A fuzzy hyperbox neural network classification method based on small sample data

Through data preprocessing and superbox construction methods, an adaptive superbox fuzzy set is generated, which solves the problem that the classification performance of fuzzy superbox neural networks in small sample data is affected by the input order, and realizes effective classification under small sample conditions and Online learning ability.

CN118673369BActive Publication Date: 2024-12-06NORTH CHINA MUNICIPAL ENG DESIGN & RES INST
View PDF 2 Cites 0 Cited by

Patent Information

Application Number
CN202411172719.3
Authority / Receiving Office
CN · China
Patent Type
Patents(China)
Current Assignee / Owner
Filing Date
2024-08-26
Publication Date
2024-12-06
Estimated Expiration
2044-08-26

AI Technical Summary

Technical Problem

Traditional fuzzy superbox neural network classification method is easily affected by the data input sequence under small sample data conditions, resulting in large errors in classification results and it is difficult to achieve effective learning and classification under limited data volume.

Method used

Through data preprocessing and hyperboxing construction, including normalization processing, data segmentation, initialization of hyperboxing, overlap detection and shrinking operations, an adaptive final hyperboxing fuzzy set is generated, reducing the impact of input sequence on the results, and in small samples Effective classification is achieved under conditions.

Benefits of technology

The superbox construction method has been improved, the problem of classification performance being affected by the order of sample input is overcome, and the ability to learn online is able to achieve better classification performance under small sample data, avoiding the problem of overfitting and poor classification results.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN118673369B_ABST
    Figure CN118673369B_ABST
Patent Text Reader

Abstract

The present invention proposes a fuzzy hyperbox neural network classification method based on small sample data. The method comprises: firstly normalizing the data set and scaling the data to a uniform range; secondly dividing the normalized data set into a training data set and a test data set; then constructing a hyperbox, taking the point with the maximum density of each category sample in the training data set as the center, defining the side length parameter to determine the size of the hyperbox, and generating an initial hyperbox fuzzy set; then performing two-step operations of overlap detection and shrinkage on the initial hyperbox fuzzy set to appropriately adjust and refine the hyperbox, and generating a final hyperbox fuzzy set; finally calculating the classification result according to the test data set. Compared with the traditional classification method, this method improves the problem that the hyperbox construction is affected by the input order, has the ability of online learning, does not need to retrain the early and new data, and can achieve better classification performance under small sample conditions.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] The invention relates to the field of neural network classification, and in particular to a fuzzy hyperbox neural network classification method based on small sample data. Background Art

[0002] Small sample data classification is a classification task when the number of samples in the data set is relatively small. In machine learning and deep learning, a large amount of data is usually required to train the model so that the model can learn enough features and rules to perform accurate classification. However, in practical applications, there is a limited number of samples, which leads to the inability to achieve applicable accuracy after model training. At this time, a small sample learning method is needed to solve the classification problem; Fuzzy hyperbox neural network is an effective algorithm for solving sample classification problems. The algorithm uses hyperbox fuzzy sets to create and store knowledge in its network structure, and can use less data to train the model, reducing the high cost associated with data collection and labeling. It has good online learning ability and classification performance, but the classification performance of traditional fuzzy hyperbox neural network is affected by the input order of data in the data set, which will affect the subsequent generated hyperbox fuzzy sets. If the training data set uses an inappropriate input order, the generated hyperbox fuzzy sets will have large errors in the classification results during the test process. Summary of the invention

[0003] The purpose of the present invention is to overcome the above technical shortcomings and deficiencies, and to provide a fuzzy hyperbox neural network classification method based on small sample data, which improves the problem that the hyperbox construction is affected by the input order, has online learning capabilities, does not require retraining of early and new data, and can achieve better classification performance under small sample conditions. The technical solution adopted by the present invention is as follows:

[0004] A fuzzy hyperbox neural network classification method based on small sample data of the present invention comprises the following steps:

[0005] Step 1: Data preprocessing: normalize the input data set to ensure that all eigenvalues ​​are scaled to a uniform numerical range, providing a standardized basis for subsequent hyperbox construction.

