A semi-supervised width learning classification method based on incremental self-training framework

Abstract

The invention discloses a semi-supervised width learning classification method based on an incremental self-training framework. The framework includes an initialization phase, in which all unlabeled data are clustered and sample deterministic sizes are used to form an ordered query sequence table, and part of the data is sequentially selected to initialize an unlabeled data pool, and the initial labeled data is used to train a base classifier, and the data in the unlabeled data pool is annotated; in an incremental self-update phase, auxiliary training data is determined according to a selection metric value and a set threshold, and is merged into a labeled training data set to form new labeled training data for updating the base classifier; at the same time, a batch of data is sequentially obtained from the query sequence table to update the unlabeled data pool and its new classification pseudo-label is predicted by the base classifier; the incremental self-update phase is repeated until the unlabeled data pool is empty. The framework of the invention reduces a large number of repeated operations, saves resources, and solves the problem of time consumption caused by iteration, making it suitable for large-scale data applications.

Description

Technical Field

[0001] The present invention belongs to the technical field of semi-supervised learning, and in particular relates to a semi-supervised width learning classification method based on an incremental self-training framework. Background Art

[0002] As an effective incremental classification method, the Broad Learning System (BLS) successfully solves the problem of many parameters and high time consumption in deep architecture models. The Broad Learning System is a single-layer network structure that obtains model parameters through mathematical solutions and efficiently completes modeling. Due to its high efficiency, Broad Learning has won the favor of a large number of scientific researchers and industry. However, this method only establishes a supervised recognition model through labeled data, which limits its applicability.

[0003] However, in real life and applications, obtaining labeled data is time-consuming and costly. In particular, with the rapid development of technologies such as the Internet of Things and cloud computing, the amount of global data has exploded. In this big data context, it is almost impossible to obtain labels for all data. Supervised learning methods such as width learning cannot utilize unlabeled samples, which is not conducive to improving model performance. Semi-supervised learning, as a compatible method, can solve the labeling obstacle in supervised learning by combining a large number of unlabeled samples with a small number of labeled samples to build a high-performance recognition model. Although the self-training semi-supervised method can achieve relatively good classification results. However, its iterative learning method greatly increases its time consumption. Especially in the context of big data, the standard self-training method has no advantage in wide application.

[0004] With the rapid development of technologies such as the Internet of Things and cloud computing, the amount of global data has exploded. In practical application problems, there is more and more large-scale data, and labeling such large-scale data is a very time-consuming, labor-intensive and material-consuming task. Therefore, how to achieve fast and efficient semi-supervised classification in large-scale data scenarios with a small number of labeled samples is a very meaningful topic. Summary of the invention

[0005] The main purpose of the present invention is to overcome the shortcomings and deficiencies of the prior art and to propose a semi-supervised width learning classification method based on an incremental self-training framework.

[0006] In order to achieve the above object, the present invention adopts the following technical solutions:

[0007] A semi-supervised width learning classification method based on an incremental self-training framework, characterized by comprising an initialization phase and an incremental self-update phase;

[0008] In the initialization phase, all unlabeled data samples are clustered and the sample deterministic size is used to form an ordered query sequence table;

[0009] The query sequence table is used to initialize the unlabeled data pool, and the initial labeled data is used to initialize the base classifier for supervised learning, and the data in the unlabeled data pool is predicted and classified to form its pseudo label, and the corresponding selection metric is calculated;

[0010] K-Means is used to cluster the unlabeled data, and a descending query list is formed according to the distance between the sample point and its cluster center, and some data is obtained from it to initialize the unlabeled data pool; in the incremental self-update stage, the auxiliary training data is determined according to the selected metric value and the set threshold and merged into the labeled training data set to form new labeled training data for updating the base classifier; at the same time, a batch of data is sequentially obtained from the query sequence table to update the unlabeled data pool and its new classification pseudo-label is predicted by the base classifier to form unlabeled data with pseudo-labels;

[0011] Repeat the incremental self-update phase until the unlabeled data pool is empty;

[0012] The BLS model is trained using the initial labeled data or the updated unlabeled data. The BLS model parameters include the number of mapping feature groups, the number of features within a group, and the number of enhanced nodes.

