Multi-label information processing method based on multi-view perturbation

Abstract

This invention relates to the fields of machine learning and data mining technology, and discloses a multi-label information processing method based on multi-view perturbation, comprising the following steps: Step 1, data collection and preprocessing; Step 2, label-specific feature generation based on LIFT; Step 3, multi-view data perturbation and feature selection; Step 4, construction of a binary classification learner and voting ensemble. In this invention, appropriate distance metrics are selected for different types of data to determine sample similarity, providing a basis for subsequent operations. By calculating the standard deviation of the feature values ​​of nearest neighbor samples to perturb the samples, and by determining nearest neighbors based on feature correlation and calculating the standard deviation of relevant statistical information to perturb the features, the diversity of data and features is increased. The ReliefF algorithm is used for feature selection, assessing the importance of features based on their differences from similar and dissimilar samples, setting thresholds to filter features, reducing computational complexity, retaining key features, and improving classification accuracy and efficiency.

Description

Technical Field

[0001] This invention relates to the fields of machine learning and data mining technology, and in particular to a multi-label information processing method based on multi-view perturbation. Background Technology

[0002] With the rapid development of information technology, labeled data is growing exponentially, containing a wealth of diverse information. In the fields of machine learning and data mining, researchers are dedicated to extracting valuable information from this labeled data, and have thus conducted a series of studies related to label classification.

[0003] Traditional single-label classification mainly covers binary and multi-class classification. Its core principle is to assign a single label from a label set to an instance. Many mature algorithms exist in this field, demonstrating excellent performance and widespread application. However, in real-world scenarios, the limitations of single-label classification are becoming increasingly apparent. In fact, an instance is often associated with multiple labels. For example, in text classification, an electronic medical record may involve multiple disease labels such as diabetes, hypertension, and coronary heart disease; in image classification, a tongue diagnosis image can simultaneously express features such as tongue color, coating, and shape, thereby inferring multiple patient signs; and in music classification, an audio clip can convey various information such as piano performance, classical music style, and Mozart's works.

[0004] The classification problem where one instance corresponds to multiple labels is defined as multi-label classification, which essentially involves associating an instance with a set of labels. Compared to single-label classification, multi-label classification faces several challenges. First, the number of labels corresponding to each instance is not fixed and is not limited to one; second, there are semantic correlations between labels, which makes the classification process more complex; furthermore, some datasets in certain domains exhibit label imbalance, further increasing the difficulty of multi-label classification tasks. These problems have prompted researchers to seek more effective multi-label information processing methods to address the complex data classification needs in the real world. This invention proposes a multi-label information processing method based on multi-view perturbation against this background. Summary of the Invention

[0005] The present invention mainly addresses the technical problems existing in the prior art and provides a multi-label information processing method based on multi-view perturbation.

[0006] To achieve the above objectives, the present invention adopts the following technical solution: a multi-label information processing method based on multi-view perturbation, comprising the following steps:

[0007] Step 1: Data Collection and Preprocessing

[0008] We collected multi-label datasets from multiple fields, including disease diagnosis datasets in the medical field, scene classification datasets in the image field, and news classification datasets in the text field.

[0009] The collected data was cleaned, and statistical imputation methods were used to fill missing values. Numerical data was filled with the mean, and categorical label data was filled with the mode. Outliers were detected and corrected using the Z-Score standardization method, and the median was used to replace outliers.

[0010] Different types of data are normalized. For image data, the minimum-maximum normalization formula is used to normalize the original pixel values ​​to [0,1].

[0011] Step 2: Generation of label-specific features based on LIFT:

[0012] The positive and negative sample sets are divided based on the association between the sample and the label. Samples related to the label belong to the positive sample set, and those not related to the label belong to the negative sample set.

[0013] The k-means clustering algorithm is used to cluster the positive and negative sample sets of each label, calculate the distance from the sample to the cluster center, determine the cluster center by minimizing the objective function, and after setting a reasonable number of clusters k, the distance from the sample to its cluster center is used as a new feature to construct a label-specific feature space.