[0006] Step 2: Dataset segmentation: segment the processed input dataset into training dataset and test dataset;

[0007] Step 3: Hyperbox construction: Initialize the hyperbox based on the data in the training data set to generate the initial hyperbox fuzzy set;

[0008] Step 3.1: Input all the data in the training data set and calculate the density of all data in each category according to the following formula, where ρ x is the density of each data, d xyis the Euclidean distance between each data and other data of the same category, x is the input training sample, y is other similar data, n is the dimension, d c is the neighborhood radius of the current data, It is a binary operation;

[0009] ;

[0010] ;

[0011] Step 3.2: Based on the density of all data calculated in step 3.1, select the point with the largest data density in each category as the center, set the side length parameter to initialize the hyperbox, and generate the initial hyperbox according to the following formula, where is a user-defined edge length parameter. Since the data is normalized in step 2, its range is [0,1]. x is the input training sample, v min is the minimum point of the hyperbox, v max It is the maximum point of the superbox;

[0012] ;

[0013] Step 3.3: Determine whether the initial hyperbox fuzzy set covers all training data. If there is uncovered data in the current hyperbox fuzzy set, continue to select the maximum density point of each category of data from the uncovered data to generate a new hyperbox until the hyperbox fuzzy set covers all data.

[0014] Step 4: Hyperbox adjustment: perform overlap detection and shrinkage operations on the generated initial hyperbox fuzzy set. The overlap detection operation is to identify and mark the overlapping areas between hyperboxes, and the shrinkage operation is to adjust the hyperbox boundaries to reduce overlap and obtain the final hyperbox fuzzy set.

[0015] Step 5: Model testing: Use the test data set to test the final hyperbox fuzzy set generated during the training process, calculate the classification results and output them.

[0016] Preferably, the data preprocessing in step 1 is to normalize the input data set to ensure that all eigenvalues ​​are scaled to a uniform numerical range, providing a standardized basis for subsequent hyperbox construction. The specific steps are as follows:

[0017] Step 1.1: Determine the normalization range and define the range of all data in the data set between [0,1];

[0018] Step 1.2: For the data that needs to be normalized, find the maximum and minimum values;

[0019] Step 1.3: According to the minimum-maximum normalization formula Calculate the normalized data, where i represents the i-th data, x min Indicates the minimum value of the data in the data set, x max Indicates the maximum value of the data in the data set.

[0020] Preferably, the data set segmentation in step 2 is to segment the processed input data set into a training data set and a test data set, and the specific steps are as follows:

[0021] Step 2.1: Set the ratio of the training data set and the test data set, and randomly extract corresponding data from the preprocessed data set to form the training data set and the test data set.

[0022] Preferably, in the fourth step of adjusting the super-box, two operations of overlap detection and shrinkage are performed on the generated initial super-box fuzzy set. The overlap detection operation is to identify and mark the overlapping areas between the super-boxes, and the shrinkage operation is to adjust the super-box boundaries to reduce overlap, so as to obtain the final super-box fuzzy set. The specific steps are as follows:

[0023] Step 4.1: Determine whether there is overlap between hyperboxes of different categories in the initial hyperbox fuzzy set. The following four overlap situations are used to determine whether there is overlap between hyperboxes of different categories. ji and w ji are the minimum and maximum points of the jth hyperbox on dimension i, respectively, v ki and w ki are the minimum and maximum points of the kth hyperbox in dimension i, respectively;

[0024] Overlapping situation 1: ;

[0025] Overlapping situation 2: ;

[0026] Overlapping case 3: ;

[0027] Overlapping situation 4: ;

[0028] Step 4.2: If there is overlap between hyperboxes of different categories detected in step 4.1, a shrinkage operation is performed to eliminate the overlapping area; where △ represents the shrinkage dimension, v j△ and w j△ are the minimum and maximum points of the j-th hyperbox in dimension △, and Respectively represent the values ​​before and after the shrink operation; v k△ and w k△ are the minimum and maximum points of the kth hyperbox in dimension △, and Respectively represent the values ​​before and after the shrink operation;

[0029] Contraction rule 1: ,but ;

[0030] Contraction Rule 2: ,but ;

[0031] Contraction Rule 3: ,

[0032] if ,but ;

[0033] on the contrary ;

[0034] Contraction Rule 4: ,

[0035] if ,but ;

[0036] on the contrary ;

[0037] Step 4.3: Determine whether there is still overlap in the processed hyperbox fuzzy set. If there is overlapping hyperbox, repeat the above steps. If not, output the final hyperbox fuzzy set.

[0038] Preferably, the model test in step 5 is performed using a test data set according to the final hyperbox fuzzy set generated during the training process, and the classification result is calculated and outputted. The specific steps are as follows:

[0039] Step 5.1: Calculate the membership function value of each hyperbox in the test data and the hyperbox fuzzy set generated by training according to the following formula; where b j is the jth superbox, is the hth input sample, b j (A h ) is the membership function value of the jth hyperbox and the input sample, a hi is the i-th value of the h-th input sample, v ji and w ji is the minimum and maximum point of the j-th hyperbox in dimension i, γ is a sensitivity parameter. h The membership function decreases when the distance to the jth hyperbox increases, and its value ranges from 0 to 1 and is defined by the user;

[0040] ;

[0041] Step 5.2: Compare the membership function values ​​calculated in step 5.1, and determine the category of the data according to the hyperbox with the highest membership;

[0042] Step 5.3: Output the final classification results based on the test data set.

