A data classification method and system based on an improved TSK fuzzy classification model

Abstract

The application relates to the technical field of data classification, and discloses a data classification method and system based on an improved TSK fuzzy classification model, which comprises the following steps: obtaining a data set to be classified, and dividing the data set to be classified into a training set and a test set; based on the training set, using a supervised enhancement type soft subspace clustering algorithm and a deep learning algorithm to respectively optimize the fuzzy rule antecedent and the fuzzy rule consequent of a pre-constructed TSK fuzzy classification model; inputting the test set into the optimized TSK fuzzy classifier model for classification processing to obtain a data classification result. The application respectively optimizes the fuzzy rule antecedent and the fuzzy rule consequent parameters of the TSK fuzzy classification model, thereby improving the adaptability of the model to high-dimensional data and complex data sets, and further improving the precision and efficiency of the model in data classification.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of image classification, more particularly, to a data classification method and system based on an improved TSK fuzzy classification model. BACKGROUND

[0002] With the development of artificial intelligence and big data technology, fuzzy systems have been widely used in classification, regression and other tasks. Among them, the Takagi-Sugeno-Kang (TSK) fuzzy system has become an important classifier due to its excellent interpretability and modeling ability for data uncertainty. The TSK fuzzy system realizes an uncertain reasoning process similar to human reasoning through the IF-THEN mode based on fuzzy rules, and performs well in classification and regression problems. Compared with other fuzzy systems, the TSK fuzzy system outputs a deterministic value, making it easier to interpret and understand, so it is widely used in practical applications. To further improve the performance of the model, existing technologies combine various optimization methods, such as optimizing the fuzzy rule antecedent through fuzzy C-means clustering (FCM) and square hierarchical fuzzy clustering (SFCM), and optimizing the fuzzy rule consequent through genetic algorithm (GA), deep convolution (DC) and support vector machine (MSVM). In addition, deep neural networks and new loss functions have been introduced in recent years to improve fuzzy rule generation and generalization performance.

[0003] However, in practical applications, the existing TSK fuzzy classification model has the following problems: high-dimensional features result in long and complex fuzzy rules, which weaken the interpretability and accuracy of the model; too many fuzzy rules can introduce redundancy and reduce generalization; in addition, the output formula based on the softmax function has a saturation phenomenon, resulting in an unbalanced activation level of fuzzy rules, with only a few fuzzy rules being frequently used, which increases the complexity of the model. The above problems result in low classification accuracy and low efficiency of the TSK fuzzy classification model in data classification. SUMMARY

[0004] To overcome the low classification accuracy and low efficiency of the existing TSK fuzzy classification model in data classification, the present application proposes the following technical solutions:

[0005] In a first aspect, the present application proposes a data classification method based on an improved TSK fuzzy classification model, comprising:

[0006] obtaining a data set to be classified, and dividing the data set to be classified into a training set and a test set;

[0007] based on the training set, using a supervised enhanced soft subspace clustering algorithm and a deep learning algorithm to optimize the fuzzy rule antecedent and the fuzzy rule consequent of the pre-constructed TSK fuzzy classification model, respectively;

[0008] The test set is input into the optimized TSK fuzzy classifier model for classification processing to obtain data classification results.

[0009] As a preferred technical solution, before the test set is input into the optimized TSK fuzzy classifier model for classification processing, the method further comprises:

[0010] The improved fuzzification operation is used to uniformly process the triggering level of the fuzzy rule, and the expression is as follows:

[0011]

[0012] wherein, is the triggering value of the fuzzy rule r , indicating the triggering level of the fuzzy rule r on the current input sample; is the dimension number of the input feature, is the value of the input sample on the d th feature dimension, is the cluster center of the fuzzy rule r on the d th feature dimension, is the variance of the fuzzy rule r on the d th feature dimension, is the maximum value in all fuzzy rule triggering values.

[0013] As a preferred technical solution, the pre-constructed TSK fuzzy classification model comprises an input layer, a fuzzification layer, a fuzzy rule layer, an inference layer, and an output layer.

[0014] The input sample is input into the fuzzification layer through the input layer;

[0015] The fuzzification layer calculates the membership value based on the feature value of the input sample, and the cluster center and feature range of the fuzzy rule;

[0016] The fuzzy rule layer triggers the corresponding fuzzy rule according to the membership value output by the fuzzification layer;

[0017] The inference layer performs fuzzy inference according to the triggered fuzzy rule to obtain the normalized triggering value of the fuzzy rule;

[0018] The output layer performs defuzzification processing according to the normalized triggering value of the fuzzy rule to obtain the final classification result.

[0019] As a preferred technical solution, the inference layer performs fuzzy inference according to the triggered fuzzy rule to obtain the normalized triggering value of the fuzzy rule , and the expression is as follows:

[0020]

[0021] Where R is the total number of fuzzy rules, Indicates fuzzy rules r Trigger value Perform exponential operations.

[0022] As a preferred technical solution, the output layer performs defuzzification processing based on the normalization trigger value of the fuzzy rule to obtain the final classification result, the expression of which is shown below:

[0023]

[0024] in, This represents the predicted value of input sample x in class c. This represents the output value of the fuzzy rule r on category c.