[0014] Step 3: Multi-view data perturbation and feature selection:

[0015] Distance metric selection: For image data, Euclidean distance based on pixel features is used, and for text data, cosine distance in word vector space is used; the number of nearest neighbors for each sample is set according to the dataset size and feature distribution, the standard deviation of the feature values ​​of the nearest neighbor set of each sample is calculated, and the samples are perturbed according to the standard deviation, which can be done by adding random noise or scaling the feature values;

[0016] Feature perspective perturbation: Calculate the correlation or similarity between features. For image features, calculate the cosine similarity between feature vectors. For text lexical features, calculate the co-occurrence frequency of words. Determine the nearest neighbors of each feature based on the correlation or similarity. Calculate the standard deviation of the statistical information of the nearest neighbor features. Perturb the features based on the standard deviation. Adjust the weights of lexical features for text data and scale the feature values ​​for image data.

[0017] Feature selection: The ReliefF algorithm is used for feature selection. The difference between a feature and similar and dissimilar samples is calculated to evaluate the importance of the feature. A threshold is set to filter features. The threshold is set according to the mean and standard deviation of the weights of all features. Features with weights higher than the threshold are selected as the final feature subset.

[0018] Step 4: Construct a binary classification learner and voting ensemble:

[0019] Multiple binary classifiers are constructed using selected feature subsets. Appropriate classification algorithms are selected and parameters are set according to the characteristics of the feature subsets and the data distribution, such as support vector machines, decision trees, and neural networks. The binary classifiers are trained using gradient descent with training data.

[0020] For test samples, predictions are made by multiple learners and then ensembled using a majority vote to combine the predictions from multiple learners.

[0021] Preferably, in the data collection and preprocessing steps, let the dataset be... Each sample It contains m features, if the features If missing values ​​exist, they are replaced with features. mean ,in, Not equal to missing values, here This indicates the number of samples with non-missing values ​​in feature j, ensuring data integrity.

[0022] Preferably, in the data collection and preprocessing steps, for the samples... The eigenvalues ​​of are represented by their Z-Score as . ,in It is the mean of feature j. It is the standard deviation of feature j, when When the value is determined to be out of the question, the median is used to replace it.

[0023] Preferably, in the LIFT-based label-specific feature generation step, let the positive sample set be... Cluster center is The distance from the sample to the cluster center By minimizing the objective function To determine the cluster centers.

[0024] Preferably, in the multi-view data perturbation and feature selection step, for two samples of the image data... and Its Euclidean distance , where m is the pixel feature dimension.

[0025] Preferably, in the multi-view data perturbation and feature selection step, the word vector representation of two samples of the text data is as follows: and The distance of the cosine , where n is the word vector dimension.

[0026] Preferably, in the multi-view data perturbation and feature selection steps, for image feature vectors... and cosine similarity , where m is the feature dimension.

[0027] As a preferred approach, in the multi-view data perturbation and feature selection steps, the total number of words in the text dataset is [number missing]. ,vocabulary and The number of times it appears simultaneously in all texts is Then their co-occurrence frequency for:

[0028] .

[0029] As a preferred method, the ReliefF algorithm calculates features in the multi-view data perturbation and feature selection steps. In the sample Weight ,in Indicates sample and In features Difference measurement on For the sample The set of nearest neighbor samples of the same type For the sample The set of nearest neighbor dissimilar samples is the absolute value of the difference for numerical features and an indicator function for whether they are the same for categorical features.

[0030] Preferably, in the steps of constructing the binary classifier and integrating voting, the parameter update formula for the binary classifier based on the mean squared error loss function is as follows: ,in For parameters, This is the learning rate.

[0031] Beneficial effects

[0032] This invention provides a multi-label information processing method based on multi-view perturbation. It has the following beneficial effects:

[0033] 1. This multi-label information processing method based on multi-view perturbation collects multi-label datasets from multiple domains, ensuring the richness and diversity of the data and providing comprehensive information for the model. During data cleaning, appropriate processing methods are adopted for missing values ​​and outliers to ensure data integrity and reduce interference from extreme values. Simultaneously, normalization processing is performed on different types of data to eliminate scale effects, laying a solid foundation for subsequent operations.

[0034] 2. This multi-label information processing method based on multi-view perturbation divides the sample and label sets into positive and negative sets according to the association between samples and labels, enabling the model to focus on learning label-specific information. The k-means clustering algorithm is used to further process the positive and negative sample sets, effectively clustering similar samples and highlighting label-related feature patterns. Transforming the distance from a sample to the cluster center into a new feature to construct the feature space helps the model better grasp label features and improve classification accuracy.

