Double-center intuitionistic fuzzy large-interval classification method with unified marble loss

Abstract

The invention relates to the technical field of electronic noses, in particular to a double-center intuitionistic fuzzy large-interval classification method with unified marble loss, which comprises the following steps: defining and optimizing a large-interval distributor model, and introducing a marginal mean value and a marginal variance to describe marginal distribution characteristics; replacing a hinge loss function with a marble loss function, and defining a hyper-parameter of a control quantile; calculating a membership degree and a non-membership degree, verifying the rationality of the new intuitionistic fuzzy strategy through a theorem, and calculating the confidence coefficient of the sample; constructing a double-center intuitionistic fuzzy large-interval classifier model of unified marble loss, deducing a dual problem, solving an optimization problem to obtain model parameters, and calculating a decision function to classify new examples; a classifier model is used, the performance of the model is evaluated through a verification set, the classifier model is deployed in practical application, and a data classification task is carried out, so that the anti-drift performance of the electronic nose system is effectively improved, the mean value and the variance are taken as optimization targets, and the relatively strong anti-interference capability is realized.

Description

Technical Field

[0001] The invention relates to the technical field of electronic noses, and in particular to a dual-center intuitionistic fuzzy large-interval classification method with unified pinball loss. Background Art

[0002] The electronic nose, also known as the artificial olfactory system, is an array of cross-sensitive gas sensors with intelligent signal processing and pattern recognition units. Currently, the electronic nose has been widely used in various fields, such as air quality monitoring, food analysis, and medical diagnosis. However, in actual working environments, the detection ability of the electronic nose may be reduced due to sensor drift caused by factors such as sensor aging, poisoning, and changes in environmental variables (such as humidity and temperature), which limits the reliability and accuracy of the electronic nose in practical applications. Therefore, it is necessary to develop effective methods to reduce the impact of sensor drift and improve the stability of the electronic nose.

[0003] Currently, extensive research is being conducted to mitigate the harmful effects of sensor drift, and they can be roughly divided into two categories: feature-based and classifier-based methods. Anti-drift classification algorithms can be further divided into two categories: target domain-free and target domain adaptive. Target domain adaptive methods usually achieve superior performance by leveraging information from the target domain. However, it is worth noting that further performance optimization requires additional information acquisition, which inevitably leads to increased resource consumption. In addition, the effectiveness of target domain adaptive methods is limited to a specific target domain. Therefore, when the target domain changes, the model needs to be retrained, which inevitably leads to additional time and cost. In contrast, target domain-free methods do not have the above problems.

[0004] Support vector machine (SVM), as a popular target-free domain classifier, is of great significance in electronic nose systems. Specifically, many studies use SVM as the basic classifier after processing electronic nose drift data using feature-based methods. However, traditional support vector machines have many defects, resulting in poor anti-drift performance when classifying electronic nose drift data after drift compensation. Summary of the invention

[0005] The purpose of the present invention is to provide a dual-center intuitionistic fuzzy large-margin classification method with unified pinball loss, aiming to solve the problem of poor anti-drift performance of traditional support vector machines when classifying drift-compensated electronic nose drift data.

[0006] To achieve the above object, the present invention provides a dual-center intuitionistic fuzzy large-margin classification method with unified pinball loss, comprising the following steps:

[0007] Define and optimize the large interval distribution machine model, introduce marginal mean and marginal variance to describe marginal distribution characteristics;

[0008] Replace the hinge loss function of the traditional support vector machine with the pinball loss function and define hyperparameters to control the quantiles;

[0009] The membership and non-membership are calculated based on the distance between the sample and the positive and negative class centers. The rationality of the new intuitionistic fuzzy strategy is verified by the theorem, and the confidence of the sample is calculated using a comprehensive scoring formula.