[0013] Use the BLS model to annotate pseudo labels on the data in the unlabeled data pool and calculate the confidence level corresponding to the unlabeled data;

[0014] Confidence ranking of pseudo-label data is performed, and auxiliary training data is determined by ranking and setting thresholds;

[0015] Determine whether there is auxiliary training data to be added to the labeled training data. If not, use the dynamic neural node adjustment mechanism to adjust the number of enhanced nodes in the BLS network structure, and jump to step S2 for a new round of training and labeling;

[0016] If yes, transfer this part of auxiliary training data to the unlabeled data pool and label the training data, obtain a batch of data from the query list to update the unlabeled data pool, jump to step S2, use the new unlabeled data pool to train the BLS model and label the unlabeled data;

[0017] When the unlabeled data pool is empty, all the unlabeled data are labeled, and the BLS model is trained using the labeled data to output the final result.

[0018] Furthermore, there are eight types of data involved in the incremental self-training framework, specifically:

[0019] Initial labeled data Unlabeled data {X U}、Unlabeled data pool Query sequence table Q, batch data Q used to update the unlabeled data pool a , pseudo-labeled data Auxiliary training data And new labeled data Where l, s, t, and q represent the number of samples of initial labeled data, unlabeled data pool, pseudo-label data, and added labeled data, respectively.

[0020] Furthermore, in the incremental self-update phase, the unlabeled data with pseudo labels are arranged in descending order according to the selection metric value to reduce the comparison and selection time of the auxiliary training data;

[0021] Furthermore, in the incremental self-update phase, a batch of data is obtained from the query sequence table, and the size of the batch of data is fixed or changes adaptively.

[0022] Furthermore, the use of K-Means to cluster unlabeled data is specifically as follows:

[0023] S11, randomly select a sample in the data set as the first initialized cluster center;

[0024] S12, calculating the distance between each sample point in the sample and the initialized cluster center, and selecting the shortest distance;

[0025] S13, select the point with the largest distance as the new cluster center by probability;

[0026] S14, repeat steps S12-S13 until k cluster centers are selected;

[0027] S15. Use K-Means algorithm to calculate the final clustering result for k cluster centers.

[0028] Further, the following steps are included: the training of the BLS model is specifically:

[0029] Initialize model parameters;

[0030] Using the model input X, calculate the feature map node Z n ≡[Z1,Z2,…,Z n ], the process is a linear mapping, where the i-th group of mapping features Z i The calculation of is as shown in formula (1), the connection weight W ei and β ei is randomly generated. For linear mapping, each set of features contains k mapping features;

[0031]

[0032] In order to increase the nonlinearity of the features in the model, the mapping node Z n Compute Enhanced Node H m ≡[H1,H2,…,H m ], the process is a nonlinear mapping, in which the enhanced feature H j The calculation of is as shown in formula (2), the connection weight W hj and β hj is randomly generated, ζ(Z,W hj ,β hj ) is a nonlinear activation function;

[0033] H j ≡ζ(Z n W hj +β hj ),j=1,2,…,m (2)

[0034] Mapping nodes represent linear features of data, and enhancement nodes represent nonlinear features of data;

[0035] Combine linear features and nonlinear features into a feature matrix A = [Z|H] and connect it to the model output layer Y;

[0036] Since the information of the model output layer Y is known, that is, the category corresponding to the data, the connection weight W between the feature layer and the output layer can be calculated. The calculation formula is as follows:

[0037] W=A -1 Y (3)

[0038] During calculation, ridge regression is used to solve the weight matrix, that is, W is solved by the optimization problem of formula (3):

[0039]

[0040] Among them, σ1=σ2=v=u=2, and the solution is:

[0041] W=(λI+AA T ) -1 A T Y (5).