[0025] As a preferred technical solution, the fuzzy rule antecedents of the fuzzy rules in the pre-built TSK fuzzy classification model are optimized using a supervised enhanced soft subspace clustering algorithm, including:

[0026] Initialize the fuzzy membership matrix of the antecedent of the fuzzy rule Cluster center matrix Weight matrix and probability matrix .

[0027] For training set data samples and their corresponding sample labels Construct the objective function of the supervised enhanced soft subspace clustering algorithm. Its expression is as follows:

[0028]

[0029] in, N The number of samples in the training set. R The number of cluster centers. D For data feature dimensions, C For the number of categories, Indicates the first n The nth sample pair r The membership degree of each cluster center, where m>1 is the fuzziness index. Indicates sample Category c The tag value, Representing cluster centers r In feature dimension d Weight on, Representing cluster centersr a probability value on a category c , for an input sample , for a cluster center n , for all data samples of a training set , for a cluster center r , for all data samples of a training set , for a cluster center , for a cluster center r , for a cluster center , for a cluster center

[0030] respectively iteratively optimize the fuzzy membership matrix , the cluster center matrix , the weight matrix and the probability matrix until the objective function satisfies the convergence condition, to obtain the optimized fuzzy rule antecedent.

[0031] As a preferred technical solution, the fuzzy membership matrix , the cluster center matrix , the weight matrix and the probability matrix are respectively iteratively optimized by the Lagrange multiplier method until the objective function satisfies the convergence condition, and the expression is as follows:

[0032]

[0033]

[0034]

[0035] .

[0036] As a preferred technical solution, the fuzzy rule consequent of the pre-constructed TSK fuzzy classification model is optimized by using a deep learning algorithm, including:

[0037] initializing the linear weight vector and the bias term of the fuzzy rule consequent;

[0038] performing batch normalization processing on the training set sample data, and applying a Sigmoid function to the normalized data for nonlinear activation;

[0039] using the training set sample data that has been normalized and nonlinearly activated to train and optimize the parameters of the fuzzy rule consequent until the optimization objective function of the fuzzy rule consequent satisfies the convergence condition;

[0040] wherein the expression of the optimization objective function of the fuzzy rule consequent is as follows:

[0041]

[0042] in, Indicates sample n Category c The true label; Indicates sample Predicted as category c The probability, This represents the normalized activation weights of the fuzzy rule r. Let represent the linear weight vector of fuzzy rule r with respect to category c. Represents a linear weight vector transpose, For the sample The nonlinear embedding results represent the samples Batch normalization Processing and activation functions The subsequent feature mapping.

[0043] As a preferred technical solution, in training and optimizing the parameters of the fuzzy rule consequent, the method further includes:

[0044] Assign a retention probability to each fuzzy rule;

[0045] In each training optimization iteration, a random number within the interval [0, 1] is generated for each fuzzy rule;

[0046] If the generated random number is less than or equal to the retention probability set for the fuzzy rule, then the fuzzy rule is retained in the current iteration; otherwise, the fuzzy rule is discarded.

[0047] Only the fuzzy rules that were not discarded were trained and their parameters were updated.

[0048] Secondly, the present invention also proposes a data classification system based on an improved TSK fuzzy classification model, applicable to the data classification method based on an improved TSK fuzzy classification model as described in any of the embodiments of the first aspect, comprising:

[0049] The partitioning module is used to acquire the dataset to be classified and divide the dataset into a training set and a test set.

[0050] The optimization module is used to optimize the fuzzy rule antecedent and fuzzy rule consequent of the pre-built TSK fuzzy classification model based on the training set using supervised enhanced soft subspace clustering algorithm and deep learning algorithm, respectively.

[0051] The classification module is used for inputting the test set into the optimized TSK fuzzy classifier model for classification processing to obtain a data classification result.

[0052] The present application has at least the following advantages:

[0053] The present application optimizes the fuzzy rule antecedent and the fuzzy rule consequent parameters of the TSK fuzzy classification model respectively by using the supervised enhanced soft subspace clustering algorithm and the deep learning algorithm, thereby improving the adaptability of the model to high-dimensional data and complex data sets. The supervised enhanced soft subspace clustering algorithm can select different feature subspaces for different fuzzy rules, reduce the influence of irrelevant or redundant features, and thereby improve the accuracy and classification precision of the fuzzy rules; the deep learning algorithm further improves the expression ability of the fuzzy rule consequent through an effective parameter optimization method (such as batch gradient descent), and ensures that the model has better generalization ability and stability when processing large-scale data sets. These optimizations work together to enable the model to achieve higher classification precision with less computing resources in the data classification task, while significantly improving the classification efficiency. BRIEF DESCRIPTION OF DRAWINGS

[0054] Figure 1 A flowchart of a data classification method based on an improved TSK fuzzy classification model is provided for the present embodiment.

[0055] Figure 2 The average normalized trigger level results on the entire data set using the MTSK fuzzification operation and without using the MTSK fuzzification operation are compared for the present embodiment provided by the application.

[0056] Figure 3 The average normalized trigger level results on the entire data set using the MTSK fuzzification operation and without using the MTSK fuzzification operation are compared for the present embodiment provided by the application.