[0010] Construct a dual-center intuitionistic fuzzy large-margin classifier model with uniform pinball loss, derive the dual problem, solve the optimization problem to obtain the model parameters, and calculate the decision function based on the model parameters to classify new instances;

[0011] Use the classifier model, evaluate the model performance through a validation set, and deploy the classifier model to actual applications to perform data classification tasks.

[0012] The specific method of defining and optimizing the large interval distribution machine model and introducing marginal mean and marginal variance to describe marginal distribution characteristics is as follows:

[0013] Select appropriate kernel functions and parameters, and initialize model parameters;

[0014] Calculate marginal means and marginal variances to describe marginal distribution characteristics;

[0015] Constructing an objective function of a large margin distribution machine to optimize the marginal mean and the marginal variance;

[0016] Using a quadratic programming method to solve the optimization problem of the large interval distribution machine;

[0017] The generalization performance of the model is evaluated based on the validation set, and the model parameters are adjusted as needed.

[0018] The hinge loss function of the traditional support vector machine is replaced by the pinball loss function, and the specific method of defining the hyperparameters for controlling the quantile is as follows:

[0019] Use the unified pinball loss function to replace the traditional hinge loss function;

[0020] Construct a quadratic programming problem for all hyperparameter values;

[0021] Solve the optimization problem of upward support vector machine using quadratic programming method;

[0022] The generalization performance of the model is evaluated based on the validation set, and the model parameters are adjusted as needed.

[0023] The specific method of calculating the membership and non-membership based on the distance between the sample and the positive and negative class centers, verifying the rationality of the new intuitionistic fuzzy strategy through the theorem, and using the comprehensive scoring formula to calculate the confidence of the sample is as follows:

[0024] A new dual-center intuitionistic fuzzy strategy is introduced to calculate membership and non-membership based on the distance between the sample and the positive and negative class centers;

[0025] The rationality of the new intuitionistic fuzzy strategy is verified by the theorem, and the confidence of the sample is calculated using a comprehensive scoring formula based on the membership and non-membership of the sample.

[0026] The specific method of constructing a dual-center intuitionistic fuzzy large-margin classifier model with a unified pinball loss, deriving the dual problem, solving the optimization problem to obtain model parameters, and calculating the decision function based on the model parameters to classify new instances is as follows:

[0027] Construct a dual-center intuitionistic fuzzy large-margin classifier model with uniform pinball loss;

[0028] Based on the classifier model, a dual problem is derived and the optimization problem is solved to obtain model parameters;

[0029] Based on the model parameters, a decision function is calculated to classify a new instance.

[0030] The invention discloses a dual-center intuitionistic fuzzy large-margin classification method with unified pinball loss, defines and optimizes a large-margin distribution machine model, introduces marginal mean and marginal variance to describe marginal distribution characteristics; replaces the hinge loss function of the traditional support vector machine with the pinball loss function, and defines hyperparameters for controlling quantiles; calculates membership and non-membership based on the distance between the sample and the positive and negative class centers, verifies the rationality of the new intuitionistic fuzzy strategy through theorem, and calculates the confidence of the sample using a comprehensive scoring formula; constructs a dual-center intuitionistic fuzzy large-margin classifier model with unified pinball loss, derives the dual problem, solves the optimization problem to obtain model parameters, and calculates the decision function based on the model parameters to classify new instances; uses the classifier to classify the new instance. The classifier model is constructed and the model performance is evaluated through the validation set. The classifier model is deployed in practical applications to perform data classification tasks. This method effectively improves the anti-drift performance of the electronic nose system. A novel dual-center-based intuitive fuzzy function is used to obtain the confidence level of each sample, which further weakens the noise sensitivity. This method uses the marginal mean and variance as new optimization objectives to ensure that all sample points contribute to the determination of the hyperplane, which enhances the robustness of the model to interference and helps to mitigate the impact of drift. The uniform pinball loss is introduced to more accurately measure the distance between sample sets, so as to better adapt to the drift of the electronic nose. This method proposes a dual-center intuitive fuzzy function to obtain the confidence of each sample. This confidence helps to better handle noise during the training process, thereby reducing the sensitivity of the model to drift and solving the problem of poor anti-drift performance of the traditional support vector machine when classifying drift data of the drift-compensated electronic nose. BRIEF DESCRIPTION OF THE DRAWINGS

[0031] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the drawings required for use in the embodiments or the description of the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying creative work.

