Multi-label classification method for text data based on label-specific features and correlation

Abstract

The present invention claims a method for multi-label classification of text data based on label-specific features and label correlation, comprising: 101 pre-processing text data, combining instance label space and feature space to construct a new instance feature space; 102 calculating the number of clusters for positive and negative instance feature spaces, and constructing label-specific features; 103 calculating correlations for paired labels in the text data set, and constructing an undirected complete graph of correlations between labels; 104 using an extended Kruskal algorithm to calculate the maximum spanning tree set in the graph; 105 selecting spanning trees from the tree set in turn, and using a binary classifier to recursively predict on the tree starting from the root node, and integrating the prediction results on the tree to obtain a text instance label set. The present invention can effectively suppress the error transmission problem in the prediction process, and adopts tree set integrated prediction, which can effectively utilize different correlations between label sets, and further improve the performance of multi-label classification.

Description

Technical Field

[0001] The present invention belongs to the field of machine learning and data mining, and in particular relates to a multi-label classification method for text data based on label specific features and label correlation. Background Art

[0002] Real-world instances often contain multiple labels. For example, a picture may contain both people and animals, and a news report may contain both political and economic categories. However, manually classifying instances requires a lot of human resources. The task of multi-label learning of text data is to learn from known instances and label sets to predict the label set of text data instances.

[0003] In the existing multi-label classification technology, the multi-label learning method with label-specific features (LIFT) and the multi-label learning method with label-specific features via clustering ensemble (LIFTACE) are usually used. In the LIFT and LIFTACE methods, the instance feature space is clustered and label-specific features are constructed by Euclidean distance measurement; then the instances are classified multiple times by combining the specific features with a binary classifier. However, both of the above methods fail to consider label relevance in the clustering and classification stages. Using label relevance can effectively improve the performance of multi-label classification. For example, if a picture contains the label ocean, the possibility that the picture contains a beach will be greatly increased. Based on traditional label specific features, the present invention uses label space and feature space simultaneously in the stage of constructing label specific features, constructs a label correlation undirected complete graph by quantifying the correlation between labels in the classification stage, and combines the extended Kruskal algorithm to mine the maximum spanning tree set in the graph, reconstructs the spanning trees in the set, and expresses the strong correlation of various different structures between labels while suppressing error propagation through tree integration, which is of great significance to improving the accuracy of multi-label classification of text data.

[0004] After searching, in the prior art, CN107133293A, an improved ML-kNN method suitable for multi-label classification, is characterized in that it includes the following steps: Step 1: obtain an original data set, the original data set includes multiple samples, each of which has multiple labels and multiple features, and the total number of samples of each type of label in the original data set is counted as the number of label samples, and the total number of samples of each type of feature in the samples of each type of label is counted as the number of feature samples, and the feature label weight is calculated according to the number of label samples and the number of feature samples, wherein each feature corresponds to a feature value; Step 2: split each sample in the original data set into multiple original single-label samples with a single label, and update the feature value of each original single-label sample according to the feature label weight to generate a first data set; Step 3: obtain a test sample to be predicted, split the test sample into single-label samples to be tested with a single label, predict the labels of the single-label samples to be tested in turn according to the first data set, and determine the label set of the sample to be tested.

[0005] This method splits the multi-label data set into multiple single-label data sets based on ML-kNN for classification prediction, and uses the feature space weighted by the eigenvalues ​​to generate the label set according to kNN. However, the predictions between multiple labels are not related in the prediction process, that is, the label correlation is not considered; in addition, the kNN algorithm used in this method is not stable in the prediction process, and different k values ​​will have a greater impact on the prediction results. This method also considers the correlation between labels in the process of clustering and classification, and combines the tree ensemble to generate the label set of the instance, making the prediction results more stable and accurate. Summary of the invention

[0006] The present invention aims to solve the above problems of the prior art. A multi-label classification method for text data based on label-specific features and relevance with improved accuracy is proposed. The technical solution of the present invention is as follows:

[0007] A multi-label classification method for text data based on label-specific features and correlations, comprising the following steps:

[0008] 101. Preprocess the text data and construct a new instance feature space by combining the instance label space and the feature space;

[0009] 102. Calculate the number of clusters in the feature space of positive and negative instances and construct label-specific features;

[0010] 103. Calculate the correlation between paired labels in the text data set and construct an undirected complete graph of correlation between labels;

[0011] 104. Use the extended Kruskal algorithm to calculate the maximum spanning tree set in the undirected complete graph of label correlation;

[0012] 105. Select the spanning tree from the tree set in turn, and use the binary classifier to perform recursive prediction on the tree starting from the root node, and integrate the prediction results on the tree to obtain the text instance label set.