[0057] Figure 4 A principle diagram of constructing a fuzzy rule antecedent framework is provided for the present embodiment.

[0058] Figure 5 A rule feature number reduction effect diagram of the SESSC algorithm on the DRD data set is provided for the present embodiment.

[0059] Figure 6 A comparison diagram of the model parameter number and the RCA value obtained without using the SESSC and using the SESSC is provided for the present embodiment.

[0060] Figure 7 A principle diagram of discarding fuzzy rules using the dropout technology is provided for the present embodiment.

[0061] Figure 8 A comparison chart of the number of model parameters and RCA values obtained when the deep learning optimization technique is not used and when the deep learning optimization technique is used for the embodiment provided by the present application.

[0062] Figure 9 A schematic diagram of part of the structure of the SESSC-DL-MTSK model provided by the embodiment of the present application.

[0063] Figure 10 A rule feature reduction effect chart on the DRD data set when the deep learning optimization technique is not used for the embodiment provided by the present application.

[0064] Figure 11 A rule feature reduction effect chart on the DRD data set when the deep learning optimization technique is used for the embodiment provided by the present application.

[0065] Figure 12 An architecture diagram of the data classification system based on the improved TSK fuzzy classification model provided by the embodiment of the present application. DETAILED DESCRIPTION

[0066] The embodiments of the present application will be described hereinafter with reference to the drawings and preferred technical solutions, and other advantages and effects of the present application can be easily understood by those skilled in the art from the content disclosed in the present specification. The present application can also be implemented or applied by means of other different specific embodiments, and each detail in the present specification can be modified or changed in various ways based on different views and applications without departing from the spirit of the present application. It should be understood that the preferred technical solutions are only for illustrating the present application, and are not intended to limit the protection scope of the present application.

[0067] It should be noted that the diagrams provided in the following embodiments only illustrate the basic concept of the present application in a schematic manner, and only the components related to the present application are shown in the diagrams, not the number, shape and size of the components when actually implemented. The actual implementation of each component can be randomly changed in terms of type, number and proportion, and the layout type of the components can also be more complex.

[0068] In the following description, a large number of details are discussed to provide a more thorough explanation of the embodiments of the present application, however, it is obvious to those skilled in the art that the embodiments of the present application can be implemented without these specific details, and in other embodiments, the known structures and devices are shown in the form of block diagrams rather than in the form of details, so as to avoid making the embodiments of the present application difficult to understand.

[0069] To facilitate understanding of the embodiments of the present invention, some concepts involved in all embodiments of the present invention will be appropriately explained before describing the various embodiments of the present invention.

[0070] The TSK Fuzzy System (TSKFS) is a classic fuzzy inference model. For TSKFS, the fuzzy rules are defined as follows:

[0071]

[0072] In the formula, It is a D-dimensional input vector in the training dataset. It is the first r In the fuzzy rule and the first d A feature-related fuzzy subset and It is the first c The consequent parameters of the fuzzy rules for each category.

[0073] The fuzzy algorithm of this invention can use different membership functions, among which the Gaussian membership function has a smooth shape, making it less sensitive to small changes in the input data and helping to reduce the sensitivity of the fuzzy system's output to input noise. Therefore, this invention employs the Gaussian membership function, characterized by... In fuzzy sets The membership degree value on is:

[0074]

[0075] In the formula, the parameter and This can be calculated using clustering algorithms. Specifically, FCM (Fuzzy C-means clustering algorithm) can be used for calculation. and It can be derived from the following formula:

[0076]

[0077]

[0078] In the formula, N The number of input data, Represents the input data in the r-th cluster. j The membership value. h is a manually adjustable scaling parameter.

[0079] The output of the TSK fuzzy classifier can be expressed by the following formula:

[0080]

[0081] in,

[0082]

[0083]

[0084] is the trigger level of the rth fuzzy rule. In order to optimize the TSK fuzzy classifier, the antecedent and the consequent of the fuzzy rule need to be parameterized.

[0085] Embodiment 1

[0086] The embodiment proposes a data classification method based on an improved TSK fuzzy classification model, as shown in Figure 1 Figure 1 is a flowchart of a data classification method based on an improved TSK fuzzy classification model provided by the embodiment, and the method comprises the following steps:

[0087] S1: Obtain a data set to be classified, and divide the data set to be classified into a training set and a test set.

[0088] S2: Based on the training set, use a supervised enhanced soft subspace clustering algorithm and a deep learning algorithm to optimize the fuzzy rule antecedent and the fuzzy rule consequent of the pre-constructed TSK fuzzy classification model, respectively.

[0089] S3: Input the test set into the optimized TSK fuzzy classifier model for classification processing to obtain a data classification result.

[0090] It can be understood that the present application uses a supervised enhanced soft subspace clustering algorithm and a deep learning algorithm to optimize the fuzzy rule antecedent and the fuzzy rule consequent parameters of the TSK fuzzy classification model, respectively, thereby improving the adaptability of the model to high-dimensional data and complex data sets. The supervised enhanced soft subspace clustering algorithm can select different feature subspaces for different fuzzy rules, reducing the influence of irrelevant or redundant features, thereby improving the accuracy and classification precision of the fuzzy rule; the deep learning algorithm further improves the expression ability of the fuzzy rule consequent through effective parameter optimization methods (such as batch gradient descent), ensuring that the model has better generalization ability and stability when processing large-scale data sets. These optimizations work together to enable the model to achieve higher classification precision with less computing resources in data classification tasks, while significantly improving the classification efficiency.