[0042] Furthermore, training the BLS model also includes the following steps:

[0043] When the characteristic matrix A is a non-singular matrix, there is no violation, so the violation value A of A is used. + To replace A -1 , A + Calculated by formula (6):

[0044]

[0045] Then, for the data X input into the model, the output of the model is:

[0046]

[0047] Calculate the pseudo label and the corresponding confidence level corresponding to the output, where the pseudo label is obtained by formula (8):

[0048] Label=Max(Y)(8)

[0049] Furthermore, the confidence is calculated using formula (9), specifically:

[0050]

[0051] Furthermore, the dynamic node adjustment strategy is based on the training accuracy ACC of the current BLS model. train Adjust the number of enhanced nodes in the model;

[0052] The number of enhanced nodes increased by Δm and the number of new nodes m ′ Calculated by formula (10) and formula (11):

[0053]

[0054] m ′ =m+Δm(11)

[0055] Among them, ACC thred is the expected accuracy, m is the original number of enhanced nodes, It is a round-up operation.

[0056] Compared with the prior art, the present invention has the following advantages and beneficial effects:

[0057] 1. The incremental self-training framework proposed in the present invention introduces the concepts of unlabeled data pool and increment. Unlike the original self-training method, all unlabeled samples are input at once. This framework incrementally inputs unlabeled data into the classifier, reducing a large number of repeated operations and saving resources. At the same time, it solves the time consumption problem caused by iteration, making it suitable for large-scale data applications.

[0058] 2. The incremental self-training framework proposed in the present invention uses a clustering method to process unlabeled samples to form a query sequence table, thereby ensuring the representativeness of the incremental input samples and the performance of the base classifier in the early stage of iteration. At the same time, the introduction of the query sequence table avoids repeated clustering and interaction in iteration, greatly reducing the time consumption of the method.

[0059] 3. The incremental self-training framework proposed in the present invention still retains the accuracy performance of the original self-training method.

[0060] 4. The semi-supervised width learning classification method of the present invention uses incremental self-training to guide width learning, so that it is expanded into a semi-supervised learning method. This method can obtain better classification results using a small number of labeled samples, thereby improving the applicability of the width learning method.

[0061] 5. The semi-supervised width learning classification method of the present invention adopts incremental self-training. In this mode, unlabeled data are added to the unlabeled data pool in batches to alleviate the pressure of large-scale data and make it adaptable to more big data application scenarios; it can efficiently solve complex classification problems in different scenarios. BRIEF DESCRIPTION OF THE DRAWINGS

[0062] Figure 1 It is a schematic diagram of the framework of the present invention;

[0063] Figure 2 is a width learning model diagram in an embodiment of the present invention;

[0064] Figure 3 This is a schematic diagram of MNIST data according to an embodiment of the present invention;

[0065] Figure 4 1 is a schematic diagram showing a comparison of learning time of MNIST data according to an embodiment of the present invention as the number of labeled samples changes;

[0066] Figure 5 1 is a schematic diagram comparing the learning time of MNIST data according to the embodiment of the present invention as the amount of unlabeled samples changes;

[0067] Figure 6 is a flow chart of the semi-supervised width learning classification method of the present invention;

[0068] Figure 7 is a training flow chart of the BLS model in the method of the present invention;

[0069] Figure 8 This is a schematic diagram of NORB data according to an embodiment of the present invention;

[0070] Fig. 9 is a diagram of the NORB iteration process of an embodiment of the present invention;

[0071] Fig.10 It is a schematic diagram of classification results of NORB data according to an embodiment of the present invention as the number of labeled samples changes;

[0072] Fig.11 It is a schematic diagram of classification results of NORB data according to an embodiment of the present invention as the amount of unlabeled samples changes. DETAILED DESCRIPTION

[0073] The present invention is further described in detail below in conjunction with embodiments and drawings, but the embodiments of the present invention are not limited thereto.

