A machine learning method and system for multi-label imbalanced data classification

Through iterative diffusion label features, dynamic downsampling and clustering technology, combined with LDA theme model and evolutionary calculation, the problem of insufficient accuracy and robustness of multi-label classification algorithm in relationships between labels and unbalanced data processing is solved, and the problem of inadequate accuracy and robustness of multi-label classification algorithms in the relationship between labels and unbalanced data processing is achieved, and the more Efficient multi-label classification prediction.

CN114612914BActive Publication Date: 2024-11-12JIANGSU UNIV OF SCI & TECH
View PDF 0 Cites 0 Cited by

Patent Information

Application Number
CN202210309385.4
Authority / Receiving Office
CN · China
Patent Type
Patents(China)
Current Assignee / Owner
Filing Date
2022-03-25
Publication Date
2024-11-12
Estimated Expiration
2042-03-25

AI Technical Summary

Technical Problem

Modern machine learning multi-label classification algorithms lack prediction accuracy and robustness when dealing with relationships between labels and unbalanced data, and lack effective solutions.

Method used

Using the combination method of multiple iterative diffusion label features, dynamic downsampling, label clustering and traditional multi-class classifiers, the multi-label data is transformed into multi-class data for prediction through LDA thesis model and evolutionary computing technology, and Macro-F1 is used. and Micro-F1 indicator evaluation results.

Benefits of technology

It improves the prediction accuracy and robustness of multi-label classification algorithms, provides a robust, flexible and efficient solution, and performs better than traditional algorithms.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN114612914B_ABST
    Figure CN114612914B_ABST
Patent Text Reader

Abstract

The present invention provides a machine learning method and system for multi-label imbalanced data classification. The construction method and system of machine learning for multi-label imbalanced data classification based on dual high-order strategy and evolutionary computing sampling method use feature-type and label-type high-order strategies to expand multi-label imbalanced data; integrate evolutionary computing methods, give a method for calculating the population equilibrium fitness of a multi-label data set, and based on this, implement dynamic downsampling operations in high-dimensional overlapping space according to the change of the average label imbalance rate IRLbl(P); convert the multi-label problem into a traditional classification problem, so that the traditional classifier directly participates in the multi-label classification in the form of dual high-order strategy. The present invention can enable the traditional multi-class classifier to directly participate in the multi-label imbalance classification in a way that considers the label relationship, and effectively improve the value of the algorithm's multi-label evaluation index F-measure.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] The present invention relates to the technical field of artificial intelligence-machine learning algorithm design, and in particular to a machine learning method and system for multi-label imbalanced data classification. Background Art

[0002] With the development of artificial intelligence technology, the design of machine learning algorithms has gradually developed towards practicality, integration, and refinement. As a machine learning algorithm, the application of multi-label classification algorithms has become more and more extensive. However, traditional multi-label classification algorithms generally adopt low-order strategies, completely ignore the relationship between labels, ignore key learning information, and have an unbalanced distribution of labels, resulting in low prediction accuracy and poor robustness of multi-label algorithms. In order to improve the overall prediction accuracy and robustness of the algorithm, it is necessary to make the multi-label classification algorithm adopt a high-order strategy that considers the relationship between labels, and have the ability to handle unbalanced data. These designs are all closely coordinated and collaborated under the unified deployment of the system.

[0003] In summary, the design of modern machine learning multi-label classification algorithms still lacks effective solutions for considering the relationship between labels, selecting high-order strategies, and improving accuracy and robustness.

[0004] Therefore, it is necessary to provide a machine learning method and system for multi-label imbalanced data classification to solve the above technical problems. Summary of the invention

[0005] The purpose of the present invention is to provide a machine learning method and system for multi-label imbalanced data classification to solve the defects caused in the prior art.