[0091] Embodiment 2

[0092] The embodiment improves the data classification method based on the improved TSK fuzzy classification model proposed in Embodiment 1.

[0093] ​In this embodiment, the improved fuzzification operation is used to uniformize the triggering level of fuzzy rules, which specifically includes:

[0094] Let be , which can be rewritten as:

[0095]

[0096] The final defuzzification operation can be calculated by the softmax function. Since the softmax function has a saturation phenomenon, the defuzzification operation also has a saturation phenomenon. This means that, with the increase of the dimension, is smaller, so the number of fuzzy rules that each input can trigger will also be smaller, which leads to the saturation of defuzzification. For example, assuming that the model has four fuzzy rules, then At this time, the activation level difference of each fuzzy rule is very small, which means that the number of fuzzy rules that can be triggered is more; when the dimension increases, , At this time, only the first fuzzy rule can be triggered, and the remaining fuzzy rules are almost not triggered.

[0097] In order to avoid the saturation problem of fuzzy rules, the present application proposes an improved fuzzification operation, named MTSK, whose expression is as follows:

[0098]

[0099] Among them, is the triggering value of fuzzy rule r , which represents the triggering level of fuzzy rule r on the current input sample; is the dimension number of input features, represents the value of the input sample on the d th feature dimension, represents the clustering center of fuzzy rule r on the d th feature dimension, represents the variance of fuzzy rule r on the d th feature dimension, represents the maximum value in all fuzzy rule triggering values.

[0100] In this embodiment, the pre-constructed TSK fuzzy classification model includes an input layer, a fuzzification layer, a fuzzy rule layer, an inference layer and an output layer;

[0101] The input sample is input to the fuzzification layer through the input layer;

[0102] The fuzzification layer calculates the membership value based on the feature value of the input sample, and the cluster center and feature range of the fuzzy rule;

[0103] The fuzzy rule layer triggers the corresponding fuzzy rule according to the membership value output by the fuzzification layer;

[0104] The inference layer performs fuzzy reasoning according to the triggered fuzzy rule to obtain the normalized trigger value of the fuzzy rule;

[0105] The output layer performs defuzzification processing according to the normalized trigger value of the fuzzy rule to obtain the final classification result.

[0106] In this embodiment, the inference layer performs fuzzy reasoning according to the triggered fuzzy rule to obtain the normalized trigger value of the fuzzy rule , which is expressed as follows:

[0107]

[0108] where R is the total number of fuzzy rules, denotes the exponential operation on the trigger value r of the fuzzy rule .

[0109] In this embodiment, the output layer performs defuzzification processing according to the normalized trigger value of the fuzzy rule to obtain the final classification result, which is expressed as follows:

[0110]

[0111] wherein, denotes the predicted value of the input sample x in the category c, denotes the output value of the fuzzy rule r in the category c.

[0112] As an example, in order to prove that the MTSK fuzzification operation proposed in the present application can alleviate the saturation problem of fuzzy rules, as shown in Figure 2 and Figure 3 , the average normalized trigger level result comparison chart of the entire data set using the MTSK fuzzification operation and not using the MTSK fuzzification operation provided by the embodiment of the present application on the Clave data set, Figure 2 the average normalized trigger level result comparison chart of the entire data set using the MTSK fuzzification operation and not using the MTSK fuzzification operation provided by the embodiment of the present application on the WDBC data set, Figure 3 Figure 2 and Figure 3 ​It can be seen that for the model without MTSK, the saturated fuzzy rules have a greater average firing level than other fuzzy rules, so the contribution of fuzzy rules to the output is significantly different, and the average normalized fuzzy rule firing level of the model using MTSK is more uniform than that of the model without MTSK, which also leads to better classification performance.

[0113] In this embodiment, the fuzzy rule antecedent of the fuzzy rule of the pre-constructed TSK fuzzy classification model is optimized by using a supervised enhanced soft subspace clustering algorithm, including:

[0114] Initializing the fuzzy membership matrix of the fuzzy rule antecedent , the cluster center matrix , the weight matrix and the probability matrix ;

[0115] For the training set data samples and their corresponding sample labels , the objective function of the supervised enhanced soft subspace clustering algorithm is constructed , the expression of which is as follows:

[0116]

[0117] wherein, N is the number of training set data samples, R is the number of cluster centers, D is the data feature dimension, C is the number of classification categories, denotes the membership of the n th sample to the r th cluster center, m>1 is the fuzzification index, denotes the label value of the sample belonging to the category c , denotes the weight of the cluster center r in the feature dimension d , denotes the probability value of the cluster center r in the category c , is the n th input sample, is the r th cluster center, is the cluster center of all data samples in the training set, is the weight of the r th cluster center, is the adjustment term parameter,

[0118] Iterative optimization of the fuzzy membership matrix Cluster center matrix Weight matrix and probability matrix until the objective function If the convergence condition is met, the optimized fuzzy rule antecedent is obtained.

