A Hybrid Sampling and Denoising Optimization Method for Imbalanced Data
By constructing an adaptive hybrid sampling and denoising optimization method, the problem of insufficient classification performance under high noise and high imbalance data is solved, achieving high accuracy and stability in fields such as medical diagnosis, and improving the robustness and generalization ability of the model.
Patent Information
- Application Number
- CN202610785749.4
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2026-06-02
- Publication Date
- 2026-11-17
AI Technical Summary
Existing hybrid sampling methods struggle to adaptively optimize sampling strategies when faced with high-noise and highly imbalanced data, leading to decreased minority class recognition capabilities, blurred decision boundaries, and insufficient model generalization performance. This makes it particularly difficult to meet the requirements of accuracy, stability, and robustness in high-risk fields such as medical diagnosis, financial fraud detection, industrial fault early warning, and network intrusion detection.
A hybrid sampling and denoising optimization method for imbalanced data is constructed. The number of samples is adaptively determined by a non-convex and non-smooth optimization model, and noise interference is suppressed by a denoising optimization model. The feature distribution of the generated samples is optimized, and KL divergence and covariance matrix constraints are combined to improve data quality and classification performance.
It effectively improves classification performance and stability in imbalanced and noisy data environments, enhances the ability to distinguish minority class samples, avoids the loss of information in majority class samples, improves the robustness and generalization ability of the model, and is applicable to a variety of classification algorithms and datasets.
Smart Images

Figure CN123115991A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of imbalanced classification technology in machine learning, specifically a hybrid sampling and denoising optimization method for imbalanced data. Background Technology
[0002] In machine learning classification tasks, class imbalance and label noise are long-standing key problems. Class imbalance typically manifests as a situation where the number of majority class samples is much greater than the number of minority class samples, resulting in a long-tailed distribution of the dataset's classes. Commonly used imbalance rates (...) This phenomenon is measured using [a certain method / method]. In high-risk applications such as medical diagnosis, fraud detection, facial recognition, and social network analysis, minority class samples often correspond to key events, and their recognition accuracy directly affects system performance and practical application value. However, traditional classification models typically optimize overall accuracy during training, which can easily lead to a shift in the decision boundary towards the minority class, thus significantly reducing the recall and F1 score of minority class samples.
[0003] On the other hand, the presence of label noise further exacerbates the above problems. Label noise interferes with the model training process, changes the gradient direction of the loss function, and reduces the model's generalization ability and prediction accuracy. When class imbalance and label noise coexist, their negative effects are superimposed or even amplified, leading to a significant decline in model performance. Therefore, how to effectively suppress the impact of noise in imbalanced data environments has become an urgent technical problem to be solved.
[0004] In existing technologies, solutions to the class imbalance problem are mainly divided into two categories: data-layer methods and algorithm-layer methods. Data-layer methods adjust the class distribution by preprocessing the training data, including undersampling, oversampling, and mixed sampling methods. Undersampling achieves class balance by reducing the number of majority class samples, but it is prone to losing important information. Oversampling improves the class distribution by increasing the number of minority class samples, with typical methods such as SMOTE and its improved algorithms, but it is prone to introducing noisy samples in overlapping class regions and blurring decision boundaries. Mixed sampling methods combine the above two strategies, which to some extent balances data distribution and information preservation, but their sampling process usually relies on empirical settings and lacks a unified optimization mechanism.
[0005] Algorithmic methods improve the classification model itself to adapt to imbalanced data, such as by introducing cost-sensitive learning, sample weighting mechanisms, or designing new classification criteria. While these methods can alleviate class bias to some extent, they typically depend on specific model structures, have poor generality, and struggle to simultaneously address noise issues.
[0006] Existing hybrid sampling methods still have the following shortcomings: First, most methods ignore the distribution characteristics of the dataset itself and uniformly regard the optimal class ratio after rebalancing as 1:1; second, when generating new samples or selecting retained samples, they fail to fully consider the local data distribution and probability characteristics, which can easily lead to generated samples intruding into the majority class region or accidentally deleting key boundary samples, thereby affecting classification performance; in addition, the noise detection mechanism of most methods is relatively simple, and they use sample deletion strategies when dealing with noisy samples and outliers, failing to make full use of the effective information contained in the original data.