[0074] Example

[0075] like Figure 1 As shown, the present invention, an incremental self-training framework, includes:

[0076] In the initialization stage, all unlabeled samples are clustered to form a query sequence table Q according to the sample certainty size; the query sequence table Q is sorted according to the sample certainty of the clustering results, and is used to initialize the unlabeled data pool and transmit unlabeled data to the unlabeled data pool.

[0077] The query sequence table is used to initialize the unlabeled data pool. At the same time, the existing labeled samples are used to initialize the base classifier for supervised learning, and the data in the unlabeled data pool are predicted and classified to form its pseudo-label, and the corresponding selection metric is calculated.

[0078] The data involved in the framework include: initial labeled data Unlabeled data {X U}、Unlabeled data pool Query sequence table Q, batch data q used to update the unlabeled data pool a , pseudo-labeled data Auxiliary training data And new labeled data Where l, s, t, and q represent the number of samples of initial labeled data, unlabeled data pool, pseudo-label data, and added labeled data, respectively.

[0079] In the framework, there is no restriction on clustering methods, and self-organizing maps, K-means, etc. can be used; sample certainty can be measured by any metric, such as Euclidean distance. The selected metric value can be any metric, such as the prediction confidence of the sample.

[0080] In this embodiment, a fuzzy c-means clustering algorithm is specifically used, which specifically includes the following steps:

[0081] Step 1: Initialize the sample membership matrix R(0) with random numbers between (0,1) to satisfy the constraints and iterate step by step;

[0082] Step 2: Calculate the cluster center V(l) of the corresponding matrix R(l);

[0083] Step 3: Update the membership matrix R(l) and compare R(l) with R(l+1). For the accuracy threshold Thred set to be greater than 0, max{|R ik (l+1)-R ik (l)|≤Thred, then R(l+1) and V(l) are what we are looking for, that is, R=R(l+1), V=V(l), and the iteration stops. Otherwise, l=l+1 and return to step 2 and repeat.

[0084] The sample certainty size adopts the distance between the sample and its cluster center, and the query sequence table is sorted in descending order according to the distance size. The selection metric for the pseudo-label sample adopts the prediction confidence level of the current classifier for the sample.

[0085] In this embodiment, the base classifier selects a width learning method such as Figure 2 , the specific steps include:

[0086] Step 1: Set the initial parameters of the model, including the mapping feature group, the number of features in the group, and the number of enhanced nodes;

[0087] Step 2: Randomly generate connection weights based on sample size and number of nodes;

[0088] Step 3: Calculate the characteristic nodes and enhanced nodes in the network;

[0089] Step 4: Connect the feature nodes and enhanced nodes to the output layer, and use the ridge regression method to calculate the output layer connection weights;

[0090] Step 5: Use the trained width learning to predict the category of unlabeled samples.

[0091] In the incremental self-update stage, auxiliary training data is determined based on the selected metric value and the set threshold and merged into the labeled training data set to form new labeled training data for updating the base classifier; at the same time, a batch of data is sequentially obtained from the query sequence table to update the unlabeled data pool and its new classification pseudo-label is predicted by the base classifier; the size of the data is fixed or changes adaptively.

[0092] Repeat the incremental self-update phase operation until the unmarked data pool is empty.

[0093] Among them, the unlabeled data with pseudo labels are arranged in descending order according to the selection metric value to reduce the comparison and selection time of auxiliary training data.

[0094] The unlabeled data pool is dynamic. Before initialization, the unlabeled data pool is empty. In step S1, a small amount of unlabeled data is obtained sequentially from the query list for initialization. In step S2, the data in the unlabeled data pool will be continuously updated. On the one hand, the pseudo-labeled data selected as auxiliary training data will be deleted from the pool. On the other hand, new unlabeled data will be obtained sequentially from the query list as a supplement during the iteration process. When the algorithm ends, the data pool is empty.