[0013] Furthermore, the text data preprocessing in step 101 is as follows: for the positive and negative instance data sets, the feature space and the label space are merged so that a plurality of remaining labels supplement the feature space to construct the positive and negative instance feature space. and

[0014]

[0015] Where D = {(x i ,Y i )|1≤i≤n} represents a multi-label dataset, x i represents the original feature space of the i-th instance, Y i Represents the multi-label set associated with the i-th instance.

[0016] Furthermore, the step 102 calculates the number of clusters in the positive and negative instance feature spaces and constructs label-specific features. The specific steps are:

[0017] 1021. To convert positive and negative instance features into label-specific features, the size of the label-specific feature space is first calculated based on the feature space. To prevent the label-specific feature dimension from exploding due to the large feature space, the smaller value of the positive and negative space sizes is taken and scaled using logarithms. Finally, the number of clusters m is obtained. k ;

[0018]

[0019] 1022. K-MEANS algorithm is used to cluster the positive and negative instance spaces, and the original feature space vector is combined with the cluster center point and the number of cluster points for measurement, from the original d-dimensional space Mapped to a d k Dimensional label-specific feature space

[0020]

[0021] in represents the label-specific feature corresponding to the i-th instance, is the Euclidean distance, Represent the positive and negative cluster centers respectively, C is the i-th cluster center sample set, |C i | is the number of center points of the i-th cluster. The Euclidean distance is scaled by the number of clusters to alleviate the label imbalance problem in multi-label.

[0022] Furthermore, the specific steps of calculating the correlation between paired labels in the text data set in step 103 and constructing an undirected complete graph of correlation between labels are as follows:

[0023] 1031.Define a correlation undirected complete graph: there is a correlation edge between any two labeled vertices in the graph and there are no self-loops or multiple edges;

[0024] 1032.Use corr(l i ,l j ) calculates the label correlation set E, and uses the multi-label set V in the dataset as j ,…,l q} is the point set in the undirected complete graph G, q is the size of the multi-label set, and the pairwise label correlation set E = {corr(l i ,l j ) i ∈V,l j ∈V} is the edge set in the undirected complete graph G;

[0025]

[0026] Where |l i | is the number of times label i appears, l ik represents the k-th dimension value in the i-th label vector, l i Represents the mean of the i-th label vector, takes the value with the smaller number of occurrences of paired labels, and combines the covariance and standard deviation of the labels in the label space to obtain the correlation between multiple labels.

[0027] Furthermore, the step 104 uses the extended Kruskal algorithm to calculate the maximum spanning tree in the graph, selects each vertex in the tree as a root to generate multiple maximum spanning trees with the same weight but different tree heights, and sorts the trees from low to high according to their heights. The specific steps are:

[0028] 1041. Define the maximum spanning tree: a spanning tree containing a connected graph of all label nodes, containing all labels in the undirected complete graph of pairwise label correlations and having an edge set that maintains the maximum sum of weights of the connected graph of correlations;

[0029] 1042. Define the extended Kruskal algorithm: adopt the greedy idea, sort the edge set in descending order according to the weight, and then recursively search the edges with the same weight to generate different tree structures until the maximum spanning tree set with the same total weight is generated;

[0030] 1043. For the maximum spanning tree set, enumerate any child node in the tree as the root node to reconstruct the tree. A maximum spanning tree will produce |V| different tree structures, where V represents the set of nodes on the tree, and sort each reconstructed tree from low to high according to the tree height.

[0031] Furthermore, the extended Kruskal algorithm:

[0032] 1). Sort the edge sets with correlation in descending order of weight.

[0033] 2). Recursively search the weighted edges: If this edge does not form a loop when added to the current tree, add the edge to the tree until all label nodes are in the tree.

[0034] 3). For the maximum spanning tree obtained in 2, enumerate each point as the root node to build multiple maximum spanning trees, and sort them from bottom to top according to their height.

