Workload identification method in presence of label noise

Abstract

The application discloses a workload recognition method with label noise, and belongs to the field of physiological signal recognition. The application firstly pre-trains a workload recognition model with three different regularization methods to reduce overfitting caused by the influence of label noise on the model; then, whether the label of the data is noisy or not is judged based on the pre-trained model, the loss of each sample is firstly judged, and then the judgment result is corrected through the feature distribution of the noisy sample and the non-noisy sample; finally, the sample judged as noisy is regarded as a no-label sample and semi-supervised learning is performed; the application proposes a label noisy learning mode aiming at the problem that the data label of workload data may be noisy, reduces the influence of label noise on the workload recognition model, increases the generalization of the model, and better guarantees the recognition accuracy of the workload recognition model.

Description

Technical Field

[0001] This invention relates to the field of physiological signal recognition, and more particularly to a method for identifying workloads with tag noise. Background Technology

[0002] Since most existing workload datasets use scales to label the data, this labeling method has two main problems: (1) from the perspective of scale design: the description of the scale is ambiguous; (2) from the perspective of the subjects: the subjects' understanding of the scale is prone to bias, and different subjects will have different understandings of the scale. Therefore, the labels obtained by labeling through scales may be inaccurate, resulting in noisy labels in the dataset. This may lead to overfitting of the model on noisy labels, thereby reducing the generalization ability of the entire model. Many researchers have tried to solve the problem of noisy dataset labels and proposed various algorithms, which have achieved good results. In particular, the algorithm based on semi-supervised learning first identifies clean samples and noisy samples, treats the noisy samples as unlabeled data, and then performs semi-supervised learning. It greatly reduces the impact of noisy data, so it has received more attention in the past two years and has high research value.

[0003] To address the issue of insufficient generalization in models trained under noisy label conditions, many scholars have attempted to propose different methods. The first type of method uses robust loss functions. These methods often design loss functions that are robust to label noise, reducing the impact of label noise. Ghosh et al. demonstrated that for classification tasks with noisy labels, Mean Absolute Error (MAE) is more robust than Cross Entropy Loss (CE). Zhang et al., addressing the issue of CE's lack of robustness to label noise, combined the robustness of MAE to noise with the training stability of CE, proposing a generalized cross-entropy loss, GCE Loss.

[0004] The second type of method is based on sample importance weighting. The basic idea of ​​this type of method is to assign lower weights to noisy samples and higher weights to clean samples, allowing the model to reduce the impact of noise. Sukhbaatar et al. reduced the weights of noisy data and increased the weights of clean data, but this resulted in the loss of a significant amount of information and a limited upper limit to generalization ability. Litany et al. treated this problem as transfer learning, where the source domain is the original data (with noisy labels), and the target domain is a subset of the clean dataset. The data in the source domain are then weighted and arranged to minimize the loss in the target domain.

[0005] In the past two years, many researchers have attempted to address this problem using semi-supervised learning methods, achieving relatively good results. This type of method first divides the training set into noisy and clean data using certain methods, treating the noisy data as unlabeled data and the clean data as labeled data, and then performing semi-supervised learning on the labeled and unlabeled data separately. Li et al. attempted to use the magnitude of the loss as an indicator to determine whether a label is noisy. Gaussian mixture models can separate noisy and clean labels, and during training, the loss keeps the model's entropy at a high level, minimizing the influence of noisy labels on the model. Further, the authors achieved good results with a semi-supervised learning method that improved the MixMatch algorithm.

[0006] Geoff et al. used another method to determine whether the labels were noisy. They calculated the AUM (Area Under the Margin) during the training process, which is the sum of the logit output of the network, to determine whether the labels of the corresponding samples were noisy. The authors discarded the data that were judged to be noisy, built a relatively clean training dataset, and trained a model with good generalization based on this dataset.

[0007] While the aforementioned algorithms have improved the generalization of label noise learning to some extent, several problems remain. First, methods other than semi-supervised learning often only mitigate the negative impact of label noise without fundamentally correcting the data. This characteristic limits their accuracy, resulting in a low upper limit. Second, while some existing semi-supervised learning algorithms have achieved good results, they still have issues in determining whether sample labels are noisy. These algorithms rely on a specific metric (e.g., the DivideMix method uses the loss magnitude) to determine data noise, and the results may still contain some noise. These methods do not correct the determination at the feature level, potentially leading to noise in the final result and affecting subsequent semi-supervised learning outcomes. Summary of the Invention

[0008] This invention addresses the problem of noise in workload data labels by proposing a workload identification method with label noise. This method aims to reduce the impact of label noise on the workload identification model, increase its generalization ability, and ensure identification accuracy.