[0007] Therefore, it is crucial to construct a unified optimization framework that adaptively determines the sampling strategy and optimizes the feature distribution of generated samples while simultaneously considering class distribution, imbalance, and noise effects, thereby improving the model's classification performance and robustness. Especially in high-risk fields such as medical diagnosis, financial fraud detection, industrial fault early warning, and network intrusion detection, existing hybrid sampling methods tend to suffer from reduced minority class recognition ability, blurred decision boundaries, and insufficient model generalization performance when faced with high-noise and high-imbalance data. This makes it difficult to meet the accuracy, stability, and robustness requirements of real-world scenarios. Therefore, it is necessary to construct a hybrid sampling optimization framework that can adaptively optimize the number of samples and jointly optimize the sample feature distribution for imbalanced and noisy data. Summary of the Invention
[0008] To solve or partially solve the problems existing in the prior art, the present invention provides a hybrid sampling and denoising optimization method for imbalanced data.
[0009] To achieve the above objectives, the technical solution of the present invention is as follows: A hybrid sampling and denoising optimization method for imbalanced data is provided, with the specific steps as follows: S1. Data Input and Parameter Initialization Obtain the original imbalanced binary classification dataset Divide it into majority class sample sets With minority class sample set Extracting the dataset Feature Dimensions And the number of samples m, the number of samples in the majority class Minority class sample size And select noise rate ; Calculate the imbalance rate , Simultaneously, calculate the mean vectors of the majority class and minority class samples respectively. The covariance matrices of the majority class and minority class samples are calculated separately as follows: .
[0010] S2. Construct a hybrid sampling optimization model Based on the features extracted by S1, with the number of oversampled samples and undersampling quantity As decision variables, the following non-convex and non-smooth optimization model is constructed: The constraints are: in, It is the objective function. This represents the minimum value of the objective function; It is an adaptive bias weight. The larger the value, the closer the two class centers are in the feature space. It is an adaptive variance weight. The larger the value, the stronger the anisotropy; It is the adaptive regularization coefficient; Tr(·) is the sum of the diagonal elements of the matrix. and The specific definitions are as follows: The objective function includes a prior bias term. Structural variance term and noise penalty items Prior bias measures the deviation between the sampled class prior and the optimal equilibrium state. Considering the anisotropy of the Gaussian distribution, a variance term is introduced, where the variance of the parameter estimates is related to the effective sample size and the trace. To prevent oversampling from amplifying noise and undersampling from causing loss of dimensional information, a noise penalty term is introduced.
[0011] S3. Solving for the optimal number of samples Under the premise of satisfying the constraints, the optimal oversampling quantity is obtained by solving the optimization model of S2 using the sequential quadratic programming method or the interior point method. With the optimal number of undersamples And calculate the optimal sampling ratio. The optimal solution exists theoretically, but it does not have an analytical expression. The proof is as follows: right Taking the first-order partial derivative, we get: This formula strikes a balance between "class equilibrium" and "noise risk." If Larger penalties will limit Growth.
[0012] right Taking the first-order partial derivative, we get: This formula shows that when anisotropy ( When it is significantly large, it should be reduced. The value of is determined to maintain the structural integrity of the feature space.
[0013] set up , Then the bias term is For bivariate functions By taking the second derivative and calculating its Hessian matrix, we can obtain: , matrix The determinant is: In the formula, To The second-order partial derivative measures the bias term. The curvature of the majority class sample size after undersampling. Positive values indicate convexity, and negative values indicate concavity; To measure the effect of a change in u on the partial derivatives of u and v (mixed partial derivatives), we first take the partial derivatives with respect to u. The degree of influence of the derivative with respect to v, i.e., the interaction between the two variables; To first take the partial derivative with respect to v, and then take the partial derivative with respect to u, and... Equality (Clairaut's theorem, mixed partial derivatives are equal under continuous conditions) also reflects the interaction between variables; The determinant of the Hessian matrix; bias term Hessian matrix exist domain The matrix is non-positive definite over most of the region. Therefore, It is non-convex, and There is no closed-form solution.