[0035] Furthermore, the step 105 selects spanning trees from the tree set in sequence, and uses a binary classifier to perform recursive prediction on the tree starting from the root node, and integrates the prediction results on the tree to obtain a text instance label set. The specific steps are:

[0036] 1051. For the multiple maximum spanning trees with different structures in step 105, select the first m k The tree is classified, and for each tree, recursive prediction is performed starting from the root node. During the prediction process, the prediction results of the ancestor node labels are added to the child node label feature space to construct a new feature space. At this time, the label feature space can be expanded to at this time Indicates the label l k New tag specific features.

[0037] Among them, ancestor(l k ) indicates label l k Ancestor node labels in the maximum relatedness spanning tree;

[0038]

[0039] 1053. Use a binary classifier to classify the label node dataset Perform classification, and then recursively classify the labels of the child nodes until all nodes in the spanning tree are classified;

[0040]

[0041] η(Y i ,l k ) indicates label l k Is it in the instance-related tag set Y i middle.

[0042] 1054. Since different root nodes will result in different tree structures, the strong correlation between the expressed labels is also different. In order to make full use of the strong correlation between labels and reduce the impact of selecting different root nodes, multiple trees are used for multi-label classification. For an unknown instance The number of tree sets is t, f kj For label l k The classifier on the jth tree; integrate the same node on different trees, and finally get the label set Yu;

[0043]

[0044] The advantages and beneficial effects of the present invention are as follows:

[0045] 1. Clustering is performed by combining instance feature space and label space to construct label-specific features. Traditional clustering methods often only cluster the feature space of instances. This is because traditional classification only targets single-label data sets, and most of the information that can be used comes only from the feature space in the data set. In multi-label classification, as the label set expands, the reasonable use of label set information can improve the clustering effect. Therefore, this patent innovatively combines the feature space of the instance with the label space to construct a new feature space, and then clusters the instances according to their positivity and negativity, so that the clustering structure can effectively utilize the similarity of the label space. When constructing label-specific features, in addition to the information of the original feature space of the label, it also contains the effective information of the remaining label space, thereby improving the performance of multi-label classification.

[0046] 2. The extended Kruskal algorithm is used in combination with the undirected complete graph of the correlation between labels to mine the correlation between labels. The traditional methods of using label correlation can be divided into three types: the first-order strategy completely ignores the correlation between labels, the second-order strategy considers the correlation between paired labels, and the high-order strategy considers the high-order relationship between labels; the second-order strategy only considers the correlation between paired labels and ignores the correlation with the label set, and the high-order strategy considers the correlation between the label set, but the correlation between the label sets considered is unique. Therefore, this patent innovatively uses the extended Kruskal algorithm combined with the undirected complete graph of the correlation between labels to mine the correlation between labels, and uses multiple maximum spanning tree structures to reflect the correlation between labels, so that while fully reflecting the correlation between labels, different tree structures are also used to express different correlations.

[0047] 3. Use tree sets for integrated prediction; arrange the tree sets in ascending order according to the tree height, use multiple maximum spanning trees to recursively predict from the root node, add the labels of the ancestor nodes to the feature space, use a binary classifier for classification, and perform integrated classification. In traditional multi-label classification, high-order strategies consider the relationship between label sets, such as the classifier chain algorithm. However, since there is often error propagation in chain predictions, if there is an error in the previous label prediction, it will affect the accuracy of the subsequent label prediction. Therefore, this patent innovatively performs recursive prediction on the tree. Since the prediction is made from the shortest tree according to the tree height, the size of the ancestor label set is limited during the prediction process, which can effectively suppress the error propagation problem in the prediction process. In addition, the use of tree set integrated prediction can effectively utilize the different correlations between label sets to further improve the performance of multi-label classification. BRIEF DESCRIPTION OF THE DRAWINGS

[0048] Figure 1 The present invention provides a flowchart of a method for multi-label classification of text data based on label-specific features and label relevance in a preferred embodiment;

[0049] Figure 2 The present invention is a schematic diagram of an integrated tree prediction combining label relevance in a multi-label classification method for text data based on label specific features and label relevance in an embodiment of the present invention. DETAILED DESCRIPTION

[0050] The following will describe the technical solutions in the embodiments of the present invention in detail in conjunction with the accompanying drawings in the embodiments of the present invention. The described embodiments are only part of the embodiments of the present invention.