[0009] The technical solution adopted in this invention is as follows:

[0010] A method for identifying workloads with tag noise, comprising the following steps;

[0011] Step 1: Pre-train the workload identification model based on the original dataset;

[0012] During pre-training, the workload labels output by the workload recognition model are smoothed, and the training loss of the samples is calculated using entropy maximization loss. When calculating the gradient based on the training loss, the gradient is clipped and the model parameters are tuned using the gradient descent algorithm.

[0013] Step 2: Based on the labels of the training dataset of the workload recognition model, perform label discrimination to distinguish between noisy datasets and clean datasets (noisy datasets);

[0014] Then, the discrimination result is corrected based on the feature distribution of the samples: the contribution ratio of the samples in gradient update is configured based on the feature distribution of the samples. The contribution ratio of densely distributed samples is large, and the contribution ratio of sparsely distributed samples is small. For example, the learning weight of sparsely distributed samples is reduced, or the learning weight of densely distributed samples is increased, or the learning weight of densely distributed samples is increased and the learning weight of samples with distribution coefficients is reduced.

[0015] Step 3: Using the noisy dataset as the unlabeled dataset and the clean dataset as the labeled dataset, supervise the learning of the workload identification model based on the correction results to obtain the final workload identification model.

[0016] Furthermore, the workload identification model includes a time-frequency feature extraction network and a classifier. The time-frequency feature extraction network is used to extract the time-domain signal features and frequency-domain signal features of the physiological signal. The time-domain signal features and frequency-domain signal features are then concatenated and fused to obtain the final signal features, which are then input into the classifier. The classifier is used to output the classification probability of the workload category to which the current signal features belong.

[0017] Furthermore, the time-frequency feature extraction network extracts the final signal features, specifically including:

[0018] The physiological signals of the sample are segmented based on a preset cutting window size, dividing the physiological signals into several signal segments.

[0019] Each segment is input into a convolutional neural network for dimensionality reduction to obtain the encoded representation of each segment;

[0020] The physiological signal encoding composed of all the segments is then input into the Transformer encoder for feature extraction to obtain the temporal features of the sample.

[0021] Frequency domain features of the sample are extracted based on the attention mechanism.

[0022] The time-domain and frequency-domain features are spliced ​​and fused to obtain the final signal features of the sample.

[0023] Furthermore, the label smoothing process specifically involves:

[0024] Encode p for each label encoding position of the sample i Set to:

[0025] If the predicted label of the sample matches the true label, then p i Let τ be a number less than 1 and greater than 0;

[0026] If the predicted label of a sample is inconsistent with the true label, then p i Set as Where C represents the number of categories identified by the workload identification model.

[0027] Furthermore, the expression for the entropy-maximizing loss is:

[0028]

[0029] Where C represents the number of categories identified by the workload identification model, p(x i ) represents the probability distribution of the i-th type of sample output by the workload identification model.

[0030] Furthermore, the gradient clipping process specifically involves: if the L2 norm of the calculated gradient is less than or equal to the first threshold `clip_value`, it remains unchanged; otherwise, the gradient is adjusted to... Here, norm(ω) represents the L2 norm of the gradient ω.

[0031] Furthermore, when correcting the discrimination results based on the feature distribution of the samples, the density of the region where the sample is located is represented by the Euclidean distance of the features.

[0032] Furthermore, in step 2, when correcting the discrimination result based on the feature distribution of the sample, the evolved K-nearest neighbor entropy is calculated based on the entropy of the probability distribution obtained after the sample is input into the K-nearest neighbor classifier.

[0033] If the K-nearest neighbor entropy of a sample is greater than a preset second threshold, the current sample is determined to be in a low-density region of the feature space; if the K-nearest neighbor entropy of a sample is less than a preset third threshold, the current sample is determined to be in a high-density region of the feature space; when correcting the discrimination result, the learning weight of samples in the low-density region is reduced and / or the learning weight of samples in the high-density region is increased.