[0006] To achieve the above object, the present invention is implemented by adopting the following technical scheme: a machine learning method for multi-label imbalanced data classification, comprising the following steps:

[0007] S1: Perform multiple iterations based on the characteristics of the multi-label dataset and finally diffuse to the characteristics of the labeled dataset;

[0008] S2: Dynamic downsampling operation is performed according to the population equilibrium fitness of the multi-label dataset;

[0009] S3: Cluster the labels according to the features of the multi-label dataset and form label clusters, and record the label numbers before clustering;

[0010] S4: Control the maximum size of each tag cluster, so that all clusters meet the condition of no more than 3 tags;

[0011] S5: Convert the multi-label distribution in all clusters into multi-category distribution and then into multiple multi-classification datasets;

[0012] S6: Call the traditional multi-class imbalanced classifier to learn and predict the transformed multi-class data set separately to obtain the predicted multi-class classification results;

[0013] S7: reconvert the predicted multi-class classification results into multi-label distribution results, and restore the original label set order after conversion;

[0014] S8: Use Macro-F1 and Micro-F1 indicators to evaluate the prediction results. Before calculating the Macro-F1 and Micro-F1 values, calculate the precision and recall according to the following formula:

[0015]

[0016] Among them, TP, True Positive: predicted to be positive, the actual is also positive FP, False Positive: predicted to be positive, the actual is negative FN, False Negative: predicted to be negative, the actual is positive TN, True Negative: predicted to be negative, the actual is also negative.

[0017] The S4 is: when the number of clusters of labels exceeds 3, cluster the clusters multiple times and split them into multiple clusters again until the size of all the obtained clusters does not exceed 3 labels.

[0018] The traditional multi-class classifier in S6 should meet the following requirements: the selected multi-class classifier can effectively distinguish at least 5 categories of classification problems when dealing with imbalanced problems; the selected multi-class classifier takes as little time as possible to process the above tasks, and the algorithm time complexity is lower than O(n3).

[0019] The S1 comprises the following steps:

[0020] S1-1: Using the LDA topic model to expand the feature set of multi-label data, the number of iterations i is pre-set, and the number of topics K is set to 2. The reason for setting the number of topics to 2 is to ensure that the resulting topic distribution is a binary distribution;

[0021] S1-2: For multi-label datasets, each instance is considered as a document, each label is considered as a word in the document, the Dirichlet distribution parameters are determined, and then introduced into the LDA topic model;

[0022] S1-3: According to the calculation rules of the LDA topic model, the instance-topic probability distribution matrix is ​​calculated. The matrix represents the probability value of each instance belonging to each topic. A binary discrete matrix is ​​generated based on the probability value to determine the topic to which each instance belongs;

[0023] S1-4: The training dataset and the test dataset have the same topic probability distribution, so first extract the training set features, and use the discrete matrix in S1-3 as the result to merge into a new multi-label dataset, and then use the traditional multi-class classifier to learn and predict the discrete matrix of the corresponding test set;

[0024] S1-5: Use the discretization matrix in step S1-3 to expand the original training set feature space, use the traditional multi-class classifier in S1-4 to learn and predict the discrete matrix of the corresponding test set to expand the test set feature space, check whether the iteration is completed, if not, go to S1-4, otherwise end.

[0025] The dynamic downsampling operation in S2 specifically includes the following steps:

[0026] S2-1: Calculate the value of the population equilibrium fitness of the multi-label data set. At this time, according to the standard processing method in the field of evolutionary computation, each sample in the data set is treated as an individual, and these individuals together constitute a population;

[0027] S2-2: Map each individual to the high-dimensional label space and the high-dimensional feature label space at the same time, and map the total N contained in the population P individuals are simultaneously placed in a connected network N, forming a network with N P A connected network of vertices N = {V, E}; at this time, the connected network forms a network with only N P A non-connected graph T = {V, E} with vertices and an empty edge set, where V represents the point set and E represents the edge set;

[0028] S2-3: According to the Kruskal principle, each time two individuals with the shortest distance in the overlapping space that are not recorded in the connection taboo table are selected, if the connected components of the system do not decrease when the two individuals are connected, the operation is canceled and recorded in the connection taboo table, and the two individuals are no longer allowed to be connected. If the connected components of the system decrease when the two individuals are connected, go to S2-4; if there are no individuals that meet the requirements to perform this operation, go to step S2-5; where the overlapping space distance is calculated as follows:

[0029]

[0030] Among them, D C is the overlapping space distance, D L (p,q),D F (p, q) are the Euclidean distances of the p-th and q-th individuals in the population after being mapped to the high-dimensional label space and the high-dimensional feature label space respectively;

[0031] S2-4: Mark the two individuals passed in S2-3 as connected; at this time, the disconnected graph T = {V, E} is updated synchronously, and the connected components of the system are also reduced with this operation; if there is an individual whose number of connected vertices exceeds 1 due to this operation among the two individuals performing the connection operation, the individual is recorded in the connection taboo table, and the individual is no longer allowed to connect with any other individual; after completion, go to step S2-3;

[0032] S2-5: Calculate the connected components at this time. Treat each interconnected individual as an individual combination, select an individual combination containing the largest number of individuals, and then select an edge with the shortest length from the individual combination. For the individuals at the two vertices of the edge, connect them with all other individuals in this individual combination in sequence and calculate the total length of the generated new edges; for individuals with a smaller total length, delete them from the population to achieve a single downsampling operation; when the total length values ​​are the same, delete the two individuals from the population at the same time;

[0033] S2-6: Calculate the value of the average label imbalance rate IRLbl(P) of each label of population P using the standard IRLbl calculation method; when the mean value IRLbl(L) of the average label imbalance rate IRLbl(P) of all labels is still higher than the preset threshold ERT, return to S2-1, otherwise end.

[0034] The algorithm for population equilibrium fitness in S2-1 is as follows:

[0035]

[0036] In the population equilibrium fitness calculation formula, EF P is the population equilibrium fitness of population P, EF L EF F are the population label equilibrium fitness and population feature equilibrium fitness, N L 、N F are the number of labels contained in the population and the dimension of the features contained in the population, respectively. L and F represent the set of labels and the set of features contained in the population, respectively. c1 and c2 are two constant coefficients. is the number of individuals with the i-th label in the population, The number of individuals in the population whose j-th eigenvalue is not 0.

[0037] In S1-2, the prior distribution of topics in each document and the distribution of words in each topic are determined according to the following formula based on the Dirichlet parameter:

[0038]

[0039] In the above formula, N is the number of instances in the dataset, K is the number of topics, θ is the prior distribution of topics in the document, which is similar to the distribution of topics in the dataset instances, and φ is the distribution of words in the topic, which is similar to the distribution of tags in the topic.

[0040] The S3 includes the following steps:

[0041] S3-1: According to the label distribution of the data set, each label is regarded as a clustering element, each instance is regarded as one dimension of the label feature, as a data set for calculating Jaccard similarity, and the label order is recorded;

[0042] S3-2: Using Jaccard similarity as the evaluation standard, calculate the Jaccard similarity between all tags, and then perform hierarchical clustering based on the calculation results, using A and B to represent the instance distribution of two tags respectively. At this time, calculate the Jaccard similarity between each tag according to the following formula:

[0043]

[0044] Where A and B represent the instance distribution of two labels respectively;

[0045] S3-3: The clustering results are formed into clusters, so as to achieve structural storage of the clustering results in the form of clusters, which is convenient for further processing of the clustering results.

[0046] The conversion process between multi-label distribution and multi-classification distribution in S5 includes: when the size of all obtained clusters does not exceed 3 labels, the multi-label distribution in all clusters does not exceed 8 types at most, and the corresponding multi-classification distribution is 8 types, that is, 2 3 In addition, the number of labels in each cluster may be different, so before training a multi-class classifier, the multi-label distribution should be uniquely matched to the multi-class distribution. The corresponding rules should be as shown in the following table:

[0047]

[0048] When the multi-label distribution in each cluster is converted to a multi-class distribution, the multi-label data feature set expanded by the LDA topic model is merged with the clusters to form a new multi-class data set, and the number of combined sets is the same as the number of clusters.