[0051] The technical solution of the present invention to solve the above technical problems is:

[0052] A multi-label classification method for text data based on label specific features and label correlation, comprising the following steps:

[0053] 101. Preprocess the text dataset and construct a new instance feature space by combining the instance label space and feature space (Formula (1)).

[0054] 102. Clustering technology is used to construct multiple cluster centers for the positive and negative instance feature spaces, and the distance between the original feature space and each point in the cluster center set is measured and combined with the number of clusters to construct label-specific features. Specifically, it includes: 1. Formula (2) is used to calculate the number of clusters, and ln logarithmic processing is used to prevent the problem of too many cluster centers caused by the large dimension of the feature space, so as to construct effective label-specific features; 2. The K-MEANS algorithm is used to cluster the positive and negative instance spaces, and the original feature space vector is combined with the cluster center point to measure the number of clusters, from the original d-dimensional space Mapped to a dk Dimensional label-specific feature space (Formula (3)).

[0055] 103. Calculate the correlation between any two labels in the text data set and construct an undirected complete graph of the correlation between labels; specifically, use the correlation corr(l i ,l j ) (Formula (4)) Calculate the correlation to obtain the label correlation set E = {corr(l i ,l j )}; Taking the multi-label set V in the dataset as {l1,…,l j ,…,l q} is the point set in the undirected complete graph G, and the pairwise label correlation set E = {corr(l i ,l j ) i ∈V,l j ∈V} is the edge set in the undirected complete graph G.

[0056] 104. The extended Kruskal algorithm is used to calculate the maximum spanning tree set in the graph; the specific steps are: 1. The extended Kruskal algorithm is used to calculate the maximum spanning tree set on the undirected complete graph of label correlation, where the tree weight is the correlation between labels, and the maximum spanning tree structure expresses the strong correlation between labels; 2. For multiple maximum spanning trees, any child node in the enumeration tree is used as the root node to reconstruct the tree. A maximum spanning tree will produce at most |V| different tree structures, and each tree is sorted from low to high according to the height of the tree.

[0057] 105. Select spanning trees from the tree set in turn, and use the binary classifier to perform recursive prediction on the tree starting from the root node, integrate the prediction results on the tree and finally obtain the instance label set; the specific steps are: 1. For multiple spanning trees with different structures and the maximum correlation, select the first m k The classification is performed on each tree. For each tree, recursive prediction is performed starting from the root node. During the prediction process, the prediction result of the ancestor node label is added to the child node label feature space to construct a new feature space. At this time, the label feature space can be expanded (Formula (5)) as follows: Among them, ancestor(l k ) indicates label l k 1. Label the ancestor node in the maximum correlation spanning tree; 2. Use a binary classifier to classify the label node dataset 3. Since different root nodes will result in different tree structures, the strong correlation between the expressed labels will also be different. In order to make full use of the strong correlation between labels and reduce the impact of selecting different root nodes, multiple trees are used for multi-label classification. The number of tree ensembles is t; the same node on different trees is integrated, and finally the label set Y is obtained u (Formula (6)).

[0058] Furthermore, the step 101 preprocesses the original data set to obtain the positive and negative instance feature space, and for the labels in the label set of the original data set, the following steps are repeated to construct the label data set: for the positive and negative instance data sets, the feature space and the label space are merged so that the plurality of remaining labels supplement the feature space to construct the positive and negative instance feature space and

[0059]

[0060] Furthermore, the step 102 uses clustering technology to construct multiple cluster centers for the positive and negative instance feature spaces, measures the distance between the original feature space and each point in the cluster center set, and constructs label-specific features in combination with the number of clusters. The specific steps are:

[0061] (1) To convert positive and negative instance features into label-specific features, we first calculate the size of the label-specific feature space based on the feature space. To prevent the label-specific feature dimension from exploding due to the large feature space, we take the smaller value of the positive and negative space sizes and scale them logarithmically. Finally, we get the number of clusters m. k .

[0062]

[0063] (2) The K-MEANS algorithm is used to cluster the positive and negative instance spaces. The original feature space vector is combined with the cluster center point and the number of cluster center points is measured. Mapped to a d k Dimensional label-specific feature space

[0064]

[0065] Furthermore, the step 103 calculates the correlation between any two labels in the text data set and constructs an undirected complete graph of the correlation between labels; specifically, the label space adopts the correlation corr(l i ,l j) calculates the pairwise label correlation set E, and uses the multi-label set V in the dataset as j ,…,l q} is the point set in the undirected complete graph G, and the pairwise label correlation set E = {corr(l i ,l j ) i ∈V,l j ∈V} is the edge set in the undirected complete graph G.