[0034] The technical solution provided by this invention brings at least the following beneficial effects:

[0035] This invention employs a semi-supervised learning workload identification mechanism that first determines the label and then corrects it to address the problem of label noise in workload identification. First, the invention warms up the neural network using raw data. Then, it divides the raw data into a clean dataset and a noisy dataset. Finally, it treats the noisy dataset as unlabeled data and performs semi-supervised learning. This invention proposes a novel physiological signal feature extraction network that utilizes three different methods to reduce overfitting to noisy data during the warm-up stage, increasing the model's robustness and improving the accuracy of subsequent label noise discrimination. Furthermore, it uses information from feature distribution to correct the label noise discrimination results, thereby increasing the generalization ability of the identification model and better ensuring the accuracy of the workload identification model. Attached Figure Description

[0036] To more clearly illustrate the technical solutions in the embodiments of the present invention, the accompanying drawings used in the description of the embodiments will be briefly introduced below. Obviously, the accompanying drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0037] Figure 1 This is a schematic diagram illustrating the processing steps of the workload identification method with tag noise provided in an embodiment of the present invention. Detailed Implementation

[0038] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be described in detail and completely below with reference to the accompanying drawings. Obviously, the described embodiments are only a part of the embodiments of this application, and not all of them. Generally, the components of the embodiments of the present invention described and shown in the accompanying drawings can be arranged and designed using different configurations. Therefore, the following detailed description of the embodiments of the present invention provided in the accompanying drawings is not intended to limit the scope of the claimed application, but merely represents selected embodiments of the present invention.

[0039] This invention provides a method for identifying workloads with label noise. The method proposes a First Discriminate Then Correct Semi-supervised Learning Workload Recognition (FDTC) algorithm to address the problem of label noise in workload identification. The overall approach consists of three stages: first, warming up the neural network with raw data; then, dividing the raw data into clean datasets and noisy datasets; and finally, treating the noisy datasets as unlabeled data to perform semi-supervised learning on the recognition model. The main work is in the first two stages, which mainly include three points: (1) proposing a new physiological signal time-frequency feature extraction network; (2) using three different methods to reduce overfitting of noisy data in the warm-up stage; and (3) using feature distribution information to correct the label noise discrimination results. First, in the warm-up stage, three methods are used to try to alleviate the model's overfitting to noisy labels, increase the model's robustness, and improve the subsequent accuracy of label noise discrimination. In addition, the discrimination results based on loss magnitude are corrected using the feature distribution of the data.

[0040] The workload identification model employed in the method proposed in this embodiment of the invention includes: a time-frequency feature extraction network (for physiological signals) and a classifier. The time-frequency feature extraction network extracts time-domain and frequency-domain signal features of physiological signals (such as electroencephalogram signals), then concatenates and fuses these features to obtain the final signal features, which are then input into the classifier. The classifier outputs the classification probability of the workload category to which the current signal features belong. See also... Figure 1 The specific implementation process of the workload identification method with labeled noise provided in this embodiment of the invention includes:

[0041] The purpose of Phase 1, the warm-up phase, is to train a model that is not overfitted to noisy data as much as possible. Therefore, the method proposed in this embodiment of the invention uses three different methods to alleviate the overfitting of the model during the entire pre-training phase, so as to increase the accuracy of subsequent noise label discrimination.

[0042] (1) Entropy maximization loss

[0043] Entropy measures the degree of disorder in a system, and its expression is:

[0044]

[0045] Where p(·) represents the probability distribution of the recognition model output, x i Let represent the i-th class of samples, and C represent the number of classes in the dataset.

[0046] Optimizing a recognition model using traditional cross-entropy loss can lead to overconfidence in the model's output, causing it to overfit to noisy data and reducing its generalization ability. Therefore, the model's output needs to be less confident to reduce overfitting to noise. This invention utilizes the definition of entropy, which measures the disorder of a system, to calculate the negative entropy of the probability distribution of the final output of the recognition model and add it to the loss function. An optimizer is then used to try to keep the entropy from becoming too small, thus reducing overfitting to label noise. The entropy maximization loss can be expressed as:

[0047]

[0048] (2) Gradient clipping

[0049] Deep neural network optimization often uses gradient descent, an algorithm based on the gradient (first derivative) of network parameters. A larger gradient results in a larger update magnitude for the neural network. A larger update magnitude increases the likelihood of the neural network overfitting to noisy labels, thus reducing its generalization ability. Therefore, it is necessary to constrain the gradient during the neural network optimization process. Gradient clipping is a method to avoid excessively large gradients during optimization. Its basic idea is that if the L2 norm of the gradient exceeds a certain threshold `clip_value`, the gradient is scaled so that its scaled L2 norm equals `clip_value`. Gradient clipping can be expressed by the formula:

[0050]

[0051] in, Let ω represent the gradient after gradient clipping, and norm(ω) represent the L2 norm of the gradient.