[0014] Decision variables domain It is a bounded closed set, and the objective function is... exist The above is continuous. Therefore, there must exist a globally optimal solution.
[0015] S4. Perform mixed sampling operation According to S3 and Perform mixed sampling on the original dataset: undersample the majority class sample set using the ClusterCentroids (Cc.), NearMiss (Nm.), or RandomUnderSampler (Rus.) methods, and delete samples after undersampling. From these samples, we obtain the extended sample set for the majority class. Oversampling is performed on the minority class sample set using methods such as SMOTE (S.), SVMSMOTE (SvS.), BorderlineSMOTE (BS.), RandomOverSampler (Ros.), or SMOTEN (SN.). The oversampled sample set then generates... A new sample is obtained, resulting in an expanded sample set for the minority class. The final reconstructed dataset is obtained. .
[0016] S5. Construct a noise reduction optimization model Dataset after hybrid sampling Includes the following information: majority class features Minority characteristics , Feature matrix Feature distribution of the majority class , Characteristic distribution ( ),as well as Characteristic distribution ( ), The features of the newly generated minority class samples follow a normal distribution, with the mean and covariance varying with... It changes through iteration. For decision variables, the specific denoising optimization model as follows: The constraint condition for minimizing the eigenvalue of the covariance matrix is: in It is the manifold coefficient; It is the penalty coefficient. make sure and Alignment, and Noise can be suppressed by controlling the sample generation location and the elimination strategy. and The specific calculation formula is as follows: In the formula, Covariance matrix Maximum eigenvalue; Covariance matrix Minimum eigenvalue; It is the square root of the condition number of the covariance matrix. This ratio is large when the original distribution of the minority class samples has strong anisotropy (i.e., it is very narrow in some directions). Increase → The increased weighting of the generated samples results in stricter penalties for deviations from the original covariance structure, thus protecting the geometric integrity of the data.
[0017] The denoising optimization model includes manifold terms. Discriminant Item and noise reduction The discriminant term maximizes the KL divergence between the minority and majority class distributions after mixed sampling, thereby expanding the decision boundary by forcing minority features away from the majority class. The manifold term prevents oversampling of the sample set. Deviating from the original distribution If the distance is too far, the denoising term includes a penalty term. , The specific formula is as follows: in, It is a noise penalty item; This represents the Mahalanobis distance. If the generated samples are close to the high center region of the majority class, they will be affected by the noise rate. The punishment is proportional to the punishment. It is an anisotropic penalty term, which ensures that the covariance structure after oversampling does not deviate from the original anisotropic structure.
[0018] S6. Optimize the feature distribution of generated samples Under constraints, the denoising optimization model is iteratively optimized, and the feature matrix of the minority class samples is updated. This allows it to gradually converge to the optimal distribution. The constraints are the minimum eigenvalue constraints of the covariance matrix in S5, including the feature boundary p and the minimum eigenvalue of the covariance matrix. .
[0019] S7. Output the optimized training dataset. The optimized generated sample feature matrix Compared with the original dataset The matrix is vertically concatenated to obtain the final training dataset, which is used for subsequent classification model training and prediction tasks.
[0020] S8, Model Training and Application The data from the training dataset obtained in S7 is input into any classification model for training to obtain a classifier, which is then used for classification tasks in real-world application scenarios.
[0021] Through the above steps, this invention realizes a unified processing flow from "sampling quantity optimization" to "sample quality optimization", effectively improving classification performance and stability in imbalanced and noisy data environments.
[0022] This invention provides a hybrid sampling and denoising optimization method for imbalanced data, which has the following beneficial effects: (1) This invention constructs a convex non-smooth optimization model for the number of samples and a unified hybrid sampling non-convex combination optimization model to adaptively optimize the number of oversampling and undersampling for different datasets, thus avoiding the instability problem caused by relying on experience to set a ratio of approximately 1:1 in traditional methods.
[0023] (2) This invention introduces a denoising optimization model based on probability distribution to effectively suppress noise interference in generated samples and improve data quality. The denoising optimization model maximizes the difference between the generated sample distribution and the majority class distribution, while constraining the distance between it and the original minority class distribution, so that the generated samples are both far away from the majority class region and kept in the minority class safe region, thereby effectively avoiding the problems of generating noisy samples and fuzzy decision boundaries.