[0035] 3. This multi-view perturbation-based multi-label information processing method selects appropriate distance metrics to determine sample similarity for different types of data, providing a basis for subsequent operations. It perturbs features by calculating the standard deviation of nearest neighbor sample feature values ​​and by determining nearest neighbors based on feature correlation and calculating the standard deviation of relevant statistical information, increasing the diversity of data and features. The ReliefF algorithm is used for feature selection, assessing the importance of features based on their differences from similar and dissimilar samples, setting thresholds to filter features, reducing computational complexity, retaining key features, and improving classification accuracy and efficiency.

[0036] 4. This multi-view perturbation-based multi-label information processing method selects an appropriate classification algorithm and sets parameters according to the characteristics of the feature subset, adapting to various types of data and improving classification performance. A binary classification learner is trained using gradient descent on the training data, and parameters are optimized to achieve accurate predictions. Test samples are predicted by multiple learners and then ensembled through voting to synthesize the results, reducing the impact of individual learner errors and enhancing prediction stability and accuracy. Attached Figure Description

[0037] To more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings in the following description are merely exemplary, and those skilled in the art can derive other embodiments based on the provided drawings without creative effort.

[0038] Figure 1 This is a flowchart of the steps of the present invention;

[0039] Figure 2 This is a flowchart of the data collection and preprocessing process of the present invention;

[0040] Figure 3 Flowchart for generating specific features of the label in this invention;

[0041] Figure 4 This is a flowchart of the multi-view perturbation and feature selection process of the present invention;

[0042] Figure 5 This is a flowchart of the integration of the binary classification learner and voting in this invention. Detailed Implementation

[0043] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0044] Example: A multi-label information processing method based on multi-view perturbation, comprising the following steps:

[0045] Step 1: Data Collection and Preprocessing

[0046] We collect multi-label datasets from a wide range of fields, such as disease diagnosis datasets in the medical field (containing multi-label information such as patient symptoms and examination results), scene classification datasets in the image field (covering multi-label images such as natural scenes and human scenes), and news classification datasets in the text field (involving news articles with multiple thematic tags such as economy and culture), to ensure the richness and diversity of the datasets.

[0047] A comprehensive data cleaning process was performed on the collected data. The first step in data cleaning was to address missing and outlier values. For missing values, statistical imputation methods were used: mean imputation for numerical data and mode imputation for categorical label data.

[0048] Given a dataset Each sample It contains m features, if the features If missing values ​​exist, they are replaced with features. The mean, that is:

[0049]

[0050] in, Not equal to missing values, here This indicates the number of samples with non-missing values ​​in feature j, ensuring data integrity.

[0051] Outliers are detected and corrected using the Z-Score normalization method. For the sample... The eigenvalues ​​of are represented by their Z-Score as follows:

[0052]

[0053] in It is the mean of feature j. is the standard deviation of feature j. When When this occurs, it is considered an outlier, and a threshold is set. For outliers, the median can be used instead to reduce the impact of extreme values ​​on subsequent analysis.

[0054] In the normalization process, adaptive conversions are performed for different data types. For image data, the original pixel value p is normalized to a specific range [0,1] using the minimum-maximum normalization formula:

[0055]

[0056] in and These are the minimum and maximum values ​​of all pixel values ​​in the image, respectively. This method eliminates the influence of different pixel scales in the image, ensures the consistency of color and texture features, and facilitates subsequent label-specific feature generation and clustering operations.

[0057] Step 2: Generation of label-specific features based on LIFT

[0058] For each label, the dataset is divided into positive and negative sets based on the association between the sample and that label. The tag set is For tags If the sample and Related (e.g., in image scene classification, the image contains the labeled scene), then Belongs to the positive sample set Otherwise, it belongs to the negative sample set. The purpose of this partitioning method is to subsequently construct a space that highlights the features of each label, enabling the model to learn label-specific information. For example, in a medical diagnostic dataset, for the label "diabetes," samples with abnormal blood glucose levels, insulin levels, and other related indicators can be classified as positive samples, while normal samples are classified as negative samples.