[0052] (3) Label smoothing

[0053] Generally, one-hot encoding (also called hard labeling) is used when training deep neural networks. One-hot encoding represents the correct class as 1 and the incorrect class as 0. Assuming a sample has a moderately heavy correct label and the dataset contains three classes, its corresponding one-hot encoding is (0, 1, 0). It can be seen that one-hot labeling makes the trained neural network overconfident, easily leading to overfitting to noisy labels. Therefore, this embodiment of the invention uses label smoothing technology to reduce the overfitting of the neural network to noise. Label smoothing does not set the encoding of the true class to 1, but rather a number τ that is less than 1 and greater than 0. It also does not set the encoding of other classes to 0, but rather... C represents the number of categories. The code p for each position... iThe following formula can be used to express this:

[0054]

[0055] Where i is the sample number and y is the true category (true label).

[0056] It should be noted that, in the implementation of this invention, workload categories are predicted based on a recognition model. The granularity (number of categories) of the workload recognition results depends on the dataset used. For example, the STEW dataset classifies workload levels into two categories, while the CONTEST dataset triggers three different levels of workload in the subjects. To mitigate model overfitting, the labels are first smoothed. Then, entropy maximization loss is used as the classification loss during model training. Finally, when backpropagating using gradient descent to optimize model parameters based on the calculated loss, gradient clipping is performed before backpropagation.

[0057] In stage 2, after the warm-up stage, it is necessary to use the trained time-frequency feature extractor (time-frequency feature extraction network) to judge the label noise of the original data. In addition to using loss to judge the labels, the method proposed in this embodiment of the invention also corrects the judgment result based on the feature distribution.

[0058] High-density region criterion: This criterion primarily applies to unlabeled data (referring to noisy labeled data obtained by loss-based discrimination methods). It states that the true label of unlabeled data falling within a high-density region (a high-density region is defined as a region in the feature space where the density of labeled samples of the same class approaches 1) is usually the same as the label of the corresponding high-density region. High-density regions are characterized by low sample mixing and high data density.

[0059] Low-density region criterion: This criterion mainly applies to labeled data (referring to noisy data with labels obtained by loss-based discrimination methods). It states that labeled data falling within low-density regions (where the density of labeled samples of the same class in the feature space approaches 0) typically have labels different from the true labels. Low-density regions are characterized by high sample mixing and low data density.

[0060] This invention uses the Euclidean distances of the K nearest neighbors to a sample point's features to represent the density of the region where the sample is located. A larger average distance indicates a lower sample density in the region, while a smaller average distance indicates a higher sample density. The entropy of the class distribution of these K nearest neighbors is used to measure the mixing degree of the samples. A larger entropy indicates a higher mixing degree in the region, while a smaller entropy indicates a lower mixing degree. Therefore, considering both of these points, this invention, based on the K nearest neighbor algorithm, considers the distance factor of the K nearest neighbors and calculates the entropy of the classification probability distribution to measure the sample distribution density near a sample. Using p... i Let p represent the probability distribution obtained after the i-th sample is input into the K-nearest neighbor classifier. Then, we can calculate this probability distribution p. i The entropy of a sample is called the K-Nearest Neighbors Entropy (KNNE). If the K-Nearest Neighbors Entropy of a sample is greater than a certain threshold, the sample can be considered to be in a low-density region of the feature space; if the K-Nearest Neighbors Entropy of a sample is less than a certain threshold, the sample can be considered to be in a high-density region of the feature space. Based on the high-density region criterion, it can be seen that when the features of unlabeled data fall into a high-density region of a certain class, the label of the unlabeled data should be the same as that class. Therefore, this embodiment of the invention proposes a first correction rule: for a certain unlabeled data, if its K-Nearest Neighbors Entropy is less than the threshold α, then the unlabeled data can be labeled with the corresponding K-Nearest Neighbors algorithm output label. (Using u...) i If we represent unlabeled data features, then the method for calculating their K-nearest neighbor entropy is as follows:

[0061] KNNE(u i = Entropy(KC(u i )),

[0062] Where KC(·) represents the K-nearest neighbor classifier, and KNNE(·) represents the K-nearest neighbor entropy.