[0119] In this embodiment, the fuzzy membership matrix is ​​iteratively optimized using the Lagrange multiplier method. Cluster center matrix Weight matrix and probability matrix until the objective function It satisfies the convergence condition, and its expression is as follows:

[0120]

[0121]

[0122]

[0123] .

[0124] As an example, since the core of the TSK fuzzy system is building a fuzzy rule base, each fuzzy rule is divided into a fuzzy rule antecedent and a fuzzy rule consequent. Therefore, the fuzzy rule antecedent must be constructed first. However, most current TSK fuzzy systems use FCM clustering algorithms, or more novel feature-weighted FCM clustering algorithms, entropy-weighted clustering algorithms, etc., which cannot effectively handle multi-dimensional feature spaces and cannot extract potential feature relationships. This invention uses the supervised augmented soft subspace clustering algorithm SESSC to process the features of the input data and train effective fuzzy rule antecedents. This embodiment uses the PyTorch library in Python to construct the fuzzy rule antecedent framework, such as... Figure 4 As shown, Figure 4 The schematic diagram of the framework for constructing fuzzy rule antecedents provided in the embodiments of the present invention specifically includes:

[0125] Calculate the sample center of all input samples The original cluster center matrix is ​​calculated using the standard FCM clustering algorithm. V Initialize according to dimensions Initialize based on the number of categories in the data samples .

[0126] Write a SESSC function using PyTorch, and implement a fuzzy membership matrix internally within the SESSC function. Cluster center matrix , weight matrix and probability matrix updating formula of the iteration code, after continuous iteration update, a trained clustering center matrix is returned when reaching the threshold value V .

[0127] This embodiment constructs a fuzzy rule base containing 20 fuzzy rules, each fuzzy rule has a different clustering center matrix. Taking the medical data DRD in the medical field as an example, the number of data sample features after using the SESSC algorithm is viewed. The DRD data set has 1151 data samples, the feature number is 19, and the category number is 2. The clustering center matrix processed by the SESSC algorithm is visualized using the matplotlib library, and whether the feature number used by each fuzzy rule is successfully simplified is viewed, as shown in Figure 5 , as shown in Figure 5 , the rule feature number reduction effect diagram of the SESSC algorithm provided by the embodiment of the application on the DRD data set can be found. In the case of 19 original feature numbers, the feature number usage of 20 fuzzy rules is less than one time, and the maximum feature number usage is only 9 features. It is proved that each fuzzy rule has achieved rule reduction, can well mine the potential relationship between features, and remove redundant features, and finally forms a simplified and efficient fuzzy rule base.

[0128] In order to prove the effectiveness of SESSC, by keeping other parameters unchanged, the complexity and classification performance of the model when not using SESSC and using SESSC are analyzed. As shown in Figure 6 , as shown in Figure 6 , the model parameter number and RCA value comparison diagram obtained when not using SESSC and using SESSC is provided by the embodiment of the application. From Figure 6 , it can be seen that after using SESSC, the model complexity of each data set is reduced to different degrees, indicating that the feature usage is different. And when the data dimension is large, such as data set Satellite, Usps, the model accuracy is higher than that without using SESSC. Although the RCA value of data sets DRD and Steel has a significant decrease, but it is caused by too few features used. Therefore, it can be concluded that SESSC can effectively reduce the number of parameters, and has better classification performance when processing high-dimensional data, and helps to build a simple and efficient fuzzy system.

[0129] In this embodiment, the fuzzy rule consequent of the fuzzy rule of the pre-constructed TSK fuzzy classification model is optimized by using a deep learning algorithm, including:

[0130] initializing the linear weight vector and bias term of the fuzzy rule consequent

[0131] The training set sample data is batch normalized, and the normalized data is applied to the Sigmoid function for nonlinear activation;

[0132] The parameters of the fuzzy rule consequent are trained and optimized using the normalized and nonlinear activated training set sample data, until the optimization objective function of the fuzzy rule consequent meets the convergence condition;

[0133] The expression of the optimization objective function of the fuzzy rule consequent is as follows:

[0134]

[0135] Among them, represents the true label of the sample n belongs to the category c ; represents the probability that the sample is predicted to be of category c , represents the normalized activation weight of the fuzzy rule r, represents the linear weight vector of the fuzzy rule r to the category c, represents the transpose of the linear weight vector , is the nonlinear embedding result of the sample , which represents the feature mapping of the sample after batch normalization processing and activation function .

[0136] BN is used to normalize the input data, which can accelerate the training of the model by reducing the internal covariance, and can alleviate the problem of gradient disappearance. For a small batch , the output of BN is:

[0137]

[0138] Among them:

[0139]

[0140]

[0141] In the formula, and are the mean and standard deviation of the small batch sample, and are parameters that can be learned during training, and their role is to slightly offset if the standardized data is not suitable, so that it becomes suitable. It is usually set to 1e-8 to avoid cases where the denominator is 0. After performing Batch Normalization (BN) on the original input data, the resulting... Nonlinear processing is performed to enhance the detection and characterization capabilities of implicit nonlinear structures and features in the data. This invention uses the sigmoid function as the nonlinear function, and the data after nonlinear transformation is... Its expression is as follows:

[0142]

[0143] Therefore, the output of the TSK fuzzy classifier can also be rewritten as:

[0144]

[0145] In this embodiment, during the training and optimization of the parameters of the fuzzy rule consequent:

[0146] Assign a retention probability to each fuzzy rule; in each training optimization iteration, generate a random number in the interval [0, 1] for each fuzzy rule; if the generated random number is less than or equal to the retention probability set for that fuzzy rule, then retain that fuzzy rule in the current iteration; otherwise, discard that fuzzy rule; only train and update the parameters of the fuzzy rules that are not discarded. Figure 7 As shown, Figure 7 This is a schematic diagram illustrating the principle of using dropout technology to discard fuzzy rules, as provided in an embodiment of the present invention. Figure 7 The second fuzzy rule was discarded.