[0095] In this embodiment, the number of a new batch of data obtained from the query sequence table to update the unlabeled data pool is fixedly set to 100.

[0096] In this embodiment, the model is tested using the public classic dataset MNIST handwritten digital images and its extended dataset EMNIST. Each grayscale image in the dataset has a pixel size of 28×28 and is divided into 10 categories, namely 0, 1, 2, 3, 4, 5, 6, 7, 8, and 9. The dataset is presented as follows Figure 3 As shown in the figure. The dataset contains 70,000 data in total. In the experiment, 1,000 labeled samples are used as the initial training set, 9,000 unlabeled samples, and 60,000 data are used for model testing. The EMNIST data contains 280,000 data, of which 240,000 are training data.

[0097] In order to explore the time consumption of the model on different amounts of unlabeled data, in this example, the initial number of labeled samples is set to 100, and it is increased by 100 each time, up to 1000. The number of unlabeled samples remains at 9000, and the number of test samples is 60000. The results are shown in Figure 2. Figure 4 In addition, in order to explore the time consumption of the model on different amounts of unlabeled data, the initial number of unlabeled samples is set to 1000 in this embodiment, and it is increased by 1000 each time, reaching a maximum of 9000. The number of unlabeled samples is still 1000, and the number of test samples is 60000. The results are shown in Figure 5 shown.

[0098] In order to further explore the performance of the present invention on data of different scales, including comparison of accuracy and time consumption, this embodiment is tested on standard MNIST and EMNIST data with different proportions.

[0099] Table 1 and Table 2 show the performance comparison on two types of data respectively.

[0100]

[0101]

[0102] Table 1

[0103]

[0104]

[0105] Table 2

[0106] like Figure 6 As shown, in another embodiment, a semi-supervised width learning classification method based on the above embodiment is also provided, comprising the following steps:

[0107] S1. Use K-Means to cluster the unlabeled data, form a descending query list based on the distance between the sample point and its cluster center, and obtain some data from it to initialize the unlabeled data pool; use the initial labeled data or the updated unlabeled data to train the BLS model, use the BLS model to annotate the data in the unlabeled data pool with pseudo labels, and calculate the corresponding prediction confidence;

[0108] Using K-Means to cluster unlabeled data specifically includes the following steps:

[0109] Step 1: Randomly select a sample in the data set as the first initialized cluster center;

[0110] Step 2: Calculate the distance between each sample point in the sample and the initialized cluster center, and select the shortest distance;

[0111] Step 3: Probably select the point with the largest distance as the new cluster center;

[0112] Step 4: Repeat steps 2 and 3 until k cluster centers are selected;

[0113] Step 5: Use the K-Means algorithm to calculate the final clustering results for the k cluster centers.

[0114] The query list is sorted according to the distance between each sample and its cluster center in the K-means clustering result. It is responsible for initializing the unlabeled data pool and iteratively feeding unlabeled data to the data pool until the query list is empty.

[0115] The unlabeled data pool changes dynamically during the iteration process. In the initialization phase, a small amount of unlabeled data is obtained from the query list; in the incremental training phase, each time the data selected as auxiliary training data is deleted from the pool, new unlabeled data is obtained from the query list as a supplement.