[0043] The advantages and technical effects of the present invention are:

[0044] The present invention proposes a fuzzy hyperbox neural network classification method based on small sample data. The method improves the traditional hyperbox construction method, retains the advantages of the traditional fuzzy hyperbox neural network, overcomes the problem that the classification performance is affected by the sample input order, can adaptively adjust the hyperbox according to the distribution characteristics of the data, and is interpretable; the method is particularly suitable for small sample data scenarios, and can perform effective learning and classification with a limited amount of data, overcoming the problem that traditional methods are prone to overfitting or poor classification results in small sample conditions. BRIEF DESCRIPTION OF THE DRAWINGS

[0045] Figure 1 It is an algorithm flow chart of the fuzzy hyperbox neural network classification method based on small sample data of the present invention;

[0046] Figure 2 is an example diagram of generating an initial hyperbox fuzzy set in the present invention;

[0047] Figure 3 is an example diagram of four superbox overlapping situations in the present invention;

[0048] Figure 4 2 are example diagrams of four super box contraction situations in the present invention. DETAILED DESCRIPTION

[0049] The embodiments of the present invention are described in detail below, wherein the same or similar reference numerals throughout represent the same or similar elements or elements with the same or similar functions. The following embodiments are exemplary and are only used to explain the present invention, and cannot be understood as limiting the present invention.

[0050] The present invention is a fuzzy hyperbox neural network classification method based on small sample data. The algorithm uses four data sets of Thyroid, Iris, Wine and Parkinsons in the international standard data set for training and testing. The data set information is shown in Table 1. The present invention takes the Thyroid data set as an example to demonstrate the entire training and testing process. The same operation will also be applied to the other three international standard data sets. Figure 1 As shown, the specific implementation plan includes the following contents and steps:

[0051] Table 1 Dataset information

[0052]

[0053] Step 1: Data preprocessing, normalize the input data set to ensure that all feature values ​​are scaled to a uniform numerical range, providing a standardized basis for subsequent hyperbox construction and classification decisions. The specific steps are as follows:

[0054] Step 1.1: Determine the normalization range, define the range of all data in the Thyroid dataset between [0,1], and give 5 features to each of the 215 records in the Thyroid dataset;

[0055] Step 1.2, find the maximum and minimum values ​​of each of the 215 records in the Thyroid dataset;

[0056] Step 1.3: According to the minimum-maximum normalization formula The Thyroid dataset is normalized, where Indicates data, x min Indicates the minimum value of the data in the data set, x max Indicates the maximum value of the data in the data set.

[0057] Step 2: Dataset segmentation: segment the processed input dataset into training dataset and test dataset. The specific steps are as follows:

[0058] Step 2.1: Divide the training data set and the test data set in a ratio of 3:2. After the division, the number of training data sets is 129, and the number of test data sets is 86. The training data set is used in the training process of steps 3 and 4, and the test data set is used in the testing process of step 5.

[0059] Step 3: If Figure 2 As shown, the hyperbox construction uses the data in the training data set as the basis to initialize the hyperbox and generate the initial hyperbox fuzzy set. The specific steps are as follows:

[0060] Step 3.1: Input all the data in the training data set and calculate the density of all the data according to the following formula, where ρ x is the density of each data, d xy is the Euclidean distance between each data and other data of the same category, x is the data, y is other similar data, n is the dimension, d c is the neighborhood radius of the current data, It is a binary operation;

[0061] ;

[0062] ;

[0063] Step 3.2: Based on the density of all data calculated in step 3.1, select the point with the largest data density in each category as the center, set the side length parameter to initialize the hyperbox, and generate the initial hyperbox according to the following formula, where is a user-defined edge length parameter. Since the data is normalized in step 2, its range is [0,1]. x is the input training sample, v min is the minimum point of the hyperbox, v max It is the maximum point of the superbox;

[0064] ;

[0065] Step 3.3: Determine whether the initial hyperbox fuzzy set covers all training data. If there is data not covered by the hyperbox fuzzy set, perform density evaluation on the uncovered data, continue to select the maximum density point of each category of data from the uncovered data to generate a new hyperbox, and add the new hyperbox to the existing hyperbox fuzzy set until the hyperbox fuzzy set covers all training data.

