and the K nearest neighbor model KNN). It represents The sample predicted by the classifier belongs to The probability of the class, is a hyperparameter that needs to be adjusted.

[0034] In fact, the above screening method is: for each sample data in the noise data set, when two classifiers predict that the same sample data has the same pseudo label and the predicted probability of the pseudo label is greater than the probability consensus threshold, it is identified as a potential minority class sample, so as to construct a pseudo minority class data set. Considering that the majority class data needs to be downsampled when dealing with the class imbalance problem, only the minority class data is screened out. The data predicted as the minority class label in the binary classification The samples are combined into a pseudo minority dataset .

[0035] The heterogeneous consensus strategy in this embodiment is different from the traditional consensus strategy. It combines the decision-making capabilities of the samples’ own characteristics and spatial characteristics in the clean sample set, and mines The potential minority class samples in the dataset further alleviate the category imbalance phenomenon of the dataset by considering the prior condition of category imbalance.

[0036] Finally, the clean dataset and the pseudo minority class dataset are integrated to obtain the final dataset with correct labels, which is used for subsequent data processing and model training.

[0037] The superiority of the method proposed in this embodiment is further verified by the following experimental comparison. The method proposed in this embodiment is compared with the noise label filtering method based on confidence learning and tested on four financial data sets with different imbalance rates.

[0038] First, the imbalance rates are compared. As shown in Table 1 below, the two strategies of this embodiment have lower imbalance rates than those obtained by confidence learning, and thus can better improve the generalization performance of the classification model.

[0039] Table 1 Comparison of imbalance rate results

[0040] Secondly, the performance improvement comparison test of the classification model based on the method proposed in this embodiment and the confidence learning method is carried out. Two classifiers, logistic regression (LR) and lightweight gradient boosting machine (LightGBM, LGBM for short), are selected for testing. The test results are shown in Table 2 below, where AUC (Area Under the Curve) is a performance evaluation index of a classification model, which is used to measure the ability of a classifier to distinguish samples of different categories, specifically the area under the ROC curve (Receiver Operating Characteristic Curve). The larger the AUC value, the better the performance of the model, that is, the model can better distinguish positive and negative samples; KS is a commonly used statistical indicator for evaluating the degree of difference between two samples or between a sample and a theoretical distribution. The larger the KS value, the higher the prediction accuracy of the model and the better the stability. In Table 2 below, the data results in bold and arrows illustrate that the method of this embodiment (i.e., this method) can enable the classifier to obtain the optimal performance index.

[0041] Table 2 Comparison of classifier performance improvement results

[0042] Embodiment 2 This embodiment provides a noise label recognition system based on prior confidence and heterogeneous consensus strategy, which specifically includes: The label sample acquisition module is used to obtain a number of labeled sample data to construct an original data set; wherein the sample data in the original data set is divided into majority class samples and minority class samples according to the number of labels; The noise label sample identification module is used to calculate the gradient contribution of each sample data in the original data set using a binary classification model based on cross-validation; taking category imbalance as a priori condition, the gradient contribution of sample data is used as a confidence metric for the annotation label, and a priori confidence strategy is adopted to identify the noise label samples in the original data set and construct a noise data set; The pseudo minority sample identification module is used to obtain a clean data set from the original data set by removing the noise data set, and use the clean data set to train a heterogeneous classifier; use the heterogeneous classifier to predict the pseudo labels of each sample data in the noise data set, and use the heterogeneous consensus strategy to screen out potential pseudo minority sample data in the noise data set based on the pseudo labels and the preset probability consensus threshold.

[0043] Embodiment 3 This embodiment provides an electronic device, including a memory and a processor, and computer instructions stored in the memory and executed on the processor. When the computer instructions are executed by the processor, the steps in the noise label identification method based on prior confidence and heterogeneous consensus strategy as described above are completed.

[0044] Embodiment 4 This embodiment also provides a computer-readable storage medium for storing computer instructions. When the computer instructions are executed by a processor, the steps in the noise label identification method based on prior confidence and heterogeneous consensus strategy as described above are completed.

[0045] The steps involved in the above embodiments 2 to 4 correspond to the method embodiment 1, and the specific implementation methods can refer to the relevant description part of embodiment 1. The term "computer-readable storage medium" should be understood as a single medium or multiple media including one or more instruction sets; it should also be understood to include any medium that can store, encode or carry an instruction set for execution by a processor and enable the processor to execute any method in the present invention.

[0046] Those skilled in the art should understand that the modules or steps of the present invention described above can be implemented by a general-purpose computer device, or alternatively, they can be implemented by a program code executable by a computing device, so that they can be stored in a storage device and executed by the computing device, or they can be made into individual integrated circuit modules, or multiple modules or steps therein can be made into a single integrated circuit module for implementation. The present invention is not limited to any specific combination of hardware and software.

[0047] The above description is only a preferred embodiment of the present invention. Although the specific implementation mode of the present invention is described in conjunction with the accompanying drawings, it is not a limitation of the protection scope of the present invention. Those skilled in the art should understand that on the basis of the technical solution of the present invention, various modifications or variations that can be made by those skilled in the art without creative work are still within the protection scope of the present invention.