[0066]

[0067] Furthermore, the step 104 uses the extended Kruskal algorithm to calculate the maximum spanning tree in the graph, selects each vertex in the tree as a root to generate multiple maximum spanning trees with the same weight but different tree heights, and sorts the trees from low to high according to their heights; the specific steps are:

[0068] (1) The extended Kruskal algorithm is used to calculate the maximum spanning tree set on the undirected complete graph of label correlations, where the tree weight is the correlation between labels and the maximum spanning tree structure expresses the strong correlation between labels.

[0069] (2) For multiple maximum spanning trees, any child node in the enumeration tree is used as the root node to reconstruct the tree. A maximum spanning tree will produce at most |V| different tree structures, and each tree is sorted from low to high according to its height.

[0070] Furthermore, the step 105 selects spanning trees from the tree set in turn, and uses a binary classifier to perform recursive prediction on the tree starting from the root node, and integrates the prediction results on the tree to finally obtain the instance label set: the specific steps are:

[0071] (1) For multiple maximum spanning trees with different structures, select the first m k The classification is performed on each tree. For each tree, recursive prediction is performed starting from the root node. In the prediction process, as shown in formula (5), the prediction result of the ancestor node label is added to the child node label feature space to construct a new feature space, where ancestor(l k ) indicates label l k Ancestor node labels in the maximum relatedness spanning tree.

[0072]

[0073] (2) For the label classification data set, it can be used Formula (6) shows that Indicates the label l k Whether to use a binary classifier for the label node dataset in the instance's label set Classify and then recursively predict the child nodes until all nodes in the spanning tree are classified.

[0074]

[0075] 1054. Since different root nodes will result in different tree structures, the strong correlation between the expressed labels is also different. In order to make full use of the strong correlation between labels and reduce the impact of selecting different root nodes, multiple trees are used for multi-label classification. The number of tree ensembles is t; the same node on different trees is integrated, and finally the label set Y is obtained u Formula (7).

[0076]

[0077] It should also be noted that the terms "include", "comprises" or any other variations thereof are intended to cover non-exclusive inclusion, so that a process, method, commodity or device including a series of elements includes not only those elements, but also other elements not explicitly listed, or also includes elements inherent to such process, method, commodity or device. In the absence of more restrictions, the elements defined by the sentence "comprises a ..." do not exclude the existence of other identical elements in the process, method, commodity or device including the elements.

[0078] The above embodiments should be understood to be only used to illustrate the present invention and not to limit the protection scope of the present invention. After reading the contents of the present invention, technicians can make various changes or modifications to the present invention, and these equivalent changes and modifications also fall within the scope defined by the claims of the present invention.

Claims

1. A multi-label classification method for text data based on label-specific features and correlation, characterized in that: The following steps are involved:

101. Preprocess the text data and construct a new instance feature space by combining the instance label space and the feature space; 102. Calculate the number of clusters in the feature space of positive and negative instances and construct label-specific features; 103. Calculate the correlation between paired labels in the text data set and construct an undirected complete graph of correlation between labels; 104. Use the extended Kruskal algorithm to calculate the maximum spanning tree set in the undirected complete graph of label correlation; 105. Select spanning trees from the tree set in turn, and use a binary classifier to perform recursive prediction on the tree starting from the root node, and integrate the prediction results on the tree to obtain a text instance label set; The text data preprocessing step 101 is: merging the feature space and the label space of the positive and negative instance data sets so that the plurality of remaining labels supplement the feature space to construct the positive and negative instance feature space and Where D = {(x i ,Y i )|1≤i≤n} represents a multi-label dataset, x i represents the original feature space of the i-th instance, Y i Represents the multi-label set associated with the i-th instance; The step 102 calculates the number of clusters in the positive and negative instance feature spaces and constructs label-specific features. The specific steps are: 1021. To convert positive and negative instance features into label-specific features, the size of the label-specific feature space is first calculated based on the feature space. To prevent the label-specific feature dimension from exploding due to the large feature space, the smaller value of the positive and negative space sizes is taken and scaled using logarithms. Finally, the number of clusters m is obtained. k ; 1022. K-MEANS algorithm is used to cluster the positive and negative instance spaces, and the original feature space vector is measured by combining the cluster center point with the number of cluster points, from the original d-dimensional space Mapped to a d k Dimensional label-specific feature space in represents the label-specific features corresponding to the i-th instance, is the Euclidean distance, Represent the positive and negative cluster centers respectively, C is the i-th cluster center sample set, |C i | is the number of the i-th cluster center point, and the Euclidean distance is scaled by the number of clusters to alleviate the label imbalance problem in multi-label; the specific steps of step 103 calculating the correlation of paired labels in the text data set and constructing an undirected complete graph of the correlation between labels are: 1031.Define a correlation undirected complete graph: there is a correlation edge between any two labeled vertices in the graph and there are no self-loops or multiple edges; 1032.Use corr(l i ,l j ) calculates the label correlation set E, and uses the multi-label set V in the dataset as j ,…,l q } is the point set in the undirected complete graph G, q is the size of the multi-label set, and the pairwise label correlation set E = {corr(l i ,l j )|l i ∈V,l j ∈V} is the edge set in the undirected complete graph G; Where |l i | is the number of times label i appears, l ik represents the k-th dimension value in the i-th label vector, Represents the mean of the i-th label vector, takes the value with the smaller number of occurrences of paired labels, and combines the covariance and standard deviation of the labels in the label space to obtain the correlation between multiple labels.

2. The text data multi-label classification method based on label specific features and correlation according to claim 1 is characterized in that: The step 104 uses the extended Kruskal algorithm to calculate the maximum spanning tree in the graph, selects each vertex in the tree as a root to generate multiple maximum spanning trees with the same weight but different tree heights, and sorts the trees from low to high according to their heights. The specific steps are: 1041. Define the maximum spanning tree: a spanning tree containing a connected graph of all label nodes, containing all labels in the undirected complete graph of pairwise label correlations and having an edge set that maintains the maximum sum of weights of the connected graph of correlations; 1042. Define the extended Kruskal algorithm: adopt the greedy idea, sort the edge set in descending order according to the weight, and then recursively search the edges with the same weight to generate different tree structures until the maximum spanning tree set with the same total weight is generated; 1043. For the maximum spanning tree set, enumerate any child node in the tree as the root node to reconstruct the tree. A maximum spanning tree will produce |V| different tree structures, where V represents the set of nodes on the tree, and sort each reconstructed tree from low to high according to the tree height.

3. The text data multi-label classification method based on label specific features and correlation according to claim 2 is characterized in that: The extended Kruskal algorithm steps: 1). Sort the edge sets with correlation in descending order of weight; 2) Recursively search the weighted edges: If this edge does not form a loop when added to the current tree, add the edge to the tree until all label nodes are in the tree; 3). For the maximum spanning tree obtained in 2, enumerate each point as the root node to build multiple maximum spanning trees, and sort them from bottom to top according to their height.

4. The text data multi-label classification method based on label specific features and correlation according to claim 2 is characterized in that: The step 105 sequentially selects spanning trees from the tree set, and uses a binary classifier to perform recursive prediction on the tree starting from the root node, and integrates the prediction results on the tree to obtain a text instance label set. The specific steps are: 1051. For the multiple maximum spanning trees with different structures in step 105, select the first m k The tree is classified. For each tree, recursive prediction is performed starting from the root node. During the prediction process, the prediction results of the ancestor node labels are added to the child node label feature space to construct a new feature space. At this time, the label feature space is expanded to at this time Indicates the label l k New tag-specific features; in ancestor(l k ) indicates label l k A binary dataset of ancestor node labels in a maximum relatedness spanning tree; 1053. Use a binary classifier to classify the label node dataset Perform classification, and then recursively classify the sub-node labels until all nodes in the spanning tree are classified; η(Y i ,l k ) indicates label l k Is it in the instance-related tag set Y i middle; 1054. Since different root nodes will result in different tree structures, the strong correlation between the expressed labels is also different. In order to make full use of the strong correlation between labels and reduce the impact of selecting different root nodes, multiple trees are used for multi-label classification. For an unknown instance The number of tree sets is t, For label l k The classifier on the jth tree; integrate the same node on different trees, and finally get the label set Y u ;

Images