[0032] Figure 1 It is a comparison chart between the traditional intuitionistic fuzzy strategy and the invented dual-center intuitionistic fuzzy strategy.

[0033] Figure 2 The present invention provides a flow chart of a dual-center intuitionistic fuzzy large-interval classification method with unified pinball loss.

[0034] Figure 3 It is a flowchart of the specific way of defining and optimizing the large interval distribution machine model and introducing marginal mean and marginal variance to describe the marginal distribution characteristics.

[0035] Figure 4It is a flowchart of the specific way to replace the hinge loss function of the traditional support vector machine with the pinball loss function and define the hyperparameters that control the quantile.

[0036] Figure 5 It is a flowchart of the specific method of calculating membership and non-membership based on the distance between the sample and the positive and negative class centers, verifying the rationality of the new intuitionistic fuzzy strategy through the theorem, and calculating the confidence of the sample using the comprehensive scoring formula.

[0037] Figure 6 It is a flowchart of a specific method of constructing a dual-center intuitionistic fuzzy large-margin classifier model with a unified pinball loss, deriving the dual problem, solving the optimization problem to obtain model parameters, and calculating the decision function based on the model parameters to classify new instances. DETAILED DESCRIPTION

[0038] Embodiments of the present invention are described in detail below, examples of which are shown in the accompanying drawings, wherein the same or similar reference numerals throughout represent the same or similar elements or elements having the same or similar functions. The embodiments described below with reference to the accompanying drawings are exemplary and are intended to be used to explain the present invention, and should not be construed as limiting the present invention.

[0039] See also Figures 1 to 6 The present invention provides a dual-center intuitionistic fuzzy large-margin classification method with unified pinball loss, comprising the following steps:

[0040] S1 defines and optimizes the large interval distribution machine model, and introduces marginal mean and marginal variance to describe marginal distribution characteristics;

[0041] Specific method:

[0042] S11 selects appropriate kernel functions and parameters and initializes model parameters;

[0043] S12 calculates marginal means and marginal variances to describe marginal distribution characteristics;

[0044] In the embodiment of the present invention, the marginal mean and marginal variance are calculated, and these two second-order statistics are used to describe the marginal distribution characteristics.

[0045] S13 constructs an objective function of a large interval distribution machine to optimize the marginal mean and the marginal variance;

[0046] In the embodiment of the present invention, an objective function of a large margin distribution machine (LDM) is constructed to optimize the marginal mean and variance.

[0047] In LDM, all sample points can affect the classification hyperplane, not just the support vector, which has better anti-interference performance than support vector machine.

[0048] LDM uses marginal mean and marginal variance to describe marginal distribution characteristics, expressed as:

[0049]

[0050] The expression of LDM is as follows:

[0051]

[0052] In particular, when λ1=λ2=0, LDM degenerates into a traditional support vector machine. Compared with support vector machines, the goal of LDM is to optimize the marginal mean and variance simultaneously, ensuring that all samples participate in the optimal hyperplane and have better anti-interference performance.

[0053] S14 uses a quadratic programming method to solve the optimization problem of the large interval distribution machine;

[0054] S15 evaluates the generalization performance of the model based on the validation set and adjusts the model parameters as needed.

[0055] S2 replaces the hinge loss function of the traditional support vector machine with the pinball loss function and defines hyperparameters to control the quantiles;

[0056] In the embodiment of the present invention, the hinge loss function of the traditional support vector machine is replaced by the pinball loss function, and a hyperparameter τ for controlling the quantile is defined. According to the value of τ, a suitable optimization problem is selected for solution to ensure the robustness of the model.