[0063] Furthermore, the low-density region criterion can be used to determine the reliability of certain labeled samples. If the K-nearest neighbor entropy of a labeled sample is higher than a threshold β, then the sample is considered to be in a low-density region. If some labeled samples are in low-density regions, then the actual label of the labeled sample may be incorrect. Based on this, embodiments of the present invention reduce the actual learning weight of all samples in low-density regions.

[0064] Based on the high-density region criterion and the low-density region criterion, this invention shows that the contribution ratio of different samples to gradient updates varies depending on their feature distribution. Therefore, the learning weights can be adjusted. For example, the actual learning weights of all samples in the low-density region can be reduced, and the specific reduction can be set based on the actual application scenario.

[0065] In stage 3, after loss-based label noise discrimination processing, labeled and unlabeled datasets are obtained. Then, based on a selected semi-supervised learning algorithm, the noisy labeled dataset is treated as unlabeled data for semi-supervised learning. This yields a recognition model for workload identification. For example, the semi-supervised learning algorithm used is the FixMatch algorithm. FixMatch requires strong and weak enhancement of the data. In this embodiment, the enhancement methods used for the physiological signals include: adding Gaussian white noise (Jitter), scaling, random permutation, random clipping, and random masking.

[0066] Transformers have shown significant advantages in processing time-series data. However, directly inputting physiological signals into the Transformer Encoder for feature extraction can lead to overfitting. To address this issue, a series of processing steps are required for the physiological signals. The input physiological signals are divided into multiple time windows, each of which undergoes dimensionality reduction via a convolutional neural network. Finally, the output encodes the signals within that window, reducing the overall model's parameters and thus mitigating overfitting.

[0067] In this embodiment of the invention, the size of the cutting window for the physiological signal is defined as W, and the step size is S. Based on this cutting window, the entire physiological signal x can be segmented. i It was divided into several segments:

[0068]

[0069] Where K represents the number of segments, S i This represents the segmentation result of sample i, i.e., the fragment set. Let j represent the j-th segment, where j = 1, 2, ..., K.

[0070] Then each segment Input to coding layer E m This yields the encoded representation of each segment: Thus, the encoded representation based on K segments Obtain the encoding of this physiological signal

[0071] The encoded physiological signal is then input into the Transformer Encoder for feature extraction, as shown in the formula:

[0072]

[0073] Where En(·) is the Transformer Encoder. The obtained time-domain features.

[0074] The entire feature extraction process of a time-domain signal can be expressed by the following formula:

[0075]

[0076] in, This represents the entire time-domain feature extraction process, including the division of physiological signal time windows and feature extraction based on Em and En.

[0077] Different physiological signals have different dominant frequencies. For example, the dominant frequency of an electrocardiogram (ECG) signal is 0.5-45 Hz, while that of an electroencephalogram (EEG) signal is 0.5-80 Hz. Therefore, noise outside the dominant frequency should be discarded, and the model should focus more on the dominant frequency. Furthermore, the importance of the dominant frequencies of physiological signals is not the same. For example, workload is more relevant to the alpha band (8-12 Hz) and theta band (4-8 Hz) of an EEG signal. Based on this, this chapter uses an attention mechanism for the frequency domain feature extraction network to make the model focus on those frequency bands that are more important to the workload, thereby improving the generalization ability of the network model. The entire feature extraction process of the frequency domain signal can be expressed by the following formula:

[0078]

[0079] in This represents a frequency domain feature extractor. This represents the extracted frequency domain features.