[0066] Step 4: Hyperbox adjustment. Perform overlap detection and shrinkage operations on the generated initial hyperbox fuzzy set. The overlap detection operation is to identify and mark the overlapping areas between hyperboxes. The shrinkage operation is to adjust the hyperbox boundaries to reduce overlap and obtain the final hyperbox fuzzy set. The specific steps are as follows:

[0067] Step 4.1: Figure 3 As shown, overlap detection determines whether there is overlap between hyperboxes of different categories in the initial hyperbox fuzzy set. If each dimension between two hyperboxes satisfies one of the following four overlap conditions, it means that there is overlap between the two hyperboxes;

[0068] Overlapping situation 1: ;

[0069] Overlapping situation 2: ;

[0070] Overlapping case 3: ;

[0071] Overlapping situation 4: ;

[0072] Step 4.2: Figure 4 As shown, the shrinking operation, if after step 4.1, it is detected that there is overlap between hyperboxes of different categories, the shrinking operation is used to eliminate the overlapping area. During the shrinking process, only one dimension of the n-dimensional hyperbox is operated. The following shrinking rule formulas correspond to the four overlapping situations in step 4.1, where Indicates the shrinking dimension;

[0073] Contraction rule 1: ,but ;

[0074] Contraction Rule 2: ,but ;

[0075] Contraction Rule 3: ,

[0076] if ,but ;

[0077] on the contrary ;

[0078] Contraction Rule 4: ,

[0079] if ,but ;

[0080] on the contrary ;

[0081] Step 4.3: Determine whether there is still overlap in the processed hyperbox fuzzy set. If there is overlapping hyperbox, repeat the above steps to shrink until there is no overlapping hyperbox. If not, output the final hyperbox fuzzy set after processing.

[0082] Step 5: Model testing: Use the test data set to test the final hyperbox fuzzy set generated during the training process, calculate the classification results and output them. The specific steps are as follows:

[0083] Step 5.1: Calculate the membership function value of each hyperbox in the hyperbox fuzzy set of the test data and the training generated hyperbox according to the following formula to determine the degree of association between the test data and each hyperbox;

[0084] ;

[0085] Step 5.2: Compare the membership function values ​​calculated in step 5.1, and determine the category of the data according to the hyperbox with the highest membership function value;

[0086] Step 5.3: After determining the category of each test data, output the final classification result.

[0087] Finally, all the parts not described in the present invention adopt mature products and mature technical means in the prior art.

[0088] In the description of this specification, reference to the terms "one embodiment", "some embodiments", "example", "specific example", or "some examples" etc. means that the specific features, structures, materials or characteristics described in conjunction with the embodiment or example are included in the embodiments or examples of the present invention.

[0089] Although the embodiments of the present invention have been shown and described, it will be appreciated by those skilled in the art that various changes, modifications, substitutions and variations may be made to the embodiments without departing from the principles and spirit of the present invention, and that the scope of the present invention is defined by the claims and their equivalents.

Claims

1. A fuzzy hyperbox neural network classification method based on small sample data, characterized in that: The following steps are involved: Step 1: Data preprocessing: normalize the input data set to ensure that all eigenvalues ​​are scaled to a uniform numerical range, providing a standardized basis for subsequent hyperbox construction. Step 1.1: Determine the normalization range, define the range of all data in the Thyroid dataset between [0,1], and give 5 features to each of the 215 records in the Thyroid dataset; Step 1.2, find the maximum and minimum values ​​of each of the 215 records in the Thyroid dataset; Step 1.3: According to the minimum-maximum normalization formula Normalize the Thyroid data set, where i represents the i-th data, x min Indicates the minimum value of the data in the data set, x max Indicates the maximum value of the data in the data set; Step 2: Dataset segmentation: segment the processed input dataset into training dataset and test dataset; Step 3: Hyperbox construction: Initialize the hyperbox based on the data in the training data set to generate the initial hyperbox fuzzy set; Step 3.1: Input all the data in the training data set and calculate the density of all data in each category according to the following formula, where ρ x is the density of each data, d xy is the Euclidean distance between each data and other data of the same category, x is the input training sample, y is other similar data, n is the dimension, d c is the neighborhood radius of the current data, χ is a binary operation; Step 3.2: Based on the density of all data calculated in step 3.1, select the point with the largest data density in each category as the center, set the side length parameter to initialize the hyperbox, and generate the initial hyperbox according to the following formula: l is the user-defined side length parameter. Since the data is normalized in step 2, its range is [0,1]; x is the input training sample, v is min is the minimum point of the hyperbox, v max It is the maximum point of the superbox; Step 3.3: Determine whether the initial hyperbox fuzzy set covers all training data. If there is uncovered data in the current hyperbox fuzzy set, continue to select the maximum density point of each category of data from the uncovered data to generate a new hyperbox until the hyperbox fuzzy set covers all data. Step 4: Hyperbox adjustment: perform overlap detection and shrinkage operations on the generated initial hyperbox fuzzy set. The overlap detection operation is to identify and mark the overlapping areas between hyperboxes, and the shrinkage operation is to adjust the hyperbox boundaries to reduce overlap and obtain the final hyperbox fuzzy set. Step 5: Model testing: Use the test data set to test the final hyperbox fuzzy set generated during the training process, calculate the classification results and output them.