[0057] Specific method:

[0058] S21 uses a unified pinball loss function to replace the traditional hinge loss function;

[0059] S22 constructs a quadratic programming problem for all hyperparameter values;

[0060] In an embodiment of the present invention, a quadratic programming problem (QPP) is constructed that is applicable to all hyperparameter (τ) values.

[0061] S23 uses quadratic programming method to solve the optimization problem of upward support vector machine;

[0062] In an embodiment of the present invention, a quadratic programming method is used to solve the optimization problem of an upward support vector machine (UP-SVM).

[0063] S24 evaluates the generalization performance of the model based on the validation set and adjusts the model parameters as needed.

[0064] In the embodiment of the present invention, the traditional support vector machine model uses hinge loss, which is related to the shortest distance between sample sets and thus has poor robustness. Based on this, the hinge loss function is replaced by the pinball loss function, which is expressed as:

[0065]

[0066] Among them, τ∈[-1,0)∪(0,1] represents a hyperparameter that can control the quantile. In Pin-SVM, τ is non-negative, and Pin-SVM is expressed as:

[0067]

[0068] The optimization problem of Pin-SVM is stated as:

[0069]

[0070] The constraints of UP-SVM are:

[0071]

[0072] Here, the value of τ can be negative.

[0073] S3 calculates membership and non-membership based on the distance between the sample and the center of the positive and negative classes, verifies the rationality of the new intuitionistic fuzzy strategy through the theorem, and calculates the confidence of the sample using a comprehensive scoring formula;

[0074] In the embodiment of the present invention, a new dual-center intuitionistic fuzzy strategy is introduced to consider the contribution of samples to positive and negative classes, and redefine the membership and non-membership functions. The rationality of the new intuitionistic fuzzy strategy is verified by the theorem, and the confidence of the sample is calculated using a comprehensive scoring formula based on the membership and non-membership of the sample.

[0075] Specific method:

[0076] S31 introduces a new dual-center intuitionistic fuzzy strategy to calculate membership and non-membership based on the distance between the sample and the positive and negative class centers;

[0077] S32 verifies the rationality of the new intuitionistic fuzzy strategy through theorem, and calculates the confidence of the sample using a comprehensive scoring formula based on the membership and non-membership of the sample.

[0078] In the embodiment of the present invention, Figure 1As shown in the figure, points x1 and x5 are at equal distances from the center of the positive class. However, x1 is closer to other classes than x5 and is more likely to be a noise point. According to the intuitive fuzzy strategy, points x1 and x5 have the same membership and non-membership, that is, they have the same confidence level, but this is inconsistent with the actual situation. Therefore, the intuitive fuzzy strategy based on a single center will lead to inaccurate confidence, resulting in a decrease in the model's anti-noise ability.

[0079] In order to solve the inherent defects of the above IF strategy, we introduce a new TC-IF strategy, which inherits the concept of local neighborhood set of the original IF strategy and considers the contribution to the positive and negative classes. Taking the positive sample xi as an example, the membership function μ(xi) is defined as

[0080] μ(x i )=1-D P (x i ) / (R P +σ),

[0081] Where D P (x i ) represents the distance between the sample score and the positive class center, R P Indicates the maximum distance from the positive sample in the training set to the center of the positive class.

[0082] The non-membership function v(xi) is defined as

[0083]

[0084] Where D m =min(D P (x i )-D N (x i ), i = 1, 2, ..., m represents the minimum distance difference, D PN Represents the distance between the positive class center and the negative class center.

[0085] To prove the rationality of the new intuitionistic fuzzy strategy, we give the following four theorems.

[0086] Theorem 1: When a sample is located at the center of the positive class, its non-membership degree is 0.