[0049] In addition, the present invention also discloses a system of a multi-label imbalanced classification machine learning framework: including a terminal device, the terminal device adopts an Internet terminal device, includes a processor and a computer-readable storage medium, the processor is used to implement each instruction; the computer-readable storage medium is used to store multiple instructions, and the instructions are suitable for being loaded by the processor and executing a machine learning method for multi-label imbalanced data classification.

[0050] According to the above technical scheme, the present invention has the following beneficial effects: the present invention provides a dual high-order strategy of feature type and label type, which provides methodological guidance for considering the relationship between labels. The LDA topic model, as a high-order strategy, extends the intrinsic connection of labels to the data feature set from the perspective of label relationship, and clustering considers related labels jointly without affecting the expansion of the former. In addition, the formation of clusters limits the distribution scale of labels to a certain extent, and the conversion of clusters enables traditional multi-class classifiers to directly participate in multi-label learning by considering the relationship between labels without complicated modifications to the algorithm. After experiments, the present invention is robust, flexible and efficient. It not only has the ability to enhance traditional algorithms, but also outperforms existing algorithms. BRIEF DESCRIPTION OF THE DRAWINGS

[0051] Figure 1 A specific flow chart of the machine learning method for multi-label imbalanced data classification provided by the present invention.

[0052] Figure 2 A schematic diagram of applying the LDA topic model provided by the present invention to a multi-label dataset.

[0053] Figure 3 This is a schematic diagram of the LDA workflow provided by the present invention.

[0054] Figure 4 A schematic diagram of a clustering tree generated by clustering label sets using the hierarchical clustering method provided by the present invention. DETAILED DESCRIPTION

[0055] In order to make the technical means, creative features, objectives and effects achieved by the present invention easy to understand, the present invention is further explained below in conjunction with specific implementation methods.

[0056] Please refer to Figure 1 , Figure 2 and Figure 3 ,in, Figure 1 A method flow chart of a method and system for constructing a machine learning method for multi-label imbalanced data classification based on a dual high-order strategy and an evolutionary computing sampling method of the present invention; Figure 2 This is a schematic diagram of the use of the first high-level strategy method provided by the present invention, the LDA topic model on a multi-label set; Figure 3This is the second high-order strategy method provided by the present invention, which is a clustering tree obtained after hierarchical clustering of the label set.

[0057] like Figure 1 As shown, a method for constructing machine learning for multi-label imbalanced data classification includes:

[0058] Step S1: According to the characteristics of the multi-label dataset, the LDA topic model is used to calculate the topic to which each instance should belong, the binary topic representation calculated by the LDA topic model is represented, and multiple iterations are performed, and then the iterated topic distribution is extended to the characteristics of the multi-label dataset;

[0059] Before using the LDA topic model to expand the feature set of multi-label data, the number of iterations i and the number of topics K should be set in advance. The reason for setting the number of topics to 2 is to ensure that the resulting topic distribution is a binary distribution.

[0060] Step S1-2: For a multi-label dataset, consider each instance as a document and each label as a word in the document, determine the Dirichlet distribution parameters, and then introduce them into Figure 2 In the LDA topic model shown;

[0061] like Figure 3 As shown, in order to meet the initial calculation conditions of the LDA topic model, the prior distribution of topics in each document and the distribution of words in each topic are determined according to the Dirichlet parameter according to the following formula:

[0062]

[0063] In the above formula, N is the number of instances in the dataset and K is the number of topics.

[0064] θ is the prior distribution of topics in the document, which is similar to the distribution of topics in the dataset instances.

[0065] φThe distribution of words in the topic, analogous to the distribution of labels in the topic in the multi-label dataset.

[0066] Step S1-3: According to the calculation rules of the LDA topic model, the instance-topic probability distribution matrix is ​​calculated, which represents the probability value of each instance belonging to each topic, and a binary discrete matrix is ​​generated according to the probability value to determine the topic to which each instance belongs;

[0067] Step S1-4: The training dataset and the test dataset have the same topic probability distribution, so first extract the training set features, and merge them into a new multi-label dataset with the discrete matrix in step 2-3 as the result. Then use the traditional multi-class classifier to learn and predict the discrete matrix of the corresponding test set.