[0024] (3) While ensuring the discrimination ability of minority class samples, this invention avoids excessive deletion of important samples of the majority class, thereby maintaining the overall data structure information; by introducing KL divergence, covariance matrix and trace operation to characterize the data structure, the method's adaptability to complex data distribution is improved.
[0025] (4) The method of the present invention is applicable to existing hybrid sampling algorithms as a general optimization framework. By receiving the sampling results or sampling parameters generated by the hybrid sampling algorithm of 15 simple combinations of sampling algorithms, it can be used in combination with various classification algorithms in the data preprocessing stage, and has strong versatility and scalability.
[0026] (5) The non-convex, non-smooth optimization model of the present invention determines the optimal number of oversampling samples. Number of undersampled samples Furthermore, by optimizing the feature distribution model based on KL divergence to reduce the noise impact in the generated samples, it can improve classification performance and generalization ability under multiple imbalanced and noisy datasets, demonstrating strong robustness. Attached Figure Description
[0027] Figure 1This is a flowchart illustrating a hybrid sampling and denoising optimization method for imbalanced noise datasets according to the present invention.
[0028] Figure 2 The images show the effect before and after noise reduction optimization in Embodiment 1 of the present invention.
[0029] Figure 3 This is a comparison chart of the sampling results of four oversampling methods and three undersampling methods optimized using the method of this invention.
[0030] Figure 4 To achieve the noise rate = 0.15 and 0.25, the Friedman average ranking of the classifier under some evaluation metrics.
[0031] Figure 5 To achieve the noise rate = 0.35 and 0.45, the Friedman average ranking of the classifier under some evaluation metrics. Detailed Implementation
[0032] The present invention will be further described in detail below with reference to specific embodiments and accompanying drawings, but the scope of protection of the present invention is not limited to the content described.
[0033] Figure 1 Figure 1 shows a flowchart of the algorithm of the present invention. The overall process of the method of the present invention includes four stages: data preprocessing, optimization of the number of mixed samples, execution of mixed sampling, and optimization of sample denoising.
[0034] Based on the method of this invention, it is applied to the Wisconsin breast cancer binary classification dataset and the BS._Cc. hybrid sampling algorithm. The first 10 rows of the dataset are shown in Table 1. In Table 1, the first row is the header, the first to ninth columns are the nine dimensions of the data (d=9), and the last column is the category label of the data, which only includes 0 and 1, corresponding to the two types of breast cancer diagnosis results. In this embodiment, the classifier DTree is selected to form the basic hybrid sampling process.
[0035] Table 1 The classifier used in this embodiment is DTree and noise rate. = 0.25 to form the basic mixed sampling process, the specific steps are as follows: S1. Obtain the original imbalanced binary classification dataset. Divide it into majority class sample sets With minority class sample set Extracting the dataset Feature Dimensions =9 and the sample size m is 683, the majority class sample size =444, Minority class sample size =239, and select noise rate =0.25; Calculate the imbalance rate. , = =1.86. Simultaneously calculate the mean vectors of the majority and minority classes separately. The covariance matrices of the majority class and minority class samples are calculated separately as follows: .
[0036] S2, based on the features extracted by S1, with the number of oversampled features. and undersampling quantity As decision variables, the following non-convex, non-smooth optimization model is constructed according to the scheme of the present invention, and model constraints are set.
[0037] S3. Under the premise of satisfying the model constraints, solve the optimization model of S2 using the sequential quadratic programming method or the interior point method, and obtain the optimal oversampling quantity under the constraint of minimizing the eigenvalue of the covariance matrix. 78 and the optimal number of undersampling The value is 91, and the optimal sampling ratio is calculated. = =1.114.
[0038] S4, obtained from S3 (78) and (91) Perform mixed sampling on the original dataset: perform undersampling on the majority class sample set using the ClusterCentroids method, and delete 91 samples after undersampling to obtain the majority class extended sample set. Oversampling was performed on the minority class sample set using the SMOTE method, resulting in 78 new samples and thus an expanded minority class sample set. The reconstructed dataset is finally obtained. .