[0087] Proof: x3 is located in Figure 1 The positive center in the middle has no discordant points around it, so the following equation can be derived

[0088] ρ(x3)=0,D P (x3)-D N (x3)=-PN. (18)

[0089] According to the knowledge of the three sides of the triangle formed by the sample, the positive center, and the negative center, when the sample is located on the extension line connecting the positive center and the negative center, D m The minimum value of is -PN. Substituting this value and formula (18) into formula (17), we can get the following result

[0090]

[0091] Theorem 2: When a sample is located at the center of the negative class, its non-membership degree is 1.

[0092] Proof: x4 in Figure 1 is located in the center of the negative class and is surrounded by inharmonious points, so the following equation can be derived:

[0093] ρ(x4)=1,D P (x4)-D N (x4) = PN (20)

[0094] Similarly, we can get the following results

[0095]

[0096] Theorem 3: When the distances between two samples and the positive class center are the same, the non-membership degrees of the two samples (not located on the connecting line of the two class centers at the same time, and one end is far away from the positive class center) are different.

[0097] prove. Figure 1 The distances of points x1 and x5 from the center of the positive class are the same, but the distances from the negative class are different, and there are no discordant points around them. So the following expression can be derived

[0098] ρ(x1)=ρ(x5)=0 (22)

[0099] D P (x1)-D N (x1)>D P (x5)-D N (x5) (23)

[0100] Similarly, we can obtain the following results

[0101]

[0102] v(x1)>v(x5) (26)

[0103] Theorem 4: If a sample and its own class center are on the same side of the perpendicular bisector of the two class centers, the non-membership of the sample is less than 0.5. On the contrary, when they are on opposite sides, the non-membership is mostly 1.

[0104] Proof: In Figure 1 In the equation, x1 is on the same side as the center of the positive class, and there are no discordant points around it. x2 is on the opposite side of the center of the positive class, and is surrounded by discordant points. So the following expression can be derived

[0105] ρ(x1)=0,ρ(x2)=1 (27)

[0106] D P (x1)-D N (x1)<0, D P (x2)-D N (X2)>0 (28)

[0107] Similarly, we can obtain the following resu1ts

[0108]

[0109] For formula (30), when there is a small amount of noise, according to formula (16), it can be inferred that μ(x2) = 0. From Theorem 2, we know that when x2 is located at the negative center, v(x2) is equal to 1. Further, we can find that when x2 is located on the line connecting the two class centers and at one end away from the negative class center, v(x2) is equal to 1. Therefore, in most cases, v(x2) is equal to 1.

[0110] After obtaining the membership and non-membership of the samples, we proposed the following comprehensive scoring formula:

[0111]

[0112] Among them, w′∈[O, 1] is the decision maker’s preference parameter, which determines the weights of membership and non-membership, and e can be adjusted to a positive number to distinguish between the two IF values.

[0113] S4 constructs a dual-center intuitionistic fuzzy large-margin classifier model with uniform pinball loss, derives the dual problem, solves the optimization problem to obtain model parameters, and based on the model parameters, calculates the decision function to classify new instances;

[0114] In an embodiment of the present invention, given a labeled sample set, the kernel function is used to map the input data to the Hilbert space, and a dual-center intuitionistic fuzzy large margin classifier (TCIFUPLMC) model with unified pinball loss is constructed. The dual problem of TCIPUPLMC is derived, the optimization problem is solved to obtain the model parameters, and the decision function is calculated based on the solved model parameters to classify the new test instance.

[0115] Labeled sample sets refer to training data sets that have been labeled with categories. These sample sets include gas sensor data collected by the electronic nose system. Each sample has a corresponding label indicating the category it belongs to. In addition, the sample set contains the source domain and the target domain. These labeled samples are used to train the TC-IFUPLMC model, enabling it to learn how to correctly classify new, unseen data.

[0116] Specific method:

[0117] S41 Constructing a dual-center intuitionistic fuzzy large-margin classifier model with uniform pinball loss;

[0118] S42 derives a dual problem based on the classifier model and solves the optimization problem to obtain model parameters;

[0119] S43 calculates a decision function based on the model parameters to classify the new instance.