[0116] like Figure 7 As shown, training the BLS model specifically includes the following steps:

[0117] Initialize model parameters, including the number of mapping feature groups, the number of features within a group, and the number of enhanced nodes;

[0118] Using the model input X, calculate the feature map node Z n ≡[Z1,Z2,…,Z n ], the process is a linear mapping, where the i-th group of mapping features Z i The calculation of is as shown in formula (1), the connection weight W ei is randomly generated. For linear mapping, each set of features contains k mapping features;

[0119]

[0120] In order to increase the nonlinearity of the features in the model, the mapping node Z n Compute Enhanced Node H m ≡[H1,H2,…,H m ], the process is a nonlinear mapping, in which the enhanced feature H j The calculation of is as shown in formula (2), the connection weight W hj and β hj is randomly generated, ζ(Z,W hj ,β hj ) is a nonlinear activation function;

[0121] H j ≡ζ(Z n W hj +β hj ),j=1,2,…,m (2)

[0122] The mapping nodes represent the linear features of the data, and the enhancement nodes represent the nonlinear features of the data;

[0123] Combine linear features and nonlinear features into a feature matrix A = [Z|H] and connect it to the model output layer Y;

[0124] Since the information of the model output layer Y is known, that is, the category corresponding to the data, we only need to calculate the connection weight W between the feature layer and the output layer. The calculation formula is as follows:

[0125] W=A -1 T (3)

[0126] During calculation, ridge regression is used to solve the weight matrix, that is, W is solved by the optimization problem of formula (3):

[0127]

[0128] Among them, σ1=σ2=v=u=2, and the solution is:

[0129] WλI+AA T ) -1 A T Y(5).

[0130] Training the BLS model also includes the following steps:

[0131] When the characteristic matrix A is a non-singular matrix, there is no violation, so the violation value A of A is used. + To replace A -1 , A + Calculated by formula (6):

[0132]

[0133] Then, for the data X input into the model, the output of the model is:

[0134]

[0135] Calculate the pseudo label and the corresponding confidence level corresponding to the output, where the pseudo label is obtained by formula (8).

[0136] Label=Max(Y)(8)

[0137] The prediction confidence is calculated using formula (9), which is:

[0138]

[0139] S2, sorting the pseudo-label data by confidence, and determining the number of pseudo-label samples to be added by sorting and setting a threshold;

[0140] S3. If there is no auxiliary training data to add to the labeled training data, the number of enhanced nodes in the current BLS network structure is adjusted using the dynamic neural node adjustment mechanism, and the process jumps to step S2 for a new round of training and labeling;

[0141] If new auxiliary training data is added, this part of data is transferred from the unlabeled data pool to the labeled training data. A batch of data is obtained from the query list to update the unlabeled data pool.

[0142] Jump to step S2, use the new unlabeled data pool to train the BLS model and annotate the unlabeled data;

[0143] S4. When the unlabeled data pool is empty, all unlabeled data are labeled, and the BLS model is trained using the labeled data to output the final result.

[0144] The dynamic node adjustment strategy is based on the current training accuracy of the BLS model ACC train The number of enhanced nodes in the adjustment model; the number of enhanced nodes increased by Δm and the number of new nodes m ′ Calculated by formula (10) and formula (11)

[0145]

[0146] m ′ =m+Δm(11)

[0147] Among them, ACC thred is the expected accuracy, m is the original number of enhanced nodes, It is a round-up operation.

[0148] In this embodiment, the public dataset NORB images are used for model testing. Each image in the dataset has a pixel size of 2×32×32 and contains five categories, including animals, humans, airplanes, trucks, and cars. The images are 3D images collected under different lighting and angles, such as Figure 8 As shown in the figure. The dataset contains 48,600 data in total, the training set contains 24,300 data of 25 objects, and the test set contains 24,300 data of 25 objects. In the experiment, 10,000 labeled samples are used as the initial training set, 14,300 unlabeled samples, and 24,300 data are used for model testing.

[0149] When comparing with other methods, the data is divided into 1000 labeled training data, 14300 unlabeled data, and 24300 test data. The network structure and parameter settings of the comparison model and STSS-BLS are shown in Table 3.

[0150]

[0151]

[0152] Table 3

[0153] With the above data allocation and parameter settings, the following Table 4 shows the classification performance of various models on this dataset, including training time and test accuracy. In particular, the results are the average results of 10 experiments. Fig. 9 As shown, the changes in accuracy and AUC during one of the training processes are shown.