[0039] S5. Dataset after hybrid sampling The information contained is used to construct a denoising optimization model and set the minimum eigenvalue constraint of the covariance matrix.
[0040] The feature representation of the generated samples is used as the feature matrix of the new samples. Step S5 optimizes the feature distribution of the generated samples to reduce the interference of noisy samples on the model. The optimization effect is as follows: Figure 2 As shown, Figure 2 In the diagram, pink, purple, and red represent minority class samples, majority class samples, and oversampled samples, respectively, while the blue plane is the classification hyperplane.Figure 2 The image on the left shows the result before noise reduction optimization. Figure 2 The image on the right shows the result after noise reduction and optimization. Figure 2 It can be seen that the sample location has shifted significantly, moving to a safer location.
[0041] S6. Under constraints, iteratively optimize the denoising optimization model and update the feature matrix of the minority class samples. This allows it to gradually converge to the optimal distribution. .
[0042] S7. Optimize the generated sample feature matrix. Compared with the original dataset The matrix is vertically concatenated to obtain the final training dataset, which is used for subsequent classification model training and prediction tasks. To verify the effectiveness of the method of this invention, the evaluation indicators of the original mixed sampling algorithm (BS_CC.) and the algorithm optimized by the method of this invention (DNNOM) were recorded under the same experimental conditions. The results are shown in Table 2. To avoid randomness, there are a total of 8 sets of results under the same conditions.
[0043] S8. Input the data from the training dataset obtained in S7 into any classification model for training to obtain a classifier, and use it for classification tasks in real-world application scenarios.
[0044] Table 2 For all six metrics above, higher values generally indicate better performance. Experimental results show that, under the same classifier, noise rate, and dataset conditions, the algorithm optimized using this invention exhibits significant improvements in Recall, AUC, F1 score, and G-means compared to the original algorithm. Specifically: (1) The number of successes in Accuracy remained the same, indicating that the overall classification accuracy was basically unchanged from the original algorithm; (2) The number of Precision successes remained the same, indicating that the misdiagnosis rate was basically unchanged from the original algorithm; (3) Increased recall indicates an enhanced ability to identify positive cases and a reduced risk of missed diagnoses; (4) An improved F1 score indicates that precision and recall have reached a better balance; (5) An increase in AUC indicates that the overall discriminative ability of the classifier has been enhanced; (6) The improvement of G-mean indicates that the model has a better ability to balance the majority and minority classes.
[0045] The comparison method for the above conclusions is to compare the method of this invention with the BS._CC method in each round in order of rounds. If the number of wins is the same as that of the BS._CC method, then it is equal (e.g., in the Accuracy metric, if the method of this invention and the BS._CC method each win four times, then it is equal). If the number of wins is greater than that of the BS._CC method, then it is an improvement.
[0046] The framework of this invention can adaptively correct problems such as fixed sampling ratio, boundary sample distortion and noise amplification in traditional hybrid sampling algorithms in the presence of label noise, thereby significantly improving the stability and generalization performance of existing algorithms in complex data environments.
[0047] The optimization framework of this invention is currently applied to fifteen hybrid sampling algorithms (composed of three undersampling methods and five oversampling methods from S4). Experimental results show that various existing hybrid sampling algorithms achieve varying degrees of performance improvement after incorporating this invention, verifying that this invention has good versatility and scalability.
[0048] Three undersampling samplers (ClusterCentroids, NearMiss, RandomUnderSampler) and four oversampling samplers (SMOTE, SVMSMOTE, BorderlineSMOTE, and RandomOverSampler) were used to sample a random 3D dataset using the methods described above. The sampling results are as follows: Figure 3 As shown. Figure 3 In the image, the four columns from left to right represent the original dataset sample distribution, the sample distribution after sampling by the corresponding sampler, the sample distribution after performing the S4 sampling operation, and the sample distribution after denoising optimization. Pink, purple, and red markers represent minority class, majority class, and oversampled samples, respectively, and the blue plane is the classification hyperplane.