[0147] As an example, to demonstrate the effectiveness of deep learning optimization techniques, this embodiment analyzes the model's complexity and classification performance with and without deep learning optimization techniques, keeping other parameters constant. Figure 8 As shown, Figure 8 This is a comparison chart of the number of model parameters and RCA values ​​obtained using deep learning optimization techniques, as provided in an embodiment of the present invention. Figure 8 It can be seen that using the model with redundant rules removed reduces the number of parameters, improves the model's generalization ability, and significantly enhances the overall classification performance of the model.

[0148] Finally, the schematic diagram of part of the structure of the improved TSK fuzzy classification model obtained in this embodiment is as follows: Figure 9 As shown, the model is named SESSC-DL-MTSK. The figure shows that the antecedent of the fuzzy rule is generated and optimized using SESSC (Supervised Enhanced Soft Subspace Clustering Algorithm), and the consequent can be generated and optimized using DL (Deep Learning Technique), and the fuzzification operation is performed using MTSK.

[0149] Example 3

[0150] This example is based on the data classification method based on the improved TSK fuzzy classification model proposed in Example 2, and medical data, food data, satellite data and interpersonal interaction data are classified and tested to verify the effectiveness of the SESSC-DL-MTSK model proposed in the application.

[0151] This example evaluates the SESSC-DL-MTSK model proposed from the aspects of classification performance and model complexity. When processing high-dimensional data and large-scale data, the model has better classification performance and robustness, ensures that the activation level of each fuzzy rule in the fuzzy rule base constructed by the model is relatively average, and does not appear only part of the activation, and at the same time of ensuring good performance, the complexity of the whole model is lower than that of the prior art model, and is more concise.

[0152] In this example, 10 classification data sets in different fields are selected from the UCI machine learning database Repository and the NTU machine learning database Lib for evaluating the proposed model algorithm. For each data set, 70% of the samples are randomly selected as the training set, and the remaining 30% of the samples are selected as the test set. Table 1 gives the detailed information of the actual data set, including the name of the data set, the number of samples and features in the data set, and the number of categories.

[0153] Table 1 Actual data set information

[0154]

[0155] In this example, the original classification accuracy (RCA), balanced classification accuracy (BCA) and Jaccard-measure are used as the three indicators, and five-fold cross-validation is used to evaluate the classification performance. The RCA is the total number of correctly classified test samples divided by the total number of test samples, the BCA is the average of the RCA of each category, and is used to evaluate the performance of the model on the unbalanced data set. Jaccard-measure is used to evaluate the robustness of processing unbalanced data sets.

[0156] Table 2 Comparison of average RCA of six algorithms on ten data sets

[0157]

[0158] Table 3 Comparison of average BCA of six algorithms on ten data sets

[0159]

[0160] Table 4. Comparison of average Jaccard index of six algorithms on ten datasets

[0161]

[0162] The average test RCA, BCA and Jaccard-measure are shown in Table 2, Table 3 and Table 4, respectively, and the maximum value (best performance) on each dataset is in bold. From Table 2, it can be seen that the proposed method is superior to other algorithms in terms of RCA index on 7 datasets, especially for datasets with dimension greater than 30. Further, Table 3 and Table 4 give the results of BCA and Jaccard-measure, and it can be seen that the proposed method also shows superiority and better robustness on imbalanced datasets. On average, the proposed algorithm is superior to FCM-based algorithms (TSK-PCA, TSK-UR-BN, LogTSK) in handling high-dimensional data.

[0163] The SESSC-DL-MTSK of the embodiment has lower model complexity. The advantage is verified from two aspects of quantitative evaluation of model complexity and intuitiveness based on fuzzy rule form.

[0164] The model complexity of the SESSC-DL-MTSK of the embodiment is calculated based on the optimal model in terms of BCA value in Table 3, and the results are shown in Table 5. From the results, it can be seen that the invented model is superior to other existing models on 6 datasets. More importantly, the model complexity of the proposed method is significantly reduced, and the average number of parameters on 10 datasets is the least.