[0120] In the embodiment of the present invention, given a labeled sample set S, the corresponding membership sample set is recorded as

[0121]

[0122] Among them, (x i ,y i ) represents a pair of sample points, score s i Represents the reliability of the sample.

[0123] In real-world scenarios, it is often observed that most real datasets exhibit linear inseparability in their original feature space, so our main focus here is to derive the TCIPUPLMC formula for linearly inseparable data. Using the kernel function, the input data is mapped into the Hilbert space through a typical nonlinear mapping. Based on the proposed TC-IF function and UP constraints, the optimization problem of TC-IFUPLMC can be expressed as

[0124]

[0125] Among them, v τ is a sign function whose value is 1 when τ is greater than 0 and -1 when τ is less than 0. Note that when τ = 0, the second constraint becomes ξ i ≥0, the UP loss function is transformed into the hinge loss function.

[0126] 1. Dual Problem of TC-IFUPLMC

[0127] In this section, we only derive v in detail. τ = 1, and v τ The derivation of =-1 is similar.τ = 1, the original QPP of TC-IFUPLMC is

[0128]

[0129] Theorem 5: The optimal solution ω* of model (34) can be concisely expressed as

[0130] ω * =φ(X)β, (35)

[0131] in, represents the coefficient vector.

[0132] According to Theorem 5, we can obtain the dual problem of the original optimization task.

[0133] Theorem 6: The dual problem of the original problem Eq. (34) is as follows

[0134]

[0135] Proof: According to Theorem 5, the following formula can be derived

[0136]

[0137] Where A = φ(X) T φ(X), A i represents the i-th column of matrix A.

[0138] Substituting formula (37) into formula (34), we can get the matrix representation of TC-IFUPLMC as follows:

[0139]

[0140] The Lagrangian function of formula (38) is expressed as

[0141]

[0142] Where δ and η are Lagrange multiplier vectors. Setting the partial derivatives of {δ, η} to 0, we can get the following formula:

[0143]

[0144] Substituting formula (40) into formula (39), the dual form of formula (38) can be expressed as

[0145]

[0146] Theorem 7: The final QPP of TC-IFUPLMC can be expressed as

[0147]

[0148] Proof: Similarly, when v τ = -1, we can get the dual problem of TC-IFUPLMC

[0149]

[0150] Combining the two dual problems, we can obtain the complete dual problem of TC-IFUPLMC

[0151]

[0152] Assuming α = δ - |τ|η, applying the above constraints (44), the range of η can be derived. Thus, the final dual form of TC-IFUPLMC is obtained (42):

[0153] After solving the QPP in model (42), the decision function of TCIFUPLMC can be derived to solve the data classification problem. According to formula (40), the coefficient can be calculated as

[0154]

[0155] Therefore, a new test instance X T The following decision function can be used to classify

[0156]

[0157] S5 uses the classifier model, evaluates the model performance through a validation set, and deploys the classifier model to actual applications to perform data classification tasks.

[0158] This method can efficiently and accurately classify gas components based on a large amount of sensor data, and improves the defects of traditional support vector machines (SVM) in drift resistance. Specifically, the TC-IFUPLMC model uses marginal mean and variance as optimization targets to improve anti-interference ability; uses a unified pinball loss function to measure the distance between sample sets to enhance robustness; and uses a dual-center intuitionistic fuzzy function to calculate the confidence level of each sample, further reducing noise sensitivity. When it comes to actual data classification processing, the present invention uses UCSD and CQU sensor drift data sets for experimental verification. Firstly, the domain regularized component analysis method (DRCA) is used to perform drift compensation on the electronic nose drift data, that is, the data of the two domains are projected into a subspace to achieve drift compensation (the centers of the two domains are closely converged to eliminate the drift), and then the TC-IFUPLMC classifier is used in this subspace to classify the drift-compensated data, which can effectively improve the anti-drift performance of the gas sensor and ensure more accurate identification of various gas components. The data processing flow of the gas sensor data in specific application scenarios is carefully designed. TC-IFUPLMC not only improves the classification accuracy, but also enhances the adaptability and stability of the system.