[0154]

[0155] Table 4

[0156] In order to explore the classification performance of the model on different amounts of labeled data, the initial number of labeled samples in this experiment is set to 1000, and then increased by 1000 each time, reaching a maximum of 10000. The number of unlabeled samples remains at 14300, and the number of test samples is 24300. The experimental results are shown in Figure 2. Fig.10 shown.

[0157] In addition, in order to explore the classification performance of the BLS model on different amounts of unlabeled data, the initial number of unlabeled samples in this experiment is set to 3000, and then increased by 1000 each time, up to 9000. The number of labeled samples remains at 2500, and the number of test samples is 24300. The experimental results are shown in Figure 2. Fig.11 shown.

[0158] The present invention guides width learning through incremental self-training and proposes a semi-supervised width learning classification method, which enables width learning to use unlabeled data and greatly expands the scope of application of width learning. At the same time, the unlabeled data in the incremental self-training method is added to the unlabeled data pool in batches to relieve the pressure of large-scale data and make it adaptable to more big data application scenarios.

[0159] It should also be noted that in this specification, terms such as "comprises", "includes" or any other variations thereof are intended to cover non-exclusive inclusion, so that a process, method, article or device including a series of elements includes not only those elements, but also includes other elements not explicitly listed, or also includes elements inherent to such process, method, article or device. In the absence of more restrictions, an element defined by the sentence "comprises a ..." does not exclude the presence of other identical elements in the process, method, article or device including the element.

[0160] The above description of the disclosed embodiments enables those skilled in the art to implement or use the present invention. Various modifications to these embodiments will be apparent to those skilled in the art, and the general principles defined herein may be implemented in other embodiments without departing from the spirit or scope of the present invention. Therefore, the present invention will not be limited to the embodiments shown herein, but rather to the widest scope consistent with the principles and novel features disclosed herein.

Claims

1. A semi-supervised width learning classification method based on an incremental self-training framework, characterized in that: The following steps are involved: S1. Use K-Means to cluster the unlabeled data, form a descending query list based on the distance between the sample point and its cluster center, and obtain some data from it to initialize the unlabeled data pool; Using the initial labeled data or the updated unlabeled data to train the BLS model, using the BLS model to label the data in the unlabeled data pool with pseudo labels, and calculating the corresponding prediction confidence; the data is image data; S2, sorting the pseudo-label data by confidence, and determining the auxiliary training data by sorting and setting a threshold; S3, determine whether there is auxiliary training data to be added to the labeled training data. If not, use the dynamic neural node adjustment mechanism to adjust the number of enhanced nodes in the BLS network structure, and jump to step S2 for a new round of training and labeling; If yes, transfer this part of auxiliary training data to the unlabeled data pool and label the training data, obtain a batch of data from the query list to update the unlabeled data pool, jump to step S2, use the new unlabeled data pool to train the BLS model and label the unlabeled data; S4, when the unlabeled data pool is empty, all the unlabeled data are labeled, and the BLS model is trained using the labeled data to output the final image data classification result; The incremental self-training framework includes an initialization phase and an incremental self-update phase; In the initialization phase, all unlabeled data samples are clustered and the sample deterministic size is used to form an ordered query sequence table; The query sequence table is used to initialize the unlabeled data pool, and the initial labeled data is used to initialize the base classifier for supervised learning, and the data in the unlabeled data pool is predicted and classified to form its pseudo label, and the corresponding selection metric is calculated; In the incremental self-update phase, auxiliary training data is determined based on the selected metric value and the set threshold and merged into the labeled training data set to form new labeled training data for updating the base classifier; at the same time, a batch of data is sequentially obtained from the query sequence table to update the unlabeled data pool and its new classification pseudo-label is predicted by the base classifier to form unlabeled data with pseudo-labels; Repeat the incremental self-update phase until the unlabeled data pool is empty; The BLS model is trained using the initial labeled data or the updated unlabeled data. The BLS model parameters include the number of mapping feature groups, the number of features within a group, and the number of enhanced nodes. The BLS model is used to annotate pseudo-labels on the data in the unlabeled data pool and calculate the confidence level corresponding to the unlabeled data.