[0068] Step S1-5: Use the discretization matrix in step S1-3 to expand the original training set feature space, use the traditional multi-class classifier in step S1-4 to learn and predict the discrete matrix of the corresponding test set to expand the test set feature space, check whether the iteration is completed, if not, go to step S1-4, otherwise end.

[0069] Step S2: In order to solve the class imbalance problem in the multi-label data set by dynamic downsampling operation, the step S2 also includes:

[0070] Step S2-1: Calculate the value of the population equilibrium fitness of the multi-label dataset. At this time, according to the standard processing method in the field of evolutionary computation, each sample in the dataset is treated as an individual, and these individuals together constitute a population. In order to calculate the value of the population equilibrium fitness of the multi-label dataset, the population equilibrium fitness of the multi-label dataset should be calculated according to the following formula:

[0071]

[0072] In the population equilibrium fitness calculation formula, EF P is the population equilibrium fitness of population P, EF L EF F are the population label equilibrium fitness and population feature equilibrium fitness, N L 、N F They are the number of labels contained in the population and the dimensions of the features contained in the population, respectively. L and F represent the set of labels and the set of features contained in the population, respectively. c1 and c2 are two constant coefficients. is the number of individuals with the i-th label in the population, The number of individuals in the population whose jth eigenvalue is not 0;

[0073] Step S2-2: Map each individual to the high-dimensional label space and the high-dimensional feature label space at the same time, and map the total N contained in the population P individuals are simultaneously placed in a connected network N, forming a network with N P The connected network N = {V, E} of vertices. At this time, the connected network forms a network with only N P A non-connected graph T = {V, E} with vertices and an empty edge set, where V represents the point set and E represents the edge set;

[0074] Step S2-3: According to the Kruskal principle, select two individuals each time that are not recorded in the connection taboo table and have the shortest distance in the overlapping space. At this time, if the connected component of the system does not decrease when the two individuals are connected, cancel this operation and record the operation in the connection taboo table, and the two individuals are no longer allowed to be connected. If the connected component of the system decreases when the two individuals are connected, go to step 3-4. If there are no individuals that meet the requirements to perform this operation, go to step 3-5. The overlapping space distance is calculated as follows:

[0075]

[0076] Among them, D C is the overlapping space distance, D L (p,q),D F (p, q) are the Euclidean distances of the p-th and q-th individuals in the population after being mapped to the high-dimensional label space and the high-dimensional feature label space respectively;

[0077] Step S2-4: Mark the two individuals passed in by step S2-3 as connected. At this time, the disconnected graph T = {V, E} is updated synchronously, and the connected components of the system are also reduced with this operation. If there is an individual whose number of connected vertices exceeds 1 due to this operation among the two individuals performing the connection operation, the individual will be recorded in the connection taboo table, and the individual will no longer be allowed to connect with any other individual. After completion, go to step S2-3;

[0078] Step S2-5: Calculate the connected components at this time. Treat each interconnected individual as an individual combination, select an individual combination containing the largest number of individuals, and then select an edge with the shortest length from the individual combination. For the individuals at the two vertices of the edge, connect them with all other individuals in this individual combination in turn and calculate the total length of the generated new edges. For individuals with a smaller total length, delete them from the population to achieve a single downsampling operation. When the total length values ​​are the same, delete the two individuals from the population at the same time;

[0079] Step S2-6: Using the standard IRLbl calculation method, calculate the value of the average label imbalance rate IRLbl(P) of each label of the population P. When the mean value IRLbl(L) of the average label imbalance rate IRLbl(P) of all labels is still higher than the preset threshold ERT, return to step S2-1, otherwise end.