[0059] For each label's positive and negative sample sets, the k-means clustering algorithm is used for clustering. The k-means algorithm aims to divide the sample set into k clusters such that the sum of the distances from each sample in each cluster to its cluster center is minimized. Let the positive sample set be... ( (Number of positive samples), cluster centers are Then the sample To the cluster center distance By minimizing the objective function This formula is used to determine cluster centers. The effect of this formula is to ensure that the clustering results reflect the distribution of samples in the feature space, grouping similar samples together to highlight feature patterns related to the label. For example, in image scene classification, for a positive sample set labeled "beach," clustering can group beach images with similar colors and textures to form a representative feature space.

[0060] After setting a reasonable number of clusters k, the distance from each sample to the cluster center is calculated to construct a label-specific feature space. The samples... To its respective cluster center distance As a new feature For each tag Its characteristic representation , where n is the total number of samples. This formula transforms the distribution information of the original samples in the high-dimensional feature space into new features related to the label, giving each label its own unique feature space, which helps the subsequent model learn the label features better. For example, for the "beach" label, in the feature space obtained through clustering, distance features can highlight beach-related feature patterns such as sand color, ocean color, and wave texture in beach images. These feature patterns help to accurately identify images containing beach scenes in subsequent steps.

[0061] In step three, the samples and features used in the sample-perspective perturbation and feature-perspective perturbation are based on the samples and features in the label-specific feature space constructed in this step. In step four, when constructing the binary classifier, a feature subset is selected using the feature representation of each label obtained in this step. These feature subsets, generated through label-specific feature generation, are more representative and discriminative, providing more valuable information to the classifier and thus improving the accuracy of the classifier's classification of each label.

[0062] Step 3: Multi-view data perturbation and feature selection

[0063] First, the distance metric needs to be chosen. For different types of data, an appropriate distance calculation method must be selected to measure the similarity between samples. In image data, due to the continuity and numerical nature of its pixel features, Euclidean distance based on pixel features is suitable. Let's consider two image samples... and ,in If the pixel feature dimension is used, the Euclidean distance between them is calculated using the following formula. :

[0064]

[0065] This formula accurately reflects the distance between two image samples in pixel space; the smaller the distance, the more similar the samples. By calculating the Euclidean distance between samples, the distribution relationship of the samples in feature space can be determined, providing a basis for subsequently identifying nearest neighbor samples.

[0066] For text data, due to its sparsity and semantic nature, it is more appropriate to use the cosine distance in the word vector space to measure the similarity between samples after converting the text into word vector representations. Let the word vector representations of two text samples be... and Where n is the word vector dimension, then the cosine distance between them is... The calculation formula is:

[0067]

[0068] Cosine distance measures the angle between two text samples in semantic space; the closer the value is to 1, the more similar the samples are. This allows us to determine the relationship between samples based on the semantic similarity of the text, which is helpful for subsequent perturbation operations.

[0069] After determining the distance metric, the number of nearest neighbors for each sample is appropriately set based on the dataset size and feature distribution. For a medium-sized image dataset, 10-15 nearest neighbors are typically chosen. One method for determining nearest neighbors is to calculate the distance between each sample and all other samples, then select the closest samples as the nearest neighbors. For a sample, calculate its distance to all other samples in the dataset, sort the distances from smallest to largest, and select the top 10-15 samples as the nearest neighbor set. The determination of nearest neighbor samples reflects the distribution of the sample in the local space, and the subsequent calculation of the standard deviation of the sample's feature values ​​is based on the feature value distribution of the nearest neighbor samples.

[0070] For each sample's nearest neighbor set, calculate the standard deviation of its eigenvalues. For image samples, the eigenvalues ​​include color, texture, and other features. Let the nearest neighbor set be... The l-th feature value of the k-th sample is Then the mean of this feature is:

[0071]

[0072] The standard deviation is:

[0073]

[0074] The standard deviation reflects the degree of dispersion of the nearest neighbor samples on this feature.

[0075] The samples are perturbed based on the calculated standard deviation. One approach is to add random noise based on the standard deviation to the sample feature values. For the color feature values ​​of image samples... ,i represents the sample number, l represents the color feature dimension, and the values ​​are added to the original values ​​to conform to a mean of 0 and a standard deviation calculated from the original values. Normally distributed random noise The perturbated sample feature values ​​are obtained.

[0076]

[0077] This allows for the generation of multiple perturbed sample sets, each with a different distribution of eigenvalues, thus increasing the diversity of the data.