[0049] from Figure 3 As can be seen, after introducing this algorithm, the data distribution generated by the sampler is significantly more compact and structurally more complete. Both oversampling and undersampling methods benefit from a more reasonable sample allocation around the true decision region, thereby reducing inter-class overlap and forming a smoother, more consistent decision boundary. These observations demonstrate that this invention effectively regulates the sampling process by guiding sample distribution and suppressing the amplification effect of noisy / informative samples.
[0050] The proposed optimization method for existing hybrid sampling algorithms is applicable to binary classification data scenarios with unbalanced class distribution and accompanying label noise, such as medical diagnosis, financial risk control, industrial inspection, network security and anomaly identification, which require high accuracy in minority class identification.
[0051] Furthermore, the robustness of this algorithm under different imbalance levels further verifies its stability. Specifically, the method of this invention was used to randomly sample 17 open-source datasets (corresponding to fields including Wisconsin breast cancer binary classification, protein localization, gene function prediction, glass classification, object recognition, abalone age prediction, web page block classification, handwritten digit recognition, letter recognition, satellite image classification, diabetes diagnosis, breast cancer survival rate, climate simulation, image segmentation, waveform recognition, radar signal classification, and credit card fraud detection). The results are as follows: Figures 4-5 As shown.
[0052] like Figure 4 for When the coefficients are 0.15 and 0.25, the benchmark algorithms AdaBoost, DTree, KNN and SVM rank on different evaluation metrics when using different classifiers for classification. Figure 4 In the diagram, light red and light blue represent the original sampling algorithm and the optimized version of this invention, respectively. = 0.15, optimized average rank; dark red and dark blue represent the original sampling algorithm and the optimized average rank of this invention, respectively. The average rank when = 0.25.
[0053] like Figure 5 for When the coefficients are 0.35 and 0.45, the benchmark algorithms AdaBoost, DTree, KNN and SVM rank on different evaluation metrics when using different classifiers for classification. Figure 5 In the diagram, light yellow and light purple represent the original sampling algorithm and the optimized version of this invention, respectively. = 0.35, optimized average rank; dark yellow and dark purple represent the original sampling algorithm and the optimized version of this invention, respectively. The average rank when = 0.45.
[0054] Figure 4 and 5 The evaluation metrics include precision, recall, area under the ROC curve, F1 score, and geometric mean. The classifier combinations used in the original sampling algorithm include the classifier combinations Bls.Cc., Ros.Nm., S.Rus., SN.Cc., and SvS.Nm. from left to right for each metric in the figure. The classifier combinations for each algorithm optimized in this invention are the classifier combinations Bls.Cc._D, Ros.Nm._D, S.Rus._D, SN.Cc._D, and SvS.Nm._D.
[0055] from Figure 4 and 5It can be seen that the benchmark algorithm exhibits significant fluctuations with changes in evaluation metrics or imbalance rate, but the method of this invention has good generalization performance under various evaluation conditions.
[0056] Experimental results show that, under different imbalance ratios (e.g., 35% and 45%), the proposed algorithm exhibits lower Friedman mean rankings in the Recall, F1, and G-mean evaluation metrics combined with the classifier. Since higher values for F1, Recall, AUC, Precision, and G-mean indicate better performance, a lower Friedman mean ranking signifies better overall performance. This result strongly demonstrates that the proposed algorithm is statistically significantly superior to the comparison algorithms, rather than merely exhibiting a random advantage on specific datasets or metrics. The algorithm demonstrates a clear ranking advantage in Recall, F1, and G-mean metrics, highlighting its strong discriminative ability and effectiveness in minority class identification.
[0057] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.