[0080] Step S3: According to the label distribution of the multi-label dataset, in order to measure the differences between labels, the binary distribution of each label belonging to each instance is characterized, and the Jaccard similarity is used as the measurement standard. The Jaccard similarity between each label is calculated according to the following formula:

[0081]

[0082] Where A and B represent the instance distribution of two labels respectively;

[0083] like Figure 4 As shown, the label clustering result is obtained to form a label cluster, and the label sequence number is recorded before clustering; in order to use the hierarchical clustering algorithm to mine and learn the potential relationship in the label, the step S2 also includes:

[0084] Step S3-1: According to the label distribution of the data set, each label is regarded as a clustering element, each instance is regarded as one dimension of the label feature, as a data set for calculating Jaccard similarity, and the label order is recorded;

[0085] Step S3-2: Using Jaccard similarity as the evaluation standard, calculate the Jaccard similarity between all tags, and then perform hierarchical clustering based on the calculation results;

[0086] Step S3-3: Cluster the clustering results and store them.

[0087] Step S4: Control the maximum size of each label cluster. For clusters with more than 3 labels, perform multiple clustering and re-split them into multiple clusters until the size of all clusters does not exceed 3 labels. During the processing, the splitting method is to re-cluster all these clusters. After the re-clustering is completed, the new clusters should replace the original clusters. After the restoration is completed, re-check the size of each cluster to see if there are still clusters with more than 3 labels. If so, re-cluster them again until all clusters meet the condition of not more than 3 labels.

[0088] Step S5: Convert the multi-label distribution in all clusters into multi-category distribution. The number of categories converted in each cluster shall not exceed 8, i.e. 2 3 . Then, the feature set expanded by the LDA topic model is combined with the converted multi-classifications to convert them into multiple multi-classification datasets;

[0089] The conversion process between the multi-label distribution and the multi-classification distribution in step S5 further includes: when the size of all the obtained clusters does not exceed 3 labels, the multi-label distribution in all clusters does not exceed 8 types at most, and the corresponding multi-classification distribution is 8 types, that is, 2 3 In addition, the number of labels in each cluster may be different, so before training a multi-class classifier, the multi-label distribution should be uniquely matched to the multi-class distribution. The corresponding rules should be as shown in the following table:

[0090]

[0091] Step S6: Call the traditional multi-class imbalanced classifier to learn and predict the converted multi-class data set separately to obtain the predicted multi-class classification results; in order to meet the requirements of the converted multi-class data set, before calling the traditional multi-class imbalanced classifier, its type requirements also include: the number of labels in the cluster is limited, the label distribution scale is reduced, but the reduced label distribution still presents an imbalanced situation, and the time complexity is slightly increased. According to the characteristics of the converted data set, the selected traditional multi-class classifier should meet the requirements of: having the ability to handle imbalanced problems and taking as little time as possible;

[0092] Step S7: Reconvert the predicted multi-class classification results into multi-label distribution results and restore them to the original label order; after the traditional multi-class classifier predicts the multi-class results, each multi-class classification result is converted back to the multi-label distribution, and the conversion process is the inverse process of S4. After the conversion is completed, merge the conversion results of each cluster and restore them to the original label set order.

[0093] Step S8: In order to calculate the Macro-F1 and Micro-F1 values, the values ​​to be calculated also include: calculating the precision Precision and the recall Recall according to the following formula:

[0094]

[0095] Among them, TP, True Positive: predicted positive, actually positive FP, False Positive: predicted positive, actually negative FN, False Negative: predicted negative, actually positive TN, True Negative: predicted negative, actually negative. Finally, the Macro-F1 and Micro-F1 indicators are used to evaluate the prediction results.

[0096] It is known from common technical knowledge that the present invention can be implemented by other embodiments that do not deviate from its spirit or essential features. Therefore, the above disclosed embodiments are only illustrative in all respects and are not exclusive. All changes within the scope of the present invention or within the scope equivalent to the present invention are included in the present invention.