[0078] Another perturbation method can be eigenvalue scaling, which multiplies the sample's eigenvalues ​​by a scaling factor based on the standard deviation:

[0079]

[0080] in, The scaling factor is used to obtain the perturbed eigenvalues.

[0081]

[0082] Calculate the correlation or similarity between features and use appropriate methods for different types of data.

[0083] For image features, the cosine similarity between feature vectors is used to measure correlation. Let the two feature vectors of the image be... and Where m is the feature dimension, the formula for calculating the cosine similarity between them is:

[0084]

[0085] The purpose of this formula is to accurately measure the directional similarity between two feature vectors; the closer the value is to 1, the more correlated the two features are. By calculating cosine similarity, the similarity relationship between image features can be determined, providing a basis for subsequently identifying feature neighbors.

[0086] For text lexical features, relevance is measured by statistically analyzing word co-occurrence frequencies. Let the total number of words in the text dataset be... ,vocabulary and The number of times it appears simultaneously in all texts is Then their co-occurrence frequency for:

[0087]

[0088] A higher co-occurrence frequency indicates a greater semantic correlation between two words. This approach reflects the correlation between text features at the lexical level, which is helpful for subsequent feature perturbation operations.

[0089] Based on the correlation or similarity between features, the nearest neighbors of each feature are determined. For text datasets with 50-100 features, the number of nearest neighbors for each feature is set to 3-5. The method for determining nearest neighbors is to sort the calculated correlation or similarity measures and select the most relevant or similar features as the nearest neighbors. For a word feature in the text data, its co-occurrence frequency with all other word features is calculated, and then the co-occurrence frequencies are sorted from largest to smallest, selecting the top 3-5 word features as the nearest neighbor set. Determining the nearest neighbors of a feature reflects the local relationship of the feature in semantics or feature space, and is of great significance for subsequent calculation of the standard deviation of the statistical information of the nearest neighbor features.

[0090] For each feature's nearest neighbor set, calculate its mean and standard deviation. For the word feature weights in the text data, let the nearest neighbor set be... The weight of the k-th feature is Then the mean of the feature weights is:

[0091]

[0092] The standard deviation is:

[0093]

[0094] Standard deviation reflects the degree of dispersion of neighboring features in terms of weights.

[0095] The features are perturbed based on the standard deviation. For the lexical feature weights in text data, the adjustment magnitude is determined based on the standard deviation. Let the adjusted lexical feature weights be... Adjust the formula to in To adjust the coefficients. This generates a perturbed feature set, increasing feature diversity. For the feature values ​​of the image data, a scaling operation is performed, with the scaling factor determined based on the standard deviation. Let the image feature values ​​be... Where i represents the sample index, l represents the feature dimension, and the scaling factor is calculated using the following formula:

[0096]

[0097] in, The standard deviation of the corresponding feature. If is the scaling factor, then the perturbed eigenvalue is .

[0098] Finally, feature selection is performed. For each label, after perturbation, multiple feature spaces are used to select features separately using the ReliefF algorithm. The ReliefF algorithm evaluates feature importance by calculating the differences between features and similar and dissimilar samples. Let the sample set be... Each sample It contains m features, for a certain feature Calculate its value in the sample The differences between the sample and its neighboring similar and dissimilar samples. First, find the sample Nearest neighbor set of similar samples and neighboring heterogeneous sample set The nearest neighbor determination method is the same as the nearest neighbor determination method used in the previous sample viewpoint and feature viewpoint perturbation. Then, the features are calculated. In the sample Weight The calculation formula is:

[0099]

[0100] in Indicates sample and In features The difference measure can be the absolute value of the difference for numerical features, or an indicator function for whether they are the same for categorical features. The purpose of this formula is to highlight features that effectively distinguish different categories by comparing the differences in features between a sample and samples of the same and different classes, ensuring that feature weights accurately reflect the importance of features for classification.

[0101] The weights of each feature are calculated using the ReliefF algorithm. Then, a reasonable threshold needs to be set. To filter features, the threshold setting process involves first calculating the mean of all feature weights. and standard deviation Then set the threshold to Where k is the adjustment coefficient. Select weights higher than the threshold. The features selected are considered relevant and necessary, as they are deemed to contribute significantly to label classification. Features with weights below a threshold are treated as redundant or irrelevant and removed, resulting in the final feature subset for each label. This approach reduces the dimensionality of the feature space, lowers the computational complexity of the model, and retains the most valuable features for classification, thereby improving the model's classification accuracy and efficiency.