Claims

1. A noise label recognition method based on prior confidence and heterogeneous consensus strategy, characterized in that: include: Obtain a number of labeled sample data to construct an original data set; wherein the sample data in the original data set is divided into majority class samples and minority class samples according to the number of labels; Using a binary classification model based on cross-validation, calculate the gradient contribution of each sample data in the original data set; Taking category imbalance as a priori condition, taking the gradient contribution of sample data as the confidence metric of the annotation label, adopting the prior confidence strategy, identifying the noise label samples in the original data set, and constructing the noise data set; Obtain a clean data set by removing the noise data set from the original data set, and use the clean data set to train a heterogeneous classifier; Heterogeneous classifiers are used to predict pseudo labels for each sample data in the noise dataset. Based on the pseudo labels and the preset probability consensus threshold, a heterogeneous consensus strategy is adopted to screen out potential pseudo minority sample data in the noise dataset.

2. The noise label identification method based on prior confidence and heterogeneous consensus strategy as claimed in claim 1 is characterized in that: Using a binary classification model based on cross-validation, the gradient contribution of each sample data in the original data set is calculated, including: Input each sample data in the original data set into a binary classification model based on cross-validation, and output the probability that each sample data is predicted to be a minority class sample; the binary classification refers to predicting that the labeled sample data is a majority class sample or a minority class sample; According to the probability of predicting the minority class sample, the cross entropy loss of the binary classification is calculated; The derivative of the cross entropy loss with respect to the probability of predicting the minority class sample is taken as the gradient contribution of the sample data.

3. The noise label identification method based on prior confidence and heterogeneous consensus strategy as claimed in claim 1 is characterized in that: Taking category imbalance as a priori condition, taking the gradient contribution of sample data as the confidence metric of the annotation label, and adopting the prior confidence strategy, we can identify the noise label samples in the original data set, including: Based on the gradient contribution of each sample data in the original data set, the average gradient contribution of majority class samples and minority class samples is calculated; Taking class imbalance as a priori condition, the expansion factor is set according to the priori condition, and the average gradient contribution of minority class samples is optimized according to the expansion factor; For majority class samples and minority class samples, screening is performed based on their corresponding average gradient contribution and preset prediction probability threshold to identify noise label samples in majority class samples and minority class samples.

4. The noise label identification method based on prior confidence and heterogeneous consensus strategy as claimed in claim 3 is characterized in that: The screening is: For the majority class samples, determine the maximum value of the average gradient contribution of the majority class samples and the preset prediction probability threshold, compare the gradient contribution of each sample data in the majority class samples with the maximum value, and if it is greater than the maximum value, it is judged as a noise label sample; For minority class samples, determine the maximum value of the average gradient contribution of the minority class samples after optimization and the preset prediction probability threshold, and compare the gradient contribution of each sample data in the minority class samples with the maximum value. If it is greater than the maximum value, it is judged as a noise label sample.

5. The noise label identification method based on prior confidence and heterogeneous consensus strategy as claimed in claim 1 is characterized in that: Two heterogeneous classifiers are trained using clean datasets, namely: Multilayer Perceptron MLP and K-Nearest Neighbor Model KNN; The pseudo labels of each sample data in the noise data set are predicted by the two trained heterogeneous classifiers to obtain the corresponding prediction probability.

6. The noise label identification method based on prior confidence and heterogeneous consensus strategy as described in claim 5 is characterized in that: Based on pseudo labels and preset probability consensus thresholds, a heterogeneous consensus strategy is used to filter out potential pseudo minority sample data in the noisy dataset, including: For each sample data in the noise data set, when two classifiers predict that the same sample data has the same pseudo label and the predicted probability of the pseudo label is greater than the probability consensus threshold, it is identified as a potential minority sample to construct a pseudo minority data set.

7. The noise label identification method based on prior confidence and heterogeneous consensus strategy as claimed in claim 1 is characterized in that: Integrate the clean dataset and the pseudo minority dataset to obtain the final dataset with correct labels.

8. A noise label recognition system based on prior confidence and heterogeneous consensus strategy, characterized in that: include: The label sample acquisition module is used to obtain a number of labeled sample data to construct an original data set; wherein the sample data in the original data set is divided into majority class samples and minority class samples according to the number of labels; The noise label sample identification module is used to calculate the gradient contribution of each sample data in the original data set using a binary classification model based on cross-validation; taking category imbalance as a priori condition, the gradient contribution of sample data is used as a confidence metric for the annotation label, and a priori confidence strategy is adopted to identify the noise label samples in the original data set and construct a noise data set; The pseudo minority sample identification module is used to obtain a clean data set from the original data set by removing the noise data set, and use the clean data set to train a heterogeneous classifier; use the heterogeneous classifier to predict the pseudo labels of each sample data in the noise data set, and use the heterogeneous consensus strategy to screen out potential pseudo minority sample data in the noise data set based on the pseudo labels and the preset probability consensus threshold.

9. An electronic device, characterized in that: The method comprises a memory and a processor, and computer instructions stored in the memory and executed on the processor. When the computer instructions are executed by the processor, the steps of a noise label identification method based on a priori confidence and heterogeneous consensus strategy as described in any one of claims 1 to 7 are completed.

10. A computer-readable storage medium, characterized in that: Used to store computer instructions, which, when executed by a processor, complete the steps of a noise label identification method based on prior confidence and heterogeneous consensus strategy as described in any one of claims 1-7.

Citation Information

Patent Citations

  • Label noise correction method based on heterogeneous integration

    CN111814883A

  • Automatic identification and cleaning method for traditional Chinese medicine tongue color noise labeling sample

    CN114037011A

  • Virtual traffic anomaly detection method and system based on weighted adaptive ensemble learning, and intelligent traffic monitoring platform

    CN114785548A

  • Image classification method and system based on label noise domain self-adaption

    CN114842267A

  • Active learning sample labeling method based on latent layer representation

    CN115937592A

Cited By

  • Data processing method and system, terminal and storage medium

    CN120011915A