2. The semi-supervised width learning classification method based on the incremental self-training framework according to claim 1 is characterized in that: There are eight types of data involved in the incremental self-training framework, specifically: Initial labeled data , unlabeled data , unlabeled data pool , query sequence table , used to update the batch data of the unlabeled data pool , pseudo-labeled data , auxiliary training data And new labeled data , where l, s, t, and q represent the number of samples of initial labeled data, unlabeled data pool, pseudo-label data, and added labeled data, respectively.

3. The semi-supervised width learning classification method based on the incremental self-training framework according to claim 1 is characterized in that: In the incremental self-update phase, the unlabeled data with pseudo labels are sorted in descending order according to the selection metric value to reduce the comparison and selection time of auxiliary training data.

4. The semi-supervised width learning classification method based on the incremental self-training framework according to claim 1 is characterized in that: In the incremental self-update phase, a batch of data is obtained from the query sequence table, and the size of the batch of data is fixed or changes adaptively.

5. The semi-supervised width learning classification method based on the incremental self-training framework according to claim 1 is characterized in that: The specific method of clustering unlabeled data using K-Means is as follows: S11, randomly select a sample in the data set as the first initialized cluster center; S12, calculating the distance between each sample point in the sample and the initialized cluster center, and selecting the shortest distance; S13, select the point with the largest distance as the new cluster center by probability; S14, repeat steps S12-S13 until k cluster centers are selected; S15. Use K-Means algorithm to calculate the final clustering result for k cluster centers.

6. The semi-supervised width learning classification method based on the incremental self-training framework according to claim 1 is characterized in that: The following steps are involved: The training BLS model is specifically as follows: Initialize model parameters; Using model input , compute feature map nodes , the process is a linear mapping, where Group Mapping Features The calculation of is as shown in formula (1), the connection weight and is randomly generated. For linear mapping, each set of features contains k mapping features; (1) To increase the nonlinearity of the features in the model, the mapping node Compute Enhanced Node , the process is a nonlinear mapping, in which the enhanced features The calculation of is as shown in formula (2), the connection weight and is randomly generated. is a nonlinear activation function; (2) Mapping nodes represent linear features of data, and enhancement nodes represent nonlinear features of data; Combine linear features and nonlinear features into a feature matrix , connected to the model output layer ; Since the model output layer The information is known, and the connection weight W between the feature layer and the output layer can be calculated. The calculation formula is as follows: (3) During calculation, ridge regression is used to solve the weight matrix, that is, W is solved by the optimization problem of formula (3): (4) in, , and we get: (5)。 7. The semi-supervised width learning classification method based on the incremental self-training framework according to claim 6 is characterized in that: Training the BLS model also includes the following steps: When the characteristic matrix A is a non-singular matrix, there is no violation, so the violation value of A is used to replace , Calculated by formula (6): (6) Then, for the data X input into the model, the output of the model is: （7） Calculate the pseudo label and the corresponding confidence level corresponding to the output, where the pseudo label is obtained by formula (8): （8）。 8. The semi-supervised width learning classification method based on the incremental self-training framework according to claim 1 is characterized in that: The confidence level is calculated using formula (9), which is: (9)。 9. The semi-supervised width learning classification method based on the incremental self-training framework according to claim 1 is characterized in that: The dynamic node adjustment strategy is based on the training accuracy of the current BLS model. Adjust the number of enhanced nodes in the model; Increased number of enhanced nodes And the new number of nodes Calculated by formula (10) and formula (11): (10) (11) in, is the expected accuracy, is the original number of enhanced nodes, It is a round-up operation.

Images