Claims

1. A machine learning method for multi-label imbalanced data classification, characterized by: The following steps are included: S1: Perform multiple iterations based on the characteristics of the multi-label dataset and finally diffuse to the characteristics of the labeled dataset; Specifically: S1-1: Use the LDA topic model to expand the feature set of multi-label data, pre-set the number of iterations i, set the number of topics K = 2, and set the number of topics to 2 to ensure that the resulting topic distribution is a binary distribution; S1-2: For multi-label datasets, each instance is considered as a document, each label is considered as a word in the document, the Dirichlet distribution parameters are determined, and then introduced into the LDA topic model; S1-3: According to the calculation rules of the LDA topic model, the instance-topic probability distribution matrix is ​​calculated. The matrix represents the probability value of each instance belonging to each topic. A binary discrete matrix is ​​generated based on the probability value to determine the topic to which each instance belongs; S1-4: The training dataset and the test dataset have the same topic probability distribution, so first extract the training set features, and use the discrete matrix in S1-3 as the result to merge into a new multi-label dataset, and then use the traditional multi-class classifier to learn and predict the discrete matrix of the corresponding test set; S1-5: Use the discrete matrix in S1-3 to expand the original training set feature space, use the traditional multi-class classifier in S1-4 to learn and predict the discrete matrix of the corresponding test set to expand the test set feature space, check whether the iteration is completed, if not, go to S1-4, otherwise end; S2: Dynamic downsampling operation is performed according to the population equilibrium fitness of the multi-label dataset; Specifically: S2-1: Calculate the value of the population equilibrium fitness of the multi-label data set, treat each sample in the data set as an individual, and these individuals together constitute a population; S2-2: Map each individual to a high-dimensional label space and a high-dimensional feature label space at the same time, and map the N contained in the population P individuals are simultaneously placed in a connected network N, forming a network with N P The connected network N = {V, E} of vertices, at this time, the connected network forms a network with only N P A non-connected graph T = {V, E} with vertices and an empty edge set, where V represents the point set and E represents the edge set; S2-3: According to the Kruskal principle, each time two individuals with the shortest distance in the overlapping space that are not recorded in the connection taboo table are selected. At this time, if the connected component of the system does not decrease when the two individuals are connected, the connection operation is canceled and the operation is recorded in the connection taboo table. The two individuals are no longer allowed to be connected. If the connected component of the system decreases when the two individuals are connected, go to S2-4; if there are no individuals that meet the requirements and can perform the connection operation, go to S2-5; where the overlapping space distance is calculated as follows: Among them, D C is the overlapping space distance, D L (p,q),D F (p, q) are the Euclidean distances of the p-th and q-th individuals in the population after being mapped to the high-dimensional label space and the high-dimensional feature label space respectively; S2-4: Mark the two individuals passed in S2-3 as connected. At this time, the disconnected graph T = {V, E} is updated synchronously, and the connected components of the system are also reduced. If there is an individual whose number of connected vertices exceeds 1 due to the operation between the two individuals performing the connection operation, the individual is recorded in the connection taboo table, and the individual is no longer allowed to connect with any other individual. After completion, go to S2-3; S2-5: Calculate the connected components at this time, treat each interconnected individual as an individual combination, select an individual combination containing the largest number of individuals, and then select a shortest edge from the individual combination. For the individuals at the two vertices of the edge, connect them with all other individuals in the individual combination in sequence and calculate the total length of the generated new edges; for individuals with a smaller total length, delete them from the population to achieve a single downsampling operation. When the total length values ​​are the same, delete the two individuals from the population at the same time; S2-6: Calculate the value of the average label imbalance rate IRLbl(P) of each label of population P using the standard IRLbl calculation method. When the mean value IRLbl(L) of the average label imbalance rate IRLbl(P) of all labels is still higher than the preset threshold ERT, return to S2-1, otherwise end; S3: Cluster the labels according to the features of the multi-label dataset and form label clusters, and record the label numbers before clustering; S4: Control the maximum size of each tag cluster, so that all clusters meet the condition of no more than 3 tags; S5: Convert the multi-label distribution in all clusters into multi-category distribution and then into multiple multi-classification datasets; S6: Call the traditional multi-class imbalanced classifier to learn and predict the transformed multi-class data set respectively to obtain the predicted multi-class classification results; the traditional multi-class classifier meets the following requirements: the selected multi-class classifier can effectively distinguish at least 5 categories of classification problems when dealing with imbalanced problems; the selected multi-class classifier takes as little time as possible to process the task, and the algorithm time complexity is less than O(n 3 ); S7: reconvert the predicted multi-class classification results into multi-label distribution results, and restore the original label set order after conversion; S8: Use Macro-F1 and Micro-F1 indicators to evaluate the prediction results. Before calculating the Macro-F1 and Micro-F1 values, calculate the precision and recall according to the following formula: Among them, TP, True Positive: predicted to be positive, the actual is also positive; FP, False Positive: predicted to be positive, the actual is negative; FN, False Negative: predicted to be negative, the actual is positive; TN, True Negative: predicted to be negative, the actual is also negative.