[0159] What is disclosed above is only a preferred embodiment of the dual-center intuitive fuzzy large-interval classification method with unified pinball loss of the present invention. Of course, this cannot be used to limit the scope of rights of the present invention. Ordinary technicians in this field can understand that all or part of the processes of the above embodiments are implemented, and equivalent changes made according to the claims of the present invention still fall within the scope of the invention.

Claims

1. A dual-center intuitionistic fuzzy large-margin classification method with unified pinball loss, characterized in that: The following steps are involved: Define and optimize the large interval distribution machine model, introduce marginal mean and marginal variance to describe marginal distribution characteristics; Replace the hinge loss function of the traditional support vector machine with the pinball loss function and define hyperparameters to control the quantiles; The membership and non-membership are calculated based on the distance between the sample and the positive and negative class centers. The rationality of the new intuitionistic fuzzy strategy is verified by the theorem, and the confidence of the sample is calculated using a comprehensive scoring formula. Construct a dual-center intuitionistic fuzzy large-margin classifier model with uniform pinball loss, derive the dual problem, solve the optimization problem to obtain the model parameters, and calculate the decision function based on the model parameters to classify new instances; Use the classifier model, evaluate the model performance through a validation set, and deploy the classifier model to actual applications to perform data classification tasks.

2. The dual-center intuitionistic fuzzy large-margin classification method with unified pinball loss as claimed in claim 1, characterized in that ； The specific method of defining and optimizing the large interval distribution machine model and introducing marginal mean and marginal variance to describe marginal distribution characteristics is as follows: Select appropriate kernel functions and parameters, and initialize model parameters; Calculate marginal means and marginal variances to describe marginal distribution characteristics; Constructing an objective function of a large margin distribution machine to optimize the marginal mean and the marginal variance; Using a quadratic programming method to solve the optimization problem of the large interval distribution machine; The generalization performance of the model is evaluated based on the validation set, and the model parameters are adjusted as needed.

3. The dual-center intuitionistic fuzzy large-margin classification method with unified pinball loss as claimed in claim 1, It is characterized by: The specific method of replacing the hinge loss function of the traditional support vector machine with the pinball loss function and defining the hyperparameters for controlling the quantile is as follows: Use the unified pinball loss function to replace the traditional hinge loss function; Construct a quadratic programming problem for all hyperparameter values; Solve the optimization problem of upward support vector machine using quadratic programming method; The generalization performance of the model is evaluated based on the validation set, and the model parameters are adjusted as needed.

4. The dual-center intuitionistic fuzzy large-margin classification method with unified pinball loss as claimed in claim 1, characterized in that ； The specific method of calculating the membership and non-membership based on the distance between the sample and the positive and negative class centers, verifying the rationality of the new intuitionistic fuzzy strategy through the theorem, and using the comprehensive scoring formula to calculate the confidence of the sample is as follows: A new dual-center intuitionistic fuzzy strategy is introduced to calculate membership and non-membership based on the distance between the sample and the positive and negative class centers; The rationality of the new intuitionistic fuzzy strategy is verified by the theorem, and the confidence of the sample is calculated using a comprehensive scoring formula based on the membership and non-membership of the sample.

5. The dual-center intuitionistic fuzzy large-margin classification method with unified pinball loss as claimed in claim 1, characterized in that ； The specific method of constructing a dual-center intuitionistic fuzzy large-margin classifier model with a unified pinball loss, deriving the dual problem, solving the optimization problem to obtain model parameters, and calculating the decision function based on the model parameters to classify new instances is as follows: Construct a dual-center intuitionistic fuzzy large-margin classifier model with uniform pinball loss; Based on the classifier model, a dual problem is derived and the optimization problem is solved to obtain model parameters; Based on the model parameters, a decision function is calculated to classify a new instance.

Images