[0080] The time-domain feature extractor and the frequency-domain feature extractor constitute the time-frequency feature extractor (i.e., the time-frequency feature extraction network). After extracting the time-domain and frequency-domain features, the two types of features are fused using a concatenation method to obtain the final physiological signal features, as expressed in the following formula:

[0081]

[0082] Then, the features of each sample, fea iThe input is fed into a classifier (such as a cascaded fully connected layer and a fully connected layer with a softmax function). Based on its output and label smoothing, entropy maximization loss, and gradient clipping, the time-frequency feature extraction network and the classifier are trained. Based on the trained model, the loss of each sample is used to make a discrimination, and then the discrimination result is corrected by the feature distribution of noisy samples and noisy samples (clean samples). Finally, the samples that are classified as noisy are treated as unlabeled samples, and the clean samples are treated as labeled samples. The recognition model (or only its classifier) ​​is semi-supervised learning is performed based on unlabeled and labeled samples to obtain the final workload recognition model.

[0083] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, and not to limit them; although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features; and these modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present invention.

[0084] The above descriptions are merely some embodiments of the present invention. Those skilled in the art can make various modifications and improvements without departing from the inventive concept of the present invention, and these all fall within the scope of protection of the present invention.

Claims

1. A method for identifying workloads with tag noise, characterized in that, Includes the following steps: Step 1: Pre-train the workload identification model based on the original dataset; The workload identification model includes a time-frequency feature extraction network and a classifier. The time-frequency feature extraction network is used to extract the time-domain signal features and frequency-domain signal features of physiological signals. The time-domain signal features and frequency-domain signal features are then concatenated and fused to obtain the final signal features, which are then input into the classifier. The classifier outputs the classification probability of the workload category to which the current signal feature belongs; During pre-training, the workload labels output by the workload recognition model are smoothed, and the training loss of the samples is calculated using entropy maximization loss. When calculating the gradient based on the training loss, the gradient is clipped and the model parameters are tuned using the gradient descent algorithm. Step 2: Based on the labels of the training dataset of the workload recognition model, perform label discrimination to distinguish between noisy datasets and clean datasets; The discrimination results are then corrected based on the feature distribution of the samples: the contribution ratio of samples during gradient update is configured based on their feature distribution, with densely distributed samples contributing a larger proportion and sparsely distributed samples contributing a smaller proportion. Step 3: Using the noisy dataset as the unlabeled dataset and the clean dataset as the labeled dataset, supervise the learning of the workload identification model based on the correction results to obtain the final workload identification model.

2. The method as described in claim 1, characterized in that, The time-frequency feature extraction network extracts the final signal features, specifically including: The physiological signals of the sample are segmented based on a preset cutting window size, dividing the physiological signals into several signal segments. Each segment is input into a convolutional neural network for dimensionality reduction to obtain the encoded representation of each segment; The physiological signal encoding composed of all the segments is then input into the Transformer encoder for feature extraction to obtain the temporal features of the sample. Frequency domain features of the sample are extracted based on the attention mechanism. The time-domain and frequency-domain features are spliced ​​and fused to obtain the final signal features of the sample.

3. The method as described in claim 1, characterized in that, Tag smoothing processing specifically includes: Encoding each label encoding position of the sample Set to: If the predicted label of the sample matches the true label, then... Set as ,in, A number that is less than 1 and greater than 0; If the predicted label of a sample is inconsistent with the true label, then... Set as Where C represents the number of categories identified by the workload identification model.

4. The method as described in claim 1, characterized in that, The expression for the entropy maximization loss is: Where C represents the number of categories identified by the workload identification model. This represents the probability distribution of the i-th type of sample output by the workload identification model.

5. The method as described in claim 1, characterized in that, The gradient clipping process specifically involves: if the L2 norm of the calculated gradient is less than or equal to the first threshold... If the gradient is unchanged, it remains the same; otherwise, the gradient is adjusted to... ,in, Gradient The L2 norm.

6. The method as described in claim 1, characterized in that, When correcting the discrimination results based on the feature distribution of the samples, the density of the region where the sample is located is represented by the Euclidean distance of the features.

7. The method as described in claim 1, characterized in that, In step 2, when correcting the discrimination result based on the feature distribution of the sample, the evolved K-nearest neighbor entropy is calculated based on the entropy of the probability distribution obtained after the sample is input into the K-nearest neighbor classifier. If the K-nearest neighbor entropy of a sample is greater than a preset second threshold, the current sample is determined to be in a low-density region of the feature space; if the K-nearest neighbor entropy of a sample is less than a preset third threshold, the current sample is determined to be in a high-density region of the feature space; when correcting the discrimination result, the learning weight of samples in the low-density region is reduced and / or the learning weight of samples in the high-density region is increased.

Images