Claims
1. A hybrid sampling and denoising optimization method for imbalanced data, characterized in that: The specific steps are as follows: S1. Obtain the original imbalanced binary classification dataset and divide it into a majority class sample set and a minority class sample set. Extract the features of the dataset. Calculate the imbalance rate of the dataset, the mean vector of the majority class sample set and the covariance matrix of the minority class sample set. S2. Based on the features of the dataset extracted in S1, a non-convex and non-smooth optimization model is constructed using the number of oversampling and undersampling as decision variables, and model constraints are set. S3. Under the premise of satisfying the model constraints, use the sequential quadratic programming method or the interior point method to solve the optimization model of S2, obtain the optimal oversampling number and the optimal undersampling number, and calculate the optimal sampling ratio. S4. Based on the optimal oversampling and undersampling quantities obtained in S3, perform mixed sampling on the original dataset: obtain the majority class extended sample set through undersampling; obtain the minority class extended sample set through oversampling; the remaining samples are obtained by subtracting the majority class extended sample set from the majority class sample set; all samples are obtained by adding the minority class extended sample set to the minority class sample set; and the reconstructed dataset is obtained by combining the remaining samples and all samples. S5. Based on the reconstructed dataset after S4 mixed sampling, construct a denoising optimization model and set the minimum eigenvalue constraint condition of the covariance matrix; S6. Iteratively optimize the denoising optimization model, update and reconstruct the minority class sample feature matrix in the dataset, so that the minority class sample feature matrix gradually converges to the optimal distribution. S7. Concatenate the sample feature matrix of the optimal distribution obtained in S6 with the original imbalanced binary classification dataset vertically to obtain the final training dataset. S8. Input the data from the training dataset obtained in S7 into any classification model for training, and then use the trained classification model for classification tasks in real-world application scenarios.
2. The hybrid sampling and denoising optimization method for imbalanced data according to claim 1, characterized in that: In S1, let's assume we obtain the original imbalanced binary classification dataset. Divide it into majority class sample sets With minority class sample set Extracting the dataset Features Includes: feature dimensions And the number of samples m, the number of samples in the majority class Minority class sample size And select noise rate ; Calculate the imbalance rate , ; Calculate the mean vector of the majority class and the minority class samples. The covariance matrices of the majority class and minority class samples are calculated separately as follows: .
3. The hybrid sampling and denoising optimization method for imbalanced data according to claim 2, characterized in that: The specific process of S2 is as follows: using the oversampling quantity and undersampling quantity As decision variables, the following non-convex and non-smooth optimization model is constructed: The model constraints are as follows: in, It is the objective function. This represents the minimum value of the objective function; It is an adaptive bias weight; It is an adaptive variance weight; It is the adaptive regularization coefficient; and The specific definitions are as follows: The objective function includes a prior bias term. Structural variance term and noise penalty items .
4. The hybrid sampling and denoising optimization method for imbalanced data according to claim 3, characterized in that: Solve the optimization model in S3 to obtain the optimal oversampling quantity. With the optimal number of undersamples Calculate the optimal sampling ratio The formula is: 。 5. The hybrid sampling and denoising optimization method for imbalanced data according to claim 4, characterized in that: The specific process of mixed sampling in S4 is as follows: based on the results obtained in S3... and Perform mixed sampling on the original dataset: undersample the majority class sample set using the ClusterCentroids, NearMiss, or RandomUnderSampler methods, and delete the samples after undersampling. From a sample, we obtain the majority class extended sample set. Oversampling is performed on the minority class sample set using the SMOTE, SVMSMOTE, BorderlineSMOTE, RandomOverSampler, or SMOTEN methods. The resulting sample is then oversampled. A new sample is obtained, resulting in an expanded sample set for the minority class. The final reconstructed dataset is obtained. .
6. The hybrid sampling and denoising optimization method for imbalanced data according to claim 5, characterized in that: S5 specifically includes the following steps: Reconstructed dataset after S4 hybrid sampling Includes the following information: majority class features Minority characteristics , Feature matrix Feature distribution of the majority class , Characteristic distribution ( ),as well as Characteristic distribution ( ), The features of the newly generated minority class samples follow a normal distribution, with the mean and covariance varying with... It changes iteratively; Construct a denoising optimization model for the decision variables. as follows: The constraint condition for minimizing the eigenvalue of the covariance matrix is: in It is the manifold coefficient; It is the penalty coefficient; Indicates a very small quantity; and The specific calculation formula is as follows: In the formula, Covariance matrix Maximum eigenvalue; Covariance matrix Minimum eigenvalue; It is the square root of the condition number of the covariance matrix; The denoising optimization model includes manifold terms. Discriminant Item and noise reduction The denoising term includes a penalty term. , The specific formula is as follows: in, It is a noise penalty item; Indicates Mahalanobis distance; It is an anisotropic penalty item.