[0165] Table 5. Comparison of model complexity of six algorithms on ten datasets

[0166]

[0167] For example, taking the DRD dataset as an example, as shown in Figure 10 and Figure 11 , the figure of rule feature number reduction effect on the DRD dataset without using the deep learning optimization technology provided by the embodiment of the application is Figure 10 , the figure of rule feature number reduction effect on the DRD dataset using the deep learning optimization technology provided by the embodiment of the application is Figure 11 , the figure of rule feature number reduction effect on the DRD dataset using the deep learning optimization technology provided by the embodiment of the application is Figure 9 and Figure 10It can be seen that only part of the features are used after applying SESSC, i.e., the length of fuzzy rules is reduced. Considering the limited ability of human cognition, the proposed method produces more interpretable results. In addition, after applying the deep learning optimization algorithm, some redundant fuzzy rules are replaced. For Figure 10 the replaced fuzzy rules are 5, 6, 7, 9, 12, and 16, while for Figure 11 the replaced fuzzy rules are 5, 6, 7, 9, 12, 13, 16, and 18. Therefore, the proposed SESSC-DL-MTSK can construct a fuzzy system with lower model complexity and improve interpretability.

[0168] Embodiment 4

[0169] As shown in Figure 12 , the embodiment proposes a data classification system based on an improved TSK fuzzy classification model, which is applied to the data classification method based on the improved TSK fuzzy classification model as described in the above embodiments, and includes a division module 100, an optimization module 200, and a classification module 300.

[0170] The division module 100 is configured to obtain a data set to be classified and divide the data set to be classified into a training set and a test set. The optimization module 200 is configured to optimize the fuzzy rule antecedent and the fuzzy rule consequent of the pre-constructed TSK fuzzy classification model based on the training set by using the supervised enhanced soft subspace clustering algorithm and the deep learning algorithm, respectively. The classification module 300 is configured to input the test set into the optimized TSK fuzzy classifier model for classification processing to obtain a data classification result.

[0171] It should be noted that the foregoing explanation of the data classification method based on the improved TSK fuzzy classification model is also applicable to the data classification system based on the improved TSK fuzzy classification model of this embodiment, which will not be described here.

[0172] In the description of the present specification, the description of the terms "one embodiment", "some embodiments", "an example", "a specific example", or "some examples" means that the specific features, structures, materials or characteristics described in connection with the embodiment or example are included in at least one embodiment or example of the present application. In the present specification, the illustrative description of the above terms is not necessarily for the same embodiment or example. Moreover, the specific features, structures, materials or characteristics described can be combined in any one or N embodiments or examples in a suitable manner. In addition, the skilled in the art can combine and combine the different embodiments or examples described in the present specification and the features of the different embodiments or examples without contradiction.

[0173] Furthermore, the terms "first", "second", or the like, are used only to describe the purpose of distinguishing between different features, and are not intended to imply or suggest relative importance of, or a hierarchy between, the indicated features. Thus, a feature defined with "first" or "second" can implicitly or explicitly include at least one of the other feature. The meaning of "N" is at least two, such as two, three, etc., unless explicitly specified otherwise.

[0174] Any process or method described in a flowchart or otherwise described herein can be understood as representing code modules, segments, or portions of code that include one or more executable instructions for implementing specific logic functions (or steps) of the process, and the various embodiments of the present application can include additional or fewer steps or processes in alternative embodiments, and the order of the steps can be changed, including the use of concurrent steps or functions, unless otherwise specifically noted. The various embodiments of the present application can be implemented in hardware, software, firmware, or any combination thereof.

[0175] It should be understood that aspects of the present application can be implemented in hardware, software, firmware, or any combination thereof. In the above embodiments, the steps or methods can be implemented in software or firmware that is stored in memory and executed by a suitable instruction execution system. As such, the steps or methods can be stored in the memory of the instruction execution system, or stored on or in a machine-readable medium such as a floppy disk, a RAM, a ROM, an erasable programmable read only memory (EPROM), EEPROM, a magnetic or optical card, a flash memory, or an optical storage device, etc. and executed by the instruction execution system. If implemented in hardware, the steps or methods can be implemented with any or a combination of the following technologies, which are all well known in the art: a discrete logic circuit having logic gates for implementing logic functions upon data signals, an application specific integrated circuit having appropriate combinational logic gates, a programmable gate array (PGA), a field programmable gate array (FPGA), etc.

[0176] Those of skill in the art would understand that information and signals can be represented using any of a variety of technologies and techniques. For example, data, instructions, commands, information, signals, bits, symbols, and chips that can be referenced throughout the above description can be represented by voltages, currents, electromagnetic waves, magnetic fields or particles, optical fields or particles, or any combination thereof.

[0177] Obviously, the above-described embodiments of the present application are only illustrative and not restrictive, and modifications and changes can be made by those skilled in the art on the basis of the above description. It is therefore not necessary to exhaust all the embodiments. Any modification, equivalent replacement, and improvement made within the spirit and principle of the present application shall be included in the scope of protection of the claims of the present application.