[0102] Step 4: Construct a binary classification learner and voting ensemble

[0103] For each label, multiple binary classifiers are constructed using a selected subset of features. Different classification algorithms have their own characteristics and applicable scenarios when constructing multiple binary classifiers based on the selected feature subset. Support Vector Machines (SVMs) are suitable for linearly separable or non-linearly separable data; their kernel function parameters determine how the data is mapped in high-dimensional space. Let the kernel function be... Adjusting the kernel function parameters can alter the shape of the classification boundary, improving classification accuracy. Decision trees classify data by constructing a tree structure; their depth parameter affects the tree's complexity and generalization ability. Shallow decision trees may lead to underfitting, while overly deep decision trees may overfit the data. Neural networks learn data features by constructing multiple layers of neurons; the number of layers and nodes determines the model's expressive power. For example, a simple neural network can have an input layer, hidden layers, and an output layer. Let's say the hidden layer has h nodes, the input layer has m nodes (corresponding to the feature subset dimension), and the output layer has one node de for binary classification. Appropriate values ​​can be set based on data complexity and feature subset characteristics.

[0104] These parameters are set based on the statistical properties of the feature subset, including mean and variance, and the data distribution, which includes whether the data is normally distributed and its sparsity. When the feature subset has high dimensionality and the data distribution is complex, a neural network is selected, and the number of layers and nodes is appropriately increased; when the data exhibits linearly separable features, an SVM with a linear kernel function is used, and its regularization parameters are adjusted.

[0105] Each binary classifier is trained using training data to optimize the algorithm and adjust the learner parameters. This embodiment uses gradient descent for a binary classifier based on the mean squared error (MSE) loss function, as shown in the following formula:

[0106]

[0107] in For real labels, The formula for updating the parameters to predict the label is:

[0108]

[0109] in, For the current parameter, Let be the learning rate. The purpose of this formula is to gradually decrease the loss function value by iteratively updating the parameters, thereby enabling the learner to predict sample labels more accurately. During training, the gradient of the loss function with respect to the parameters is calculated based on the feature values ​​of the samples and the true labels. Then, the parameters are updated according to the above formula until the stopping condition is met.

[0110] For the test sample, it is input into the binary classifier corresponding to each label to obtain the prediction result. Each learner judges whether the sample belongs to the label based on its learned model and outputs a prediction result of 0 (negative class) or 1 (positive class).

[0111] For each label's prediction result, a majority vote ensemble is performed. For a given label of a sample, k binary classifiers participate in the vote, where... The learner predicts it as a positive class. Each learner predicts the class as negative, and .if If the result is positive, the sample is determined to belong to that label; otherwise, it is not. Voting ensemble can combine the prediction results of multiple learners, reducing the impact of errors from individual learners and improving the stability and accuracy of predictions.

[0112] The foregoing has shown and described the basic principles, main features, and advantages of the present invention. Those skilled in the art should understand that the present invention is not limited to the above embodiments. The embodiments and descriptions in the specification are merely illustrative of the principles of the invention. Various changes and modifications can be made to the invention without departing from its spirit and scope, and all such changes and modifications fall within the scope of the present invention as claimed. The scope of protection of this invention is defined by the appended claims and their equivalents.