2. The machine learning method for multi-label imbalanced data classification according to claim 1, characterized in that: The S4 is: when the number of clusters of labels exceeds 3, cluster the clusters multiple times and split them into multiple clusters again until the size of all the obtained clusters does not exceed 3 labels.

3. The machine learning method for multi-label imbalanced data classification according to claim 1, characterized in that: The algorithm for population equilibrium fitness in S2-1 is as follows: In the population equilibrium fitness calculation formula, EF P is the population equilibrium fitness of population P, EF L EF F are the population label equilibrium fitness and population feature equilibrium fitness, N L 、N F are the number of labels contained in the population and the dimension of the features contained in the population, respectively. L and F represent the set of labels and the set of features contained in the population, respectively. c1 and c2 are two constant coefficients. is the number of individuals with the i-th label in the population, The number of individuals in the population whose j-th eigenvalue is not 0.

4. The machine learning method for multi-label imbalanced data classification according to claim 1, characterized in that: In S1-2, the prior distribution of topics in each document and the distribution of words in each topic are determined based on the Dirichlet parameter, which is determined according to the following formula: In the above formula, N is the number of instances in the dataset, K is the number of topics, θ is the prior distribution of topics in the document, that is, the distribution of topics in the dataset instances, and φ is the distribution of words in the topic, that is, the distribution of labels in the topic.

5. The machine learning method for multi-label imbalanced data classification according to claim 3, characterized in that: The S3 includes the following steps: S3-1: According to the label distribution of the data set, each label is regarded as a clustering element, each instance is regarded as one dimension of the label feature, as a data set for calculating Jaccard similarity, and the label order is recorded; S3-2: Using Jaccard similarity as the evaluation standard, calculate the Jaccard similarity between all tags, and then perform hierarchical clustering based on the calculation results. Use A and B to represent the instance distribution of two tags respectively. Calculate the Jaccard similarity between each tag according to the following formula: Where A and B represent the instance distribution of two labels respectively; S3-3: The clustering results are formed into clusters, thereby achieving: structural storage of the clustering results in the form of clusters, which facilitates further processing of the clustering results.

6. The machine learning method for multi-label imbalanced data classification according to claim 1, characterized in that: The conversion process between multi-label distribution and multi-class distribution in S5 includes: when the size of all obtained clusters does not exceed 3 labels, the multi-label distribution in all clusters does not exceed 8 types at most, and the corresponding multi-class distribution is 8 types, that is, 2 3 ; In addition, the number of labels in each cluster may be different, so before training the multi-class imbalanced classifier, let the multi-label distribution uniquely correspond to the multi-class distribution, and the corresponding rules should be as shown in the following table: When the multi-label distribution in each cluster is converted to a multi-category distribution, the multi-label data feature set expanded by the LDA topic model is merged with the clusters to form a new multi-classification data set, and the number of combined sets is the same as the number of clusters.

7. A system for a multi-label imbalanced classification machine learning framework: characterized in that, It includes a terminal device, which is an Internet terminal device, including a processor and a computer-readable storage medium, the processor is used to implement each instruction; the computer-readable storage medium is used to store multiple instructions, and the instructions are loaded by the processor and executed according to a machine learning method for multi-label imbalanced data classification as described in any one of claims 1 to 6.