Claims

1. A data classification method based on an improved TSK fuzzy classification model, characterized in that, The method comprises the following steps: obtaining a data set to be classified, and dividing the data set to be classified into a training set and a test set; based on the training set, using a supervised enhanced soft subspace clustering algorithm and a deep learning algorithm to optimize the fuzzy rule antecedent and the fuzzy rule consequent of a pre-constructed TSK fuzzy classification model respectively, comprising: initializing the linear weight vector and the bias term of the fuzzy rule consequent; performing batch normalization processing on the training set sample data, and applying a Sigmoid function to the normalized data for nonlinear activation; training and optimizing the parameters of the fuzzy rule consequent using the normalized and nonlinearly activated training set sample data until the optimization objective function of the fuzzy rule consequent meets the convergence condition; wherein the expression of the optimization objective function of the fuzzy rule consequent is as follows: wherein, denotes a sample belongs to the class c with true label y; denotes a sample is predicted to belong to the class c with probability p, N is the total number of samples, C is the total number of classes, denotes the normalized activation weight of fuzzy rule r, denotes the linear weight vector of fuzzy rule r for class c, denotes the transpose of the linear weight vector , is the nonlinear embedding result of sample , denotes the feature map of sample after batch normalization processing and activation function ; wherein the pre-constructed TSK fuzzy classification model comprises an input layer, a fuzzification layer, a fuzzy rule layer, an inference layer and an output layer; input samples are input to the fuzzification layer through the input layer; the fuzzification layer calculates membership values based on the feature values of the input samples, and the cluster centers and feature ranges of the fuzzy rules; the fuzzy rule layer triggers the corresponding fuzzy rules according to the membership values output by the fuzzification layer; the inference layer performs fuzzy reasoning according to the triggered fuzzy rules to obtain normalized trigger values of the fuzzy rules; and the output layer performs defuzzification processing according to the normalized trigger values of the fuzzy rules to obtain the final classification result; using an improved fuzzification operation to uniformly process the trigger level of the fuzzy rule, and the expression is as follows: wherein, is a trigger value of a fuzzy rule r representing a trigger level of the fuzzy rule r on the current input sample; is a number of dimensions of input features, is a value of the input sample on the d th feature dimension, is a cluster center of the fuzzy rule r on the d th feature dimension, is a variance of the fuzzy rule r on the d th feature dimension, is a maximum value among all fuzzy rule trigger values; inputting the test set into the optimized TSK fuzzy classifier model for classification processing to obtain the data classification result. 2.The data classification method based on the improved TSK fuzzy classification model according to claim 1, characterized in that, The inference layer carries out fuzzy inference according to the triggered fuzzy rules to obtain normalized trigger values of the fuzzy rules The expression is as follows: where R is the total number of fuzzy rules, denotes an exponential operation on the trigger value of the fuzzy rule r .​ 3.The data classification method based on the improved TSK fuzzy classification model according to claim 2, characterized in that, The output layer performs defuzzification processing according to the normalized trigger values of the fuzzy rules to obtain the final classification result, and the expression is as follows: wherein, represents the predicted value of the input sample x on the class c, represents the output value of the fuzzy rule r on the class c. 4.The data classification method based on the improved TSK fuzzy classification model according to claim 1, characterized in that, using a supervised enhanced soft subspace clustering algorithm to optimize the fuzzy rule antecedent of the fuzzy rule of the pre-constructed TSK fuzzy classification model, comprising: initializing a fuzzy membership matrix for a fuzzy rule antecedent , a cluster center matrix , a weight matrix , and a probability matrix ; For the training set data samples and their corresponding sample labels , the objective function of the supervised enhanced soft subspace clustering algorithm is constructed , the expression is as follows: wherein, are the adjustment term parameters, N is the number of training set data samples, R is the number of cluster centers, i.e., the total number of fuzzy rules, D is the dimension number of input features, C is the number of classification categories, denotes the membership degree of the n th sample to the r th cluster center, m > 1 is the fuzzification index, denotes the label value of the sample belonging to the category c , denotes the weight of the cluster center r in the feature dimension d , denotes the probability value of the cluster center r in the category c , is the n th input sample, is the r th cluster center, is the cluster center of all data samples of the training set, is the weight of the r th cluster center; Iterative optimization of the fuzzy membership matrix Cluster center matrix Weight matrix and probability matrix until the objective function If the convergence condition is met, the optimized fuzzy rule antecedent is obtained. 5.The data classification method based on the improved TSK fuzzy classification model according to claim 4, characterized in that, The fuzzy membership matrix is ​​iteratively optimized using the Lagrange multiplier method. Cluster center matrix Weight matrix and probability matrix until the objective function The convergence condition is met. 6.The data classification method based on the improved TSK fuzzy classification model according to claim 2, characterized in that, in the training and optimization of the parameters of the fuzzy rule consequent, the method further comprises: allocating a reserved probability to each fuzzy rule; generating a random number in the interval [0, 1] for each fuzzy rule in each training and optimization iteration; if the generated random number is less than or equal to the reserved probability set for the fuzzy rule, the fuzzy rule is retained in the current iteration, otherwise, the fuzzy rule is discarded; only the fuzzy rules that are not discarded are trained and the parameters are updated.

7. A data classification system based on an improved TSK fuzzy classification model, applied to the data classification method based on an improved TSK fuzzy classification model according to any one of claims 1 to 6, characterized in that, The method comprises the following steps: an acquisition module is configured to obtain a data set to be classified, and divide the data set to be classified into a training set and a test set; an optimization module is configured to use a supervised enhanced soft subspace clustering algorithm and a deep learning algorithm to optimize the fuzzy rule antecedent and the fuzzy rule consequent of a pre-constructed TSK fuzzy classification model based on the training set; a classification module is configured to input the test set into the optimized TSK fuzzy classifier model for classification processing to obtain the data classification result.

Images