Claims

1. A multi-label information processing method based on multi-view perturbation, characterized in that, Includes the following steps: Step 1: Data Collection and Preprocessing We collected multi-label datasets from multiple fields, including disease diagnosis datasets in the medical field, scene classification datasets in the image field, and news classification datasets in the text field. The collected data was cleaned, and statistical imputation methods were used to fill missing values. Numerical data was filled with the mean, and categorical label data was filled with the mode. Outliers were detected and corrected using the Z-Score standardization method, and the median was used to replace outliers. Different types of data are normalized. For image data, the minimum-maximum normalization formula is used to normalize the original pixel values ​​to [0,1]. Step 2: Generation of label-specific features based on LIFT: The positive and negative sample sets are divided based on the association between the sample and the label. Samples related to the label belong to the positive sample set, and those not related to the label belong to the negative sample set. The k-means clustering algorithm is used to cluster the positive and negative sample sets of each label, calculate the distance from the sample to the cluster center, determine the cluster center by minimizing the objective function, and after setting a reasonable number of clusters k, the distance from the sample to its cluster center is used as a new feature to construct a label-specific feature space. Step 3: Multi-view data perturbation and feature selection: Distance metric selection: For image data, Euclidean distance based on pixel features is used, and for text data, cosine distance in word vector space is used; the number of nearest neighbors for each sample is set according to the dataset size and feature distribution, the standard deviation of the feature values ​​of the nearest neighbor set of each sample is calculated, and the samples are perturbed according to the standard deviation, which can be done by adding random noise or scaling the feature values; Feature perspective perturbation: Calculate the correlation or similarity between features. For image features, calculate the cosine similarity between feature vectors. For text lexical features, calculate the co-occurrence frequency of words. Determine the nearest neighbors of each feature based on the correlation or similarity. Calculate the standard deviation of the statistical information of the nearest neighbor features. Perturb the features based on the standard deviation. Adjust the weights of lexical features for text data and scale the feature values ​​for image data. Feature selection: The ReliefF algorithm is used for feature selection. The difference between a feature and similar and dissimilar samples is calculated to evaluate the importance of the feature. A threshold is set to filter features. The threshold is set according to the mean and standard deviation of the weights of all features. Features with weights higher than the threshold are selected as the final feature subset. Step 4: Construct a binary classification learner and voting ensemble: Multiple binary classifiers are constructed using a selected subset of features. A suitable classification algorithm is selected and parameters are set according to the characteristics of the feature subset and the data distribution. The binary classifiers are trained using gradient descent with the training data. For test samples, predictions are made by multiple learners and then ensembled using a majority vote to combine the predictions from multiple learners.

2. The multi-label information processing method based on multi-view perturbation according to claim 1, characterized in that: In the data collection and preprocessing steps, let the dataset be... Each sample It contains m features, if the features If missing values ​​exist, they are replaced with features. mean ,in, Not equal to missing values, here This indicates the number of samples with non-missing values ​​in feature j, ensuring data integrity.

3. The multi-label information processing method based on multi-view perturbation according to claim 1, characterized in that: In the data collection and preprocessing steps, for the samples The eigenvalues ​​of are represented by their Z-Score as . ,in It is the mean of feature j. It is the standard deviation of feature j, when When the value is determined to be out of the question, the median is used to replace it.

4. The multi-label information processing method based on multi-view perturbation according to claim 1, characterized in that: In the label-specific feature generation step based on LIFT, let the positive sample set be... Cluster center is The distance from the sample to the cluster center By minimizing the objective function To determine the cluster centers.

5. The multi-label information processing method based on multi-view perturbation according to claim 1, characterized in that: In the multi-view data perturbation and feature selection step, for two samples of image data and Its Euclidean distance , where m is the pixel feature dimension.

6. The multi-label information processing method based on multi-view perturbation according to claim 1, characterized in that: In the multi-view data perturbation and feature selection step, the word vectors of two samples of the text data are represented as follows: and The distance of the cosine , where n is the word vector dimension.

7. The multi-label information processing method based on multi-view perturbation according to claim 1, characterized in that: In the multi-view data perturbation and feature selection steps, for image feature vectors and cosine similarity , where m is the feature dimension.

8. The multi-label information processing method based on multi-view perturbation according to claim 1, characterized in that: In the multi-view data perturbation and feature selection steps, the total number of words in the text dataset is... ,vocabulary and The number of times it appears simultaneously in all texts is Then their co-occurrence frequency for: 。 9. The multi-label information processing method based on multi-view perturbation according to claim 1, characterized in that: In the multi-view data perturbation and feature selection steps, the ReliefF algorithm calculates features. In the sample Weight ,in Indicates sample and In features Difference measurement on For the sample The set of nearest neighbor samples of the same type For the sample The set of nearest neighbor dissimilar samples is the absolute value of the difference for numerical features and an indicator function for whether they are the same for categorical features.

10. The multi-label information processing method based on multi-view perturbation according to claim 1, characterized in that: In the steps of constructing a binary classifier and integrating voting, the parameter update formula for the binary classifier based on the mean squared error loss function is as follows: ,in For parameters, This is the learning rate.

Images