2. The fuzzy hyperbox neural network classification method based on small sample data according to claim 1 is characterized by: The data set segmentation in step 2 is to segment the processed input data set into a training data set and a test data set. The specific steps are as follows: Step 2.1: Set the ratio of the training data set and the test data set, and randomly extract corresponding data from the preprocessed data set to form the training data set and the test data set.

3. The fuzzy hyperbox neural network classification method based on small sample data according to claim 1 is characterized in that: The hyperbox adjustment in step 4 performs two operations of overlap detection and shrinkage on the generated initial hyperbox fuzzy set. The overlap detection operation is to identify and mark the overlapping areas between hyperboxes, and the shrinkage operation is to adjust the hyperbox boundaries to reduce overlap, so as to obtain the final hyperbox fuzzy set. The specific steps are as follows: Step 4.1: Determine whether there is overlap between hyperboxes of different categories in the initial hyperbox fuzzy set, based on the following four overlap situations: vji and wji are the minimum and maximum points of the jth hyperbox on dimension i, and vki and wki are the minimum and maximum points of the kth hyperbox on dimension i; Overlap case 1: v ji <v ki <w ji <w ki ; Overlap case 2: v ki <v ji <w ki <w ji ; Overlap case 3: v ji <v ki <w ki <w ji ; Overlap case 4: v ki <v ji <w ji <w ki ; Step 4.2: If there is overlap between hyperboxes of different categories detected in step 4.1, a shrinkage operation is performed to eliminate the overlapping area; where Δ represents the shrinkage dimension, v jΔ and w jΔ are the minimum and maximum points of the j-th hyperbox in dimension Δ, and Respectively represent the values ​​before and after the contraction operation; vkΔ and wkΔ are the minimum and maximum points of the kth hyperbox on the dimension Δ, and Respectively represent the values ​​before and after the shrink operation; Contraction rule 1: v jΔ <v kΔ <w jΔ <w kΔ ,but Contraction rule 2: v kΔ <v jΔ <w kΔ <w jΔ ,but Contraction rule 3: v jΔ <v kΔ <w kΔ <w jΔ , If (w kΔ -v jΔ )<(w jΔ -v kΔ ),but on the contrary Contraction rule 4: v kΔ <v jΔ <w jΔ <w kΔ , If (w kΔ -v jΔ )<(w jΔ -v kΔ ),but on the contrary Step 4.3: Determine whether there is still overlap in the processed hyperbox fuzzy set. If there is overlapping hyperbox, repeat the above steps. If not, output the final hyperbox fuzzy set.

4. The fuzzy hyperbox neural network classification method based on small sample data according to claim 1 is characterized in that: The model test in step 5 is to use the test data set to test the final hyperbox fuzzy set generated during the training process, calculate the classification results and output them. The specific steps are as follows: Step 5.1: Calculate the membership function value of each hyperbox in the test data and the hyperbox fuzzy set generated by training according to the following formula; where b j is the jth superbox, A h =(a h1 , a h2 , …, a hn )∈I n is the hth input sample, b j (A h ) is the membership function value of the jth hyperbox and the input sample, ahi is the i-th value of the h-th input sample, vji and wji are the minimum and maximum points of the j-th hyperbox on dimension i, and γ is a sensitivity parameter. h The membership function decreases when the distance to the jth hyperbox increases, and its value ranges from 0 to 1 and is defined by the user; Step 5.2: Compare the membership function values ​​calculated in step 5.1, and determine the category of the data according to the hyperbox with the highest membership; Step 5.3: Output the final classification results based on the test data set.

Citation Information

Patent Citations

  • Fatigue driving fusion detection method based on soft computing

    CN101746269A

  • Ammonian online soft measuring method for dynamic modularized nerve network

    CN102662040A